EVALUATION (XIV)
Studies in Applied Electromagnetics
and Mechanics
Series Editors: K. Miya, A.J. Moses, Y. Uchikawa, A. Bossavit, R. Collins, T. Honma,
G.A. Maugin, F.C. Moon, G. Rubinacci, H. Troger and S. A. Zhou
Volume 35
Previously published in this series:
Vol. 34. S. Wiak and E. Napieralska Juszczak (Eds.), Computer Field Models of
Electromagnetic Devices
Vol. 33. J. Knopp, M. Blodgett, B. Wincheski and N. Bowler (Eds.), Electromagnetic
Nondestructive Evaluation (XIII)
Vol. 32. Y. K. Shin, H. B. Lee and S. J. Song (Eds.), Electromagnetic Nondestructive
Evaluation (XII)
Vol. 31. A. Tamburrino, Y. Melikhov, Z. Chen and L. Udpa (Eds.), Electromagnetic
Nondestructive Evaluation (XI)
Vol. 30. S. Wiak, A. Krawczyk and I. Dolezel (Eds.), Advanced Computer Techniques in
Applied Electromagnetics
Vol. 29. A. Krawczyk, R. Kubacki, S. Wiak and C. Lemos Antunes (Eds.), Electromagnetic
Field, Health and Environment Proceedings of EHE07
Vol. 28. S. Takahashi and H. Kikuchi (Eds.), Electromagnetic Nondestructive Evaluation (X)
Vol. 27. A. Krawczyk, S. Wiak and X.M. Lopez Fernandez (Eds.), Electromagnetic Fields in
Mechatronics, Electrical and Electronic Engineering
Vol. 26. G. Dobmann (Ed.), Electromagnetic Nondestructive Evaluation (VII)
Vol. 25. L. Udpa and N. Bowler (Eds.), Electromagnetic Nondestructive Evaluation (IX)
Vol. 24. T. Sollier, D. Prmel and D. Lesselier (Eds.), Electromagnetic Nondestructive
Evaluation (VIII)
Vol. 23. F. Kojima, T. Takagi, S.S. Udpa and J. Pv (Eds.), Electromagnetic Nondestructive
Evaluation (VI)
Vol. 22. A. Krawczyk and S. Wiak (Eds.), Electromagnetic Fields in Electrical Engineering
Vol. 21. J. Pv, G. Vrtesy, T. Takagi and S.S. Udpa (Eds.), Electromagnetic Nondestructive
Evaluation (V)
Vol. 20. Z. Haznadar and . tih, Electromagnetic Fields, Waves and Numerical Methods
Vol. 19. J.S. Yang and G.A. Maugin (Eds.), Mechanics of Electromagnetic Materials and
Structures
Volumes 1 6 were published by Elsevier Science under the series title Elsevier Studies in
Applied Electromagnetics in Materials.
Edited by
Tomasz Chady
West Pomeranian University of Technology, Szczecin, Poland
Stanisaw Gratkowski
West Pomeranian University of Technology, Szczecin, Poland
Toshiyuki Takagi
Tohoku University, Japan
and
Satish S. Udpa
Michigan State University, USA
All rights reserved. No part of this book may be reproduced, stored in a retrieval system,
or transmitted, in any form or by any means, without prior written permission from the publisher.
Publisher
IOS Press BV
Nieuwe Hemweg 6B
1013 BG Amsterdam
Netherlands
fax: +31 20 687 0019
e mail: order@iospress.nl
LEGAL NOTICE
The publisher is not responsible for the use which might be made of the following information.
Preface
This volume contains selected papers from the fifteenth International Workshop on
Electromagnetic Nondestructive Evaluation, which was held in Szczecin, Poland, from
June 13 to 16, 2010.
Previous ENDE Workshops have been held in: London, United Kingdom (1995);
Tokyo, Japan (1996); Reggio Calabria, Italy (1997); Chatou, France (1998); Des
Moines, United States (1999); Budapest, Hungary (2000); Kobe, Japan (2001); Saar
brcken, Germany (2002); Paris, France (2003); East Lansing, United States (2004);
Iwate, Japan (2006); Cardiff, United Kingdom (2007); Seoul, Korea (2008); Dayton,
United States (2009).
The aim of the workshop, organized by the West Pomeranian University of Tech
nology, Szczecin, Poland and the Japanese Society of Maintenology, was to bring to
gether scientists from universities and research institutions conducting indepth re
search into the basics of electromagnetic nondestructive evaluation (ENDE) on the one
hand, and engineers presenting practical problems and industrial applications on the
other.
Ninety nine participants from eleven European countries and from Algeria, Austra
lia, Brazil, China, India, Japan, Korea, and the United States, were officially registered.
Eighty papers were presented in all, among them five invited papers, namely:
1. D.C. Jiles, .P. Mierczak, Y. Melikhov, Detection of Surface Condition in
Ground Steel Components Using Magnetic Barkhausen Measurements,
2. S. Honda, Pipe Wall Thickness Inspection with Current Driven Thermal
Method,
3. J.S. Knopp, M. Blodgett, J. Calzada, E. Lindgren, C. Buynak, J. Aldrin, Com
putational Methods in ENDE: Revolutionary Capability for the Sustainment of
Aerospace Systems in the 21st Century,
4. Z. Chen, T. Hwang, L. Wang, S. Tian, N. Yusa, Investigation on the Features
of the Electric Conductivity Around a Stress Corrosion Crack,
5. S.C. Chan, R. Grimberg, J.A. Hejase, Z. Zeng, P. Lekeakatakunju, L. Udpa,
S.S. Udpa, Development of a Nonlinear Eddy Current Technique for Estimat
ing Case Hardening Depths.
Short versions of all the contributions have been published in the Book of Ab
stracts, and reviewed and accordingly revised full papers have been accepted and are
now included in this volume: Electromagnetic NonDestructive Evaluation (XIV) pub
lished by IOS Press in the series Studies in Applied Electromagnetics and Mechanics.
In closing, we would like to thank the authors, session chairs, and reviewers for
conscientiously executing their duties to maintain the high scientific quality of the pa
pers published in this volume. We believe that the readership of this book will find the
included papers interesting and inspiring.
Organized by
West Pomeranian University of Technology, Szczecin, Poland
in Cooperation with
Japan Society of Maintenology, Tokyo, Japan
CoSponsors
Japan Society of Maintenology, Tokyo, Japan
Oddzia Szczeciski Stowarzyszenia Elektrykw Polskich, Szczecin, Poland
Federacja Stowarzysze NaukowoTechnicznych NOT w Szczecinie, Poland
ZAPOL grupa reklamowa, Szczecin, Poland
Technika Obliczeniowa, Krakw, Poland
Standing Committee
F. Kojima Japan
I. Altpeter Germany
J. Bowler USA
N. Bowler USA
T. Chady Poland
Z. Chen China
D. Jiles UK
J. Knopp USA
D. Lesselier France
K. Miya Japan
G.Z. Ni China
J. Pavo Hungary
G. Pichenot France
G. Rubinacci Italy
S.J. Song Korea
T. Takagi Japan
A. Tamburrino Italy
L. Udpa USA
S.S. Udpa USA
viii
Organizing Committee
Honorary Member: Rector ZUT Wodzimierz Kiernoycki
Chairman: Tomasz Chady
CoChairman: Stanisaw Gratkowski
Ryszard Sikora
Stefan Domek
Andrzej Brykalski
Members:
Piotr Baniukiewicz
Wojciech Chlewicki
Pawe Frankowski
Justyna Joczyk
Jacek Kowalczyk
Krzysztof Kujawski
Pawe Lesiecki
Przemysaw opato
Lech Napieraa
Marzena Olszewska
Grzegorz Psuj
Tomasz Pietrusewicz
Krzysztof Stawicki
Barbara Szymanik
Marcin Zikowski
ix
Contents
Preface v
T. Chady, S. Gratkowski, T. Takagi and S.S. Udpa
Conference Organization vii
List of Participants ix
Keynote Lecture
Material Characterization
Introduction
Electric potential dierence(EPD) method is widely used for diagnosing and mon
itoring ow conduits [1]. It applies large electric currents between a pair of elec
trodes on pipe wall surface, and the electrical potentials are measured with mul
tiple potential electrodes. Since the localized pipe wall thinning or a wall crack
deforms the potential eld, we can monitor and/or diagnose pipe wall condition
under test.
The method is also required to evaluate pipe wall thinning in larger span. We
have studied the prospect of the method to evaluate pipe wall thinning. Based
on the work on nonintrusive resistance thermometer for fast breeder reactor[2,3],
theoretical analysis and the exact solution of the electrical potential eld in pipe
wall have been proposed[4], and 2D approximation is eective for small pipe wall
thickness.
We propose in the paper to use temperature distribution driven by Joule heat
with the electric current other than electrical potential dierence. Under thin
wall approximation, 2D electric and thermal problems are analyzed and exact
solutions are given.
1 Corresponding Author: Satoshi Honda, Faculty of Science & Technology, Keio University,
1.1. 3D Analysis
Let the electric potential inside the uid be 1 (, , z), and the one inside the
pipe wall, 2 (, , z), where cylindrical coordinates are adopted with pipe axis as
z coordinate. Electrical current is fed between a pair of rectangular electrodes of
bwbw on the outer surface (b, 0, 0). If we let 1 and 2 be electrical conductivities
of the uid and the pipe wall, the following potential problem is formulated.
2 = 0, a<<b (1)
1 = 0, <a (2)
2 i0
= rect(; w) rect(z; bw), (3)
=b 2
2 1 2
1 rho=a = 2 =a 2 , 1 = 2 , (4)
=a =a =a
and i0 , the driving current density. In order to solve the problem, we introduce
the potential distribution 0 for the bulk cylinder of the radius b:
0 = 0, <b (6)
0 i0
= rect(; w) rect(z; bw), (7)
=b 2
1
G = ( ) ( )(z z ), (8)
G
=0 (9)
=b
m Km (b)
G(, , z; , , z ) = cos m(
) Km (
) (b) m
I (
)
m=0
2 2 0 Im
Im () cos (z z ) d
(10)
S. Honda / Pipe Wall Thickness Inspection with Current Driven Thermal Method 5
where Im () and Km () denote mth order modied Bessel functions of rst and
second kinds, respectively, and m , Neumann factor. This Greens function gives
the potential 0 for the simplied model:
i0
0 = G(, , z; b, , z ) rect( ; w) rect(z ; bw)bd dz
S 2
mw
i0 m Im () bw
= cos m sinc (b)
cos z sinc d,
2 m=0 2 2 2 0 bIm 2
(11)
Due to the linearity of the original potential problem, we assume the following
solution:
2 = 0 + 2
m
= 0 + cos m [dm Km () + cm Im ()] cos z d, (12)
m=0
2 2 0
m
1 = cos m fm Im () cos z d (13)
m=0
2 2 0
Since 0 has already accounted for the current feeding electrodes, the boundary
condition for the potential 2 reads as follows:
2
= 0. (14)
=b
From three boundary conditions (4) and (14), we obtain three independent equa
tions for the coecients cm , dm , fm of series expansions, which evaluate the coef
cients.
When we put a pair of current feeding electrodes on = 0, , the exact form
of the potential distribution is,
1
Im () Am ()
2 = cos m (b)
+ Wm () cos z d (15)
2 0 bIm Bm ()
m:odd
1 Km (b)
Am () Im (a) + Im (a) Km () Im () (16)
b b Im (b)
Im (a)
Bm () + [Im (b)Km (a) Im (a)Km (b)]
Im (a)
+ [Im (a)Km (b) Im (b)Km (a)] (17)
For the convenience, let us normalize the variables here in order to evaluate surface
potential, as b, a/b.. In the limit case of the conductivities being
6 S. Honda / Pipe Wall Thickness Inspection with Current Driven Thermal Method
0 1 2 and of very thin pipe wall, the potential distribution on the wall
reduces to,
mw
z
i0 2 w
2 (b, , z) = cos msinc Q m () sinc cos d
2 b 2 2 0 2 b
m:odd
(18)
Im () I () 1
Qm ()
2m () K ()I ()
(19)
Im () Im () Km ()Im m m
ba
Since = 1 is small, dropping higher order term of in Taylor series
b
expansion of each function of Qm gives the following simplication.
Im () Im () Im () 1
Qm () = ()
()
()I () K ()I ()]
Im 2 Im [Km m m m
1 a 1
= = 2 (20)
2 2
( + m ) b a + m2
At the reduction to this nal form, we have adopted formulas for modied Bessel
functions. From Eq.(20), the surface potential distribution is also reduced to,
a i0 2 mw
2 (b, , z) = 2 cos m sinc
b a b 2 2
m:odd
w z
1
sinc cos d (21)
0 2 + m2 2 b
Let the pipe wall thickness be t b a. If we make the size w of the current
feeding electrode innitesimally small, sinc(, w) approaches to unity. Therefore,
in this limit case surface potential can be evaluated as follows.
z
i0 2 1
2 (b, , z) = 2 cos m 2 2
cos d
2 t 0 +m b
m:odd
z
1 i0 cosh cos
= log zb (22)
4 2 t cosh + cos
b
The denominator, 2 t, implies that we can model thinning pipe wall as the region
with reduced conductivity. Consequently, reduced 2D problem can be analyzed
in the complex plane,
Z = X + iY, z X, b Y (23)
/
z/b
From the gure, we conclude that the current distribution is uniform in the region
far from the current feeding electrodes.
Joule heat generated by the driving current varies with wall pipe thickness, or
conductivity. Thus the resulting temperature distribution changes with the wall
thickness. In order to analyze the resulting temperature distribution, Joule heat
distribution is modeled beforehand.
We model the thinning wall region as small circle with the radius of a at the
origin in whole 2D plane far from the current feeding electrodes. Then Joule heat
distribution is obtained from the following 2D potential problem:
1 = 0, a<r (26)
2 = 0, r<a (27)
1 2
1 r=a = 2 r=a , 1 = 2 , (28)
r r=a r r=a
where the same symbols 1 , 2 as in the previous section are used for describing
the electric potentials on the innite 2D plate. As in the case of 3D analysis, we
8 S. Honda / Pipe Wall Thickness Inspection with Current Driven Thermal Method
expand the potentials in series of ( ar )n , cos n, sin n, and the boundary condi
tions determine the coecients of the series expansion to give the following simple
solutions.
1 a2
1 = E0 r cos E0 cos + c0 , ra (29)
1+ r
2
2 = E0 r cos + c0 , r<a (30)
1+
The resulting Joule heat reads,
P1 = j 1 E 1 = 1 (grad 1 ) (grad 1 )
2 4
1 a2 1 a
= 1 E0 1
2
2 cos 2 + , (31)
1 + r2 1+ r4
4
P2 = j 2 E 2 = 2 (grad 2 ) (grad 2 ) = 1 E02 , (32)
(1 + )2
This Joule heat generates the temperature distribution u according to the
following heat transfer equation:
u
= 2 u 2 u + P (33)
t
a2 2 a
4
P = H(r a)P0 2 2 cos 2 + 4 ,
2
(34)
r r
t
1
T (t) = exp (t t )
2
dt (38)
0 4(t t )2 t t
1
(r2 2rr cos( ) + r2 )
42
Since the stationary state is of main interest, we evaluate the integration with t
to innity and change the variable p = t t :
1 1 1
T () = exp 2 p dp = 2K0 (2 )
420 p p 4 2
1 2 2
= K0 r 2rr cos( ) + r . (39)
2
2
P0
u (r, ) u(r, , ) = K 0 2 2 cos( ) + 2
0 a 2 2
a2 2 a
4
2 2 cos 2 + 4 d d
2
(40)
integration with respect to remains only for n = 0, 2. And nally, after the
evaluation of the last integral with respect to , we obtain,
1 1
u (r, ) = P 0 2
aJ0 () K 1 (a) + a 3
K 0 (
)d
2 a 3
1
P0 2a2 cos 2J2 ()K1 (a), (r < a) (42)
2
10 S. Honda / Pipe Wall Thickness Inspection with Current Driven Thermal Method
1 2 3 1
u (r, ) = 2 P0 aJ0 () K1 () + a K0 ( )d
3
1
P0 2a2 cos 2J2 ()K1 ()
2
1 1
+ 2 P0 K0 () aJ1 (a) J1 () + a
2 4
J0 ( )d
a 3
1 1 1
2 P0 2a cos 2K2 ()
2
J1 (a) J1 () , (r a) (43)
a
An example shape of the Joule heat distribution is shown in Fig.2(a), and the
corresponding temperature distribution, in Fig.2(b).
Figure 2. Stationary temperature and heat distributions are numerically evaluated from Eqs.(42)
& (43).(a)Joule heat distribution. (b)3D plot. First quadrant is shown due to the symmetry.
3. Conclusion
References
[1] A. M. Pritchard and P. Webb Use of the FSM Technique in the Laboratory to Measure
Corrosion Inhibitor Performance in Multiphase Flow Proc. CORROSION/98, paper no.8
[2] H. Yamasaki, S. Honda, M. Ueda, A. Endou, M. Fueki A Novel Nonintrusive Resistance
Thermometer for Fast Breeder Reactor Proc. SICE2003 Annual Conf. Fukui pp 519 /
523(2003)
[3] H. Yamasaki, S. Honda, M. Ueda, A. Endou, M. Fueki A Novel Nonintrusive Resistance
Thermometer for Fast Breeder Reactor Trans. SICE, vol.43, No.9, pp 756 / 761(2007)
[4] S.Honda Thinning Pipe Wall Evaluation by using Electric Potential Dierence Method
Proc. SICE2008 Annual Conf. Tokyo pp 520 / 529(2008)
Modeling and Inverse Problems
This page intentionally left blank
Electromagnetic Nondestructive Evaluation (XIV) 13
T. Chady et al. (Eds.)
IOS Press, 2011
2011 The authors and IOS Press. All rights reserved.
doi:10.3233/978160750750513
Keywords. Eddy current, Giant magnetoresistive sensor, sensor tilt, multi layered
structure
Introduction
The detection of hidden cracks around fastener sites in multilayered structures poses a
major challenge in the aviation industry [1]. Although, eddy current testing methods
have been used extensively in this application, the skin depth or penetration depth of
the fields has limited their use in detecting 2nd and 3rd layer defects under fastener
heads. With the development of magnetoresistive (MR) sensors, eddy current
inspection system combined with MR sensors is finding increasing use in detecting
directly the magnetic field associated with the induced currents. An eddy current  giant
magnetoresistive (GMR) sensor system has been developed and shown to be effective
in detecting subsurface cracks in a multilayer structure. A schematic of the sensor, with
linear multiline currents and GMR sensor on the line of symmetry, is shown in Figure 1.
In the presence of fasteners and/or cracks, a nonzero normal magnetic field is
produced at the center of the source coil and is picked up by the GMR sensor. The
GMR signal comprises measurement of normal component of magnetic flux density
associated with induced eddy currents in the sample [2].
*
Corresponding Author: udpal@egr.msu.edu
14 G. Yang et al. / Modeling and Signal Processing Sensor Tilt in Eddy CurrentGMR Inspection
Figure 1. The EC GMR sensor: (a) the excitation coil with a GMR sensor and the tested sample; (b)
schematic of operation of sensor for crack detection.
In a practical inspection scenario, variations in several experimental parameters
such as lift off, top and bottom layer conductivity, edge effect can alter the measured
signal and hence the POD of a subsurface defect. This paper presents the modeling of
defect signals in the presence of sensor tilt during the measurement process. Further
the signals generated are used to derive discriminatory features that enhance the
detection probability of subsurface defects in the presence of sensor tilt. The detection
of detects based on the asymmetry of two lobes of the rivet signal using modelbased
methods has been reported in [34]. In these studies, the source current is modeled as
an infinite coil, obviating the need for modeling the scanning operation which is a time
intensive process. However modeling of sensor tilt is rendered complex because of the
need to model a finite source coil and the remeshing at each scan position of the
sensor.
This paper presents the application of a finite element formulation using reduced
magnetic vector potential and electric scalar potential [56] for simulating the multi
line coil motion across the sample without remeshing the source coil which results in
significant computational savings. The feature based defect detection and classification
using multicomponent signal features that are insensitive to sensortilt effect validate
the EC  GMR sensor system for the application considered.
A multiline coil producing uniform eddy current across the test rivet structure,
sketched schematically in Fig. 1 (a), is employed as the current source. In the absence
G. Yang et al. / Modeling and Signal Processing Sensor Tilt in Eddy CurrentGMR Inspection 15
of any discontinuity, the symmetry of the coil geometry results in zero normal
component of magnetic field at the line of symmetry on the center of the source coil. In
the presence of a rivet and/or crack, the uniform magnetic field is distorted to yield the
nonzero normal component response. A GMR sensor is placed at the center location to
measure this normal flux density Bz due to the perturbation in induced current. The
analysis of nonzero signal measured by the GMR sensor can be used to detect and
characterize cracks around rivet sites.
The sample consisting of three aluminum layers riveted together is studied in this
paper. The rivets are drilled through three layers and cracks of different sizes are
machined in the 3rd layer plate around rivet sites. A single frequency excitation current
is applied to the multiline coil and the changes in magnetic field are collected by the
GMR sensor.
Finite element formulation of 3D eddy current testing problems is typically based on
magnetic vector potential A and electric scalar potential V. However this formulation
requires generating a mesh for the excitation foil/coil together with the test sample and
the domain has to be remeshed for each coil position. This can result in excessive
computation time and error. Finite element model based on a formulation using the
reduced magnetic vector potential and the electric scalar potential
( Ar V Ar formulation) has been proposed to simulate the ECGMR inspection
signals [56]. The advantage of this formulation is that the source coil is not meshed
and hence there is no need for remeshing the source current during scanning. The
mesh and system matrix remain unchanged with different coil motion. The
implementation of the system matrix preconditioner is performed only once during the
entire scan.
The formulation model defines the reduced magnetic vector potential by the
decomposition of the magnetic flux density B into two parts:
B = Bs + Br = 0 H s + Br (1)
where Bs and Hs are the flux density and field density due to the source in free space,
and Br is the flux density due to induced current or magnetization. Considering the field
equations together:
Bs = 0 , Hs = Js , Br = B Bs = 0 ,
A s = Bs , A r = Br (2)
1 r r
Hs = J s (r) d (3)
4
r r
3
16 G. Yang et al. / Modeling and Signal Processing Sensor Tilt in Eddy CurrentGMR Inspection
0 J s (r)
As =
4 r r
d (4)
where is the volume of the current source; 0 is the freespace permeability. r and r
denote the coordinates of observation and source points, respectively.
Correspondingly, the governing equations using this formulation can be written as:
2 A r + j A r + V = (1 vr ) H s j A s
( jA r + V ) = jA s (5)
(N jA + N V )d + N [ j (A
i r i
i r + A s ) V ] n d
= N j A d
i s (6)
where N i ( i= 1,2, N with N the number of nodes) are shape functions, the
solution domain, conductor surface, and n the unit outward normal vector of . By
neglecting the surface integral in (6), we implicitly set the normal component of
induced currents on conductor surface to 0.
The cost of the new formulation is that besides H s , H s in presence of magnetic
materials and A s must be evaluated. However, evaluating these analytical values takes
much less time than that required for performing the incomplete factorization in each
coil position. The model has been validated by comparing model predicted signal and
experimental signal as shown in Figure 2. The Figure 2 (a) presents the model
validation in terms of defective rivet detection. The geometry of this sample is two
layers (6mm and 4mm thickness) with a 2nd layer radial crack of length 5mm at a rivet
site. The 1D signal measured in a scan across the rivet in X direction is plotted and the
same trend and asymmetry is observed in both simulated and experimental signals. The
Figure 2 (b) presents validation of the model for a cylindrical geometry; air core
pancake probe inspection of a tube with an OD axial notch at 100KHz single frequency
excitation.
G. Yang et al. / Modeling and Signal Processing Sensor Tilt in Eddy CurrentGMR Inspection 17
0.9
0.8
0.7
0.6
(a) 0.5
0.4
0.3
0.2
(b)
Figure 2. Model validation: comparison of 1D model simulated signal (red) with experimental measurement
(blue) (a) 3rd layer crack in multilayer sample (b) OD axial notch in a tube.
The FE model using reduced vector potential formulation was used to conduct a
systematic study of the effect of increasing sensor tilt on the defect signal. Three layer
Aluminum sample, each layer of 1 mm thickness, with rivets and 3rd layer subsurface
defect was modeled. Two radial defects of 3mm length and 0.1mm width and two
different depths of 20%, and 30% wall thickness were modeled with simulation
frequency of 2kHz. The sensor was moved in a linear scan across the rivet and defect
and the GMR signals measuring the normal and tangential components were generated.
In order to separate the effect of sensor tilt and defect, four different cases were studied,
namely, i) no crack ; no sensortilt, ii) no crack ; with sensortilt, iii) crack; no sensor
tilt and iv) crack ; with sensortilt. The tilt parameter was varied from 00 to 50 for two
different defects.
18 G. Yang et al. / Modeling and Signal Processing Sensor Tilt in Eddy CurrentGMR Inspection
The Figure 3 shows the geometry of sensor tilt, where the sensor measures both
normal and tangential components in contrast to measuring only normal component
when there is no tilt. Hence, if two components of the magnetic field were measured,
the magnitude of the tangential component will contain information about sensor tilt
and the normal component will contain defect information. Both components of the
magnetic fields were measured and studied. The signals ( Bz and Bx components) due
0
to 20% depth subsurface defect with 1 sensortilt are shown in Figure 4. The real part
of Bx component is seen to be sensitive to defect but insensitive to sensor tilt, whereas
the imaginary part of Bz component is seen to be more sensitive to sensor tilt. This
offers the potential of detecting defects in the presence of sensor tilt using two
component signals.
As presented in the Figure 4, the real part of Bx component and imaginary part of
Bz component carry discriminatory information about presence of defect and sensor tilt.
In particular, the peak magnitude and slope of peak to peak variation of these signal
components have been selected as useful features for classification. The peak
magnitude of real part of Bx component is extracted as feature F1 and the slope of peak
to peak variation of imaginary part of Bz component represents feature F2. These
features are illustrated in Figure 5. The two defined features were calculated from the
two defect signals with 10 sensor tilt shown in Figure 4. The features are plotted in the
feature space in Figure 6. It is seen that the 4 cases considered ((i) no crack ; no sensor
tilt, ii) no crack ; with sensortilt, iii) crack; no sensortilt and iv) crack ; with sensor
tilt ) are clearly separated in the feature space. Similarly the signals from the two
defects (depth 20%) and (depth 30%) with 30 sensor tilt were calculated using the
simulation model. The features from these signals are calculated and plotted in Figure 7.
It is seen that as the sensor tilt increases, the 4 cases described in section 2.1 are still
distinguishable but the separation is lesser that what seen in Figure 6.
G. Yang et al. / Modeling and Signal Processing Sensor Tilt in Eddy CurrentGMR Inspection 19
Figure 6. Features derived from signals in Figure 4 plotted in the two dimensional feature space.
20 G. Yang et al. / Modeling and Signal Processing Sensor Tilt in Eddy CurrentGMR Inspection
Figure 7. Features derived from simulated signals with 3 0 sensor tilt , plotted in the two dimensional feature
space.
3. Conclusion
An efficient finite element model using reduced magnetic vector potential formulation
is presented. The model was validated and used for studying the effect of varying
degrees of sensor tilt on ECGMR signals. Features in GMR output signals were
extracted and their ability to distinguish defect signals from sensortilt signals were
demonstrated for two different tilt angle values. More extensive validation of the
approach on experimental signals is under way. A more powerful approach involving
invariance transformation using independent component analysis (ICA) is also being
investigated for transforming the measured signal so that the transformed signal is
invariant to sensor tilt. A sensor system measuring two components of the magnetic
field will be built to for experimental validation of the approach.
References
[1] Aging of U.S. Air Force Aircraft, National Research Council Final Report, Publication NMAB 488 2,
National Academy Press, Washington, D.C., 1997
[2] N. V. Nair, V. R. Melapudi, H. R. Jiminez, X. Liu, Y. Deng, Z. Zeng, L. Udpa, T. J. Moran, and S. S.
Udpa, "A GMR based eddy current system for NDE of aircraft structures," IEEE Transactions on
Magnetics, vol. 42, no. 10, 2006, pp. 3312 3314
[3] J. C. Aldrin, J. S. Knopp, "Method for crack characterization with noise invariance for eddy current
inspection of fastener sites, " Review of Progress in Quantitative Nondestructive Evaluation, vol. 25,
2006, pp. 315 322
[4] X. Liu, Y. Deng, Z. Zeng, L. Udpa, and J. S. Knopp, "Model based inversion technique of GMR signal
using element free Galerkin method," Proceedings of the Annual Review of Progress in Applied
Computational Electromagnetics, 2008, pp. 221 226
G. Yang et al. / Modeling and Signal Processing Sensor Tilt in Eddy CurrentGMR Inspection 21
[5] Z. Zeng, X. Liu, Y. Deng, and L. Udpa, "Reduced magnetic vector potential and electric scalar potential
formulation for eddy current modeling," Przeglad Elektrotechniczny, vol. 83, no. 6, 2007, pp. 35 37
[6] Z. Zeng, L. Udpa, S. S. Udpa, and M. S. C. Chan, "Reduced magnetic vector potential formulation in the
finite element analysis of eddy current nondestructive testing," IEEE Transactions on Magnetics, vol.
45, no. 3, 2009, pp. 964 967
22 Electromagnetic Nondestructive Evaluation (XIV)
T. Chady et al. (Eds.)
IOS Press, 2011
2011 The authors and IOS Press. All rights reserved.
doi:10.3233/978160750750522
Abstract. A non iterative MUSIC type imaging algorithm for time harmonic
eddy current testing of a damaged conductive plate is considered. The method is
based (i) on a rst order asymptotic formulation of the secondary magnetic eld
observed outside the plate and induced by volumetric defects within it, which is
valid when those are small enough with respect to the skin depth at the frequency of
operation, and (ii) on the singular value decomposition of the multistatic response
matrix that can be collected using an appropriate source and sensor array. Herein,
the asymptotic eld formulation is validated by comparing the results with those
provided by means of the CIVA platform in a realistic case. Further, the singular
value pattern and MUSIC images are investigated, for a small spherical void placed
at different locations inside a plate. A Gaussian distributed noise is added to the
asymptotic data during the initial investigation, to alleviate inverse crime at least
in part (more complex arrangements of source/sensor arrays and defects will be
considered with external data in a later stage). Yet, these preliminary results as ex
hibited indicate the potential of the method to locate small defects in conductive
materials in a typical eddy current testing conguration
Keywords. Eddy current testing, non iterative inversion, MUSIC type algorithm,
Detection of small defects
Introduction
1. Sourcesensor arrangement
Figure 1. The arrangement for the eddy current MUSIC type algorithm investigation is sketched: a 7 7
(or 3 5 inside the dark rectangle) bobbin array above a non magnetic conducting plate affected by a small
spherical defect (with radius aj ) inside the search domain somewhere under the array.
(n)
coils). Similarly, the primary electric eld E0 (xj ) inside the search region may be com
puted by using the appropriate dyadic Greens function Gem t (xj , r) of the conguration.
The small size of the defect keeps the asymptotic formulation of the secondary magnetic
eld valid, as is seen next. Otherwise, it could be argued that the overall dimensions of
the array are not that large (side of 3s ); yet number and location of coils as hypothesized
remain realistic in view of the technology whilst most of the anomalous eld due to the
assumed defect under it appears to be collected (in practice also an array will be moved
step by step in either direction to cover a wider zone).
2. Mathematical formulation
The MUSIC algorithm herein is based upon a rstorder asymptotic formulation of the
secondary magnetic eld outside the plate induced by volumetric defects within it, which
is valid when those are small enough with respect to the skindepth at the frequency of
operation, and on the Singular Value Decomposition (SVD) of the socalled Multistatic
Response (MSR) matrix collected using, e.g., the bobbin array.The method is briey
described below for the particular case of ECT. A detailed description of the MUSIC
algorithm in the smalldefect framework for the full Maxwells equations is found for the
general case in [4], with [5] as an illustration in a stratiedspace case similar with ours
save made in the propagative (microwave) regime here only diffusion occurs.
Considering P small spherical voids centered at xj and with radii aj s , the asymp
(n)
totic formulation of the secondary magnetic eld Hs (r) may be formed as
T. Henriksson et al. / NonIterative MUSICType Algorithm 25
P
(n) j (n)
Hs(n) (r) =H (n)
(r) H0 (r) = i2 Gme
r (r, xj ) Mj , Vj E0 (xj ) ,
j=1
2
(1)
where n is the transmitter index (centered at rn ). Letting I be the identity tensor, the
polarization tensor Mj for such void spherical defects (j = 0 ) is given by
j 3(j 2 ) 4a3j
Mj , Vj = Vj I, where Vj  = . (2)
2 22 + j 3
The Gmer (r, xj ) term is the dyadic Greens function between the defect and the
receiving bobbin array; it satises the following differential equation
Gme 2 me
r (r, xj ) k Gr (r, xj ) = I(r xj ), (3)
considering the defect as an electric impulse current, with usual magnetictype boundary
conditions, k freespace (air) wavenumber.
Since the radius of each bobbin array element is bn s the dyadic Greens func
tion Gem
t (xj , rn ) between the n transmitting coil of the array placed at rn and the
th
(n)
defect can be used to compute the primary magnetic eld E0 (xj ) as
(n)
E0 (xj ) = Gem
t (xj , rn ) z, (4)
where the differential equation satised for a magnetic impulse current source is
Gem 2 em
t (xj , rn ) k Gt (xj , rn ) = I(xj rn ), (5)
From the secondary magnetic eld the Multistatic Response matrix A is given as
P
A= G me em
r (xj ) Mj G t (xj ), (6)
j=1
(7)
G em 2 em em
t (xj ) = k [Gt (xj , r1 ) z, . . . , Gt (xj , rN ) z] .
t
M and N are the numbers of receiving and transmitting array elements, respectively
(they are not necessarily equal), t denoting the transpose. Since the MSR matrix has the
dimension N M is it possible to apply the Singular Value Decomposition (SVD)
26 T. Henriksson et al. / NonIterative MUSICType Algorithm
A = UV , (8)
Pr = (I Us Us ), (9)
Pt = (I Vs Vs ). (10)
1 1
W(x) = + , (11)
Pr (G me
r (x) a)2 Pt (G em
t (x) a)2
where a = (1, 1, 1) [4]. Other values of a could be considered, taking different eld
components into account, how to achieve the best choice of a needing further analysis.
3. Results
The results presented next validate the framework of analysis considering only one spher
ical void inside the plate in account with the material shown in section 1.
letting In and Im the currents in the emitting coil n and the receiving coil m (consid
(m)
ered as a ctitious emitter), E0 (r) the primary eld due to the coil m and Jj (r) the
volumetric eddycurrent in Vj due to the source n [2]. Since the radius of the bobbin
element bn s , the zpolarized magnetic eld from the asymptotic
formulation should
(n)
be equivalent to the CIVAcomputed impedance variation, Hs (rm ) z Znm .
To verify this, a smaller 3 5 bobbin array (depicted in Figure 1) is used with the void
(a = s /10) at the centre of the search region (s /2 depth). The spherical void is mod
eled as a cube discretized by 5 5 5 cells in the CIVA model, which gives a RMS error
less than 0.2% on the computed results, compared to using a higher number of cells.
T. Henriksson et al. / NonIterative MUSICType Algorithm 27
The comparison in magnitude between the zpolarized magnetic eld from the
asymptotic formulation and the impedance variation from CIVA is illustrated in Figure 2,
6 different coils in the bobbin array transmitting (specied as bobbin 1 3 and 8 10 in
Figure 1). The asymptotic results () t very well with the CIVA results () in all cases
except when the transmitting coil is right over the void (Figure 2 (d)). In this case indeed,
(8)
the primary eld E0 (x1 ) is very small at the center of the defect. Since the asymptotic
formulations are only using this central value, the result is quite different from the one
yielded by carrying out the volume integral. However, except this peculiar point, it ap
pears that the asymptotic formulation and the more bruteforce approach via a CIVA tool
provide very close results.
Once the asymptotic formulation has been validated, it can thus be used to compute the
MSR matrix A for the 7 7 bobbin array. To minimize the inverse crime a Gaussian
distributed complex noise is added onto the data (SNR = 10 dB) as
Anoise = A + , (13)
The same void is to be retrieved at two different depths (s /4 and 3s /4). In Fig
ure 3 both the singular value distribution and the MUSIC image are depicted, for both
positions. As expected, the void is associated with two nonzero singular values (much
larger than all others, which should ideally be zero, and for all what matters very small
compared to the two main ones), where the amplitudes are decreasing with the depth
of the defect. However in this investigation the SNR of the data has been kept constant
between both positions, which might not be so realistic.
The MUSIC images are isosurface plots of W(x) with an isolevel of 70%. The
position of the void appears well retrieved for both locations inside the search region,
which is also the case (examples not shown) everywhere inside the search region. How
ever, with increasing depth the reconstruction along the zaxis gets more blurred (less
powerful signal is acquired, and the aperture angle under which the defect is seen is
getting smaller and smaller).
4 4
x 10 x 10
1 1
0.8 0.8
0.6 0.6
0.4 0.4
0.2 0.2
0 0
2 4 6 8 10 12 14 2 4 6 8 10 12 14
(a) (b)
4 4
x 10 x 10
1 1
0.8 0.8
0.6 0.6
0.4 0.4
0.2 0.2
0 0
2 4 6 8 10 12 14 2 4 6 8 10 12 14
(c) (d)
4 4
x 10 x 10
1 1
0.8 0.8
0.6 0.6
0.4 0.4
0.2 0.2
0 0
2 4 6 8 10 12 14 2 4 6 8 10 12 14
(e) (f)
Figure 2. Comparison in magnitude between the variation of impedance computed via CIVA ( ) and the
z polarized magnetic eld from the asymptotic formulation with a computed primary eld E0 using Gem t (),
along the 15 (3 5) receiving coils when transmitting from (a) coil 1, (b) coil 2, (c) coil 3, (d) coil 8, (e) coil
9, (f) coil 10.
formulation using a method of moments. Preliminary results for a single small spherical
void have been presented, as illustration of the good potential of the method. The position
of the void is indeed successfully retrieved along all three axes, with high accuracy even
in presence of strong noise, whatever its location within the plate.
Future work should involve more complex arrangements of the bobbin array and
several small defects, and also multifrequency data. The inverse crime should be fur
ther minimized by constructing the MSR matrix from CIVA data. More complex defects
T. Henriksson et al. / NonIterative MUSICType Algorithm 29
Figure 3. Singular value distribution and images for two positions of the void (a s /10), with the 7 7
bobbin array setup. (left) at (0.3, 0, 0.3)s and (right) at (0.3, 0.3, 0.3)s . The exact voids are displayed
as transparent spheres as well.
should be investigated in order to see whether their size can be approached to an ex
tent via a determination of their polarization tensor, main emphasis being however on
the investigation of cracklike defects, for which the theoretical machinery might still be
incomplete (which does not preclude useful numerical experimentation).
References
[1] D. Dos Reis, M. Lambert, and D. Lesselier. Eddy current evaluation of three dimensional defects in a
metal plate. Inverse Problems, 18:1857 1871, 2002.
[2] J. F. P. J. Abascal, M. Lambert, D. Lesselier, and O. Dorn. 3 D eddy current imaging of metal tubes by
gradient based, controlled evolution of level sets. IEEE Trans. Magn., 44:4721 4729, 2008.
[3] S. Bilicz, M. Lambert, and S. Gyimthy. Kriging based generation of optimal databases as forward and
inverse surrogate models. Inverse Problems, 26(7):074012, July 2010.
[4] H. Ammari, E. Iakovleva, D. Lesselier, and G. Perrusson. Music type electromagnetic imaging of a
collection of small three dimensional inclusions. SIAM J. Sci. Comput., 29:674 709, 2007.
[5] S. Gdoura, D. Lesselier, P. C. Chaumet, and G. Perrusson. Imaging of a small dielectic sphere buried in
a half space. ESAIM:Proceedings, 26:123 134, 2009.
[6] CIVA: State of the art simulation platform for NDE. http://www civa.cea.fr.
30 Electromagnetic Nondestructive Evaluation (XIV)
T. Chady et al. (Eds.)
IOS Press, 2011
2011 The authors and IOS Press. All rights reserved.
doi:10.3233/978160750750530
Introduction
1
Corresponding author. Email address: chlewi@zut.edu.pl
W. Chlewicki et al. / Identication of Defects in 3D Space Using Computer Radiography System 31
1. System description
Projection data have been acquired using the system presented in Figure 1. The system
consists of a Xray source CP1201, a digital detector ShadoBox1024, a manual
positioning system, a frame grabber and a PC workstation. In the near future the
manual positioning system will be replaced with motorized one.
The acquisition of radiographic projections has been performed along a linear
trajectory. The Xray source is moved above the object whose position is fixed in space.
Such an image acquisition scheme is an equivalent to limited angle tomography which
was shown by Zhou et al. [1] and confirmed by Gondrom et al. [2]. Their systems,
however, are characterized by a linear translation of the object instead of a motion of
the detector like in our system.
N
pi winvn( k )
v(jk +1) = v (jk ) + n =1
N
wij
w
n =1
2
in
(1)
The root mean square (RMS) contrast has been used to calculate the contrast of the
Region of Interest (ROI) containing the inclusion for all slices. It is defined as the
standard deviation of the pixel intensities. Maximum of the Crms stands for the inclusion
being in focus:
W. Chlewicki et al. / Identication of Defects in 3D Space Using Computer Radiography System 33
N 1 M 1
( I (i , j ) I (i , j ) )
1 2
C RMS = (2)
MN i=0 j =0
where: Crms is the root mean square contrast function, M and N stand for width and
height of the ROI respectively and I is a map of pixel's intensity of the ROI. The RMS
contrast does not depend upon the spatial frequency content or the spatial distribution
of contrast in the image.
2. Experiments
2.1. Phantom
Seven projections of size (1024x1024) have been acquired in the manner described in
the above. The voltage of the Xray tube has been set to 80kV, the current being
equated to 1 mA. The integration time of the detector has been set to 4s. During the
acquisition process the source is moved along the linear trajectory, the range of linear
translation of the detector and the object being (28cm, 28cm). In such a way, 7
radiographic projections have been taken at different source positions.
The acquired projection has been input for the reconstruction of the 1024x1024x1024
volume. Some of the significant slices for depth identification are displayed in Figure 3.
Left hand side inclusion is in focus in slice number 271 whereas in the other
crossections it appears blurred. The right hand inclusion is "in focus" in slice number
146. This proves the abilities of the system to localize the depth of the occurrence of
certain objects within the investigated volume. In Figure 4 the process of depth
identification is visually presented.
Figure 3. Crossections of number 101, 121, 146, 201, 271 and 406 (from left to right) within the
reconstructed volume. The maximum of sharpness for each inclusion is clearly visible making it possible to
localize the depth of each inclusion.
34 W. Chlewicki et al. / Identication of Defects in 3D Space Using Computer Radiography System
3. Conclusions
The ability of the system to localize the depth of the occurrence of objects within the
investigated volume has been experimentally confirmed. It can be concluded that the
diagnostic value of digital radiography systems could be significantly improved via
only minor extensions of the systems themselves. Further work is required in order to
assess the impact of changes of the linear translation range on the quality of the pseudo
reconstructed volume [5].
Acknowledgments
The work was conducted in the framework of the research project Identification of
structural heterogeneities in mechanical systems, supported by the Polish Ministry of
Science and Higher Education. Grant no. 1141/B/T02/2009/36 (20092012).
W. Chlewicki et al. / Identication of Defects in 3D Space Using Computer Radiography System 35
References
[1] J. Zhou, M. Maisl, H. Reiter, W. Arnold, Computed laminography for material testing. Appl. Phys. Lett.
68, 3500.
[2] Gondrom S., Zhou J., Maisl M., Reiter H., Kroning M., Arnold W.: Xray computed laminography
approach of computed tomography for applications with limited access. Nuclear Engineering and Design
190, 1999, pp. 141147.
[3] Kak A., Slaney M.: Principles of Computerized Tomographic Imaging. Philadelphia: SIAM, 2001.
[4] A. H. Andersen, Algebraic Reconstruction in CT from Limited Views, IEEE Trans. Med. Img., vol. 8,
no.1, 1989, pp. 5055.
[5] B. Li, G. B. Avinash, R. Uppaluri, J. W. Eberhard, B.E.H. Claus, The impact of acquisition angular
range on the zresolution of *radiographic tomosynthesis, International Congres Series 1268, 2004, pp.
1318.
36 Electromagnetic Nondestructive Evaluation (XIV)
T. Chady et al. (Eds.)
IOS Press, 2011
2011 The authors and IOS Press. All rights reserved.
doi:10.3233/978160750750536
Introduction
Fastener holes in aircraft act as stress concentrators and are common sites for fatigue
cracking. If left undetected, the growth of such fatigue cracks can lead to catastrophic
structural failure. Considerable effort has been devoted to the development of
nondestructive evaluation (NDE) techniques which can detect cracks in fastener holes
with high reliability. Along with ultrasonic testing, eddycurrent NDE is foremost
among these techniques.
The theory underpinning eddycurrent NDE for cracks in cylindrical geometries
such as fastener holes continues to be developed, including recent approaches based on
boundary element methods [1]. However, unlike planar geometries, for which there are
highly effective analytical approximations valid in the limit of small electromagnetic
skin depth [24], there are currently no similar highly developed thinskin models for
cylindrical geometries which can be used to complement finiteelement and boundary
element methods.
In this paper, a thinskin model for eddycurrent NDE of cracks in a fastener hole
is presented based on an extension of the Lewis extended surface impedance boundary
conditions [4] to a cylindrical geometry. In Section 1, a closedform expression for the
change in coil impedance Z is presented for the simple case of a coaxial bobbin coil in
a borehole containing a long crack with uniform depth and opening. The coil is
assumed to be sufficiently remote from the ends of the hole that the hole can be treated
1
Corresponding Author. S K Burke, Maritime Platforms Division, DSTO, 506 Lorimer St., Fishermans
Bend, Vic 3207 Australia. Email: steve.burke@defence.gov.au
S.K. Burke / A ThinSkin Model for EddyCurrent NDE of Cracks in a Borehole 37
as a borehole, i.e. infinite in length. The theoretical predictions are compared with
experimental results in Sections 23 and the significance of the work is discussed.
1. Theory
H = (1)
2 = 0. (2)
In the limit of small skin depth ( = [2/()] ) the normal component Hn of the free
space magnetic field at the surface of an uncracked conductor satisfies the surface
impedance boundary condition [6]
x y
Figure 1. Eddycurrent induction by a coaxial source coil in a borehole containing a long rectangular slot.
38 S.K. Burke / A ThinSkin Model for EddyCurrent NDE of Cracks in a Borehole
H n
= ( n) H n (3)
n
where n is the outward normal to the conductor surface and = (1+i)/(r). From
Eqs.(1) and (3) the surface impedance boundary condition for an uncracked borehole
using the coordinate system shown in Fig.1 is
2 1
= ( + ) , r = a. (4)
r 2
r r
Clearly Eq.(4) is not valid in the presence of a crack and must be modified. By
considering the flux leaving the mouth of the crack, Lewis [4] derived an extended
surface impedance boundary condition for a planar conductor containing a crack.
Extending this approach to cylindrical geometries, the extended surface impedance
boundary condition for the cracked borehole in Fig.1 can be written as,
2 ( e ) 1 ( e ) 2
+ ( + ) = H n( i ) ( ), r = a, (5)
r 2 r r a
and links the fields within the crack (the interior fields, denoted by the superscript i)
and the fields within the borehole (the exterior fields denoted by the superscript e). In
deriving Eq.(5), and in the following sections, it is initially assumed that the crack has
zero opening. The effect of crack opening is introduced in Section 1.4.
( e ) = ( s ) + ( r ) , a r r2 , (6)
of the known source potential (s) of the isolated coil in free space [7] and the potential
(r) arising from induced currents in the cracked conductor.
A representation for the potential (r) can be constructed by solving Eq.(2) in
cylindrical coordinates using the method of separation of variables. If the origin of the
polar coordinate system is chosen to lie at the center of the coil, then
( r ) (r , , z ) = sin( z ) ( r ) (r , ; ) d , r a, (7)
0
( r ) (r , ; ) = c0 I 0 ( r ) + 2 cn I n ( r ) cos(n ), (8)
n =1
where In are modified Bessel functions and the carat denotes the Fourier sine transform
S.K. Burke / A ThinSkin Model for EddyCurrent NDE of Cracks in a Borehole 39
f (r , ; ) = sin( z ) f (r , , z ) dz . (9)
0
Note also that (r) is an odd function of z and an even function of . The unknown
coefficients cn = cn() are to be determined by application of the extended surface
impedance boundary condition Eq.(5) once a suitable representation for the potential
within the crack has been derived.
It is convenient to describe the fields within the crack using the Cartesian coordinate
system in Fig.1. In the thinskin limit (i) satisfies the 2D Laplace equation [4]
2 (i ) 2 (i )
+ = 0, a x a + b. (10)
x 2 z 2
( i ) ( e )
= , y = 0, x = a. (11)
z z
( i )
= 0, x = a + b. (12)
x
The solution of the boundary value problem defined by Eqs.(10) (12) can be solved
through the use of Fourier transforms [9]. The resulting Fourier sine transform of the
interior potential takes the form
1.3. Solution
Closed form solutions for can now be derived using the representations obtained in
sections 1.11.2. The extended surface impedance boundary condition Eq.(5) linking
the exterior and interior potentials can be written in terms of sine Fourier transforms as
( e ) 2 ( e )
a 2 + 2 a 2 ( e ) = 2aH n(i ) ( ), r = a, (14)
r 2
where the term involving the second derivative with respect to r has been eliminated
using Eq.(2). From Eqs. (1), (9) and (13) it follows that
40 S.K. Burke / A ThinSkin Model for EddyCurrent NDE of Cracks in a Borehole
Substituting Eqs.(6), (8) and (15) into Eq.(14) and simplifying gives the condition
s + c0 0 + cn n cos(n ) =
n =1
(16)
2 a tanh( b) [ ( s ) + c0 I 0 ( a) + 2 cn n cos(n )] ( ),
n =1
( s ) (r , 0; )
s = 2 a 2 ( s ) (a, 0; ) a 2 ,
r r =a (18)
n = a I n ( a ) (n + a ) I n ( a).
2 2 2 2
Note that in Eq.(17), nd = N/[2h (r2 r1)] denotes the coil turn density and
r2
I ( r1 , r2 ) = x I1 ( x) dx. (19)
r1
It is useful first to consider the solution to Eq.(16) in the absence of the crack
b = 0. In this case c0 = s/0 and cn = 0 for n 0 as the fields are axisymmetric. Hence,
from Eqs.(6) and (8), the Fourier sine transform of the scalar potential on the surface of
the uncracked borehole in the thinskin limit is
It can be shown that the thin skin expression Eq.(20) agrees with the exact expression
[7] for an uncracked borehole to order 1/ as required.
In the presence of the crack, Eq.(16) can be solved for the unknown coefficients cn
by multiplying both sides of the equation by cos(m) where m = 0,1,2 , integrating
over the interval and applying the cosine orthogonality conditions. Solving
the resultant simultaneous equations and using Eq.(20) gives the final coefficients
s a tanh( b)
c0 = + ( u ) (a, 0; ),
0 0 [1 + tanh( b) G ( )] (21)
cn = ( s + c0 0 ) / n , n 1,
a
G=
I 0 ( a ) / 0 + 2
n =1
I n ( a ) / n ,
(22)
1
= ,
n = q I n ( ) / I n ( ) + n 2 + 2
tanh( b) G ( )
( e ) (a, 0; ) = ( u ) (a, 0; ) [1 ]. (23)
1 + tanh( b) G ( )
Combining Eq.(1) and Eq.(23) and using the properties of the Fourier transform, the
tangential component of magnetic field at the mouth of the crack can be written
2 cos( z ) tanh( b) G ( ) (u )
H z ( a, 0, z ) = H z( u ) (a, 0, z ) (a, 0; ) d , (24)
0 1 + tanh( b) G ( )
where the first term is the tangential field for an uncracked borehole and the second
term is the field scattered by the crack. Note for simple thinskin models [5] the
scattering term in Eq.(24) is omitted: an approach termed the Born approximation.
An expression for Z in the thinskin limit due to a crack with finite opening u
can now be obtained using the knowledge of the fields obtained in the proceeding
subsections. The derivation closely parallels that in previous work for the related case
of an infinitely long rectangular slot in a halfspace [2,3,5,9] and relies on the
application of surface impedance boundary conditions to the general Auld Z formula
[5] for a defect in a conductor. The higherorder Kahn effect contributions to Z are
also included by assuming that the curved corner fields take the same form [2] as for a
90 edge. Following this approach, it can be shown that
40 g
I 2 Z =
0 1 + tanh( b) G ( )
[ ( u ) (a, 0; )] 2 d , (25)
g = i u g f + (1 + i) r ( g f 12 u g 0 ) + 12 r 2 g k ,
42 S.K. Burke / A ThinSkin Model for EddyCurrent NDE of Cracks in a Borehole
g f = tanh( b),
g 0 = 2 [1 sech( b)], (26)
g k = 4 (1 + ) / + sech( b).
2 2
2. Experiment
Experiments were performed to test the validity of the theory presented in Section 1.
The experiments were carried out using a series of aircored coils and a set of 10 mm
diameter bolthole specimens containing simulated cracks. The coil parameters are
given in Table 1.
The test specimens were prepared by drilling and reaming a series of through holes
in thick 2024 Al alloy plate. The electrical resistivity of the plate was 6.00.05 cm
and the plate thickness (hole length) was 24.6 mm. The simulated cracks were
introduced into the reamed holes by electrodischarge machining a slot using a thin
wire electrode. The specimen parameters are given in Table 2.
Coil impedance measurements were made over the frequency range 5kHz400kHz
using a low frequency impedance analyser (HP4192A). The change in coil impedance
due to the slot Z was determined by measuring the difference in impedance when the
coil is centered in a cracked bolthole and when the coil is centered in an uncracked
reference hole with the same diameter. The impedance data were corrected for the
effects of stray capacitance using a procedure described previously [8].
In Fig. 2, the experimental values of Z are compared with the theoretical predictions
obtained using Eq. (25). The results are plotted as a function of skin depth (rather than
frequency) to illustrate the structure of the data clearly. Also shown in Fig. 2 are the
predictions of the Born approximation, for which the scattering term in Eq.(25) is
neglected by setting G()=0.
As shown in Fig.2a, the calculated values of the coil inductance change L are
in good agreement with experiment for small slot depths (b 6 mm ) but less so for
larger depths where the scattering from the crack is greater. Nevertheless, even for
these deeper slots (up to b = 10 mm examined here), good agreement is obtained at the
smallest skin depths within the range. In all cases, the model predictions are superior to
those of the Born approximation for which the scattering is neglected entirely.
S.K. Burke / A ThinSkin Model for EddyCurrent NDE of Cracks in a Borehole 43
12
45
Experiment
Calculated b=10 mm
40 10
Born Experiment
35 Calculated
Inductance Change (H)
Resistance Change ()
8 Born
30
b (mm) 4 mm
25 6
10
20
8 4
15 6 1.3 mm
4 2
10
3 (a) (b)
5 1.3 0
0
0.0 0.2 0.4 0.6 0.8 1.0 1.2 1.4 0.0 0.2 0.4 0.6 0.8 1.0 1.2 1.4
Skin Depth (mm) Skin Depth (mm)
Figure 2. Change in (a) coil inductance and (b) coil resistance due to the slot as a function of skin depth.
The corresponding results for the coil resistance change are shown in Fig.2b. The
agreement between the model predictions and experiment is good for all slots at small
skin depth but becomes markedly poorer with increasing skin depth. This is likely due
to the assumption in Eq.(25) that the planar Kahn effect term (which contributes to
R to a greater extent that L) can be adopted for a cylindrical geometry. The model
predictions for R overall are again superior to those based on the Born approximation.
Although the experiments were performed using Al alloys, the model is applicable
to magnetic steels for which r 1. Future experimental work on magnetic materials
would be valuable, particularly as this provides access to significantly smaller skin
depths. The analysis can also be extended to more complex defect types, such as a pair
of longitudinal cracks located on opposite sides of the borehole, and to more complex
coil configurations, e.g. where the coil axis is normal to the borehole surface.
The present model, in which the scattering is incorporated using extended surface
impedance boundary conditions, is superior to the Born approximation and represents
the most complete thinskin theory for eddycurrent NDE of cracks in boreholes to date.
It does not, however, achieve the typical accuracy obtained for planar geometries for
which the scattering from the crack has been treated more completely by Harfield and
Bowler [23]. Further theoretical work is clearly required but the extension of the
HarfieldBowler approach to cylindrical geometries does not appear straightforward.
References
[1] J.R. Bowler and T.P. Theodoulidis, IEEE. Trans. Magn. 45 (2009), 10121015.
[2] N. Harfield and J.R. Bowler, J.Appl.Phys. 82: (1997), 45904603.
[3] J. R. Bowler and N. Harfield, IEEE. Trans. Magn. 34 (1998) 515523.
[4] A.M. Lewis, J.Phys.D: Appl. Phys. 25 (1992), 31926.
[5] B.A. Auld and J.C. Moulder, J.Nondestr.Eval. 18 (1999), 336.
[6] A. Nethe, R. Quast and H. Stahlmann, IEEE.Trans.Magn. 34 (1998), 33313334.
[7] C.V. Dodd, W.E. Deeds and J. W. Luquire, Int.J.Nondestr.Test. 1 (1969), 233293.
[8] D.J. Harrison, L.D. Jones and S.K. Burke, J.Nondestr.Eval. 15 (1996), 2134.
[9] S.K.Burke, J.Appl.Phys., 76 (1994), 30723080.
44 Electromagnetic Nondestructive Evaluation (XIV)
T. Chady et al. (Eds.)
IOS Press, 2011
2011 The authors and IOS Press. All rights reserved.
doi:10.3233/978160750750544
Introduction
When applying an electrical current flow within a given material using induction or
injection techniques, variation of its spatial electrical conductivity distribution will
impact the distribution of the electrical currents and their corresponding magnetic fields
within and outside the tested sample, respectively. Significant conductivity changes can
be observed in case of defect regions within the homogeneous material structure such
as cracks, voids, shrink holes and inclusions. Among the common electromagnetic
Nondestructive Testing methods Magnetic Particle Testing (MPT) has the lions share
in routine testing applications. In comparison to Eddy Current Testing (ECT) MPT can
be regarded as a pure surface inspection technique, whereas ECT provides a view at a
significant depth of several mm below the samples surface, when using adapted testing
equipment.
The first use of MR sensors and other sensitive field sensors in NDT, like the
SQUID [13], in the mid90s opened the door for a new NDT research branch. An
AMR sensor in ECTprobes was already investigated at the Federal Institute for
Materials Research and Testing in 1995 [4]. It was the advent of a new magnetic micro
systems generation after the discovery of the Giant Magneto Resistance effect (GMR)
in 1988 that paved the way for new detection concepts in ECT as well as in flux
leakage testing. This boom in GMRresearch was also supported by the Nobel Prize in
Physics in 2007 for the discovery of the GMR effect by Peter Grnberg and Albert
Fert. Currently GMR sensors are being increasingly proposed to detect small defects
M. Kreutzbruck et al. / Magnetic Response Field of Spherical Defects 45
with a remarkable signaltonoise ratio and high spatial resolution [58]. The results of
this work meet this new research approach, in which the magnetic flux density B is
detected directly by a field sensor instead of an induction coil which measures the
corresponding field derivative with respect to time B/t. There exist a number of
inclusionrelated NDT problems such as tantalum inclusions hidden in niobium plates
used for superconducting resonators or testing superconducting cables with their
complex structure of micro filaments. The latter also involves the estimate of their size
and position within the wires crosssection, which is vital for a reliable quality check.
The problem posed by inclusions also needs to be addressed when testing aluminum
laser welds involving spherical shaped and nonconducting pores representing a
limiting case of the inclusion problem.
A remarkable amount of theoretical work has been done in electromagnetic testing
by the Eddy Current Testing community. Several approaches were devised by Bowler
et al. to calculate cracks, voids and other inhomogeneities and their impact on a sensing
coil [9]. However, these works are mainly concerned with zero conductivity defects
and the changes in the detection coils impedance, and do not address the magnetic flux
density.
1. Field calculations
The most standard FEM codes for ECT problems use the so called [A, VA]
method, in which vector potential A and scalar potential V has to be found by solving
the differential equation system [19, 20] for each node in the FEM mesh. From both of
these quantities the electrical current density and the corresponding magnetic flux
density can then be calculated.
For the generation of a homogeneous current flow both an ultralow frequency
Eddy Current (EC) model and an injection model we used as well. The latter uses two
potential boundary conditions at the left and the right edge of the sample applying a
voltage and a corresponding homogeneous current flow. The EC method and the
injection method yield the same results in most cases. It turned out that the numerical
accuracy of the solver for alternating fields was somewhat higher than that of the dc
solver. For the calculation of the response field at great distances it was therefore
preferred to calculate using ultralow frequencies of MHz to simulate the dccase.
Figure 1. left. The surface mesh of a metal plate showing the surface current density induced by a flat
circular excitation coil. Right: Model of inclusions with different sizes (50 m 800 m in dia.), incorporated
at the center of the niobium plate.
46 M. Kreutzbruck et al. / Magnetic Response Field of Spherical Defects
calculated only the vertical component of the magnetic flux density Bz, which usually is
then detected by a field sensor. For inclusions showing a higher conductivity than that
of the matrix, the currents will be focused into the inclusion, resulting in higher current
densities at the front and back of the inclusion where the current flow direction is
parallel to the normal vector of the inclusion surface. This is shown in Fig. 2, left or the
case of a tantalum inclusion hidden in a niobium matrix.
For low conductivity inclusions the situation is the opposite. Here the current is
forced to flow around the inclusion and a minimum current density can be observed at
the front and back (abutting face) with respect to the direction of the homogeneous
current flow. The maximum of the distortion current strength occurs at the inclusion
host interface along the inclusions crosssection, which is aligned perpendicular to the
current direction and runs through the center of the inclusion. If it is assumed that the
currents run along the ydirection, this interface circle would pass through the
following points in Fig. 2, left: (0.4, 0, 0), (0, 0, 0.4), (0.4, 0, 0), (0, 0, 0.4). In the
Figure 2. left: Eddy current distortion of a 2 mm deep circular tantalum inclusion (0.8 mm in dia.) located in
a planar niobium sheet. Eddy current distortion in the x y plane. The current flows along the y axis. The
conductivity of tantalum is somewhat higher than that of niobium, leading to increased distortion currents at
the front side along the y axis at (0, 0.4, 0) and (0, 0.4, 0) and to reduced current densities at (0.4, 0, 0) and
( 0.4, 0, 0). Right: Current density j (jx+jy+jz)0 5 in the vicinity of the inclusion at the x/y plane at z 0 (cut
plane through the inclusions centre). Uniform current flow passes along x axis. Zero conductivity inclusion
(200 m in dia.) hosted in a Ti matrix ( 2.34 MS/m).
M. Kreutzbruck et al. / Magnetic Response Field of Spherical Defects 47
Figure 3. Current density j (jx+jy+jz)0 5 in the vicinity of the inclusion (200 m in dia.) along the x axis at
z 0 (line through the inclusions centre). Uniform current flow passes along x axis. Left: upper curve (blue,
dashed) represents Ti inclusion ( 2.34 MS/m) hosted in a Cu matrix ( 59.6 MS/m) and lower curve
(red, solid) represents air inclusion hosted in an Al matrix ( 37.7 MS/m). Right: upper curve (red)
represents Cu inclusion ( 59.6 MS/m) hosted in a Ti matrix ( 2.34 MS/m) and lower curve (blue)
represents air inclusion hosted in a Ti matrix.
We now take a look at the magnetic fields generated by the distortion currents and
quantitatively determine how different inclusion conductivity and different host
materials affect the current distribution and the corresponding magnetic field response.
As materials for host and inclusion, Cu, Ti, Al and air with the following conductivity
values have been used: titanium ( = 2,34 MS/m), niobium ( = 6,93 MS/m), and
aluminum ( = 37,7 MS/m).
5
I < I > M
2.5
Ta
B z [T]
0
2.5
5
Nbmatrix, 6.93 MS/m
7.5
Timatrix, 2.34 MS/m
10
3 4 5 6 7 8 9
10 10 10 10 10 10 10
conductivity inclusion [S/m]
Figure 4. left: Strength of the Magnetic field variation above the sample as a function of the inclusions
conductivity. Two different host conductivities were used (black: Ti, 2.34 MS/m, grey: Nb, 6.93 MS/m).
Inclusion diameter: 200m, Sensor to inclusion separation: 5 mm. Each inclusion is located 2 mm below the
surface. Solid lines represent the results of the fit function using Eq. 3. Right: Example of the field
distribution for low and high conductivity inclusions.
In Figure 4 the maximum field magnitude (Bz, peaktopeak amplitude) is shown
for different compositions as a function of inclusion conductivity. In the case of
inclusion conductivity smaller than that of the tested material, the induced currents
which otherwise would pass undisturbed through the volume of the inclusion are
perturbed and have to flow around the inclusion. The corresponding vertical
component of the magnetic field is displayed in Fig. 4, top right (the effect of low
inclusion conductivity is shown to the left from the zero point). The situation is
opposite when the inclusion conductivity is greater than that of the tested material.
Then the currents flow from the host into the inclusion, leading to a change in the
direction of circulation. The circular perturbation currents that once flowed clockwise
now flow anticlockwise (Fig. 4 right, on the right from the zero point and Fig. 4, right
bottom). Consequently, the sign of the magnetic field is changed from positive field
values to negative ones. When the conductivity of the inclusion and matrix are the
same, this will obviously result in a vanishing response field and no perturbation
currents will occur (zero point, Fig. 4, left).
One can observe that for very low inclusion conductivity the defect response does
not change significantly if the conductivity of the inclusion is further reduced. The
same saturation effect occurs for very high inclusion conductivity. By comparing the
field values for very low and very high defect conductivities one can observe an
anomaly, already seen when dealing with the current distribution. In the limit of high
M. Kreutzbruck et al. / Magnetic Response Field of Spherical Defects 49
inclusion conductivity one finds a field response exactly twice as high as that of an
inclusion with negligible conductivity. We define as the ratio between the inclusions
conductivity I and matrix conductivity M as follows:
I
(1)
M
The numerical data for any M in the limit at infinity and zero conductivity show
the following relation:
lim Bz
2 (2)
lim Bz
0
The response field Bz of an inclusion with excellent electrical transport properties,
say silver, in a low conductivity metal matrix will be higher than that of an air inclusion
(pore) located in the same matrix carrying the same homogenous current density. The
reason for these phenomena can be explained by the limited inclusion volume. The
defect volume is limited with respect to the much larger volume of the host. Assuming
that all currents are pushed out of the inclusion, it is not possible to increase the
response field by lowering the inclusion conductivity. It is therefore reasonable to
assume that for the virtually infinite volume of the matrix significantly more current
can be passed through the inclusion if its conductivity is high enough. Of course, there
also exists an optimum path for the currents flowing from the host into the inclusion
and back again. The simulation reveals that this geometrical condition causes a
response field which is a factor of 2 larger than that of the zero conductivity case. A fit
of the numerical data shows that
M 1
Bz ~ I (3)
I 2 M 2
When using the FEM results (dots in Fig. 4, left) we found the error of this
analytical expression (solid line in Fig. 4, left) to be below 2%. Note that these results
are independent of the inclusions size or the distance of the sensor above the samples
surface.
distances are at least twice the inclusions radius. Closer examination of the data shows
that the interface of the two regimes < 2 and = 2 is located exactly at the inclusion
host interface. The simulations also confirm that Bz as function of the inclusion radius r
scales exactly with a r law. This also can be seen in Fig. 5, left, where the magnetic
field response is distinctly increased in the presence of larger inclusions. Furthermore,
the graph shows that a larger inclusion diameter will shift the regime of < 2 towards
higher values. Independent of the inclusion size, the = 2 regime invariably is entered
at the interface of inclusion and host.
Figure 5. Fall off characteristics calculated for different inclusion diameters. Magnetic field variation
as a function of the sensor to inclusion separation (left) and ratio of Bz(z/2)/ Bz(z) as function of the
dipole to sensor separation (right). Inclusion with zero conductivity (air) located in a titanium matrix
( 2.34 MS/m).
Size dependence and spacing between sensor and inclusion are in good accordance
with the results of G. Sepulveda et al. [21]. They used spheroidal shaped cracks with
variable conductivity to analytically calculate the magnetic field that would be
measured by a magneto sensor. Due to the complexity of the problem, the solution is
given only for the dcproblem and is a firstorder approximation. The general solution
is proportional to the Legendre functions of the first kind and consists of odd Legendre
polynomials. One can use several boundary conditions to determine the coefficients,
but due to the limited number of boundary conditions, only first terms of the
polynomial expansion can be used. The spheroidal approach is hereby a kind of
approximation, which makes an analytical calculation of the current distribution
possible. However, when turning to the magnetic field, unfortunately no unique
mathematical expression can be derived, leading also to finite element approaches.
Thus no expression of the field strength for different inclusion conductivity was
presented.
Our FEM results are in good agreement with the current and field distribution of
G. Sepulveda, who also found that the magnitude of the far field distances much
further than the prolate spheroids diameter falls off as 1/R2 at any given direction.
Implementing the findings of sections 3 and 4 we now present a simple analytical
expression for varying conductivity in the inclusion and the matrix, which fits the
numerical data extremely well for all types of isotropic materials:
M. Kreutzbruck et al. / Magnetic Response Field of Spherical Defects 51
0 r3 1
Bz pp
j0 (4)
1.354 z2 2
To provide a broad data set, we calculated the peaktopeak magnetic flux density
Bz pp for a number of different inclusions. The sensortoinclusion separation was varied
in the regime ranging from z = 2 mm to z = 6 mm. The radius of the inclusions was
varied between 50 m and 400 m. Finally for the inclusion conductivity we used
values between 0 MS/m and 59.9 MS/m (Cu) and for the conducting host we used
values ranging from the conductivity of Ti to that of Cu. As a result, Bz pp varies
between 30 pT and 7 T, covering almost 6 orders of magnitude. Despite this high
dynamic field range, the constant k = 0/1.354 9.28 107 Tm/A only varies in a range
of a few percent (see Fig. 6).
FIGURE 6. Variation of k for different inclusion types. Left axis: absolute value of k. Right axis: Relative
error assuming k to be 9.28 10 7. Ti: 2.34 MS/m, Al: 37.7 MS/m, Cu: 59.9 MS/m. Sample 1 4 :
i 0 , M Ti , zi 3,5,4,5 mm, r in m; sample 6 9: i 0, M Al , zi 3,4,3,5 mm, r in m; sample 11
14: i 0, M Ti , zi 2,3,4,5 mm, r 50 m; sample 16 18: i 0, M Ti , zi 2,3,4 mm, r 100 m;
sample 20 22: i 0, M Ti , zi 3,4,6 mm, r 200 m; sample 24 26: i 0, M Ti , zi 2,4,5 mm, r
400 m; sample 28 32: i in MS/m, M Ti , zi 5 [mm], r 100 m; sample 33 36: i, M , zi 4,5,4,5 mm,
r 100 m.
5. Conclusion
addressed as to how these results can be transferred to alternative defect types like
porosities and cracks. We found in further investigations that the qualitative behaviour
of the magnetic field does not change significantly for more elongated flaws. Even for
pure rectangular cracks with high aspect ratio the magnetic field above the samples
surface can be interpreted as being generated by a dipolelike source with its 1/z
characteristics  a result also found by Sepulveda. However, for the absolute signal
strength we found a stronger influence of the crosssection and the current direction
compared to the influence of the crack width. This relation still has to be further
investigated. Meanwhile there exist a series of NDT activities, with which the proposed
analytical expression can be confirmed by experimental data [2,8]. A systematic
validation of these theoretical results will be a future task.
References
[1] M. v. Kreutzbruck, J. Trll, M. Mck, C. Heiden and Y. Zhang Trans. Appl. Supercond.7, 3279, IEEE,
1997.
[2] M. Mueck, C. Welzel, F. Gruhl, M. v Kreutzbruck, A. Farr, F. Schoelz, Physica, C 368 96 99 (2002).
[3] H. J. Krause and M. v. Kreutzbruck, Physica C 368 70 79 (2002).
[4] M. Stoppel, D. Filbert, H. M. Thomas, Research report at Technical University Berlin and Federal
Institute of Materials Research and Testing, May 1995.
[5] C. H. Smith, R. W. Schneider, T. Dogaru, S. T. Smith, Rev. Prog. in QNDE, 22, 419 426, (2003).
[6] B. Wincheski, M. Namkung, Rev. Prog. Quant. Nondestr. Eval. 509, 465 (2000).
[7] K. Allweins, M. v. Kreutzbruck, G. Lembke, J. Appl. Phys. 97, 10Q102 (2005).
[8] M. Kreutzbruck, H. Bernau, K. Allweins, TMTECHNISCHES MESSEN , 75 (9) p. 477 484 (2008).
[9] J. R. Bowler, J. Appl. Phys. 75 (12) 8128 (1994).
[10] J. R. Bowler, S.A. Jenkins, L.D.Sabbagh, and H.A. Sabbath, J. Appl. Phys. 70 (3) 1107 (1991).
[11] N. Harfield, Y.Yoshida and J. R. Bowler, J. Appl. Phys. 80 (7) 4090 (1996).
[12] A.P.Raiche , J.H. Coggon, Geophys J. R. Astron. Soc. 42, 1035 (1975).
[13] J. R. Bowler, J. Appl. Phys. 86 (11) 6494 (1999).
[14] S. M. Nair and J.H. Rose, J. Appl. Phys. 70 (4) 1924 (1991).
[15] C.T. Tai, Dyadic Geenss Functions in Electromagnetic Theory (Intex, Scranton, 1971).
[16] P.E. Wannamaker, G.W. Hohmann, and W.A. SanFilipo, Geophysics 49, 60 (1984).
[17] B. J. Roth, N. G. Sepulveda, and J.P Wikswo, Jr., J. Appl. Phys. 65 (1) 361 (1989).
[18] R.E. Beissner, J. Appl. Phys. 60 (1) 352 (1986).
[19] A. Kost, Numerische Methoden in der Berechnung elektromag. Felder, Springer Verlag (Berlin) (1994).
[20] S.C. Brenner, L.R. Scott, The Mathematical Theory of Finite Element Methods, Springer Verlag, 2002.
[21] N. G. Sepulveda, J. P. Wikswo, Jr., J. Appl. Phys. 79 (4) 2122 (1996).
Electromagnetic Nondestructive Evaluation (XIV) 53
T. Chady et al. (Eds.)
IOS Press, 2011
2011 The authors and IOS Press. All rights reserved.
doi:10.3233/978160750750553
Introduction
System health monitoring (SHM) and condition based maintenance (CBM) are of high
interest nowadays in order to sustain safety, reliability and quality of various processes.
Nondestructive evaluation plays accordingly the key role in SHM and CBM where not
only reliable detection but also precise estimation of dimensions of a detected anomaly
is required. Enhancing nondestructive evaluation methods is therefore very important
for accomplishing their challenging missions.
Different physical principles are utilised for the nondestructive inspection and
evaluation of materials. Eddy current testing (ECT) is one of the widely utilized
electromagnetic methods. It originates from the electromagnetic induction phenomena
and its principle underlies in the interaction of induced eddy currents with structure of
an examined body. Perturbations in eddy currents distribution due to a presence of an
anomaly are detected. Indisputable advantages of the method such as high inspection
speed, high sensitivity for surface breaking cracks, versatility account for gradually
1
Corresponding Author: Ladislav Janousek, Department of Electromagnetic and Biomedical
Engineering, Faculty of Electrical Engineering, University of Zilina, Univerzitna 1, 010 26 Zilina, Slovak
Republic; E mail: janousek@fel.uniza.sk
54 L. Janousek et al. / Decreasing Uncertainty in Size Estimation of Stress Corrosion Cracking
1. Numerical Model
y SUS316L
wc
x
lc
c
dc
= 1.4 MS/m
r= 1
10
The numerical calculations are performed using the edgeelement code. Three
parameters of the crack are varied in other to simulate various structures of real cracks.
The crack depth dc is changed from 1 mm up to 10 mm with a step of 1 mm, its width
wc is adjusted to five values ranging from 0.2 mm until 1.0 mm with a step of 0.2 mm
and the conductivity of cracked region c is set to 0, 1, 2, 5, 10% of the base material
conductivity. Length of the crack is kept at a constant value of lc = 10 mm for all the
cases. Frequency of the harmonic driving signal is adjusted to f = 10 kHz. An ECT
probe scans 1 mm above the plate surface. Influences of the crack parameters on the
ECT response signals are studied. The results are presented and discussed in the
following sections.
2. Standard Approach
The standard selfinductance absolute pancake probe is employed for the inspection of
the specimen with crack at first. The circular coil is positioned normally regarding the
surface of plate. Its axis is parallel to the zaxis of the coordinate system shown in
Figure 1. Inner diameter of the coil is 4 mm, the outer one has a value of 8 mm and a
height of the winding is 2 mm.
05 04
1 mm 1 mm
3 mm 3 mm
5 mm 03 5 mm
04
02
absolute value [mV]
imaginary [mV]
01
03
02
01
02
01
03
0 04
20 15 10 5 0 5 10 15 20 04 03 02 01 0 01 02 03 04
probe position [mm] real [mV]
a) absolute value vs. probe position b) signal in complex plane
Figure 2. Crack response signal influence of crack depth
56 L. Janousek et al. / Decreasing Uncertainty in Size Estimation of Stress Corrosion Cracking
The inspection is realised in a standard way; one dimensional scanning with the
probe over the crack centre along its length is performed. The response signal is
calculated for the various parameters of the crack. Figure 2 shows the response signal
for the nonconductive crack with depths of dc = 1, 3, 5 mm and a width of
wc = 0.2 mm. Similar results are displayed in Figure 3. The crack has a depth of
dc = 3 mm, a width of wc = 0.2 mm and conductivities of c = 0, 5 and 10% of the base
material conductivity. Influence of the crack width on the response signal for partially
conductive crack with a conductivity of c = 5% of the base material conductivity and a
depth of dc = 3 mm is shown in Figure 4. These results are gained for the crack widths
of wc = 0.2, 0.6, 1.0 mm.
04 04
0% 0%
5% 5%
10% 03 10%
03 02
absolute value [mV]
imaginary [mV]
01
02 0
01
01 02
03
0 04
20 15 10 5 0 5 10 15 20 04 03 02 01 0 01 02 03 04
probe position [mm] real [mV]
a) absolute value vs. probe position b) signal in complex plane
Figure 3. Crack response signal influence of crack conductivity
04 04
0 2 mm 0 2 mm
0 6 mm 0 6 mm
1 0 mm 03 1 0 mm
03 02
absolute value [mV]
imaginary [mV]
01
02 0
01
01 02
03
0 04
20 15 10 5 0 5 10 15 20 04 03 02 01 0 01 02 03 04
probe position [mm] real [mV]
a) absolute value vs. probe position b) signal in complex plane
Figure 4. Crack response signal influence of crack width
It can be seen that the response signal is the complex variable of all the considered
crack parameters. Figure 5 demonstrates quite high uncertainty in depth evaluation of
partially conductive cracks. In this case only the maximum points of the response
signals are plotted for the crack with three depths of dc = 1, 3, 5 mm and all the
considered widths and conductivities. As it can be observed the crack signals of the
crack with the three different depths are highly overlapped and thus under or over
estimation of the crack depth can occur with high probability. Especially, the
underestimation can bring serious consequences in many applications.
L. Janousek et al. / Decreasing Uncertainty in Size Estimation of Stress Corrosion Cracking 57
0
1 mm
3 mm
5 mm
01
imaginary [mV]
02
03
04
05
0 01 02 03 04 05
real [mV]
The vector lines of the eddycurrent density must be closed themselves in an inspected
material. Thus, any electromagnetic anomaly at present influences the path of eddy
currents in quite complicated manner and creates perturbations in the resulting
electromagnetic field. The previous section clearly showed that the degree of
uncertainty in depth evaluation of a partially conductive crack is quite high when only
one component of the perturbation field is sensed; however, the eddy currents provide
more information.
The authors propose sensing all the three spatial components of the perturbation
electromagnetic field to increase the information rate of ECT response signals.
Numerical results are provided to support the proposal. The same coil as in the
previous section drives the eddy currents in the specimen. Three circular coils of same
dimensions, i.e. the outer diameter equals to 4 mm, the inner one is set to 2 mm and the
height of winding is 0.5 mm, sense the response signals independently. The coils are
positioned perpendicularly to each other to sense all the three spatial components of the
perturbation field. All the coils are centred at one point. The numerical simulations are
performed under similar conditions as described in the previous section; however two
dimensional scanning over the cracked region is realized. The response signal gained
with the detecting coil oriented along the xaxis according to the coordinate system
shown in Figure 1 is denoted as the x component. Accordingly, the response signals
sensed with the detection coils oriented along the yaxis and zaxis are labelled as the
y component and the z component, respectively.
Figure 6 displays the response signals for the nonconductive crack with a width of
wc = 0.2 mm and a depth of dc = 3 mm. As it can be seen, each component provides
some information about the inspected defect; however, x and y components have
lower level than the z component.
58 L. Janousek et al. / Decreasing Uncertainty in Size Estimation of Stress Corrosion Cracking
0.05
0.03 0.04
0.02 0.03
0.02
0.01
0.01
0 0
20 20
15 15
10 10
5 5
20 15 0 20 15 0
10 5 5 y [mm] 10 5 5 y [mm]
0 10 0 10
5 10 15 5 10 15
x [mm] 15 20 20 x [mm] 15 20 20
a) x component b) y component
absolute value [mV]
0.3
0.2
0.1
0
20
15
10
5
20 15 0
10 5 y [mm]
5 0 10
5 15
x [mm] 10 15 20
20
c) z component
Figure 6. Crack response signal gained using proposed detector
Particular numerical results are presented in Figure 7. Only maximum points of the
sensed signals are shown in the plots. Signals of the x and the y components are
magnified by factor 5 to be able to display them in a same plot together with the z
component because of the reason mentioned above. Figure 7a) presents influence of the
crack depth on the response signals amplitudes. Only signals amplitudes of the non
conductive crack with a width of wc = 0.2 mm and depths dc ranging from 0 to 9 mm
with a step of 1 mm are depicted. Figure 7b) displays the results for various values of
the crack conductivities; c = 0, 1, 2, 5, 10% of the base material conductivity. The
other parameters of the crack are adjusted to the following values wc = 0.2 mm,
dc = 3 mm in this case. The influence of the crack opening on the crack signals
amplitudes is shown in Figure 7c). The partially conductive crack with a relative
conductivity of c = 5% and a depth of dc = 3 mm is considered here. Its width wc is
varied from 0.2 mm up to 1.0 mm with a step of 0.2 mm.
It can be observed from the results shown in Figure 7 that each component behaves
in slightly different manner when changing parameters of the crack, i.e. its depth, width
and conductivity.
It seems from Figure 7a) that the x component is more sensitive to the crack depth
comparing to the other two components. It even does not saturate so quickly with
increasing depth of the nearside surface crack. The reason can be explained from the
nature of eddy current lines changes due to a presence of crack. The x component is
especially formed by the perturbation field arising from eddy currents flowing around
tips of the crack that is higher than in the other two axes.
On the other side, Figures 7b) and 7c) show that the conductivity and the width of
crack influences more the y component than the two other components as a sensitivity
L. Janousek et al. / Decreasing Uncertainty in Size Estimation of Stress Corrosion Cracking 59
of the y axis coil to the perturbation field arising from eddy currents flowing through a
partially conductive crack is highest comparing to the sensitivity of the pickup coils in
the other two axes.
It can be concluded that sensing all of the three spatial components of the
perturbation field can increase information rate of the ECT response signals. This is
especially advantageous when partially conductive cracks, like SCCs, can appear in an
inspected structure and dimensions of the cracked region need to be evaluated. Degree
of uncertainty in depth estimation of such cracks can be reduced by employing the
proposed approach.
x component 0
x component
0.4 y component y component
z component 0 05 z component
01
0.2 c
imaginary [mV]
imaginary [mV]
0 15
dc
0 02
0 25
0.2
03
0 35
0.4
04
0.4 0.2 0 0.2 0.4 0 0 05 01 0 15 02 0 25 03 0 35 04
real [mV] real [mV]
a) influence of crack depth b) influence of crack conductivity
0
0.05
0.1
wc
imaginary [mV]
0.15
0.2
0.25
0.3
0.35 x component
y component
z component
0.4
0 0.05 0.1 0.15 0.2 0.25 0.3 0.35 0.4
real [mV]
Conclusion
above a crack along its length. Consequently a new inspection approach was proposed
in the paper. It comes from the fact that more information about a detected crack can be
gained by sensing all of the three spatial components of the perturbation
electromagnetic field. Two dimensional scanning is required in this case. It was shown
that the response signals sensed with three coils that are oriented perpendicularly to
each other behave in slightly different way when changing the parameters of crack.
This is advantageous as degree of uncertainty in depth evaluation especially of partially
conductive cracks, as stress corrosion cracking, can be reduced in this way. Further
work is going to be concentrated on experimental verification.
Acknowledgements
This work was supported by the Slovak Research and Development Agency under the
contract No. APVV019407. This work was also supported by a grant of the Slovak
Grant Agency VEGA, project No. 1/0308/08.
References
[1] N. Yusa, H. Huang, K. Miya, Numerical evaluation of the ill posedness of eddy current problems to size
real cracks, NDT&E International 40 (2007), 185 191.
[2] G. Rubinacci, A. Tamburino, S. Ventre, Fast numerical techniques for electromagnetic nondestructive
evaluation, Nondestructive Testing and Evaluation 24 (2009), 165 194.
[3] N. Yusa, Development of computational inversion techniques to size cracks from eddy current signals,
Nondestructive testing and evaluation 24 (2009), 39 52.
[4] Z. Chen, N. Yusa, K. Miya, Some advances in numerical analysis techniques for quantitative
electromagnetic nondestructive evaluation, Nondestructive testing and evaluation 24 (2009), 69 102.
[5] B.P.C. Rao, An artificial neural network for eddy current testing of austenitic stainless steel welds,
NDT&E International 35 (2002), 393 398.
[6] M. Rebican, Z. Chen, N. Yusa, L. Janousek, K. Miya, Shape reconstruction of multiple cracks from ECT
signals by means of a stochastic method, IEEE Transactions on Magnetics 42 (2006), 1079 1082.
[7] N. Yusa, L. Janousek, M. Rebican, Z. Chen, K. Miya, N. Dohi, N. Chigusa, Y. Matsumoto, Caution
when applying eddy current inversion to stress corrosion cracking, Nuclear Engineering and Design
236 (2006), 211 221.
[8] N. Yusa, K. Miya, Discussion on the equivalent conductivity and resistance of stress corrosion cracks in
eddy current simulations, NDT&E International 42 (2009), 9 15.
Electromagnetic Nondestructive Evaluation (XIV) 61
T. Chady et al. (Eds.)
IOS Press, 2011
2011 The authors and IOS Press. All rights reserved.
doi:10.3233/978160750750561
Konstanty M. GAWRYLCZYK
West Pomeranian University of Technology in Szczecin, Poland
Department of Electrotechnology and Diagnostics
Abstract. Semi discrete method is known since 80s of XXth century. The method
provides analytical solution in time, so the time stepping may be omitted.
Comparing to usual finite elements method in time, this method seems not to be
numerically effective, because produced matrices are dense. Because of this reason,
it was rather rarely used. However to carry out sensitivity analysis with adjoint
models [1] solutions in forward and backward time have to be obtained. Both time
points should coincide. For space discretization finite elements are used, as usual.
The semi discrete method allows to determine analytically the continuous solution
for any given time of analysis. In this paper evaluation of this method for different
kinds of excitation shapes is presented. The sensitivity analysis is applied for
inverse task of recognition of conductivity distribution in non destructive testing
of materials using eddy currents.
Introduction
1. FETS analysis
sA
2 A 0 i t
f t
. (1)
st
Approximating (1) with finite elements and applying timedependent elements the
solution with well known Finite Element Time Stepping method is obtained, leading
to the following system of linear equations:
K M A ti1
t
(2)
M
f ti1
1
f ti
1
K A ti
,
t
with: K stiffness matrix, M mass matrix, t time step, i  time index and
0,5 1 defines the differential scheme of the time stepping method. The effective
solution of (2) is as long possible, as the time step t remains constant. In other case,
matrix decomposition should be renewed.
If there is the need to calculate the sensitivities of measured voltage versus
material parameters, adjoint models are used [1], which appear in Tellegen theorem.
The quantities relating to adjoint models are denoted by the index (+). Time moments
from the analysis of adjoint model should coincide with that of original model. Because
the adjoint model is analyzed applying reverse time, fulfillment of this condition may
lead to excessive time steps and very long computational time. This is the reason, why
the semidiscrete method has been elaborated. It allows to calculate the solution for any
desired time moments, not only equidistantly placed.
A(t)
[ K ]{A(t)} + [ M ] = { f (t)}. (3)
t
sAt (t)
<K >\At (t)^<M > 0 , (4)
st
which solution is
K.M. Gawrylczyk / SemiDiscrete TimeDomain Sensitivity Analysis for Cracks Recognition 63
The constant C depends on initial value A(0), which is zero in our case, and also
on the excitation shape. The solutions for different excitations f are shown below.
3. Unitstep excitation
However the current exciting a coil can not take a form of unitstep 1(t), it is handy
approximation of the real shape of current. The steadystate response of the magnetic
vector potential As in this case has the form:
1 1
\As (t)^< K > \ f (t)^<K > 1 t
. (6)
and the semidiscrete solution for vector magnetic potential takes the form:
Because of mass matrix inversion, (8) cannot be applied for the whole region.
Outside the conducting region the mass matrix contains zeroes on main diagonal and
for this reason becomes singular. The region has to be divided our into a conductive
part 1 and a nonconductive part 2:
s
K K12 A1 t
M11 0 A1 t
0
11 st . (9)
K 21 K 22 A2 t
0 0 f2
0
1
<K12 ><K22 > \ f2 (t)^, or: <Kc >\A1 (t)^<M11>\sA1 (t)/st^\ fc (t)^. (10)
The similarity of (10) to (3) allows for exploitation of solution (8). The correctness
of described method was proved using very simple 2Dmodel (Fig. 1). The model
consists of conducting region, air and linear excitation current. The material parameters
are given on the picture.
64 K.M. Gawrylczyk / SemiDiscrete TimeDomain Sensitivity Analysis for Cracks Recognition
4. Harmonic excitation
sA(t)
<K >\A(t)^<M > \I m sin(t)^ . (11)
st
1
\As^ <K > j<M >
\^,
(12)
1
\As (t)^imag <K > j<M >
\^exp(jt) .
1
\A(t)^imag As exp(jt)
imag As
exp t<M > <K >
.
1
(14)
Obtained formula was tested using the same model as this shown in Fig.1. In Fig.3
there are solutions obtained with FETS, with proposed semidiscrete method and the
steadystate shape of potential for node number 1. i means the timemoment index.
The solution for rectangular impulse was achieved as superposition of solutions for two
unitstep impulses:
The solution for single, sinusoidal current pulse was derived in the same manner, using
superposition of two solutions given by Eqn.(14):
1
t<M > <K >
imag As e jt As
e if tbT ,
\A(t)^ (16)
1 1
tT
< M > <K >
imag As
e As
e < > < > otherwise.
t M K
The comparison done for test models reveals good conformity of solutions
obtained using semidiscrete method versus this from classical FEtimestepping.
T T
( J 0 ( ) E ( t ) ) dV dt = E ( t ) E ( ) dV dt.
+ +
(17)
0V 0V
Integration of the product of electric intensity vectors over the finite elements of
first order requires the Gaussian quadrature of the second order. Because Gaussian
points are located on the element borders in this case, the products containing
combinations of nodal values should be calculated. To obtain all necessary products,
the following matrix R is defined:
( ) {i }
T T
R = E + ( ) E (t ) dt = [ M11 ] exp t [ K c ][ M11 ]
T 1 1
c
0 0
(18)
( exp ( [ K ][ M ] ) {i })
T
[ M11 ]
1 1
c 11
+
c dt
where
is the size of finite element. The matrix R can be rewritten as follows:
( ) {i }{i }
T
R = [ M11 ] exp t [ K c ][ M11 ]
1 1 + T
0 c c
(19)
[ M11 ] exp ( [ K ][ M ] ) dt,
1 1
c 11
K.M. Gawrylczyk / SemiDiscrete TimeDomain Sensitivity Analysis for Cracks Recognition 67
1
because the term [ M11 ] exp [ Kc ][ M11 ] ( 1
) is symmetric. A drawback of the
{ }
T
established definition is that R is singular (it contains a singular term [ I c ] = {ic } ic+ ).
For better numerical efficiency, analytical integration would be essential. Since simple
disentangling of exponential functions in (19) is impossible, two following methods for
this purpose are proposed.
1
exp( X ) Y exp( X ) = Y + [ X ,Y ] + [ X , [ X ,Y ]] + 1 X , [ X , [ X ,Y ]] + ... ,
2! 3! (20)
where: [ X ,Y ] = X Y Y X .
In our case:
By applying the Zassenhaus formula for two good commuting matrices, the result is
Because the identity matrix commutes well with all other matrices, the identity
matrix [1] multiplied by an arbitrarily large coefficient C is added and subtracted:
( )
T
R = [ M11 ] exp t [ K c ][ M11 ]
1 1
0
(23)
(
{ I c } [ M11 ] + C [1] C [1]) exp ( [ K ] [ M ] ) dt.
1 1
c 11
[ M ] exp ( t [ K ] [ M ] ) C [1]
T 1 1
11 c 11
0 (24)
exp (t [ K ][ M ] ) exp ( T [ K ] [ M ] ) dt,
1 1
c 11 c 11
(
where: LA = ln {I c } [ M11 ] + C [1] .
1
)
Now, the exponents in above formula can be disentangled and integrated:
( ( )
R = T [ M11 ] exp T [ K c ][ M11 ] + LA C exp T [ K c ][ M11 ]
1 1
( 1
)). (25)
This formula allows to calculate sensitivity components for the righthand side of
sensitivity equation (17).
The terms of sensitivity matrix are derived from sensitivity equation (17). V
means the differences between the measured and actually simulated coil voltage.
Solving (26) as linear equations system the corrections to conductivity in finite
elements are obtained. It can be called quasiGaussNewton algorithm. The process
is nonlinear and must be repeated iteratively. In the case of poor convergence
truncated singular value decomposition of the sensitivity matrix is applied, as described
in [1]. The process starts with initial configuration, in our case with a conductivity
distribution without cracks. The search regions were applied with number of elements j
lower then the number of time steps i, so the equations system (26) was over
determined. The excitation current had the shape of unitstep, so the solution (8) was
exploited. After each iteration the conductivity in search region was corrected with .
For the next iteration the sensitivity S was evaluated anew.
When the coil moves over the crack, number of equations i consists of number of
time steps multiplied by the number of coil positions (10 steps in examples below).
K.M. Gawrylczyk / SemiDiscrete TimeDomain Sensitivity Analysis for Cracks Recognition 69
The examples below (Figs. 4, 5) show conductivity distribution inside tube wall
estimated after 15 iterations. The input data for assumed crack shape were taken not
from measurement, but from simulation utilizing the FEMmodel with coarser
discretization. It gave the similar effect as a presence of noise provided by real
measurement. The original flaw had nonzero conductivity of = 0,1[MS/m]. The size
of test flaws is shown in Figs.4 and 5. The solutions were carried out using modified
vector potential u = Ar and cylindrical symmetry.
Figure 4. Rectangular flaw shape and conductivity distribution obtained after 15 iterations.
Conclusions
References
[1] K. M. Gawrylczyk, M. Kugler, Time domain sensitivity analysis of electromagnetic quantities utilizing
FEM for the identification of material conductivity distributions, COMPEL, 25 (2006), 589 598.
[2] K. M. Gawrylczyk, M. Kugler, Semi discrete time domain sensitivity analysis of electromagnetic field,
COMPEL, 28 (2009), 1138 1348.
[3] R. Sridhar, R. Jagannathan, On the q analogues of the Zassenhaus formula for disentangling exponential
operators, Journal of Computational and Applied Mathematics 160 (2003), 297 305.
[4] D. N. Dyck, D. A. Lowther, E. M. Freeman, A Method of Computing the Sensitivity of Electromagnetic
Quantities to Changes in Materials and Sources, IEEE Transactions on Magnetics, Vol. 30, No. 5
(1994), 3415 3418.
[5] J. W. Bandler, Q. S. Cheng, S. A. Dakroury, A. S. Mohamed, M. H. Bakr, K. Madsen, J. Sondergaard,
Space mapping: the state of the art, IEEE Transactions on Microwave Theory and Techniques, vol.52, 1
(2004), 337 361.
[6] R. K. Amineh, S. Koziel, N. K. Nikolova, J. W. Bandler, J. P. Reilly, A space mapping methodology for
defect characterization from magnetic flux leakage measurement, IEEE Transactions on magnetics, vol
44, 8 (2008), 2058 2065.
Electromagnetic Nondestructive Evaluation (XIV) 71
T. Chady et al. (Eds.)
IOS Press, 2011
2011 The authors and IOS Press. All rights reserved.
doi:10.3233/978160750750571
Introduction
the outputs with relatively few points can be encompassed. This being made, a classical
radial basis function (RBF) interpolation [5] with a thin plate spline kernel function is
applied. The adaptive database interpolated by RBF is the metamodel that is used to feed
the stochastic inverse method.
The performances of the metamodelbased inversion method are illustrated for two
congurations, each corresponding to one benchmark: a circular tube workpiece affected
by groove defects (two parameters are to be retrieved) and a plate with parallelepiped de
fects (three parameters are to be retrieved). For each conguration, the method is applied
on many simulated signals and a couple of measured data extracted from the benchmarks.
Let us assume that tests are performed on generic pieces. Constitutive materials and ge
ometrical dimensions of both workpieces and probes are known, at a given frequency of
operation. The input parameters are the ones describing the defect inside the workpiece.
Then the response of the awed piece to the probe is computed via an integral volume
formulation (see for example [6]), and a variation of impedance is obtained as an output.
The CIVA simulation software [7] is used for this purpose. If P are the input parame
ters and y is the measured or simulated output impedance, we denote y = L(P). Other
methods could obviously be preferred for carrying out the forward simulations. For our
purpose, the forward simulations are only used to generate the database.
The aim of the adaptive database is to get, at the end, input parameter sets with equivalent
simulated output impedance variations, which are such that there exists a higher density
of inputs where the output variations are faster.
The database is created iteratively in order to achieve this goal. This adaptive
database generation is described in some more detail in [4]. First, a coarse mesh is gen
erated. Then, at each edge of each simplex of the mesh, simulations are compared to
linearly interpolated data. If the interpolation error , dened below, is larger than a
prescribed stopping criterion, a node is added in the middle of the longest edge of the
simplex. An example of a 2D mesh is shown in Figure 1. The interpolation error is
expressed as
y y22
= (1)
y22
where y and y are the exact and the interpolated data, respectively.
Once all calculations have been performed, the database is a set (Pi , yi )i=1, ,N
where N is the size of the database, each vector P being of dimension D (number of
parameters to describe a defect), and each vector y being of dimension M (number of
measurement points for one conguration).
The generation of the adaptive database certainly is a crucial step for metamodel
based optimization. Depending upon the formulation of the interpolation error, different
R. Douvenot et al. / Metamodel as Input of an Optimization Algorithm 73
Parameter 2
00 1
Parameter 1
Figure 1. Example of a 2D adaptive database.
behaviors of the interpolated data can be observed (a better accuracy on low or high
amplitude signals for instance) [8].
The aim of interpolating the database is to create a metamodel. The operator L is ap
proached via an expression depending on the data in the database, this involving little
computational effort compared to a forward ECT simulation.
Once the database has been created, an RBF interpolation is used to create the meta
model. That is, when an input P is given, the output y can be approximated by the
quantity
N
y = L(P) = wi K(P, Pi ), (2)
i=1
where K is a socalled kernel function, and where the weights wi are computed dur
ing a training step. In the following, the thin plate spline kernel function [9], which is
commonly used for interpolation, is chosen (Eq. 3). However, other kernels could be
preferred without stringent differences being expected.
The thin plate spline kernel can be expressed as
P1 P2 22 ln (P1 P2 2 ) if P1 =
P2
K(P1 , P2 ) = (3)
0 if P1 = P2
The stochastic optimization is carried out by Balanced PSO [10] as a variant of the clas
sical PSO method [11]. It is a metaheuristic based on swarm intelligence. Its principle
is that particles moving in the input space while sharing information can nd the global
minimum.
Having an impedance data y (the objective function), particles (usually about 20
[10]) are positioned in the input space with given initial speeds. For each one, a cost
function is calculated, and the particles are moved towards likelier positions. The parti
cles share information with some of them (generally 3 or 4) to know which particle has
obtained the minimum cost function. Thus, the updated speeds of the particles depend
74 R. Douvenot et al. / Metamodel as Input of an Optimization Algorithm
on their current speeds, weighted by an inertia w, on their best personal position, called
personal best p, and on the best position known by the particle, called social best s.
At iteration k, and along each dimension d [1, . . . , D], the positions x of the
particles are updated according to the formula:
Results are discussed for two standard ECT congurations: a tube and a plate.
90 90
80 80
70 70
100  Depth (%)
60
Depth (%)
60
50 50
40 40
30 30
PSO likelihood
20 PSO parameters 20
True parameters
10 10
2 4 6 8 2 4 6 8
Length (mm) Length (mm)
(a) GE40 (b) GI10
Figure 2. Results of the metamodel based PSO inversion on measured data GE40 and GI10.
the inverse problem reduces to a regression on two continuous parameters. The normal
ized quadratic error q = (Pd Pd )2 /Pd2 is computed for each test case and is averaged
on the 200 test cases.
For the internal defects, the average normalized quadratic error q is 1.5 % for the
depth and 1.9 % for the length. For the external ducts, the error q is 0.6 % for the depth
and 0.5 % for the length. Since the metamodel is used as the input of the optimization
algorithm, these results are obtained in less than 2 seconds on a 3 GHz personal computer.
Two specic results are detailed. They are results on real data corresponding to this
conguration and given in [13]. Measurements are avaiable for two defects: GE40 is an
external defect with a 40 % depth, and GI10 is an internal defect with a 10 % depth.
Both have the same length of 1 mm. Figure 2 shows the results on these two sets of
measured data. Darker areas correspond to the likeliest parameter values, the lighter ones
to less likely parameter values. The approximation of the likelihood is obtained using
the computations carried out during the inversion. The circles correspond with the true
parameters of the defects, and the squares give the likeliest ones retrieved by PSO.
The approximation of the likelihood gives all the sets of parameters for which the
signal due to the reconstructed defect is close to the measured one (the darkest area).
Since different defects can lead to similar signals, comparing the measured signal with
the one due to the more likely defect is not relevant for the purpose of making a decision
analysis tool.
The results obtained by the metamodelbased PSO inversion with this congura
tion appear very good in a reasonable time. The result obtained on GI10 is particularly
promising because it is a very low signal, harder to invert. Comparisons of this method
with support vector machines on this data can be found in [8].
80 80
100  Depth (%)
40 40
20 20
0 0
5 10 15 0.2 0.4 0.6 0.8 1
Length (mm) Width (mm)
14
12
PSO likelihood 10
Length (mm)
PSO parameters
True parameters 8
6
Figure 3. Results of the metamodel based PSO inver
4
sion on measured data F5 (depth 100 %, length 2 mm,
and width 0.1 mm). 2
ness 1.25 mm, and length 2 mm. The probe scans a 8 mm per 12 mm surface. The defect
position is assumed to be known, at the center of the scanned zone.
The defect is a rectangular notch with varying length, width, and depth. Its depth is
considered between 10 and 100 % of the tube thickness, its length is taken between 0.5
and 15 mm, and its width between 0.1 and 1 mm.
The average normalized quadratic error q is 0.6 % for the depth, 1.0 % for the width,
and 3.2 % for the length of the defect. These results are obtained with an average time of
5.5 s on the same computer as in section 2.1.
So, in general, the accuracy of the inversion is good. Two measurements correspond
ing to this benchmark have been processed. F5 corresponds to a defect of depth 100 %,
length 2 mm, and width 0.1 mm. F7 corresponds to a defect of depth 40 %, length 10
mm, and width 0.1 mm. The two inverted results are displayed in Figures 3 and 4, re
spectively. Both results are quite similar: the inverted parameters are acceptable yet not
perfectly retrieved. However, the true parameters are inside the likeliest areas. In par
ticular, it is difcult to precisely decide upon the depth of crack F5 or upon the width
of the crack F7. These examples show how undetermined cases can be identied by the
inversion method.
If other inversions were to be carried out on the same data, and since a stochastic
inversion is non repeatable, slightly different results could be found (minimum of the
cost function). However, the likeliest area should be approximately the same. So, the
inversion tool is appropriate for decision analysis.
R. Douvenot et al. / Metamodel as Input of an Optimization Algorithm 77
80 80
100  Depth (%)
40 40
20 20
0 0
5 10 15 0.2 0.4 0.6 0.8 1
Length (mm) Width (mm)
14
12
PSO likelihood 10
Length (mm)
PSO parameters
True parameters 8
6
Figure 4. Results of the metamodel based PSO inver
4
sion on measured data F7 (depth 40 %, length 10 mm,
and width 0.1 mm). 2
3. Conclusion
Acknowledgements
The authors thank CEA (Commissariat lnergie Atomique) and the COFREND work
ing group Eddy current modelization for providing the measured benchmark data.
References
[1] L. Udpa and S. Udpa. Eddy current defect characterization using neural networks. Mater. Eval., 48:342
347, 1990.
[2] A. Bernieri, L. Ferrigno, M. Laracca, and M. Molinara. Crack shape reconstruction in eddy current
testing using machine learning systems for regression. IEEE Trans. Instrum. Meas., 57:1958 1968,
2008.
[3] J. Pv and S. Gyimthy. Adaptive inversion database for electromagnetic nondestructive evaluation.
NDT & E Int., 40:192 202, April 2007.
[4] G. Franceschini, M. Lambert, and D. Lesselier. Generation of adaptive database results synthesis.
Technical report, Laboratoire des Signaux et Systmes, January 2009.
[5] M. D. Buhmann. Radial Basis Functions: Theory and Implementation. Cambridge University Press,
Cambridge, 2003.
[6] J. R. Bowler, S. A. Jenkins, Sabbagh L. D., and Sabbagh H. A. Eddy current probe impedance due to
volumetric aw. J. Appl. Phys., 75:8128 8137, 1991.
[7] CIVA: State of the art simulation platform for NDE. available at http://www civa.cea.fr, accessed the
8th of June 2010.
[8] R. Douvenot, M. Lambert, and D. Lesselier. Particle optimization with metamodel for crack character
ization. In Proceedings of the 2010 URSI Electromagnetic Theory Symposium (EMTS 2010), 2010. 4
pp., to appear.
[9] F. L. Bookstein. Principal warps: Thin Plate Splines and the decomposition of deformations. IEEE
Trans. Pattern Anal. Mach. Intell., 11:567 585, 1989.
[10] M. Clerc. The mythical balance or when Particle Swarm Optimisation does not exploit. Unpublished, 6
pp., available at http://clerc.maurice.free.fr/pso/Balanced_PSO/Balanced_PSO.pdf, 2008.
[11] M. Clerc. Particle Swarm Optimization. ISTE, London, 2006.
[12] P. Rocca, M. Benedetti, M. Donelli, D. Franceschini, and A. Massa. Evolutionary optimization as
applied to inverse scattering problems. Inverse Problems, 25:123003, 2009. 41pp.
[13] C. Reboud, G. Pichenot, and S. Mahaut. 2008 ECT benchmark results: Modeling with CIVA of 3D
aws responses in planar and cylindrical workpieces. In 35th Annual Review of Progress in Quantitative
Nondestructive Evaluation (QNDE  2008), Chicago, 20 25 july 2008. 8 pp.
[14] COFREND benchmark. available at http://www.cofrend.com/modelisationcf/GT_CF_COFREND_
GB. html, accessed the 8th of June 2010.
[15] R. R. Barton. Simulation optimization using metamodels. In Proceedings of the 2009 Winter Simulation
Conference, pages 230 238, Austin, TX, December 13 16 2009.
Electromagnetic Nondestructive Evaluation (XIV) 79
T. Chady et al. (Eds.)
IOS Press, 2011
2011 The authors and IOS Press. All rights reserved.
doi:10.3233/978160750750579
Introduction
The response of an eddy current probe can currently be modelled in the CIVA non de
structive testing (NDT) simulation platform (wwwciva.cea.fr) by eddy current technique
(ECT) simulation for planar and cylindrical geometries. The developed ECT simulation
models are mainly based on the volume integral method using the dyadic Green formal
ism, for which semianalytical representations are known for canonical geometries. The
motivation of this work is to develop a treatment addressing complex 3D geometries (see
Figure 1(a)), which responds to real industrial interest. It aims at modelling the response
of an eddy current probe on complex, homogeneous and nonmagnetic 3D geometries
affected by a aw (see Figure 1(b)).
(a) (b)
Figure 1. (a) Imaginary part of the surface density current J for f = 105 Hz and = 106 S/m (results for
1062 edges) and (b) a complex 3D geometry affected by a aw.
fullspace Green function gki (r, r ) and dyad Gki (r, r ) where ki is the wavenumber in
vacuum (i = 0) or in the medium (i = 1). It is based on transmission conditions on the
surface of the object S given by
n0 E0 = n1 E1 and n0 H0 = n1 H1 on S, (1)
Ai Ainc
i = Ascat
i on S. (2)
Let M(r) and J(r) be the surface magnetic and electric densities of current dened by
the electric and magnetic scattered elds are given on S by the integral formulations
1
Escat (r) = n i M i (r) + gki (r, r ) Mi (r )dr
i
2 S
jk0 Gki (r, r ) Ji (r )dr
S (3)
1
Hscat (r) = Ji (r) ni gki (r, r ) Ji (r )dr
i
2 S
i
jk0 Gki (r, r ) Mi (r )dr
i S
1
Gki (r, r ) = [I + ]gki (r, r ).
ki2
We then insert (2) and (3) in the transmission conditions (1), and solve the problem for
M and J by employing a Galerkin variant of the method of moments [1].
A surface discretization with bilinear quadrilateral nite elements is used (see Figure 2).
The surface currents are expressed as:
Nquad
M(r) Mq (rq )
q=1
where Mq is the value corresponding to the qth quadrilateral and Nquad the number of
quadrilateral on S. On each element, the surface currents are expressed as:
4
Mq (rq ) = q
Mm fm (rq (u, v))
m=1
q
where Mm are the current values corresponding to the mth local edge of each element
and fm (rq ) are the basis functions described as
1
f1 (rq (u, v)) = (1 + u)aqu (u, v)
JSq
1
f2 (rq (u, v)) = q (1 u)aqu (u, v)
JS
1
f3 (r (u, v)) = q (1 + v)aqv (u, v)
q
JS
1
f4 (r (u, v)) = q (1 v)aqv (u, v),
q
JS
r(u, v) r(u, v)
au (u, v) = , av (u, v) = ,
u v
JS = au av .
A Galerkin variant of the method of moments, using bilinear basis functions, is employed
for discretizing the surface integral equations, leading to the linear system, with matrix
Z and primary eld Minc (r) and Jinc (r).
M bM
Z = (4)
J bJ
Accurate 3D simulations of eddy currents may lead to a matrix Z of huge size (higher
than the capacity of a standard PC). Consequently, solving the linear system using a
direct matrix solver is prohibitively expensive in computing time and memory resources.
This difculty is circumvented with the use of an iterative solver (typically GMRES) to
nd the surface currents M(r) and J(r) by minimizing
M bM
Z J
J b
with the use of the Fast Multipole Method (FMM) [2], which accelerates the compu
tation of products of the matrix Z with the surface currents [M(r) , J(r)] and which
avoid the explicit evaluation of the matrix Z. The response of an eddy current probe is
nally expressed with the Auld reciprocity theorem, which involves the electric current
M(r) and the magnetic current J(r). The FMM has been generally used for problems
with dielectric objects without attenuation. In this work, we are dealing with low fre
quency electromagnetic problem and conductive regions, which are typical conditions
under which ECT are applied, for the application of FMM.
The FMM is based on the subdivision of the region of space containing S into cubic
clusters. The double integrals arising from the Galerkin discretization are then divided
into near interactions (involving pairs of adjacent clusters) and far interactions (involving
pairs of nonadjacent clusters). The former are computed using classic integral equation
methods, while the latter are computed using a multipole expansion of gk1 (r, r ) and
Gk1 (r, r ). On the basis of preliminary numerical tests, it has been found that the most
suitable FMM formulation in the medium (k1 ) for ECT application is the decomposition
in diagonal form given by
T. Lim et al. / Fast Multipole Method for 3D Electromagnetic Boundary Integral Equations 83
where R = r r , O and O are the centers of two nonadjacent clusters and the
transfer function T L (s, O O) is dened in terms of the Legendre polynomials Pl and
the spherical Hankel functions of the rst kind hl by
(1)
T L (s, O O) =
(2l + 1)j l hl (k1 O O)Pl (s O O) (6)
0lL
Table 1. Gain of computation time of FMMs LF/HF compared to the explicit calculation for different values
of the wavenumber used in NDT applications
k = 525+525j k = 1664+1664j k = 3441+3441j k = 5256+5256j
FMM LF 50 x faster 8 x faster  
FMM HF 50 x faster 20 x faster 10 x faster 4 x faster
The regular part is integrated using classical methods and integrations involving the sin
gular part are performed using Duffy transformations [3].
Different multipole expansions exist, called High Frequency (HF) or diagonal form, and
Low Frequency (LF). As low frequency occurs in NDT applications, one might think
that the LF expansion is the most suitable multipole expansion. However, FMM has been
generally used for problems with a real wavenumber (dielectric object without attenu
ation) whereas problems involve a complex wavenumber in NDT applications, due to
the conductivity of the object. Preliminary numerical tests have been done to evaluate
computing time with explicit calculation and FMMs HF/LF in NDT.
Comparison of computation time is shown for two far clusters with 100 elements
in each one. So 10,000 interactions have to be computed in explicit calculation and 201
interactions for FMM. Even if less interactions is needed with FMM, the time spent to
calculate the integration in Equation 5 has to be evaluated to compare the methods.
One can remark in Table 1 that the gain of computation time of FMM LF compared
to the explicit calculation is only signicant for low value of wavenumber. On the other
hand, the gain of computation time of FMM HF is signicant for almost all values of
wavenumber involved in NDT applications. So even if in NDT applications, we are deal
ing with low frequency electromagnetic problem, the most suitable multipole expansion
for NDT is the expansion called HF (in the medium). It is possible to explain theoretically
this result by analysing the wavenumber expression with and without attenuation.
wavenumber expression with attenuation:
k 2  =  2 0 (0 + j/) 0
k 2 = 2 0 0
For example, in NDT applications using low frequency, for = 0.7 MS/m and f =
100kHz, we have
k 2  = 5.5 105 .
But in comparison, without attenuation, we need f = 35GHz to get the same value of
k 2 . Thats why HF expansion is the most suitable multipole expansion for NDT (low
frequency electromagnetic problem but with conductive regions). A rst study has been
done to dene the conditions under which the calculation FMM works properly in the
NDT applications. For example, gure 4 shows the accuracy of the FMM HF expansion
as a function of the truncation level L and the integer Q (such that the integration over
the unit sphere appearing in (5) is effected using (Q + 1)(2Q + 1) quadrature points),
T. Lim et al. / Fast Multipole Method for 3D Electromagnetic Boundary Integral Equations 85
Figure 4. Accuracy of the FMM HF expansion based on L and Q for k1 = 1662 + 1662jm1 and centers
of far clusters distance of 2a with a = 102 m
for k1 = 1662 + 1662jm1 and centers of far clusters distance of 2a with a = 102 m.
The recommended truncation level is
Qopt = 1.5Lopt .
Preliminary tests have been done on planar geometries (see Figure 5) and curved ge
ometries (see Figure 6 depicting the computed distribution of the current density J on a
cylindrical sample). The response of an eddy current probe in the presence of a complex
object affected or not by a aw, leads to the computation of the variation of impedance
Z of the probe in the presence of the object:
Z = Z0 + Z,
where Z0 is the impedance of the probe in free space (i.e. in the absence of the object).
For canonical geometries, for example a plate, it is possible to calculate this variation
of impedance with the Dodd and Deeds approach [4] or by using the Auld reciprocity
theorem [5] once the linear system (4) is solved for J and M:
I 2 Z = J E0 M H0 ds. (8)
S
The equivalence of the two approaches has been numericaly checked by expressing
surface currents J and M and primary eld E0 and H0 on the surface of the object with
Dodd and Deeds calculation. Therefore, Equation 8 can be used since the linear system
(Equation 4) has been resolved.
In this paper, we report a modelling study aiming at simulating the response of an eddy
current probe on complex, homogeneous and nonmagnetic 3D geometry affected by a
aw. The Galerkin variant of the method of moments is employed for the surface integral
equations with bilinear basis functions to solve the linear system (Eq. 4) by using an
iterative solver together with the Fast Multipole Method. Preliminary validation tests
have been done on planar geometries. Numerical validations on canonical geometries
are in progress by comparison with CIVAs results. Ongoing work concerns also the
extension of the approach in order to take into account a contrast of conductivity inside
the defect.
References
[1] B. C. Usner, K. Sertel, M. A. Carr and J. L. Volakis, Generalized VolumeSurface Integral Equation for
Modeling Inhomogeneities With High Contrast Composite Structures, IEEE Trans. Antennas Propag.,
Vol. 54, No. 1, Jan. 2006
[2] G. Sylvand, La mthode multiple rapide en lectromagntisme: Performances, Paralllisation, Appli
cations, PhD Student, Ecole Nationale des Ponts et Chausses, Juin 2002.
[3] J. DEla, L. Battaglia and M. A.Storti, Full Nulmerical Quadrature in Galerkin Boundary element
methods, CIMEC, INTEC, CONICET, 2008.
[4] C.V Dodd, W.E. Deeds and J.W. Luquire, Integral Solutions to some Eddy Current Problems, Interna
tional Journal of Nondestructive Testing, Vol.1, pp. 2990, 1969.
[5] B.A. Auld, F.G. Muennemann and M. Riaziat, Quantitative Modelling of Flaw Responses in Eddy Cur
rent Testing, Nondestructive Testing, Vol. 7, 1984.
Electromagnetic Nondestructive Evaluation (XIV) 87
T. Chady et al. (Eds.)
IOS Press, 2011
2011 The authors and IOS Press. All rights reserved.
doi:10.3233/978160750750587
Abstract. This paper deals with the development of a fast numerical semi
analytical model dedicated to the simulation of the nondestructive testing of fer
romagnetic tubes by using the magnetic ux leakage method(MFL). Taking into
account the characteristic of the ferromagnetic material, the BH curve, the numer
ical model is based on the generalized boundary element method which implies
the evaluation of two magnetic ctitious scalar densities. Integral equations are dis
cretized via the Galerkins method and the particularity of this paper lies in the
implementation of high order basis functions for the interpolation of the scalar un
knowns. Simulated data provided by the numerical model are compared to FE data
in order to show the validity of the numerical approach.
Keywords. NDT by magnetic ux leakage, Magnetostatic regime, Nonlinear
media, Galerkins method, High order basis functions.
Introduction
The detection of aws in steel pipes by using the magnetic ux leakage technique con
sists in magnetizing the pipe and due to the presence of a defect, a fraction of the ux
lines leaks from the pipes wall in the air inside or outside the pipe. This leakage ux
is detected by a magnetic sensor or an induction coil. In the Vallourec group, a world
leader manufacturer of steel pipes, most of NDT units use MFL for testing ferromag
netic pipes. In order to improve the performances of aw detection, specically for inner
aws, the CEA LIST and the Vallourec Research Aulnoye (VRA) are aiming to develop
1 Corresponding Author: Denis Prmel, CEA LIST; Email: denis.premel@cea.fr
88 E.A. Fnaiech et al. / Computation of the Magnetostatic Field in Nonlinear Media
a fast 3D numerical model for the simulation of MFL systems. In a rst step, the frame
work addresses an experimental system specically dedicated to the detection of longi
tudinal aws. A simplied 2D geometry derived from the second magnetic MFL bench
mark problem proposed by the World Federation of NDE centers [1] has been previously
studied [2] considering a linear regime. This means that at low magnetization, the ferro
magnetic material behaves like a linear material. In this case, the semianalytical model
based on the boundary element method (BEM) has been implemented and numerically
validated. This preliminary work could give satisfaction for linear model but we truly
aim to improve the capabilities of detection for different kinds of defects. For the detec
tion of external defect, a low level of magnetization is usually sufcient but in order to
detect an inner defect or buried defects, we have observed that it is preferable to increase
signicantly the level of magnetization [2]. In this case, the main difculty in the compu
tation of the magnetostatic eld within the pipe of ferromagnetic material lies in the non
linearity of the BH curve characterizing the material constituting the magnetizing cir
cuit and the pipe. Each material can also be characterized by its own specic BH curve.
The framework of this present work consists in extending the rst semianalytical model
for computing the magnetostatic eld in the pipe considering the nonlinear regime. For
solving a nonlinear magnetostatic problem, the generalized boundary element method
(BEMG) is widely used since the statement of the problem is determined by two equiv
alent scalar potentials: the surface charge density and the volume charge density [3,4,5].
This numerical approach has the capability to compute the response of an arbitrary de
fect in the pipe. Nevertheless, this paper is focused on the numerical validation of the
semianalytical model in the nonlinear regime and without any aw.
Let us consider some nite domain occupying a nite volume V , bounded by a surface
S and made of a ferromagnetic material. The nonlinear problem is considered: the BH
curve associated to the material is introduced in the model. r (r) stands for the rela
tive permeability. Some encircling coils generate a source magnetic eld Hs which can
be calculated by using BiotSavart law [6]. The total magnetic eld is splitted into the
sum of the source eld and the unknown induced eld produced by ferromagnetic re
gions: H(r) = Hs (r) + Hi (r). The total eld is related to the magnetization vector
M(r) = (r 1) H(r). The resolution of the modeling problem amounts to determine
the induced eld Hi (r) produced by ctitious magnetic sources. The estimation of the
surface and volume magnetic charges is performed by using the extended boundary el
ement method implying boundary (surface in 3D) elements and surface (volume in 3D)
elements. This approach is briey described in the following section.
The development, based on Maxwells equations and satisfying the condition
B = 0, leads to a system of two equations combining a surface integral equation and
a volume integral equation [4]. The forward problem consists in evaluating two scalar
functions: the surface charge density denoted by (r) = M(r) n(r) and the volume
charge density denoted by (r) = M(r). When solving the nonlinear case, the
surface charge density does not vary signicantly compared to the linear case, so the
convergence of the iterative process cannot be reached. To overcome this problem, it is
necessary to solve the forward problem in two steps. The rst one consists in assuming
E.A. Fnaiech et al. / Computation of the Magnetostatic Field in Nonlinear Media 89
that the magnetic permeability is innite in such a way that the volume charges density
disappears. In this case, the surface charge density denoted by must be computed as
follows:
1
(r) + n G(r, r ) (r ) ds = n(r) Hs (r) (1)
2 S
with the normal vector n(r) pointing outwards from the volume V , G(r, r ) denotes the
Greens function in the free space and n G is dened by n G(r, r ) = n(r)G(r, r ).
Then, in the nonlinear case, we consider that the total surface charge density results
from the addition of (r) and a correction term such as:
A new system of two coupled integral equations leads to the computation of the differ
ence surface charge and the volume charges density :
(r)
+ n G(r, r ) (r ) ds +
2(r) S
(r) 1
n G(r, r )(r ) d = (r) (3)
V 2(r)
(r) + ln(r (r)) G(r, r )((r ) + (r )) ds +
S
ln(r (r)) G(r, r ) (r ) d = lnr (r) Hs (r) (4)
V
r (r)1
is dened by (r) = r (r)+1 . Then, the total induced magnetic eld is given by:
Hi (r) = G(r, r ) (r ) ds + G(r, r ) (r ) d (5)
S V
The two coupled equations (4) and the observation equation (5) need to be discretized in
order to solve numerically the forward problem.
Before starting to solve the 3D problem, we chose to study the 2D model. The rst step
involves approximating the geometry by an appropriate mesh which leads to the dis
cretization of the integral equations. Each contour of the piece is divided into curvilinear
elements and each surface is approximated by a number of quadrilateral elements (with
8 or 9 nodes). So, the coordinates of any point on each curvilinear segment (quadrilateral
element) are dened from the coordinates of the nodes using the shape functions:
90 E.A. Fnaiech et al. / Computation of the Magnetostatic Field in Nonlinear Media
n
n
x= xi Ni (), y= yi Ni () for curvilinear elements (6)
i=1 i=1
nq nq
x= xi Ni (, ), y= yi Ni (, ) for quadrilateral elements (7)
i=1 i=1
where Ni (), Nqi () are the shape functions usually used in Finite Element Methods. n
and nq are respectively the number of nodes associated to each curvilinear element and
to each quadrilateral element.
The unknowns of the problem, the surface charge density and the volume charge density
translated into the 2D problem, are interpolated by using a set of basis functions. Two
interpolation techniques are implemented: the rst one uses isoparametric elements and
the second one is based on higherorder basis functions.
For this description, the number of unknowns i (i ) corresponds to the number of nodes
dening the contour (the surface).
where M , (Mq and Nq ) are respectively the order of Legendre polynomials used for the
description of the density on an elementary curvilinear element and for the description
of the density on one elementary quadrilateral element. Pm (), Pn () are respectively
the Legendre polynomials and the Legendre polynomials modied [7], Cm and Cn are
some scaling factors. The total number of unknowns (the number of coefcients m ) is
equal to (M + 1) Ne where Ne is the number of curvilinear elements describing the
contour of the ferromagnetic region. The number of unknowns (the number of coef
cients mn ) is equal to (Mq + 1) (Nq + 1) Neq where Neq is the number of elements
dening the surface of the ferromagnetic domain.
These two techniques of interpolation are compared in the following in order to test
their efciency and we expect to reduce the computation time and the memory space
with high order basis functions.
E.A. Fnaiech et al. / Computation of the Magnetostatic Field in Nonlinear Media 91
To avoid the difculty to dene the surface charge density at any point of the mesh, we
have used the Galerkin method. The matrix system is nally obtained by:
A A b
= (10)
A A g
with
fi (r)
A = fi (r) + fj (r)n G(r, r )dk de (11)
e 2(r) k
A = fi (r) fqj (r)n G(r, r )dSk de (12)
e Sk
A = fqi (r) fj (r)lnr (r) G(r, r )dk dSe (13)
Se k
A = fqi (r) fqi (r) + fqj (r)lnr (r) G(r, r )dSk dSe (14)
Se Sk
(r) 1
b= fi (r) (r)de (15)
e 2(r)
g= fqi (r) lnr (r) Hs (r) + (r )lnr (r) G(r, r )dk dSe
Se k
(16)
where f and fq denote respectively the basis functions associated to contours and the ba
sis functions associated to surfaces, e and k denote respectively the indices of the obser
vation element and the source element. If we use isoparametric elements, the shape basis
functions fi and fj are associated to the nodes indexed respectively i and j. If we use
higher order basis functions, the indexes i and j stand for the order of Legendre polyno
mial series. Due to the fact that the magnetic relative permeability depends on the local
intensity of the magnetic eld, this matrix system is solved iteratively. At each iteration,
the magnitude of the magnetic eld is evaluated in each quadrilateral element consider
ing a number of observation points. Then, the gradient of the logarithm of the magnetic
relative permeability is deduced numerically by using spline interpolation functions.
92 E.A. Fnaiech et al. / Computation of the Magnetostatic Field in Nonlinear Media
3. Numerical validation
3.1. System
the rst one, we aim to validate the numerical results provided by the semianalytical
model and other simulated ones obtained by using a nite element software. The second
numerical experiment aims at highlighting the interest of the use of higher order basis
functions rather than the use of shape basis functions of second order.
The mesh of the tube contains a total number of elements equal to Neq = 200. The
inner radius of the tube is Rin = 48 mm and the outer radius of the tube is Rout = 88
mm. The current density driving the two coils is xed at Js = 1.1 105 A/m2 . This level
of magnetization allows one to go beyond the linear part of the BH curve. Keep in
mind that in this numerical experiment, the unknown scalar functions are interpolated by
using isoparametric elements. The radial component and the tangential component of the
magnetic eld are computed on a circumferential line, outside the tube with a distance
of 2 mm from the outer surface. The angle of each computation point varies from 45 to
135 . All simulated data are compared in Figure 2(a) and 2(b).
Figure 3(a) displays a cartography of the intensity of the magnetic ux density inside
the tube. Figure 3(b) extracts some specic values on a radial line at = 60 from
R = 47.5 mm to R = 87.5 mm.
The good agreement between simulated data shows the numerical validity of the
semianalytical proposed model. In the following, we look for the improvement of the
accuracy of the numerical model by using higher order basis functions.
E.A. Fnaiech et al. / Computation of the Magnetostatic Field in Nonlinear Media 93
(a) Hr (b) H
Figure 2. The radial and the tangential components of the magnetic eld provided by FEM and IEM.
In any geometry which contains a geometric corner, we expect to observe very huge
variations of the eld at the vicinity of the corner [8]. In such a situation, it is required
to rene the mesh around the corner but in this case, the number of unknowns can grow
up strongly. This seems to be very harmful for the computational time and the mem
ory space. In order to give a better approximation of theses variations, we preferably
choose to use higher order basis functions, keeping in mind that the geometry remains
well approximated with a coarse mesh. In order to show the interest in the increasing of
order, a coarse mesh with Neq = 75 is applied for both interpolations techniques (shape
functions of second order and high order functions). Some reference numerical results
are nevertheless obtained with a ne mesh with shape functions of second order. Fig
ure 4(a) and Figure 4(b) show the radial and the tangential components of the magnetic
eld along a radial line inside the tube and located at = 60 .
These numerical results show really the interest of high order basis function. Indeed,
this interpolation method provides better capabilities for representing the local variations
of the eld without the need of a rene mesh.
94 E.A. Fnaiech et al. / Computation of the Magnetostatic Field in Nonlinear Media
(a) Hr (b) H
Figure 4. The radial and the tangential components of the magnetic eld: A comparison between numerical
results obtained by shape functions of second order and by higher order basis functions.
This paper is focused on the development of a 2D numerical model for solving a spe
cic nonlinear magnetostatic problem dedicated to NDT MFL systems. First numerical
results show a good agreement between IE data provided by the semianalytical model
and FE data. Two semi analytical numerical models based on two different techniques
of interpolation have been implemented. Thus, the advantage of using high order basis
functions is clearly illustrated. It seems to be necessary to build up a ne mesh for ap
proximating with a good accuracy the complex geometry and to ensure a good evalua
tion of the unknown scalar densities, implicitly the intensity of the magnetic eld, it is
interesting to obtain a better description of the unknowns with high order basis functions.
Thanks to this numerical approach, any BH curve can be introduced since we solve the
non linear problem by an iterative procedure. In order to pursue this work, specic nu
merical problems due to numerical singularities remain to overcome. Moreover, we hope
to give a better approximation of more complex geometries without using a ne mesh
but by using NonUniform Rational Basis Splines NURBS. In close future work, we
aim to simulate the second magnetic MFL benchmark problem, a specic NDT system
proposed by the World Federation of NDE centers [1].
References
[1] Z. Zeng, Y. Tian and S. Udpa, "Finite Element Modeling of the World Federations Second MFL Bench
mark Problem", Review of Progress in Quantitative Nondestructive Evaluation, 23(700), (2004), 1553
1559.
[2] EA. Fnaiech, D. Prmel, C. Marchand and B. Bisiaux, "A fast numerical model for predicting magnetic
ux leakage signals due to defects in a steel pipe by using an integral equation approach", Proceedings
of the 8th International Symposium on Electric and Magnetic Fields (EMF2009)
[3] B. Krstajic, Z. Andelic, S. Milojkovic, S. Babic, and S. Salon, "Nonlinear 3D magnetostatic eld calcu
lation by the integral equation method with surface and volume magnetic charges". IEEE Transactions
on Magnetics, 28(2), (1992),10881091.
[4] W. Haa, A. Buchau, F. Groh and W.M. Rucker, "Efcient integral equation method for the solution of
3D Magnetostatic problems". IEEE Transactions on Magnetics, 41, (2005), 14081411.
[5] J. McWhirter, J. Oravec, and R. Haack, "Computation of magnetostatic elds in threedimensions based
on fredholm integral equations", IEEE Transactions on Magnetics, 18, (1982), 373378.
E.A. Fnaiech et al. / Computation of the Magnetostatic Field in Nonlinear Media 95
[6] I.R. Ciric, "Surface source models and formulas for the magnetic eld of polygonal cross section con
ductors". IEEE Transactions on Magnetics, 24(6), (1988), 31323134.
[7] E. Jorgensen, J L. Volakis, P. Meincke, "Higher order hierarchical legendre basis functions for electro
magnetic modeling", IEEE Transactions on Antennas and Propagation, 52(11),(2004), 29852995.
[8] F. Groh, W. Haa, A. Buchau and W.M. Rucker, "Field strength compuation at edges in nonlinear mag
netisatics", The international Journal for Computation and Mathematics in Electrical and Electronic
Engineering, 23(3),(2004),662669.
96 Electromagnetic Nondestructive Evaluation (XIV)
T. Chady et al. (Eds.)
IOS Press, 2011
2011 The authors and IOS Press. All rights reserved.
doi:10.3233/978160750750596
Introduction
In the application of the Volume Integral Method (VIM) for the simulation of eddy cur
rent defect inspections we employ the following integral equation for the calculation of
the electric eld in the presence of a void defect embedded in a multilayered conductive
halfspace [1]
E(r) = E(0) (r) + i0 G(rr ) p(r )dr (1)
V0
where E(r) is the electric eld with the defect present, E(0) (r) is the electric eld with
the defect absent (incident eld), G(rr ) is the electricelectric dyadic Greens function
dened as the eld response to a unit point source (electric dipole), p(r ) is the unknown
electric dipole density, V0 is the defect volume and primed and nonprimed position
vectors refer to source and eld points respectively.
When, however, the defect has the form of a crack, the very narrow crack opening
can make the use of (1) somewhat problematic because a dense discretization is required
across the gap in order to get accurate results. The situation is even worse in the limit of
a vanishing gap which describes the ideal crack case, when the crack is represented by
just a surface impenetrable to the eddy current ow. In such a case of a narrow or ideal
1 Corresponding Author: Theodoros Theodoulidis, University of Western Macedonia, Department of
Mechanical Engineering, Bakola & Sialvera, 50100 Kozani, Greece; Email: theodoul@uowm.gr.
T. Theodoulidis / Efcient Computation of Eddy Current Crack Signals 97
crack, instead of using (1), we can replace it with a simpler integral equation if we take
into account the very fact that at the crack surface the normal component of the eddy
current density vanishes. After imposing this condition we obtain [2]
Jn(0) (r ) = k 2 Gnn (r r )p(r )dS (2)
S0
where S0 is the crack surface and subscript n denotes the direction normal to the crack
surface. Since the discretization of the crack is reduced from a volume to a surface, the
BEM is used instead of VIM. The model based on the solution of (2) has been recently
developed to achieve high computational efciency. In [3], where a literature review of
research work regarding the model is presented, further improvements are suggested on
the original numerical scheme in [2] regarding computation of the incident eld, Greens
function computation and singularity treatment. In [4], an alternative approach is sug
gested for the Greens functions computation and for the simulation of cracks with shapes
other than rectangular. During all these developments, the computation of the Greens
function, i.e. the Greens function matrix ll time was identied as the bottleneck in the
calculations. In the present work, we study the subject from a different perspective and
compute these functions in the spatial domain. This is possible by employing the Dis
crete Complex Image Method (DCIM) which in turn is based on the efcient evaluation
of the Sommerfeld integrals. The method employs the Generalized Pencil of Function
(GPOF) method to approximate the Sommerfeld integral expression for the kernel with
a series of complex exponential functions and hence cast them in closed form [5], [6].
This is similar to the method presented in [4] but with a basic difference that will be
explained later. Results are provided for a semielliptical crack and they are compared to
benchmark experimental results. The same conguration is also modelled with a numeri
cal method (a 3DFEM commercial package) in order to identify the differences between
the two methods.
1. Greens function
Consider Figure 1 which shows a cylindrical coil above a narrow slot in a homogeneous
conductive and nonmagnetic halfspace. The coil is excited by a time harmonic current
varying as the real part of I exp(it). The crack surface is discretized in small rectan
gular elements and a constant dipole density is assumed over each element. The constant
dipole density is equivalent to representing the dipole distribution with pulse functions.
The integral equation is then transformed to a linear system of equations by imposing the
vanishing current constraint at the middle points of the elements, which is the denition
of point matching and is equivalent to the use of delta functions as testing functions. The
solution of the integral equation (2) yields the dipole density on the crack surface, which
acts as a secondary eddy current source. This source determines the coil impedance due
to the crack. By using a reciprocity relation, we can calculate directly the crack signal
(impedance change due to the crack) through [2]
1
Z = 2 Jx(0) p dV0 (3)
I
V0
98 T. Theodoulidis / Efcient Computation of Eddy Current Crack Signals
Coil
y
x
Crack x
where is the halfspace conductivity. In this paper we are interested in the calculation
of the term Gxx used in (2) by nding the xcomponent of the electric eld produced
in the halfspace by an elementary electric dipole oriented also in the xdirection. The
Greens function consists of two parts.
cos2 2 2
+ e(z+z ) J0 () d
4
0
sin2 cos2 2 2
+ e(z+z ) J1 () d (7)
4
0
where Jn denotes the n order Bessel function. Eq.(7) has to be integrated over the source
cell volume. In particular, the integration over the zdimension of the source cell, i.e.
over z , can be done analytically. Assuming a cell of z extent we obtain
k2 1 z
G(r)
xx dz
= sinh e(z+z ) J0 () d
2 + 2
0
cos2 2 2 z
+ sinh e(z+z ) J0 () d (8)
2 2
0
sin2 cos2 2 2 z
) 1
+ sinh e(z+z J1 () d
2 2
0
Now we seek an analytical calculation of the above integrals. For that we utilize the
following identities
z eikR
e J0 () d = (9)
R
0
1 z eikR eikz
e J1 () d = (10)
ik
0
2 2 z
sinh = c2,n es2,n (12)
2 n
we end up with a closed form expression for the reected part of the Greens function
k2 eikR1,n cos2 eikR2,n
G(r)
xx dz c1,n + c2,n
2 n R1,n 2 n
R2,n
Discrete Complex Image is now obvious. The Greens function consists of images lo
cated at a complex distance Rj,n (compare to (5)). Eq.(13) has to be integrated over the
x and ydirections. This can be done numerically or even analytically if rst the term
eikR2,n /Rn is Taylor expanded as an mth degree polynomial in Rn . Such an approach
is followed in [8] and leads to considerable time saving. The error introduced with such
an expansion is given by a closedform expression depending on the approximating poly
nomial degree and the ratio of the cell dimensions and thus it is controlled by the number
of terms in the Taylor series. In our case, having set x = y = 0 (eld point
at the center
of the cylindrical coordinate system), we make the substitution = x 2 + y 2 and
= arctan (y /x ) and integrate numerically (13) in order to obtain Gxx for (2).
The approximation in (11)(12) is possible with the GPOF algorithm [5]. Note that
the algorithm has to be performed in each equation just once since it does not depend
on the specic distance between the source and the eld cells. In [4], the GPOF algo
rithm was used for approximating the Sommerfeld integral kernel with exponentials of
the real variable. Such an approach results in an expression that can be evaluated ana
lytically over the volume of the source cell but it requires multiple GPOFs for each z + z
combinations. This is the basic difference of the present approach and the one in [4].
2. Results
Here we present results for a benchmark problem described in [9]. A coil is moved
above and along a semielliptical crack and the impedance change due to the crack (real
and imaginary part) is recorded with respect to the coil position. The semielliptical
notch has length 22.1 mm, max depth 8.61 mm and width 0.33 mm. The host material
is Aluminum with conductivity 22.5 MS/m. The cylindrical coil has inner radius 2.51
mm, outer radius 7.38 mm, height 4.99 mm, 4000 wireturns and liftoff 0.491mm. The
signals are presented for 16 frequencies that cover a wide range from the very low to the
very high. In order to present them all in a common diagram, they are all normalized to
the isolated coil inductive reactance X0 = L0 with a theoretical value of L0 =100.47
mH and an experimental one L0 =100.45 mH. The term low and high frequency is
dened by the crack depth to skin depth ratio. For example at 50 kHz this ratio is 18.15
and the case is dened as a thinskin limit case or a high frequency case.
Fig.2 compares the normalized complex impedance plane crack signals obtained
from the BEM model to experimental data and also to numerical results from a 3DFEM
T. Theodoulidis / Efcient Computation of Eddy Current Crack Signals 101
0.035
0.03 f [Hz]
250
356
0.025 507
721
1027
1462
0.02
2081
2963
4218
0
X/X
0.015 6006
8550
12172
0.01 17329
24670
35121
50000
0.005
0.005
0.02 0.015 0.01 0.005 0 0.005 0.01
R/X 0
Figure 2. Impedance plane crack signals for the 16 frequencies shown on the right of the diagram. Comparison
of theoretical BEM (lines) to experimental (dots) and FEM (circles) results.
3. Conclusions
We have presented a method for rapidly computing the Greens function matrix, which is
the slowest part in the computation of crack signals via method of moments. Impedance
predictions compare very well to experimental measurements as well as to FEM calcu
102 T. Theodoulidis / Efcient Computation of Eddy Current Crack Signals
lations. There are many ways to further extend the scope of the model. For example, it
can incorporate various coil types or coils under conditions of tilt [11], since their effect
enters through only the incident eld.
Acknowledgements
The work is funded by CEA LIST in the framework of the collaborative project CIVA
MONT 2012.
References
[1] S. Paillard, G. Pichenot., M. Lambert and H. Voillaume, Eddy current modelling for inspection of riveted
structures in aeronautics, Electromagnetic Nondestructive Evaluation (X), S. Takahashi and H. Kikuchi
(Eds.), IOS Press, (2007), 2532.
[2] J.R. Bowler, Eddy current interaction with an ideal crack. I. The forward problem, J. Appl. Phys. 75
(1994), 81288137.
[3] T. Theodoulidis, N. Poulakis and A. Dragogias, Rapid computation of eddy current signals from narrow
cracks, NDT&E Int. 43 (2010), 1319.
[4] T. Theodoulidis, Developments in efciently modelling eddy current testing of narrow cracks, NDT&E
Int. 43 (2010), 591598.
[5] Xing Lingling, Rapid calculation of eddy current eld Greens function using the matrix pencil method,
NDT&E Int. 42 (2009), 8591.
[6] M.I. Aksun, A robust approach for the derivation of closedform Greens functions, IEEE Trans. Mi
crowave Theory and Techniques 44 (1996), 651658.
[7] A. Banos, A dipole radiation in the presence of a conducting half space, Pergamon Press, 1966.
[8] L. Alatan, M.I. Aksun, K. Mahadevan and M.T. Birand, Analytical evaluation of the MoM matrix ele
ments, IEEE Trans. Microwave Theory and Techniques 44 (1996), 519525.
[9] D.J. Harrison, L.D. Jones and S.K. Burke, Benchmark problems for defect size and shape determination
in eddycurrent nondestructive evaluation, J. NDE 15 (1996), 2134.
[10] Comsol 3.5a Users Guide, 2009.
[11] T.P. Theodoulidis and E.E. Kriezis, Eddy current canonical problems (with applications to nondestruc
tive evaluation, TechScience Press, 2006.
Electromagnetic Nondestructive Evaluation (XIV) 103
T. Chady et al. (Eds.)
IOS Press, 2011
2011 The authors and IOS Press. All rights reserved.
doi:10.3233/9781607507505103
Abstract. In this work for the first time we validate experimentally a fast (real
time) non iterative imaging method for eddy current tomography of conductive
materials. The presented method is based on the monotonicity principle applied in
the low frequency limit where the skin depth is great with respect to the relevant
dimension of the problem under test. The experimental tests prove the
effectiveness of the monotonicity based imaging method. Moreover, we prove that
the inversion can be carried out without resorting the direct problem solver.
Introduction
Figure 1. The planar surface to be investigated (specimen) together with a probe made by an array of seven
coils and a rectangular defect.
1
Corresponding Author: Ass. EURATOM/ENEA/CREATE, DAEIMI, Universit di Cassino, v. G. Di Biasio,
43 Cassino 03043 Italy; e mail: tamburrino@unicas.it.
104 S. Ventre et al. / Experimental Validation of a Fast NonIterative Imaging Algorithm
the problem of predicting the measured quantities for a given electrical resistivity and
source current, have been extensively studied. The main applications of the direct
problem are: (i) the computer aided prediction of the performances of existing probes,
(ii) the design of adhoc optimized probes and (iii) the solution of the inverse problem
through model based inversion methods. Numerical formulations based on finite
element method have been studied and developed by the authors for both volumetric
[1] and zero thickness inclusions [2].
A key role in eddy current tomography is the reconstruction of the resistivity
profile of the specimen under test, starting from the measurement data, through the
solution of an inverse problem. The identification of the conductivity distribution
inside a material is seriously affected by the inherently illposed and nonlinear nature
of the eddy currents inverse problem (see [3][5] for mathematical issues). First and
foremost, the illposedness translates into low sensitivity measurements, i.e. different
resistivity distributions produce similar measurements and, in addition, the nonlinearity
requires sophisticated inversion algorithms.
The imaging methods can be divided in iterative and noniterative (direct).
Usually the iterative algorithms are based on the GaussNewton method. Among the
iterative methods it is worth mentioning quadratic and higher order approximation of
the forward operator, deterministic and stochastic algorithms, precalculated database,
statistical methods, Total Variation regularization, Level Set regularization, etc. (see
[6][24] and references therein).
As main drawbacks, the iterative methods require the solution of the direct
problem for several assigned tentative shapes of the inclusion and this can be very
expensive in terms of computational cost. Moreover, the convergence cannot be
guaranteed.
On the other hand noniterative methods have attracted a lot of interest because
they provide a test for evaluating if a point of the domain (or a subregion) is part or not
of the anomaly, regardless other points (or subregions). The test is very cheap from the
computational viewpoint because no iterations are necessary.
In this paper we present the first experimental validation of the fast and non
iterative method based on the monotonicity property of the real part of the measured
impedance matrix. This method was initially developed by the authors for elliptic
problems such as electrical resistance tomography [25], and later extended to parabolic
problems such as the eddy current testing in both low frequency (large skin depth) [26]
and high frequency (skin depth negligible with respect to the relevant dimension) [27
28].
Numerical simulations of the direct problem have been exploited to check
numerically if a given probe was able to guarantee appropriate reconstructions. The
monotonicity imaging method has been applied to identify the resistivity profile of
several benchmarks.
The paper is organized as follows. The monotonicity is proven for a proper
unknowndata mapping in Sections 1, the inversion method is presented in Section 2,
an experimental validation of the monotonicity imaging method is presented in Section
3 and finally in Section 4 the conclusions are drawn.
1. Monotonicity principle
In this section we briefly describe the monotonicity principle for the real part of
the measured impedance matrix (self and mutual impedances between pairs of coils of
S. Ventre et al. / Experimental Validation of a Fast NonIterative Imaging Algorithm 105
the array shown in figure 1). In the low frequency limit, the impedance matrix admits
the following expansion:
D = R0 + jL0 + PD + j PD + O( )
Zcoil 2 (2) 3 (3) 4
(1)
that is valid as long as the frequency is enough low so that the 4th and higher order
terms in (1) are negligible. We mention that all matrices appearing in (1) are real and
symmetric.
Matrix PD(2 ) , which contributes to the real part of the impedance matrix, plays a key
role because of its monotonicity property [26]. In particular it is possible to show that if
the resistivity of the considered material increases pointwise, then the ohmic power
dissipated in the conductor decreases and consequently PD(2 ) decreases in the sense that
will be clarified in the following. For twophase materials the monotonicity can be
stated as [26]:
Equation (2) forms the basis for the method to solve the inverse problem (see [26]).
Let V be the (unknown) subset of D where the resistivity is a (the resistivity in D\V is
b). The inverse problem consists in retrieving V.
Let us consider a generic (and known) test domain k. From (2) it follows that:
Figure 2. The conductive domain D subdivided in elementary regions together with an anomaly V (grey
pixels) and a test region k (black pixel).
106 S. Ventre et al. / Experimental Validation of a Fast NonIterative Imaging Algorithm
(2 )
In order to test (3) we need to compute the eigenvalues of P(2k) PD for checking if
this symmetric matrix is positive semidefinite or not. However, only a noisy version
~
PD(2 ) = PD(2 ) + E , PD(2 ) being the noiseless matrix and E the noise matrix, can be
measured and it can be processed as described in [26]. Similarly, the test matrices P(2k)
may be affected by either measurement errors (if measured) or numerical errors (if
~ ~ (2 )
numerically computed). Therefore, the eigenvalues of P(2k) PD can be different
(2 )
from the eigenvalues of P(2k) PD and this may alter in an unpredictable way the sign
of the eigenvalues with smaller magnitude, i.e. this may alter the reconstruction.
~ ~ (2 )
In order to deal with this situation, we quantify how much the matrix P(2k) PD is
positive semidefinite through the socalled sign index sk defined as:
j =1
k, j
s k = M
(4)
k , j
j =1
~ ~ (2 )
where k , j is the jth eigenvalues of the matrix P(2k) PD and M is the number of the
available eigenvalues. We notice that sk is closer to 1 when almost all relevant
~ ~ (2 )
eigenvalues are nonnegative (sk=1 when P(2k) PD is rigorously positive semi
definite).
In addition, from perturbation theory [29], it follows that the eigenvalues that may
change their sign due to the presence of noise, are only those having a magnitude
smaller than the Euclidean norm of the noise. When the magnitude of the noise is
known or estimated, these eigenvalues can be removed from the summations in (4).
Finally, we highlight that the dimension (number of rows and columns) of the
~ ~ (2 )
matrix P(2k) PD is small because it is equal to the number of coils of the array that,
usually, is made by few elements (few tens). Therefore, a very low computational cost
(2 )
(computation of the eigenvalues of P(2k) PD ) is required to compute (4) for
performing test (3).
3. Experimental results
The experimental tests we describe in the following consist of an array made by two
coils mounted in a fully automated imaging system composed by a scanning system, a
LCR meter and a PC controlling the whole acquisition and inversion process (figure 3).
The specimen under test is a printed circuit boards (PCB, thickness 1.6mm) presenting
copper islands , having different size and shapes, to be imaged. The copper islands
(thickness 35mm) are union of 5mm5mm elementary domains. These elementary
domains are taken as the k during the imaging process.
The sensor, that has been designed through extensive numerical simulations, is
made by two pancake coils. The first coil (internal diameter=5mm, external
diameter=10.5mm, height=6.5mm, number of turns=700) contains internally the
S. Ventre et al. / Experimental Validation of a Fast NonIterative Imaging Algorithm 107
The first test is a printed circuit board where the copper forms the letters IP
(Inverse Problem) as shown in figure 4 (left).
The inversion, that is shown in Figure 4 (right) is error free. This is because the
experimental noise level is smaller than the threshold (50 m) found by the
aforementioned numerical simulations involving the inversion of synthetic data.
Figure 3. The two coils composing the array (left). The smaller coil is inserted into the bigger one. Block
diagram of the measurement system (right).
108 S. Ventre et al. / Experimental Validation of a Fast NonIterative Imaging Algorithm
Figure 4. The specimen under test (left) and its reconstruction (right). The white pixels represent the
conductive pixels. The pixel dimensions are 5mm5mm.
The second test is a doubleface printed circuit board (see Figure 5). It consists of a
double sided PCB (copper, thickness 35m) as shown in figure 5 (top). The
measurements are taken from only the top side of the PCB.
In this case the imaging algorithm requires the test domains on both sides on the PCB.
The test domains on a side of the PCB provide the reconstruction for the related side.
For instance, the test domains on the top side provide an error free reconstruction of the
top side (figure 5, left column). On the other hand, the test domain in the bottom layer
provide as reconstruction the union of the pixels in the top and bottom sides (figure 5,
right column). This unexpected result can be easily explained by considering that it
~ ~
results P(2k), bottom P(2k), top and, therefore, it is trivial to prove that the test domains
related to the bottom side provide a reconstruction that is the union of the pixels from
the top and the bottom sides.
4. Conclusions
A fast inversion method for inverting eddycurrent testing data has been
experimentally validated for the first time. Numerical simulation of the whole imaging
process (not reported here for the sake of brevity) provide noise level threshold of
about 50 m such that for noise level below such threshold the reconstruction is error
free. The time required to form the image in a single pixel is about 0.1ms. The imaging
algorithm is fully nonlinear and, therefore, can treat arbitrary shapes and topologies.
Moreover, the measured data have been processed by means of premeasured data and
without resorting to the numerical solution of the direct problem that, as well known, is
timeconsuming.
S. Ventre et al. / Experimental Validation of a Fast NonIterative Imaging Algorithm 109
Acknowledgements
(a) (b)
(c) (d)
Figure 5. The specimen under test: the top side (a) directly under the probe, and the bottom layer (b).
Reconstruction obtained by exploiting the test domain located on the top side (c) and reconstruction obtained
by using the test domain located on the bottom side (d). For this latter inset the white pixels represent the
pixels of the bottom side whereas the grey pixels represent the pixels of the top side.
References
[1] M. Morozov, G. Rubinacci, A. Tamburrino, and S. Ventre, Numerical Models with Experimental
Validation of Volumetric Insulating Defects in Eddy Current Testing, IEEE Trans.on Magnetics, vol.
42, pp. 1568 1576, 2006.
[2] R. Albanese, G. Rubinacci, and F. Villone, An integral computational model for defect simulation and
detection via eddy currents, J. Comp. Phys., vol.152, pp. 736 755, 1999.
[3] D. Colton and L. Paivarinta, The uniqueness of a solution to an inverse scattering problem for
electromagnetic waves, Arch. Ration. Mech. Anal., vol. 119, pp. 59 70, 1992.
[4] V. Isakov, Uniqueness and stability in multidimensional inverse problems, Inverse Problems, pp. 579
621, 1993.
[5] M. Yamamoto, A mathematical aspect of inverse problems for non stationary Maxwells equations, Int.
J. Appl. Electromagn. Mech., pp. 77 98, 1997.
110 S. Ventre et al. / Experimental Validation of a Fast NonIterative Imaging Algorithm
[6] W. Q. Yang (Guest Editor), Special feature: Electrical Tomography Industrial and Medical,
Measurement, Science and Technology, vol. 21, 2001.
[7] R. A. Williams and Beck M S (eds), Process Tomography: Principles, Techniques and Applications,
Oxford: ButterworthHeinemann, 1995.
[8] M. Soleimani and A. Tamburrino, Shape reconstruction in magnetic induction tomography using
multifrequency data, Int. Jour. of Information and System Sciences. vol. 2, pp. 343 353, 2006.
[9] W. Yin and A. J. Peyton, A planar EMT system for the detection of faults on thin metallic plates,
Measurement Science and Technology vol 17, pp. 2130 2135, 2006.
[10] A. Pirani M. Ricci R. Specogna A. Tamburrino and F. Trevisan, Multi frequency identification of
defects in conducting media, Inverse Problems, vol. 24, 2008.
[11] A. Tamburrino, S. Ventre and G. Rubinacci, Electrical resistance tomography: complementarity and
quadratic models, Inverse Problem, vol. 16, pp. 1585 1618, 2000.
[12] R. Pierri and A. Tamburrino, On the local minima problem in conductivity imaging via a quadratic
approach, Inverse Problems, vol. 13, pp. 1547 68, 1997.
[13] G. Rubinacci, A. Tamburrino and S. Ventre, A differential formulation based on a perturbative
approach to solve the ECT inverse problem, Comput. Methods Appl. Mech. Eng. vol. 169 pp. 407 24,
1999.
[14] D. Dos Reis, M. Lambert and D. Lesselier, Non destructive evaluation of 3 D voids in a metal plate,
Inverse Problems, vol. 18, pp. 1857 71, 2002.
[15] D. Premel and A. Baussard, Eddy current evaluation of three dimensional flaws in flat conductive
materials using a Bayesian approach, Inverse Problems, vol. 18, pp. 1873 89, 2002.
[16] J. R. Bowler, Thin skin eddy current inversion for the determination of crack shapes, Inverse Problems,
vol. 18, pp. 1891 905, 2002.
[17] B. Luong and F. Santosa, Quantitative imaging of corrosion in plates by eddy current methods, SIAM J.
Appl. Math., vol. 58, pp. 1509 31, 1998.
[18] V. Monebhurrun, B. Duchne and D. Lesselier, Three dimensional inversion of eddy current data for
nondestructive evaluation of steam generator tubes, Inverse Problems, vol. 14, pp. 707 24, 1998.
[19] Z. Badics, J. Pv, H. Komatsu, S. Kojima, Y. Matsumoto and K. Aoki, Fast flaw reconstruction from
3D eddy current data, IEEE Trans. Magn., vol. 34, pp. 2823 2828, 1998.
[20] Li Yue, L. Udpa and S.S. Udpa, Three dimensional defect reconstruction from eddy current NDE
signals using a genetic local search algorithm, IEEE Trans. Magn., vol. 40, pp. 410 417, 2004.
[21] W. Yin, S. J. Dickinson and A. J. Peyton, Imaging the continuous conductivity profile within layered
metal structures using inductance spectroscopy, IEEE Sensors J., vol 5, pp. 161 166, 2005.
[22] S. M. Nair and J. H. Rose, Reconstruction of three dimensional conductivity variations from eddy
current (electromagnetic induction) data, Inverse Problems, vol. 6 pp. 1007 30, 1990.
[23] O. Dorn and D. Lesselier, Level set methods for inverse scattering some recent developments, Inverse
Problems, vol. 25, pp. 67 131.
[24] Q. N. Jin, On the iteratively regularized Gauss Newton method for solving nonlinear ill posed
problems, Math. Comput., vol. 69, pp. 1603 23, 2000.
[25] A. Tamburrino and G. Rubinacci, A new non iterative inversion method for electrical impedance
tomography, Inverse Problems, pp. 1809 1829, 2002.
[26] A. Tamburrino and G. Rubinacci, Fast Methods for Quantitative Eddy Current Tomography of
Conductive Materials, IEEE Trans. Magn., vol. 42, pp. 2017 2028, 2006.
[27] G. Rubinacci, A. Tamburrino, S. Ventre, Eddy current imaging of surface breaking defects by using
monotonicity based methods, ACES Journal, vol.23, no. 1, pp. 46 52, 2008.
[28] A. Tamburrino, S. Ventre, G. Rubinacci, Recent developments of a Monotonicity Imaging Method for
Magnetic Induction Tomography, Inverse Problems, vol. 26, no. 7, pp. 074016, 2010.
[29] Golub G H and Van Loan C F 1996 Matrix Computations (Johns Hopkins Studies in Mathematical)
Sciences 3rd Edition, Johns Hopkins University Press, Baltimore, USA.
Electromagnetic Nondestructive Evaluation (XIV) 111
T. Chady et al. (Eds.)
IOS Press, 2011
2011 The authors and IOS Press. All rights reserved.
doi:10.3233/9781607507505111
Numerical Evaluation of
Microwave Testing for Pipe Thinning
Yasutomo SAKAI, Noritaka YUSA, Hidetoshi HASIZUME
Department of Quantum Science and Energy Engineering, Tohoku University
Abstract. This study evaluates the applicability of using microwaves for the
inspection of pipe wall thinning by using numerical simulations and experiments.
The configuration considered in this study is an axisymmetric one consisting of a
straight brass tube and coaxial cables. After the agreement between
threedimensional numerical simulations and experimental result is confirmed,
experiments to locate wall thinning are carried out. Clear signals due to the wall
thinning are obtained. The time of flight of microwaves agrees with the
theoretical prediction, which demonstrates the possibility of locating wall
thinning from microwave signals.
1. Introduction
Rapid inspection for piping systems used in chemical or power plants is one of the
most important issues from the viewpoint of reducing maintenance costs. Conventional
methods such as ultrasonic and eddy current testing provide local information, which
results in a long inspection period for large piping systems. Due to this fact, another
nondestructive testing method for evaluating a large area at once is highly demanded.
Currently, to solve these problems, guided waves are mainly studied[1][2][3].
Guided waves propagate in pipe walls longitudinally, which enables inspection of a
broad area. However, there are some problems, such as the reflections at the
connections of pipes. Another approach is to utilize microwaves[4][5][6][7]. The basic
concept of this method is to regard the pipes as a waveguide and to detect defects from
transmission and reflection of microwaves propagating inside the pipes. Since
microwaves propagate through the free space surrounded by the metal pipes with little
attenuation, connections of pipe do not affect signals so significantly.
Recent studies have shown through numerical simulations that the presence of
wall thinning affects transmitted energy of microwave, and analyzing the signals in the
frequency domain enables one to evaluate the shape of wall thinning[8]. The studies,
however, include several problems from the viewpoint of practical applications. That is,
the model was quite simplified. In addition, the locations of wall thinning were
assumed to be known in advance and only the intensity of the spectrum of measured
signals were analyzed. Therefore, the purpose of this study is to evaluate the
applicability of microwave nondestructive inspection to the detection and evaluation of
wall thinning in more practical configurations. This study conducts threedimensional
numerical simulations and experimental verifications. Obtained signals are analyzed in
time domain to correlate the time of flight of the signal with the location of wall
thinning.
112 Y. Sakai et al. / Numerical Evaluation of Microwave Testing for Pipe Thinning
The principle of the microwave testing is to evaluate wall thinning from the transmitted
and reflected energy of microwave as mentioned above. The time of flight (TOF) of
reflection, which is calculated from the group velocity of the microwaves, enables one
to locate discontinuities such as wall thinning.
Fig. 1 shows the configuration used in both the numerical simulation and the
experiment. It consists of a brass tube modeling the piping, and coaxial cables and
connectors to inject microwaves into the tube. The coaxial cables are extended into the
pipes with a length of 5 mm to suppress the microwave reflections occurring at the end
of the cable. Microwaves are generated and the reflected energies are measured at port
1. The microwaves propagate as TEM mode inside the coaxial cables, and are
transformed into TE and TM modes inside the tube.
This study considered two lengths of a pipe, L=120 and 1120 mm, which are
termed as the short and the long range model hereafter. The short range model is
evaluated both in numerical simulations and experiments, to confirm the validity of the
method as well as to evaluate the behavior of microwaves propagating inside the tube.
The long range model is only dealt with by the experiments due to limitations of
computational resources needed to consider the model. Table 1 shows the parameters of
the two models. The numerical simulations shortened the length of coaxial cables, Lc,
into 20 mm to decrease computational resources.
The experiment system is composed of semirigid cables (K118 Anritsu),
microwave connectors (K101, Anritsu) several brass pipes and brass plates. The brass
plates and the semirigid cable are soldered. A short brass tube with a larger diameter is
utilized to simulate wall thinnings. That is, actually the brass tube shown in Fig. 1
consists of three tubes connected with flanges. Microwaves are generated and
measured by a network analyzer (Agilent PNA, E8363B). Flexible coaxial cables are
connected from the network analyzer to semirigid cables. Data is obtained in
frequency domain and transformed to time domain using inverse Fourier
transformation.
The numerical simulations are conducted using commercial software COMSOL
Multiphysics and its RF Module. The governing equation adopted is
( )
rot r1 rotE k 02 r j E = 0 ,
0
(1)
where
k 0 = 0 0 2 .
2
(2)
Subscript 0 denotes physical quantities in vacuum. Tube wall was modeled as a
conductor with a zero thickness and impedance boundary condition,
0 r
j 0 r n H + E ( E n )n = 0 , (3)
0 r j
is imposed on the surface of the tube wall. In the analysis, tetrahedral quadratic vector
elements are employed. Table 2 summarizes parameters used in the simulations. It
should be noted that the numerical simulations are threedimensional ones, whereas the
configurations are axisymmetric. The cutoff frequencies of TE11, TM01 and TM02
modes in the pipe are 9.25, 12.1, and 19.3 GHz, respectively.
Y. Sakai et al. / Numerical Evaluation of Microwave Testing for Pipe Thinning 113
3. Results
Fig. 2 shows comparison of the numerical and experimental results of the amplitude of
the reflected energy in the short range model. The figures show good agreement, which
Figure 1 Analysis and experimental system (unit mm)
$PSOLWXGH> @
$PSOLWXGH> @
)UHTXHQF\>*+]@ )UHTXHQF\>*+]@
Figure 2. Comparison between simulations (left) and experiments (right)
validates both the simulations and the experiments. The small differences between the
two results are mainly caused by the flexible cables not considered in the numerical
simulations. Other numerical simulations confirmed that TM01 is dominant inside the
pipe, which are not shown in this paper.
Fig. 3(a) shows the results of numerical simulations in the short range model.
As reflections at the connection part are relatively large, the measured data did not
show significant difference. The signals without any defects are subtracted from those
114 Y. Sakai et al. / Numerical Evaluation of Microwave Testing for Pipe Thinning
with defects to extract and to evaluate the effect of the defects quantitatively. The
subtracted signals are presented in Fig. 3 (b), which shows the reflections caused by
wall thinning. Fig. 4 shows the results of experiments in the short range model. The
0.40 0.40
QV QV
0.30 0.30
0.20 0 20
amplitude[]
amplitude[]
0.10 0.10
0.00 0.00
0.10 0.10
0.20 0 20
0.30 0.30
0.40 0.40
0.0 1.0 2.0 3.0 4.0 0.0 1.0 2.0 3.0 4.0
time[ns] time[ns]
(a) Raw data of pipes without wall thinning(right: non wall thinning, left: wall thinning depth 2mm)
0.10 0.10
0.54ns 0.61ns 0.54ns 0.61ns
0.05 0.05
amplitude[]
amplitude[]
0.00 0.00
0.05 0.05
0.10 0.10
0.0 1.0 2.0 3.0 4.0 0.0 1.0 2.0 3.0 4.0
time[ns] time[ns]
(b) differences(left: depth 2 mm, right: depth 1 mm)
Figure 3 Reflected microwaves in the short range model, obtained by the numerical simulations
0.20 0.20
2.34ns 2.41ns 2.41ns 2.34ns
0.15 0.15
0.10 0.10
amplitude[]
amplitude[]
0.05 0.05
0.00 0.00
0.05 0.05
0.10 0.10
0.15 0.15
0.20 0.20
0.0 1.0 2.0 3.0 4.0 5.0 0.0 1.0 2.0 3.0 4.0 5.0
time[ns] time[ns]
Figure 4 Reflected microwaves in the short range model, obtained by the experiments
(left depth 2 mm, right depth 1 mm)
Y. Sakai et al. / Numerical Evaluation of Microwave Testing for Pipe Thinning 115
0.10 0.10
amplitude[]
amplitude[]
0.05 0.05
0.00 0.00
0.05 0.05
0.10 0.10
0.15 0.15
0.20 0.20
0 2 4 6 8 10 12 0 2 4 6 8 10 12
time[ns]
time[ns]
Figure 5 Reflection of wall thinning (LD=290mm, LD=790mm, both depth 2mm)
4. Conclusion
In this study, microwave testing for piping systems was evaluated both by
threedimensional finite element simulations and experiments. Signals obtained in
frequency domain were converted into time domain to locate wall thinnings from their
time of flight. After the validity was confirmed in a short range model, practical
experiments using 1120mm long brass pipe were carried out. The results confirmed the
validity of the method.
Acknowledgement
This work was supported by the Tohoku University Global COE Program on World
Center of Education and Research for Transdisciplinary Flow Dynamics
Reference
[1] D. N. Alleyne, M. J. S. Lowe, P. Gawley, The reflection of guided waves from circumferential notches in
pipes, Journal of Applied Mechanics 65 (1998), 635641.
[2] P. D. Wilcox, A rapid signal processing technique to remove the effect of dispersion from guided wave
signals, Ultrasonics, Ferroelectrics, and Frequency Control 50 (2003), 419427.
[3] JoonHyun Lee, SeungJoon Lee, Application of lasergenerated guided wave for evaluation of
corrosion in carbon steel pipe, NDT&E International 42 (2009), 222227.
[4] Kavoos Abbasi , Study of Microwave Nondestructive Technique to Detect Crack and Predict Its
Location in Piping System, Tohoku University Doctoral Dissertation, (2007).
[5] Y. Ju, L. Liu, M. Ishikawa. Quantitative evaluation of wall thinning of metal pipes by microwaves.
Materials Science Forum 614 (2009), 111116.
[6] K. Sugawara, Development of inspection method for the defect in pipes using microwave, Tohoku
University Master thesis, (2002), (in Japanese)
[7] T. Shibata, H Hashizume, S. Kitajima, K. Ogura, Experimental study on NDT method using
electromagnetic waves, Journal of Materials Processing Technology 161 (2005), 348352.
[8] Y. Sakai, N. Yusa, S. Ito, H. Hashizume, Numerical analysis of microwave NDT applied to piping
inspection, Proceeding of The 13th Asiapacific Conference on Nondestructive Testing, (2009), 230.
Electromagnetic Nondestructive Evaluation (XIV) 117
T. Chady et al. (Eds.)
IOS Press, 2011
2011 The authors and IOS Press. All rights reserved.
doi:10.3233/9781607507505117
Abstract. A calculation method for the simulation of the Eddy Current Testing
(ECT) signal of multiple cracks is implemented on a general purpose video card.
NVIDIAs parallel computing solution, the Compute Unied Device Architecture
(CUDA) was used as a shell for the implementation on both hardware and software
sides. Possibilities inherent in this technology are shown through an example and a
detailed implementation focusing on data parallel execution is discussed. Different
realizations of the same algorithm are inspected and a comparison from the view of
efciency is given.
Introduction
The mathematical model of several different types of defects is given in the form of inte
gral equations. The numerical solutions of such integral equations might be obtained us
ing the Method of Moments. The main computational task in such solutions is the evalu
ation of the kernel of the integral equation at numerous locations (either spatial or spatial
frequency locations) inside the integration domain. Most part of such calculations in a
given location can be carried out independently of the calculations related to other loca
tions. Consequently the kernel can be evaluated simultaneously in the desired locations.
As a result, parallel computation of this can be easily designed.
A number of approaches are known for accelerating an application by paralleliz
ing it. Among others, one may use supercomputers for high performance computing, or
could build up a Beowulf cluster from ordinary machines. The application of a target
hardware for solving time consuming tasks is also a possibility, like for example the
project GRAPE (GRAvity piPE) for nbody computations [6]. It is easy to see that all of
these solutions have drawbacks, especially from the nancial point of view.
In this paper a parallel computing method is described aiming to give an extremely
fast evaluation of the signal of a thin crack using a CUDA capable Graphic Processing
Unit (GPU).
1 Corresponding Author: Imre Kiss, Budapest University of Technology and Economics, Goldmann Gyrgy
Since graphics hardwares reached notable power, there always have been attempts to use
their computational capacity for highperformance computations. One of the main causes
preventing the spreading of programs utilizing graphics cards were the complexity how
these target devices could be programmed. GPUs were primarily designed for opera
tion on graphical objects, hence long standing programming languages (e.g. OpenGL,
DirectX) were hard to be used for solving general, nongraphical related problems.
NVIDIA recently introduced its CUDA (Compute Unied Device Architecture)
technology [1] which issued in a general purpose graphics hardware. The new design
unied the workow from shader levels to a general purpose core (the term unied is
originated here), which serves the needs of both traditional graphical and general purpose
utilizations. Consequently, a hardware design consisting of hundreds of processor cores,
and capable to execute thousands of threads simultaneously have become available, that
can be programmed in a general purpose programming language (a slightly modied
version of the regular C language).
Efforts have already been made at different elds of engineering to utilize gain from the
application of GPU based parallel computation. Application elds dealing with large,
independent data structures are the primary targets of the utilization. For example in
molecular dynamics, the movement of numerous particles perfectly ts into the main
idea behind parallelization by treating each particle in a separate thread. In numerical
algebra, GEMM routines in the BLAS library give an example on parallel dense matrix
multiplication. Many other references could be enumerated from the elds of medical
imaging, uid dynamics, seismic imaging. A good summary on the application elds can
be found in [5].
2.1. Integral equation model for the calculation of the signal of multiple thin cracks
Assume the arrangement shown in Figure 1 where C innitesimally thin cracks are lo
cated close to each other in a conducting specimen. The surfaces and the surface normal
vectors of the cracks are Sc and nc (c = 1, 2, . . . , C), respectively.
ECT probe
0 , 0 n1
n3
n2
... nC
S1 S2 S3 ... SC
Surface Cracks
Vs : Specimen Surface Cracks
Figure 1. Eddy current probe above a conductor specimen containing innitesimally thin planar cracks.
I. Kiss et al. / Parallelization of Crack Signal Calculation Using CUDA 119
The eddy current signal of several innitesimally thin planar cracks can be calculated
by using the extension of the integral equation model of the single thin crack [2]. Each
crack is represented by a surface current density pc (c = 1, 2, . . . , C) concentrating on
the surface Sc of the crack. The unknown pc functions can be obtained by the solution of
the following system of integral equations [3]:
C
0 = Ecn
i
(rc ) j0 gkc (rc r )pk (r ) dr
k=1
k=c Sk
j0 lim gcc (rr )pc (r ) dr , rc Sc , c = 1, 2, . . . , C, (1)
r
rc
Sc
where Ecni
(rc ) is the socalled incident eld, that is the normal component of the elec
tric eld due to the interaction of the probe and the specimen without the presence of
the cracks. gkc (rr ) = nc Ge (rr ) nk (k, c = 1, 2, . . . , C), where Ge (rr ) is
the electricelectric Greens function assuming the geometry of the given specimen. rc
denotes the vector approaching the location vector rc from both sides of the surface Sc .
Knowing the solution of (1) the ECT signal, that is the impedance variation of the
probe, Z, is calculated as,
C
1
Z = 2 i
Ecn (r)pc (r) dr, (2)
I c=1
Sc
In this part we will show that the most timeconsuming calculations during the solution
of the integral equation (1) can be done parallelly if the method of moments is used
for the discretization of (1). For the sake of simplicity assume that the only one crack
(C = 1) is located in a plate specimen. The surface of the plate is in the xy plane of the
Descartes coordinate system and the crack is in the x = 0 plane (n1 = x, where x is the
unit vector of the x coordinate direction).
Assume that the surface current density is approximated by a properly chosen func
tion series as,
M
N
p1 (y, z) pmn fym (y)fzn (z) y, z S1 , (3)
m=1 n=1
For example, pulse approximating (fym (y), fzn (z)), and testing (tkl (y, z)) functions are
used in [2], linear approximating and testing functions are applied in [3], while sinusoidal
type global approximating and pulse testing functions are used in [4].
120 I. Kiss et al. / Parallelization of Crack Signal Calculation Using CUDA
M
N
pmn [j0
m=1 n=1 S1
lim g11 (x, y, zx = 0, y , z )fym (y )fzn (z ) dy dz tkl (y, z) dydz =
xx
S1
k = 1, 2, . . . , K,
= i
E1n (x = 0, y, z)tkl (y, z) dydz; (5)
l = 1, 2, . . . , L.
S1
A p = b, (6)
where
$ %
A (k1)L+l,(m1)N +n is the normal component of the electromagnetic eld due
to a surface current distribution described by fym (y)fzn (z) and tested by the test
ing functions tkl (y, z),
(b)(k1)L+l is the incident eld tested by the testing function tkl (y, z),
$ %
p (m1)N +n = pmn is the unknown coefcients of the surface current distribu
tion.
One can see that the most timeconsuming task, the calculation of the elements of A
can be calculated independently from each other. This is the reason why the paralleliza
tion of the outlined solution can really decrease the calculation time needed to obtain the
crack signal. Note also that the same conclusion can be derived if multiple cracks are
considered as the solution of (1).
A CUDA program mainly consists of two parts. There is a so called host program,
roughly speaking the function of which is to initialize the GPU for the computation while
itself runs on the CPU only. The program running on the device (GPU) is usually called
the kernel. As the kernel can operate only on data stored in the onboard memory of the
device, all data must be transferred to the device prior to the computation. Similarly, the
host program cannot directly access the onboard device memory, hence the result of the
computations must be transferred back the computers main memory.
These transfers are often act as the bottleneck of the whole computation. If the
amount of data to be transferred is large, then due to the relatively slow PCIE bus sys
tem, even with an excessively fast GPU the overall walltime would increase, and the
parallel execution would be pointless.
I. Kiss et al. / Parallelization of Crack Signal Calculation Using CUDA 121
Since the available memory on todays average GPUs is limited in 12 GB, one may think
that this memory constraint is just preventing any practical application, as time consume
often arises from the need of manipulating large data sets. To overcome on this problem,
one must not use of long standing algorithms: algorithms that do not need the whole
data set to be stored in the memory all at once are ideally suited for this task. An other
approach could be the decomposition of the original problem to smaller ones that t into
the memory constraint.
The overall speedup strongly depends on both data transfer and computation times.
A commonly applied technique is to overlap data transfer with computation, which can
be approached from CPU and/or GPU side. Namely, if a socalled asynchronous data
transfer is applied, the control is immediately returned to the initializer of the transfer.
Since latest graphic cards are capable to overlap kernel execution and memory transfers,
partitioning the GPU memory to nonoverlapping parts, the kernel may operate on one
part, while the other(s) could be simultaneously loaded. As a result, one could keep both
the CPU and the GPU busy, signicantly decreasing the inertia of the data transfer.
Output of the compiler regarding the GPU code is in a pseudoassembly language (PTX),
which is generally understandable by any CUDA capable GPU. Compilation of this code
to the actual hardware instruction set is taking place at execution time by the device
driver, therefore portability and backwardcompatibility is also ensured.
CUDA applies a Single Instruction Multiple Thread (SIMT) execution model, which
is a variant on the Single Instruction Multiple Data (SIMD) architecture. At any given
clock cycle, each streaming processor executes the same instruction, but operates on
different data without the need of packing and unpacking the original data like in the
case of SIMD (SSE2). Due to the SIMD architecture, thread divergence in a given core
causes threads being in an inactive path to be executed, but not evaluated: all threads
walk through all paths, signicantly increasing the execution time. One may also note
that thread divergence is only critical if occurs on numerous cores. A detailed overview
on the topic can be found in [7].
The original program computing the ECT signal of multiple cracks (based on the for
mulations in Section 2.2) was written for a single core CPU. Most of the calculations
are carried out analytically in the spatial frequency domain where the singularity of the
kernel can be easily handled. Elements of matrix A in (6) are obtained in kn km points
of the spatial frequency domain followed by an analytical integration in the z direction.
Spatial boundaries are determined adaptively, the kn km sampling points are distributed
uniformly within the studied frequency domain. The main concept of the implementation
can be found in [3] and [4], while this paper will focus on the design of parallelization.
Generally, the rst step during the design of parallelization is to obtain the execu
tion times for the different parts of the program. In the given case, the assembly of the
matrix elements is performed in subroutine 4, which requires about 45% of the total ex
122 I. Kiss et al. / Parallelization of Crack Signal Calculation Using CUDA
ecution time. The rest of the work is carried out in 3 other subroutines, where certain
coefcients of the basis and testing functions related to different spatial positions are pre
calculated. These 3 subroutines requires about 15% of the total computation time each.
Certain initialization process and postprocessing is neglected for clarity. Fig. 2.a shows
the schematic owchart of the original program.
After careful inspection of the program code, subroutine 4 is found the only part
which can be efciently parallelized, since each matrix element can be computed in
dependently of the others. In the case of subroutine 13, numerous conditional state
ments are evaluated, resulting in divergent execution paths, which prevents the efcient
execution on the SIMD architecture. Regretting this fact, these subroutines were also
reformulated for data parallel execution, but obtained walltimes conrmed theoretical
considerations.
Subroutine 4 will be parallelized and outsourced to the GPU while the rest of the pro
gram will be executed on the CPU. The execution path for the original single core CPU
program is shown in Fig. 2.a: serial execution of subroutine 14 is carried out organized
into a loop, in each cycle of which a spatial frequency component is computed. Note that
the dots correspond to the neglected initialization procedure.
Figure 2. Execution paths of the CPU and the two different CUDA models.
To replace subroutine 4, a CUDA kernel was carefully designed, performing the par
allel computation over the different spatial frequency sampling points. Since the avail
able memory on the GPU is less than that is required to store all the precomputed coef
cients, the problem must be partitioned to smaller parts. Each partition contains a subset
of the coefcients belonging to several spatial frequency points, the kernel will act on.
Based on our investigations, the GPU performs the computation over the sampling
points signicantly less time (with about 2 magnitudes) than the computation of the
coefcients takes on the host side. As a weird consequence, it become more important
how the data is transferred to the GPU, as it is comparable to CPU computation time.
Two CUDA programs were designed and implemented to investigate the effect of
the different transfer methods. These differ only in the way how the coefcients are
I. Kiss et al. / Parallelization of Crack Signal Calculation Using CUDA 123
transferred to the GPU, while executing the same kernel. Both programs compute the
required coefcients on the host side, enclosed in a loop containing subroutine 13. In
Fig. 2.b, a semiserial program model is shown. After a certain number of coefcients are
computed in the host loop, a single asynchronous memory transfer occurs, followed by
the invocation of the kernel. This enables the host loop (CPU) to keep computing during
the time consuming transfers. The size of the subset to be transferred is equal to the total
available memory on the GPU.
The second program is shown in Fig. 2.c, and is based on the idea that memory trans
fer time can only be hidden if the CPU is busy with the computation of coefcients during
the transfers. In order to achieve such a case, precomputed coefcients are transferred in
smaller sets. The size of a subset to be transferred was determined to take insignicantly
small time, hence its amount is limited in several MBs. In the given example, a dataset
corresponding to km sampling points is chosen for such purpose. The kernel is invoked
on the transferred data when the memory of the GPU is full up after numerous transfers.
4. Results
The CPU and the two CUDA programs were executed on a HP XW8600 workstation
with a total of 32 GB memory, 2 QuadCore Intel Xeon X5450 processors and an
NVIDIA GTX 480 video card.
20 GB
10 GB
5 GB
2 GB
1 GB
transferred data
a)
7 MB
3 MB
1 MB
transfer in each cycle
total amount of transfers
80
CUDA semi serial
Sub_4 run time [s]
60
CUDA overlapping
b) CPU
40
20
0
6
speedup []
5
c) 4
3
2
1 200 400 600 800 1000 1200 1400 1600 1800 2000
km = kn = N
Figure 3. a) amount of data for the overlapping transfer method (note that values on the y axis are shown in
logarithmic scale); b) execution times of subroutine 4; c) measured relative speed up.
124 I. Kiss et al. / Parallelization of Crack Signal Calculation Using CUDA
Fig. 3.a shows the amount of data transferred in each cycle, as well as for the total
problem in the overlapping case for different number of sampling points (note that the
total amount of GPU memory is only 1.5 GB for the given hardware). Runtimes with
respect to subroutine 4 are shown in Fig. 3.b. Based on the amount of data transferred to
the GPU (Fig. 3.a) and calculating with an average of 1 GB/s transfer rate, it can clearly
be seen that a signicant portion of time is spent on data transfer, rather than on (kernel)
computation.
The semiserial CUDA computational model performed a bit worse than that of the
CPU. The second CUDA model, applying overlapping data transfer and CPU computa
tion, performed signicantly better, preceding both the CPU and the semiserial CUDA
model. It must be emphasized that the kernel function itself is the same in both program
models, only the applied data transfer technique varies. The results indicate that the ex
ecution time of the kernel of the implemented algorithm is negligible compared to data
transfer time. The acceleration achieved by the overlapping CUDA model is shown in
Fig. 3.c.
5. Conclusion
In this paper the possibility of using CUDA for recoding already existing single proces
sor algorithms/programs was discussed. In the presented example by parallelizing only a
fraction (subroutine 4 constitutes only about 10%) of the whole source code, a remark
able speedup could be achieved.
Drawbacks of the required memory transfers between host and device memory are
shown together with some commonly applied techniques, that could be used to hide the
inertia of the transfer. Application of overlapping memory transfers with CPU computa
tions shown remarkable performance gain compared to the semiserial data processing
model. Although it would be possible to overlap the data transfers with kernel executions
as well, in the present case due to the negligible time of kernel execution, similar results
would be obtained. For problems having more intensive (time consuming) computations
in the kernel, a tradeoff in the amount of transferred data must be found to achieve
optimal overall performance.
References
!"#$%#&'%" ( '
) *+%#&'%"
'HSDUWPHQWRI(OHFWULFDODQG&RPSXWHU(QJLQHHULQJ)DFXOW\RI(OHFWULFDO
(QJLQHHULQJ:HVW3RPHUDQLDQ8QLYHUVLW\RI7HFKQRORJ\6LNRUVNLHJR
6]F]HFLQ3RODQG
.H\ZRUGV 6
. .
7
7
).7
7

,QWURGXFWLRQ
6
. .
/6 0
4 )

 5
,
5 .
;
;7 ,,
7 ,
,
5 )
3 ' ,
$ +
 /$+0

).
1 1 5 ,2
,

,
= ,,= 53

5 5
)
,
= )
;3 #
55
1
57
,
5
,, ,
1
3 "
) 
,5  ,
)
(
9,
> !
;);7 ,
5
9,
7
 5
7 &
8 .
 5
7 ';; ?@7 @A ?(?
'227 8
7 32
;);B2
33,

4 0 1
)
,= 5

,
15 ,
,= 5
1
3
D ,1
1
2
4
,
.
E(7
:F3
.
7 ) .
.
7
,

,
5
4 1
5
5
 )
,
,3
 5
).
5
5 1
,, 1 5
,
3
) .
,7 33 $
2 57
2
57
. 5 /
)
; ,

5 5
0

/ 0 E?GHF3 , ).
5
7

7
.
3 " ,,
.
,2
55
7

 )
,
1 ,
3
. .
, 5 1
;
).3
"
,, ) )
55
,
5
5
5
,
7
,
7 33 8 8
$ 9
) 2
 /'I0
,
2 ).7 )
,.
.
) .
 5 5
E(G@F3 '
,
7
 $
2 5 )
1
3 1
. 1 ,

,

3 ' 5
)
) 5
5 ,
).
7 ) ,,
,
3 "
5
7
;)
,,
,3
1XPHULFDOPRGHOGHVFULSWLRQ

5
)
<
,
5
1
5
3
)  u% /(0
330(0$7
.

8 8
/8803 ;
],
5

 .
4
27 ] K A7
,
5
5 )
;
5>
,3
<
1
5 1 
E:F>
:SIW
, A :SIW ( 7 5 W d 1 I
, 1 /?0
A7 5 W ! 1 I
)LJXUH #
5
<
7
54 ,
L I M(HAA I27 ,A :A 7 1 H

3
128 M. Ziolkowski and S. Gratkowski / EMATs Modeling in Time Domain
,
7
5
<
1
3
7
,
.
1 2 1
(7 : ?7 ,
.
3
7DEOH
,
5HPDQHQWIOX[GHQVLW\RIPDJQHW%U ( EF
&XUUHQW, :A EF
6SHFLPHQPDJQHWLFUHODWLYHSHUPHDELOLW\U ( EF
6SHFLPHQFRQGXFWLYLW\ H3@N(A@ E'OF
7DEOH
,
<RXQJ
VPRGXOXV( :N(A(( E8F
3RLVVRQ
VUDWLR A3??
'HQVLW\ @MHA E;O?F
7DEOH *
3ODWHGLPHQVLRQV (<( EF
3ODWHWKLFNQHVV (A EF
&RLOVHPLD[HV P < (3H EF
0DJQHWGLPHQVLRQV H < H < (A EF
/LIWRII A3H EF
7
5
1 9
,
5
) 1 3
. 1
2
.
,
3
5
.
1
5
5
1
,
3 6<
7
1
,
5

7
)
5
) ,
>
5
). 4
7
,
3
,
.
1 ,
1
$
2 53 ?H )
,
,
<,

W K A3(H / ,
5
).0 [7 \ ]
7 ,
.
3
)LJXUH ,
[
,
5
).
W A3(H 3
M. Ziolkowski and S. Gratkowski / EMATs Modeling in Time Domain 129
)LJXUH ,
\
,
5
).
W A3(H 3
)LJXUH ,
]
,
5
).
W A3(H 3
,
57
 1
55
). ,,3
Q
;
57 1
).7 <
3
).
)
5
).3
 EHF3
0HDQGHU /LQH&RLO(0$7
1
,3 '
1 ) 3 P3
130 M. Ziolkowski and S. Gratkowski / EMATs Modeling in Time Domain
,3
)LJXUH ,
[
,
5
).
W A3(H 3
)LJXUH ,
\
,
5
).
W A3(H 3
M. Ziolkowski and S. Gratkowski / EMATs Modeling in Time Domain 131
)LJXUH ,
]
,
5
).
W A3(H 3
" 7
) ). 3 6<
,
)
)
,
 ,
,7 )
5
,
). 5
3
7UDQVLHQWDQDO\VLV
 ,,
.7
)
).=
53 ,
1
5
5
5

3 57
5
,
).
,
3 & ,,
,,

)
5
5
,
<
,
/
7
0
, . ,
/547 ).10 E(F3
" 3 (A
: ,

1 ) 5 <,
 \ ]
,
;
[</ ? 9
07
; 5 88 3
,
) 1
)>
x
, W K HN(A@ EF7 <
W< K SN(AS EF
x
, [ K A3A:A? EF
132 M. Ziolkowski and S. Gratkowski / EMATs Modeling in Time Domain
)LJXUH :
5 \ ,
/
5
0 ] ,
/
0 5
88 7 54 I M(3H E;I2F7 1 H 
L ,11
, . ; 1
;
3
#
,
3 (A ,11
, . ,
. 1
; / 1
;
03 T
;)
7 5
, 1
)

5
,
). 3
&RQFOXVLRQVDQGIXUWKHUZRUN
)
, 5 2
). ,
7
,
. 1 ,
,,3 5
, 4
5
.
 ,
2 ,
,
3
.
1
)
5

1 1
1
5
)
5
/33 5 ,
0
3
,
1 5
<1
,
1
55
,
543 "
1
55
). ,,3 "
5
5
3 ,, )
1
).= :
,1
). 3 I).7
4
5
.
)

,

3
"
<
);
)
5 5
$FNQRZOHGJPHQW
5HIHUHQFHV
E(F '3 '7 3 + I3 *7 * &. '
6 9
6
,, 5
4 #,
7 <,
5
8 5
9#'#$ 95 :AAR7
3
E:F 3 !
;);7 +3 C7 3'3
7 "3 +54 3 Q> .
,
5 6.
'
+
* &. /'+*&0 4 5 $
8
",
7 T"6 :AA@ 957
(M :A ',
1 :AA@7 *
)7
%3
E?F I33 7
> 8,
7 ,
,,
7 8
7 D
3 (S7 &383 +363
37 87 6) U;7 (R@R3
ESF 3 %;7 3 %C7 I3 U %3
7
'
5 8, &
7 Q
5 .
7 D
3 7
Q, '
 5
 :A(AO:A((3
EHF 3 I I3 #7 (0$76IRU6FLHQFHDQG,QGXVWU\7 %
) 81
7 :AA?3
EPF U3 !7 33 I; Q3$3 +7 6.
5 9;
* '
8
7
+.) 5 8 V
. 6
. .
7 D
3 7 "
5 87
:AAP3
E@F 3 ' '3 7 6
 5
1 7 +.) 5
8 V
. 6
. .
7 D
3 7 "
5 87 :AA(3
134 Electromagnetic Nondestructive Evaluation (XIV)
T. Chady et al. (Eds.)
IOS Press, 2011
2011 The authors and IOS Press. All rights reserved.
doi:10.3233/9781607507505134
Abstract. Eddy Current Testing (ECT) is a standard technique in industry for the
detection of surface breaking aws in ferromagnetic materials such as steels. In this
context, simulation tools can be used to improve the understanding of experimental
signals, optimize the design of sensors or evaluate the performance of ECT proce
dures. CEA has developed for many years semianalytical models embedded into
the simulation platform CIVA [1] dedicated to nondestructive testing. Following
a previous work carried out at the laboratory in the case of one aw located in a
cylindrical ferromagnetic part, the developments presented herein address the case
of a aw located inside a planar and ferromagnetic medium. The theoretical ap
proach will be presented, as well as comparisons between simulation data obtained
from the literature.
Keywords. Integral formulation, Dyadic Green Functions, ferromagnetism, cracks,
eddy current testing
Introduction
Eddy current testing (ECT) is a standard technique in industry for the detection of surface
breaking aws in ferromagnetic materials such as steels. In this context, simulation tools
can be used to improve the understanding of experimental signals, optimize the design
of sensors or evaluate the performance of ECT procedures. CEA has developed for many
years semianalytical models embedded into the simulation platform CIVA [1] dedicated
to nondestructive testing.
Following a previous work [2] carried out at the laboratory in the case of one aw
located in a cylindrical ferromagnetic part, the developments presented herein address
the case of one aw located inside a planar [3] and ferromagnetic medium. Simulation
results are obtained through the application of the Volume Integral Method (VIM) [4].
This approach has proved its efciency when considering canonical geometries, mainly
1 Corresponding Author Email: Chiara.Zorni@cea.fr
C. Zorni et al. / ECT of Ferromagnetic Materials: Modelling of Flaws in a Planar Medium 135
due to the fact that, in those cases, analytical expressions of dyadic Green operators are
available in the spectral domain. While only one integral equation, involving either the
electric or the magnetic eld, is needed to describe the nonmagnetic case completely, in
the ferromagnetic case two coupled integral equations have to be solved.
Therefore, when considering the ECT of a single aw, a system of two integro
differential equations is derived from Maxwell equations. The numerical resolution of
the system is carried out using the classical Galerkin variant of the Method of Moments
[5]. Finally, the probe response is calculated by application of the Lorentz reciprocity
theorem [6].
The resolution has been generalized to the ECT simulation of M aws located in a
planar stratied medium. The theoretical approach will be presented, as well as compar
isons between simulation results and measured data obtained from the literature [7].
1. Theoretical formulation
Let us consider the general conguration described in Figure 1: a planar, stratied and
ferromagnetic medium consisting of N linear, isotropic and homogeneous layers. The
rst and the last layers are two air halfspace of permeability 0 and permittivity 0
and a time harmonic source with implied timedependence ejt is placed in the rst
layer (with = 2f the angular frequency). The others k = 2, . . . , N 1 layers are
characterized by their conductivity k , their permeability k = 0 kr (kr, being the k th
layer relative permeability) and the complex permittivity k = 0 + jk 1 . One or
more 3D bounded material(s) is contained in l = n, . . . , m layers and is characterized
m
by its volume = l=n l and its conductivity ld and permeability dl .
The approach consists in calculating the electric and magnetic elds, E and H re
spectively, in the volume of the defect without considering the defect contribution.
Then, the variations of the elds induced by the presence of the defect are computed
in . The volume integral formulation is derived by applying the Greens theorem [4]
to the Maxwells equations and involves the dyadic Greens functions for a multilay
ered medium [3]. Let us introduce the ctitious electric and magnetic total (Eq. (1)) and
incident (Eq. (2)) currents as:
J0k (r) = k (r) E0k (r) , M0k (r) = k (r) H0k (r) (2)
136 C. Zorni et al. / ECT of Ferromagnetic Materials: Modelling of Flaws in a Planar Medium
where k (r) = k kd (r) and k (r) = k dk (r) are the electric contrast and
the magnetic contrast functions respectively and Ek (r) and Hk (r) the total and E0k (r)
and H0k (r) the incident electric and magnetic elds in the k th layer. A system of two
integrodifferential equations gives the expression of the electric and magnetic elds in
the k th layer as:
m
Jk (r) = J0k (r) jk k (r) Gee
kl (r, r ) Jl (r ) dr
l=n l
m
jk (r) Gem
kl (r, r ) Ml (r ) dr
(3)
l=n l
m
Mk (r) = M0k (r) k (r) Gme
kl (r, r ) Jl (r ) dr
l=n l
m
2
k k (r) Gmm
kl (r, r ) Ml (r ) dr , (4)
l=n l
where Gee me mm em
kl (r, r ), Gkl (r, r ), Gkl (r, r ) and Gkl (r, r ) stand for the dyadic Greens
functions and correspond to the electric or magnetic eld response at the point r of an
unit point current source of electric or magnetic nature placed in r and are solution of
the Helmholtz equation. Thanks to the duality principle, only the dyads Gee
kl (r, r ) and
me
Gkl (r, r ) have to be computed and the two others are respectively obtained through:
Gem me
kl (, ) = Gkl (, ) (5)
Gmm ee
kl (, ) = Gkl (, ) (6)
Finally the response of the probe, in this case a single emitting/receiving coil, is
given by:
m & '
1
Z = 2
E0k (r ) Jk (r ) + jHk (r ) M0k (r ) dr (7)
I k
k=n
e r
1.25 mm 100 1 MS m1 2.3 mm
Table 1. Geometrical and physical parameters of the ferromagnetic layer.
L w d
6 mm 0.2 mm [40 %; [100 %] e
Table 2. Description of the defect affecting the layer.
outer diameter inner diameter height number of turns liftoff feeding current
3.2 mm 1.2 mm 0.8 mm 140 0.5 mm 1 mA
Table 3. Description of the acquisition conguration.
A rst comparison between our semianalytical model and data from the literature [7] is
presented for the conguration described in the previous section. A ferromagnetic plate
with an outer defect of 40 % of the plates thickness is considered. The number of
voxels N used to discretize the volume of the defect is 20 20 6 on x, y and z axis
respectively.
The result of this simulation is compared with a niteelements model result from
literature [7]. In Figure 2 the impedance variations obtained with the two different models
are superposed and a good agreement between the models is observed.
6
x 10
0
8
iterature
model
9
6 5 4 3 2 1 0
Real part [ohm] x 10
6
(a) (b)
Figure 2. Simulated ECT conguration (a) and the corresponding results (b) for a 40 % depth outer aw.
Signals show the impedance variations corresponding to the passage of the emitting/receiving coil over the
aw. The result is compared with a nite element result from the literature [7]
(a) (b)
Figure 3. Equivalent ECT conguration in cylindrical geometry (a) and the corresponding angular extension
(b). Geometrical and physicals parameters are given in Table 4.
the inner and outer diameters of the tube (given in Table 4a) have been chosen large
enough with respect to those of the acquisition probe. The geometrical parameters of the
corresponding 100 % defect and 40 % inner and outer defects are reported in Table 4b.
To validate the similarity between the planar and the cylindrical congurations both
simulated results are compared in the case of an outer defect (see Figure 4). As expected
the two Z have the same shape and the same order of magnitude.
Two others congurations have been then simulated in planar and cylindrical ge
ometries: an inner defect of 40 % (Figure 5) and a throughwall crack (Figure 6). Those
results show a good agreement between the two ferromagnetic models even if, for the
100 %crack, a discrepancy between the two curves exists and might be caused by the
two different geometries.
C. Zorni et al. / ECT of Ferromagnetic Materials: Modelling of Flaws in a Planar Medium 139
6
x 10
7
cylindrical model
planar model
8
8 6 4 2 0
Real part [ohm] 6
x 10
(a) Comparison between the pla (b)
nar conguration and the cylindrical
one.
Figure 4. Results for a 40 % depth outer aw. Signals show the impedance variations corresponding to the pas
sage of the emitting/receiving coil over the aw. Comparison with the equivalent case in cylindrical geometry
of CIVA.
5
x 10
6
Imaginary part [ohm]
10
12
14
16 cylindrical model
planar model
10 5 0 5
Real part [ohm] 5
x 10
(a) (b)
Figure 5. Results for a 40 % depth inner aw. Signals show the impedance variations corresponding to the pas
sage of the emitting/receiving coil over the aw. Comparison with the equivalent case in cylindrical geometry
of CIVA.
A model has been developed in order to simulate the response of the ECT probe, consist
ing in a set of emitting and receiving coils, to the presence of a aw in a planar stratied
and ferromagnetic medium. Simulations results and comparisons with the literature and
140 C. Zorni et al. / ECT of Ferromagnetic Materials: Modelling of Flaws in a Planar Medium
5
x 10
0
10
cylindrical model
planar model
12
14 12 10 8 6 4 2 0
Real part [ohm] x 10
5
(a) (b)
Figure 6. Results for a 100 % depth aw. Signals show the impedance variations corresponding to the pas
sage of the emitting/receiving coil over the aw. The result is compared with a nite element result from the
literature. Comparison with the equivalent case in cylindrical geometry of CIVA.
the equivalent cylindrical model of CIVA carried out in the case of a single layer and a
single emitting/receiving coil show a good agreement between the simulated signals.
Experimental validations are currently in progress and will take into account also
the stratied case and the multiple aws aspect. The perspective for this work is the
extension of the model to simulate the response of a magnetic probe containing GMR
elements.
References
[1] CIVA: State of the art simulation platform for NDE. http://wwwciva.cea.fr.
[2] A. Skarlatos, G. Pichenot, D. Lesselier, M. Lambert, and B. Duchne. Electromagnetic modeling of a
damaged ferromagnetic metal tube via a volume integral formulation. IEEE Trans. Mag., 44:623632,
2008.
[3] Sverine Paillard, Grgoire Pichenot, Marc Lambert, and Hubert Voillaume. Eddy current modelling for
inspection of riveted structures in aeronautics. In S. Takahashi and H. Kikuchi, editors, Electromagnetic
NonDestructive Evaluation (X), Studies in Applied Electromagnetics and Mechanics, pages 2532. IOS
Press, Amsterdam, 2007.
[4] Weng Cho Chew. Waves and Fields in Inhomogeneous Media. WileyIEEE Press, NewYork, 1999.
[5] R.F. Harrington. The method of moments in electromagnetics. J. Electromagn. Waves Appl., 1:181
200(20), 1987.
[6] B. Auld, F. Muennemann, and M. Riaziat. Research Techniques in Nondestructive Testing, volume 7,
chapter Quantitative modelling of aw responses in eddy current testing, pages 3776. 1984.
[7] Haoyu Huang, Toshiyuki Takagi, and Tetsuya Uchimoto. Fast numerical calculation for crack modeling
in eddy current testing of ferromagnetic materials. J. Appl. Phys., 94(9):58665872, 2003.
[8] A. Abubakar and P. M. van den Berg. Iterative forward and inverse algorithms based on domain integral
equations for threedimensional electric and magnetic objects. J. Comp. Phys., 23:236262, 2004.
Electromagnetic Nondestructive Evaluation (XIV) 141
T. Chady et al. (Eds.)
IOS Press, 2011
2011 The authors and IOS Press. All rights reserved.
doi:10.3233/9781607507505141
Abstract. Eddy current models have matured to such a degree that it is now
possible to simulate realistic nondestructive inspection (NDI) scenarios. Models
have been used in the design and analysis of NDI systems and to a limited extent,
model based inverse methods for Nondestructive Evaluation (NDE). The science
base is also being established to quantify the reliability of systems via Model
Assisted Probability of Detection (MAPOD). In realistic situations, it is more
accurate to treat the input model variables as random variables rather than
deterministic quantities. Typically a Monte Carlo simulation is conducted to
predict the output of a model when the inputs are random variables. This is a
reasonable approach as long as computational time is not too long; however, in
most applications, introducing a flaw into the model results in extensive
computational time ranging from hours to days, prohibiting Monte Carlo
simulations. Even methods such as Latin Hypercube sampling do not reduce the
number of simulations enough for reasonable use. This paper presents the
Probabilistic Collocation Method as a non intrusive alternative to other
uncertainty propagation techniques.
Keywords: eddy current, Polynomial Chaos, Probabilistic Collocation Method
Introduction
1
Air Force Research Laboratory (AFRL RXLP), Wright Patterson AFB, OH 45433, USA, Email:
Jeremy.Knopp@WPAFB.AF.MIL
142 J.S. Knopp et al. / Efcient Propagation of Uncertainty in Simulations via the PCM
detection (MAPOD) protocols and modelbased inversion schemes should not depend
on a revolution in computational efficiency. One technical capability necessary for the
realization of MAPOD and deployable modelbased inverse methods is the propagation
of uncertainty in input variables. In this work, variable refers to a measurable quantity,
and parameter refers to a quantity that must be estimated such as liftoff.
Stochastic numerical methods have been developed to propagate uncertainty
efficiently without relying solely on advances in computational power. Significant
advances in stochastic computation have been made especially in the last 20 years.
This paper briefly traces the development of stochastic numerical methods pertinent to
NDE modeling and simulation. The probabilistic collocation method (PCM) is
introduced and applied to two eddy current forward models.
b
( f g )w = f ( x) g ( x) w( x)dx = 0 (1)
a
The inner product of functions f(x) and g(x) is zero if they are orthogonal with w(x).
The only polynomial functions that satisfy this for the case where the weighting
function is equal to the PDF of a standard normal distribution are Hermite polynomials.
Ghanem and Spanos combined this concept with the finite element method to
simulate problems in solid mechanics [6]. For nonGaussian processes, the
convergence is not optimal, but fortunately there are other orthogonal polynomial sets
for other general types of random variables that converge exponentially [7].
In work mentioned thus far, the formulations are often intrusive, meaning that
alternation of deterministic code is necessary for implementation. A nonintrusive
method was introduced in the mid1990s [8]. Essentially, a reduced order polynomial
form of the model is derived with minimal simulations. The input values for these
simulations are derived from the roots of orthogonal polynomials which are dependent
only on the input distributions. Hermite polynomials are associated with Gaussian
distributions as described in the preceding paragraph. The weighting function for the
J.S. Knopp et al. / Efcient Propagation of Uncertainty in Simulations via the PCM 143
2. Case Study 1.
L1 ( A) = A 2.03
L2 ( A) = A2 4.06 A + 4.0376
L3 ( A) = A3 6.09 A2 + 12.2127 A 8.0609 (2)
4 3 2
L4 ( A) = A 8.12 A + 24.5111A 32.5917 A + 16.1041
L5 ( A) = A5 10.15 A4 + 40.9312 A3 81.9626 A2 + 81.4899 A 32.1796
H 1 ( ) =
H 4 ( ) = 4 6 2 + 3
H 2 ( ) = 2 1 (3)
H 5 ( ) = 5 10 3 + 15
H 3 ( ) = 3 3
Note that since the mean of the Gaussian distribution is 5, parameter B is translated by
5.
B = 5 + H1 ( ) (4)
Next, a first order model might be used to approximate the real model. In this
case, there are three unknowns X0, X1, and X2 as shown in equation 5. Since we are
interested in both the resistance and reactance, this process is done separately for each.
144 J.S. Knopp et al. / Efcient Propagation of Uncertainty in Simulations via the PCM
(a) (b)
Figure 1. Change in the (a) real and (b) imaginary part of the impedance due to the
plate and flaw.
Three simulations are required to solve for the three coefficients. Four pairs of
collocation points for the inputs of the simulations are available from the roots of the
2nd order polynomials. Three pairs are selected to solve for the three unknowns. L 1
and H1 are simply evaluated at the three pairs of collocation points. This is all that is
needed for the a 1st order approximate model. Then a MonteCarlo simulation can be
conducted on this simple model to calculate the PDF of the resistance and reactance.
To check the accuracy of this approximation, the next higher order collocation points
are necessary. The roots of the 3rd order orthogonal polynomials are used and six
simulations using the full model are conducted and compared with the approximate
model. Ultimately nine simulations are necessary for construction and evaluation of
the 1st order approximate model. Fortunately the same six simulations used for the
error check of the 1st order model can be used to construct the 2nd order approximate
model. The 4th order orthogonal polynomials are then used to check the accuracy of the
2nd order approximate model requiring eight more simulations for a total of fourteen
simulations. The authors have generally found that a 3rd order approximate model with
one interaction term is usually sufficient to predict output PDFs with good accuracy
for forward eddy current simulations, but there may be situations where higher order
approximations with more interaction terms are needed.
Y = X 0 + X 1 L1 ( A) + X 2 H1 ( B) (5)
The form of the 3rd order reduced model as described in equation 6 has eight
coefficients, one of which is an interaction term.
Y = X 0 + X 1 L1 ( A ) + X 2 H 1 ( ) + X 3 L 2 ( A ) + X 4 H 2 ( )
(6)
+ X 5 L 3 ( A ) + X 6 H 3 ( ) + X 7 L1 ( A ) H 1 ( )
To solve for the eight unknowns, eight simulations must be run. The roots of the
4th order Legendre and Hermite polynomials are used to select the input values for A
and B in these simulations. This is done separately for the real and imaginary
components of the impedance. Once again, everything is known except for the X
coefficients. The resulting system of equations is then solved and the coefficients are
J.S. Knopp et al. / Efcient Propagation of Uncertainty in Simulations via the PCM 145
listed in Table 1 for both the real and imaginary components of the impedance.
Incidentally, X0 is the expected value, which is one of the primary benefits of using
orthogonal polynomials. As expected, the interaction term X7 is close to zero in both
cases.
To check the error of this third order approximation, the roots of the 5 th order
polynomials are used to select collocation points. Ten more simulations are needed for
comparison with the reduced order model predictions. The results of the simulations
and the results of the reduced model are displayed in Table 2.
The residual is defined as the difference between the full and reduced model in
equation 7, and the sum of squares of the residuals is defined in equation 8. The
relative sum of squares of the residuals is used to quantify the error in the
approximation. The relative sum of squares of the residuals is calculated by simply
dividing by the expected values as shown in equation 9. These quantities are summed
over the model results for each collocation point.
Y Y (7)
i
2
(8)
ssr = i
n
ssr
rssr = (9)
E[Y ]
146 J.S. Knopp et al. / Efcient Propagation of Uncertainty in Simulations via the PCM
The errors for the real and imaginary parts are given by: ssr(resistance) =
2.9428104, rssr(resistance) = 3.0359104, ssr(reactance) = 0.0020, and
rssr(reactance) = 8.4790104.
Now there are two 3rd order polynomials with good accuracy for the real and
imaginary components of the impedance. Monte Carlo simulations can now be run
with ease on this 3rd order approximate model. The PDF for the real and imaginary
component of the impedance is shown in Figure 2(a) and 2(b) respectively.
6 25
5
2
4
15
PDF
P DF
3
1
2
05
1
0 0
0 85 09 0 95 1 1 05 11 27 26 25 24 23 22 21 2
Resistance Reactance
(a) (b)
Figure 2. PDF of the (a) real and (b) imaginary component of the impedance.
3. Case Study 2.
A second case study is presented for the characterization of a gap between multiple
plates, representing the problem of characterizing corrosion at the faying surface of a
multilayer aircraft structure. A diagram of the problem is shown in Figure 3 including
material properties and dimensions. In this hypothetical study, the thickness and
conductivity of the layer are treated as random variables. The conductivity of the top
plate is assumed to be a normal random variable with a mean value of 1.876107 S/m
and standard deviation of 3.384106 S/m as shown in Figure 4(a). The liftoff is defined
as a uniform random variable with the range of 0.55 +/ 0.05 mm as shown in Figure
2(b). VIC3D was used to generate the simulated results for the study. The gap
between the plates was fixed at 0.3 mm and the frequency for the study was set to 4.0
kHz. Both 1st order and 2nd order PCM models were evaluated for this case requiring
only 3 and 6 model calls respectively. Monte Carlo simulations were also run with
100,000 model calls used in the study.
1 ^
air ^
2 ^
600 200
160
500 140
150
120
400
100
300 100
80
200 60
40 50
100 MC
20
PCM
0 0 0
15 2 25 05 0 52 0 54 0 56 0 58 06 0 085 0 09 0 095
(S/m) x 10
7 (a) dz (mm) (b) R (c)
g
300 0 704 0 704
0 706 0 706
250
0 708 0 708
200
0 71 0 71
X
0 714 0 714
100
0 716 0 716
50 MC 0 718 0 718
PCM 0 72 0 72
0
07 0 71 0 72 0 73 0 74 0 09 0 091 0 092 0 093 0 09 0 091 0 092 0 093
X R R
(d) (e) (f)
Figure 4. (a) Conductivity and (b) liftoff input distributions with ouput (c) resistance
and (d) reactance output 1D distributions and (e) joint distribution calculated using
PCM. (f) A joint distribution calculated using Monte Carlo simulations is provided for
comparison.
Results for the 2nd order PCM model are presented in Figures 4(c)(f). The results
in Figures 4(c)(d) show that the resistance is basically distributed uniformly and the
reactance is basically distributed as a normal random variable. This relationship is
expected based on the different effect of liftoff and conductivity changes in the
impedance plane. A comparison of the joint distribution calculated using both PCM
and Monte Carlo methods are shown in Figures 2(e) and 2(f) respectively. By
observation, good agreement was achieved using PCM with Monte Carlo methods
while only requiring 6 model calls here. Values for the error between the PCM and
Monte Carlo simulations for the 1st order PCM model, ssr and rssr, were 43.58 and
0.3177 respectively. Error values for the 2nd order PCM model, ssr and rssr, were
15.49 and 0.1129 respectively. Thus, the 2nd order model provides a 64.5% reduction
in the error between the PCM and Monte Carlo simulations with respect to the 1 st order
model. In conclusion, the PCM approach is highly efficient at propagating parameter
variation through models with minimal model calls.
4. Summary
more input variables and also for inverse problems. The robustness of this method for
inverse problems will likely be much more sensitive to choice of order and interaction
terms.
5. Acknowledgements
Funding was provided in part, by the Air Force Office of Scientific Research.
Many thanks to Matt Cherry for reviewing the manuscript.
References
[1] J. S. Knopp, J. C. Aldrin, and K. V. Jata, Computational Methods in Eddy Current Crack Detection at
Fastener Sites in Multi Layer Structures, Nondestructive Testing and Evaluation, 24 (2009), 103 120.
[2] A. Skarlatos, C.G. Pascaud, G. Pichenot, G. Cattiaux and T. Sollier, Modelling of steam generator tubes
inspection in the proximity of support plates area via a coupled finite elements volume integral
method approach, in Electromagnetic Non Destructive Evaluation (XII), Studies in Applied
Electromagnetics and Mechanics, Y. K. Shin, H. B. Lee, and S. J. Song, (Eds.). Amsterdam, IOS Press
(2009), 51 58.
[3] H. A. Sabbagh, J. C. Aldrin, R. K. Murphy, and E. H. Sabbagh, Application of Model Based Inversion
to Eddy Current NDE of Heat Exchanger Tubing, in Electromagnetic Non Destructive Evaluation
(XII), Studies in Applied Electromagnetics and Mechanics, Y. K. Shin, H. B. Lee, and S. J. Song,
(Eds.). Amsterdam, IOS Press (2009), 26 33.
[4] N. Wiener, The homogeneous chaos. American Journal of Mathematics, 60 (1938), 897 936.
[5] R. H. Cameron and W. T. Martin, The orthogonal development of non linear functionals in series of
Fourier Hermite functionals. Annals of Mathematics, 48 (1947), 385 392.
[6] R. G. Ghanem and P. D. Spanos, Stochastic finite elements: a spectral approach. Springer Verlag, New
York, 1991.
[7] D. Xiu and G. E. Karniadakis, The Wiener Askey polynomial chaos for stochastic differential
equations, SIAM J. Sci. Comput., 24 (2002), 619 644.
[8] M.A. Tatang, W.W. Pan, R.G. Prinn, and G.J. McRae, An efficient method for parametric uncertainty
analysis of numerical geophysical model, J. Geophy. Res., 102 (1997), 21925 21932.
[9] www.compumag.org/jsite/images/stories/TEAM/problem15.pdf
Electromagnetic Nondestructive Evaluation (XIV) 149
T. Chady et al. (Eds.)
IOS Press, 2011
2011 The authors and IOS Press. All rights reserved.
doi:10.3233/9781607507505149
Abstract. Eddy current techniques are useful for the detection and the characteriza
tion of aws in conductive structures. The aw signal results from the measurement
of the local punctual value of the perturbed magnetic eld above the metal surface
in the vicinity of the defect by using some magnetic eld sensor (Giant Magneto
Resistance sensor (GMR) or Giant Magnetic Impedance sensor (GMI). This paper
aims to describe a semianalytical numerical model integrated into the CIVA plat
form which is able to compute the aw signal due to a given notch embedded in
a planar stratied media. The inducer may be chosen in a list of available exciting
coils into CIVA. Some numerical experiments and a comparison between simulated
data and experimental data show the validity of the new developments.
Keywords. Eddy current non destructive testing, Greens dyad formalism, Giant
MagnetoResistance sensor, Giant MagnetoImpedance sensor
1. Introduction
Eddy current techniques are very useful for the detection and the characterization of
aws in conductive structures. Eddy currents are induced in the workpiece under test by
carrying a sinusoidal current in one or a set of exciting coils. According to the operat
ing frequency, eddy currents induced in a thin "skin" near the surface of the conductive
media interact with a aw in such a way that currents ow around the ends of the defect
and down the faces of the defect. The deviation of eddy currents results in the perturba
tion of the magnetic eld above the metal surface and in the vicinity of the defect. The
awed signal is analyzed in order to locate and size surface defects. Considering Non
Destructive Techniques (NDT), the variations of the magnetic eld can be detected by
the changes in the impedance of the exciting coil as it is commonly achieved in Eddy
Current (EC) methods. The aw signal is thus due to the global variation of the per
turbed magnetic eld. An alternative NDT method, Alternating Current Field Measure
ment (ACFM) consists in analyzing the aw signal provided by some magnetic eld sen
sor as Giant MagnetoResistance (GMR) or Giant Magneto Impedance sensors [1,2,3,4].
1 Corresponding Author: Denis Prmel, CEA LIST; Email: denis.premel@cea.fr
150 D. Prmel and G. Pichenot / Computation of the Magnetic Field
The aw signal results from the measurement of local, punctual values of the surface
magnetic eld. This signal is obtained by using a tiny inductive sensor or a magnetic
eld sensor [5].
ACFM technique is very useful for sizing any crack with a simple geometry. The
prediction of crack parameters (i.e. the length, the opening and the depth) from the mag
netic eld is of great interest in practice. The resolution of this inverse problem requires
preliminarily solving efciently the forward problem, i.e. to be available to compute the
components of the magnetic eld resulting from the interaction of the crack with eddy
currents generated by some appropriated inducer. This paper aims to describe a numer
ical model based on the volume integral approach [6,7,8] which allows one to compute
the aw signal due to a given notch embedded in a planar stratied media. The inducer
may be chosen arbitrarily from a list of exciting coils [9,10] available into CIVA [11] (a
NDT numerical platform dedicated to several NDT techniques). In ACFM techniques,
rectangular coils or current foils are preferably used in order to generate a uniform cur
rent ow orientated perpendicularly to the length of the aw. This numerical model is
rstly numerically validated by using two kinds of numerical models. These ones differ
from the observation equation which is used for computing the response of the magnetic
sensor. Then, some validation experimental results are presented. These new develop
ments, translated into new functionalities in the last version of CIVA allow one to con
duct some parametric studies such as the liftoff of the magnetic sensor, the tilt of the
inducer, and other parameters which could be able to inuence the sizing of the defect by
using one or two components of the magnetic eld. This work precedes also some work
dedicated to the inversion problem [12].
order to have the capability to compute the punctual values of each component of the
perturbed magnetic eld above the awed material. So, two approaches are considered
for obtaining the numerical solutions of the forward problem. These two approaches are
used for the numerical validation of the new functionalities. Numerical results used for
the numerical validation are compared in section 4 while a real experiment is proposed in
section 5. Finally experimental data are compared with simulated data in order to show
the experimental validation. A short conclusion follows in section 6.
3. The Volume Integral Method and the specic dyad for the problem
The computation of the perturbed magnetic eld above the plate requires computing
rstly the electrical internal eld in a nite domain surrounding the aw. Each layer
may include one nite volume containing an object and the total internal eld in each
layer k results from the interaction from the primary eld Epk (r) in the layer due to the
driving probe and the perturbation eld due to a set of elementary aws, each one being
considered in a single layer l. The total electric eld in each layer, denoted by Ek (r), is
given by:
N
Epk (r) Gkl (r, r ) [l (r )] El (r ) d
(ee)
Ek (r) = j0 (1)
l=1 aw
The contribution is null if there is no aw in the lth layer ((r ) = l ). The superscript
(ee) means that we compute the components of the electrical eld due to an unit point
source which is a solution of:
In this last equation, the observation point r is assumed to be in the layer k while the
source point r is in the layer l. kl stands for the Kronecker symbol and I is the unit
dyad. According to the quasistatic regime, the wave number kk2 in the k th layer is given
152 D. Prmel and G. Pichenot / Computation of the Magnetic Field
by kk2 = j0 k . The Greens dyads satisfy as usually the appropriated boundary condi
tions at the innity in the transverse directions and at the interfaces between two different
layers in the normal direction to the planar surface of the stratied medium. The analyt
ical expression of the Greens dyad corresponding to a multilayered planar medium is
given in [13].
The total internal eld in each layer appears inside and outside the integral. This
state integral equation must be solved numerically by the Moment Method (MoM) [14].
Knowing the total electrical eld in each layer, the perturbed magnetic eld can be com
puted above the plate, in the air region 1, by an observation equation which introduces
the Greens dyad G1k (r, r ). The superscript (me) means that the components of the
(me)
The total perturbed magnetic eld takes into account all the contributions of each
layer k:
N
G1k (r, r ) [k (r )] Ek (r ) d
(me)
B1 (r) = 0 (4)
k=1 aw
This equation has been implemented into CIVA. In order to validate these new de
velopments, some other numerical data have been computed by using another numerical
code which has been previously validated. Let us consider that the components of the
magnetic eld can be estimated by the e.m.f induced in a tiny pickup coil when the sec
tion of the pickup coil tends towards zero. In this case, the response of three receiving
pickup tiny coils is given by the changes in the mutual impedance and can be obtained
via the reciprocity theorem:
N
I1 I2 Z = [l (r)] El (r) Epl (r) d (5)
l=1 aw
where Epl (r ) stands for the primary eld that would be induced by the receiving coil
in the awed region, in the lth layer, if it was driven by a current I2 , and I1 is the
driving current of the exciting coil. The total electrical eld El (r) in the lth layer comes
from the numerical resolution of equation (1). According to the orientation of the pick
up receiving coil, one can obtain an approximation of one component of the magnetic
eld by using the FaradayLorentz law. The numerical results which can be obtained by
these two approaches can be now compared in order to show the validity of these new
functionalities.
4. Numerical Validation
Figure 2 displays the overview and the front view of an ECT conguration consisting of
a non magnetic slab constituted by four layers. The thickness of each layer is xed at
D. Prmel and G. Pichenot / Computation of the Magnetic Field 153
0.8 mm, the values of the conductivities of each layer expressed in MS/m are given by
1 = 0.1, 2 = 0.5, 3 = 1, 4 = 10. The operating frequency is chosen at 100 kHz and
let us consider two notches. The rst one is breaking the surface of the rst layer, it is 2
mm length, 0.3 mm depth and its opening is 0.1 mm. The second aw is embedded at the
top of the third layer, its length is equal to 4 mm, its depth is 0.5 mm and its opening is
also 0.1 mm. These two aws are differently orientated, the rst one is orientated along
the X axis while the second one is orientated along the Y direction. The eddy current
(EC) probe is constituted by two exciting rectangular coils. For each one, the inner width
is 4 mm , the inner length is 8 mm and the height is equal to 0.1 mm. The number of
turns is 5 and nally the width of the current foil is 1 mm.
Figure 2. An overview (on the left) and a front view (on the right) of an ECT conguration used for a numerical
validation.
The EC probe is scanning along three lines in the X direction and three values of
the Y position have been chosen in order to verify the validity of the results in the two
directions of scanning. Figure 3 and 4 show a comparison between the two numerical
approaches resulting from the implementation of equations (4) and (5). The real and the
imaginary parts of the Bx , Bz components are respectively shown. Three curves are dis
played corresponding to three values of the Y position of the probe. These numerical re
sults show a good agreement between the two numerical approaches. In the next section,
an experimental validation is achieved.
5. Experimental validation
Experimental data have been obtained from an experimental setup which is depicted in
Figure 5. Let us consider a non magnetic slab made of Inconel 600, the conductivity is
assumed to be 1.02 MS/m. The operating frequency is 100 kHz. Eddy current are in
duced in the slab by a rectangular coil which is etched on a exible lm. The inner width
of the coil is 23 mm, the inner length is 62 mm and the height is equal to 35 m. The
number of turns is 100 and nally the width of the current foil is 7 mm. The distance from
the bottom of the exciting coil to the slab is about 2 mm. A Giant Magneto Resistance
sensor (GMR) measures the Bx component of the magnetic eld. The EC experimental
signal is obtained by subtracting the response of the probe in the awed region and the
154 D. Prmel and G. Pichenot / Computation of the Magnetic Field
signal without the aw. The distance from the GMR to the target is about 0.1 mm but
this value can be affected by an uncertainty due to the encapsulation of the sensitive part
of the sensor in the chip. The thickness of the slab is 1.55 mm and we consider a rst
notch for a preliminary procedure denoted by "calibration" . Indeed, in most industrial
applications, due to uncertainties in the experimental system, the measured EC signal
has to be calibrated with respect to a reference conguration. At this stage, we choose a
rst notch of 10 mm length, 0.93 mm depth and the opening is 0.1 mm. The EC probe is
scanning a line along the length of the defect and nally simulated data and experimental
data are normalized in order to t the maximum value of the magnitude of the EC signal.
The shape of the two signals are compared in Figure 6(a). In this gure, the real and the
imaginary part of the signals are represented, the same signals are also displayed in the
impedance plane diagram. As it is expected, the comparison shows a very small discrep
ancy between simulated data and experimental data due to, probably, local variations of
the liftoff during the movement of the probe. The complex value of the normalization
constant is xed for all the following results of comparison. Figure 6(b) displays the EC
signals when the probe is scanning perpendicularly to the defect.
In the last experiment, let us consider two identical aws separated by a gap of 3
mm (See Fig. 8(a)). The length of each defect is 10 mm, the opening is 0.1 mm and
the depth is 0.76 mm. Figure 7(a) and 7(b) display two cartographies of the EC signals
D. Prmel and G. Pichenot / Computation of the Magnetic Field 155
Figure 5. A scheme of the experimental setup and an example of the EC probe including the GMR Sensor.
Only one rectangular coil is activated for obtaining experimental data.
(a) the probe scans a line parallel to the notch (b) the probe scans perpendicularly to the notch
collected when the probe is scanning above the awed region. Figure 8(b) shows a slice
view when the probe is moved along a line parallel to the X axis.
6. Conclusion
The Volume Integral Method (VIM) and the Greens dyadic formalism has been applied
in order to compute the response of a magnetic sensor due to an arbitrary number of aws
in a stratied planar structure. The new functionalities which have been developed are
able to compute the punctual value of the perturbed magnetic eld above a awed region.
Two numerical approaches have been compared in order to show the validity of the new
developments and some comparison results between simulated data and some experi
156 D. Prmel and G. Pichenot / Computation of the Magnetic Field
(a) The experimental setup. (b) the probe scans perpendicularly to the notch
Figure 8. A comparison if the EC signals due to two close aws when the probe is scanning along a line
parallel to the notch.
mental data conrm the validity of the global modeling approach. These developments
will be available in the new version of the CIVA platform and some new developments
are today already engaged for considering magnetic materials of constant permeability.
References
[1] R. F. Mostafavi and D. MirshekarSyahkal, "AC Fields Around Short Cracks in Metals Induced by
Rectangular Coils", IEEE Transactions on Magnetics, 35(3), (1999), 20012006.
[2] M. Ravan, S.H.H. Sadeghi and R. Moini, "Neural network approach for determination of fatigue crack
depth prole in a metal, using alternating current eld measurement data", IET Sci. Measurement Tech
nology, 2(1), (2008), 3238.
[3] R. K. Amineh, M. Ravan, S. H. Sadeghi, and R. Moini, "Removal of Probe Liftoff Effects on Crack
Detection and Sizing in Metals by the AC Field Measurement Technique", ", IEEE Transactions on
Magnetics, 44(8), (2008), 20662073.
[4] R. K. Amineha,1, M. Ravan, S.H. Sadeghi R. Moini, "Using AC eld measurement data at an arbi
trary liftoff distance to size long surfacebreaking cracks in ferrous metals", NDT&E International, 41,
(2008), 169177.
[5] B. Marchand, C. Zorni, J.M. Decitre and O. Casula, "Recent Developments of Eddy currents
probes",ENDE 2009, to be published.
[6] J.R. Bowler and L.D. Sabbagh and H.A. Sabbagh, "A Theoretical and Computational Model of Eddy
Current Probes incorporating volume integral and Conjuguate Gradient Methods", IEEE Transactions
on magnetics, 25(3), (1989), 26502664.
[7] J.R. Bowler and S.A. Jenkins and L.D. Sabbagh and H.A. Sabbagh, "Eddy Current Probe Impedance
due to a volumetric aw", Journal of Applied Physics, 70(3), (1991), 11071114.
[8] J.M. Decitre, D. Prmel and M. Lemistre, "3D Modelling of a Magnetooptic imager by a dyadic Greens
functions approach", Review of quantitative Nondestructive Evaluation, 22, ed. D.O. Thompson and
D.E. Chimenti, IOP, (2003), 695702.
[9] T. Theodoulidis, G Pichenot , "Integration of tilted coil models in a volume integral method for realistic
simulations of eddy current inspections", Electromagnetic NonDestructive Evaluation (XI), Studies in
Applied Electromagnetics and Mechanics, A. Tambourino, Y. Melikhov and L. Udpa, Amsterdam, IOS
Press, 2008.
[10] T.P. Theodoulidis and E.E. Kriezis, Impedance evaluation of rectangular coils for eddy current testing
of planar media, NDT E International, 35(6), (2002), 407414.
[11] wwwciva.cea.fr.
[12] D. Prmel and A. Baussard, "Eddy current evaluation of threedimensional aws in at conductive
materials using a Bayesian approach, Inverse problems, 18(6), (2002), 18731889.
[13] W.C. Chew,"Waves and elds in inhomogeneous media", IEEE Press (2nd edition), Piscataway,1995.
[14] R.F. Harrington, "The method of moments in electromagnetics", Journal of Electromagnetic Waves and
Applications, 1(3), (1987), 181200.
Electromagnetic Nondestructive Evaluation (XIV) 157
T. Chady et al. (Eds.)
IOS Press, 2011
2011 The authors and IOS Press. All rights reserved.
doi:10.3233/9781607507505157
Abstract. In this paper the authors focus on detection of wire image quality
indicators, which are widely applied in industrial radiography. Two algorithms for
IQI detection are proposed. The first algorithm is intended for fast and rough
detection of position of IQIs. It is based on evaluation of correlation between the
radiographic image of IQI (together with a plastic case) and the fragments of the
analyzed radiogram. The second algorithm uses the Radon transform applied to the
region of interest (ROI) obtained from the first algorithm to extract the exact
position of wires as well as information on visibility of each wire.
Introduction
1
Corresponding Author: Piotr BANIUKIEWICZ, West Pomeranian University of Technology, al Piastow 17,
70 310 Szczecin, Poland, email: baniuk@zut.edu.pl
158 P. Baniukiewicz and R. Sikora / Automatic Detection and Identication of IQIs in Radiograms
cases, the weld area is inhomogeneous, which is the result of processes which take
place during metal cooling. This is a potential source of unwanted signals and
disturbances that affect all methods of weld testing.
Through the years the technology that utilizes penetrating rays has been the most
important nondestructive testing (NDT) method. Fast and accurate xray converters that
have been developed recently open new possibilities for xray applications and make
this technique as flexible as never before. Digital radiography joins modern digital
image processing algorithms with traditional xray testing method. Together with the
fast development of digital radiography, there arise new nondestructive testing
technologies based on it. One of the biggest problems of assuring the demanded quality
of a product is the need of analyzing a large number of radiograms. In shipbuilding
industry this number can reach even thousands. All radiograms must be analyzed by a
qualified operator in order to find those welds which do not fulfill very restrictive
safety norms. This process takes a lot of time and effort but it tends to be fully
automated.The advanced computerized systems for radiogram analysis are still under
development [1] and the decision taken by a human being is still considered as the most
reliable and trustworthy. The process of radiogram analysis consists of a few stages
which are strictly regulated by adequate norms. The first stage is always accepting or
rejecting the radiogram according to its quality. Too dark, too bright, blurred or noised
radiograms do not provide trustworthy information on internal structure of the
specimen. In this paper the authors present an algorithm for detecting Image Quality
Indicators (IQIs), which are commonly used for obtaining the quality of radiogram.
This algorithm can be applied in the first stage of radiogram analysis process and it
stands as a part of the automatic radiogram analysis system (ISARView) which is being
developed by the authors. The aim of this procedure is only to find IQI in the
radiogram and pass its position to other procedures implemented it the system for
further analysis (e.g. accepting or rejecting radiogram because of its quality). The main
procedure has been implemented in C++ and optimized for 64bit environment.
In radiography, one usually has a wide choice in the parameters of the radiographic
technique. Most of them are regulated by norms but there is still margin for changes
that can be done making the radiogram better or worse. For example, one has a choice
of sourcetospecimen distance, a choice of film type, a choice of film density, a choice
of Xray kilovoltage, Xrays or gammarays etc. The resulting radiograph therefore can
vary between high quality and poor quality; quality in this case meaning the ability to
detect small critical flaws such as cracks [2]. There are a lot of norms, both national
and international, issued for particular applications as well as for a general purpose,
giving advice or instructions on the choice of technique details, in order to ensure good
quality radiographs. The usual method for measuring image quality in radiography is
the use of Image Quality Indicators (IQIs). The quality of a radiographic image can be
estimated in terms of three factors:
Image unsharpness
Image contrast
Image noise
P. Baniukiewicz and R. Sikora / Automatic Detection and Identication of IQIs in Radiograms 159
Figure 1. Exemplary image quality indicators of one wire (a). The effects of increasing geometric
unsharpness in a duplex type IQI. A pair where the individual wire images are merged (c) and one can
no longer see the two separate (b) wires stand for the unsharpness measure.
Using IQI it is possible to obtain all these three parameters and assess the quality of
picture. There are four main types of IQIs. Those are: the wire type, the step/hole type,
the holeinplaque type and the duplex wire type. Here, only the wire type and the
duplex wire type are considered. A typical wire IQI contains six or seven straight wires
of the same or similar material to the specimen having various diameters. The wires are
held parallel to one another in a plastic mount with appropriate identification symbols.
The double wire indicator is quite similar to wire type one. It consists of a series of
pairs of wires of high density material (tungsten and platinum) where each pair of wires
of diameter d is spaced at a distance d apart. Typical construction details and
dimensions are specified in norms [3],exemplary indicators are depicted in Figure 1.
The quality of the radiogram is defined based on the visibility of the wires. The duplex
wire IQI measures only unsharpness as it is explained in Figure 1.
2. Problem statement
The most important problem relating to IQI detection is that the low signal amplitude
of the IQIs contributes to their poor visibility. Additionally, a typical radiogram
consists of nonlinear trend that results from nonhomogeneity of the specimen and its
nonuniform thickness. The amplitude of the trend signal is much greater than the
intensities of particular wires. Thus, the process of the automatic determination of the
Figure 2. Exemplary real radiograms (fragment with IQI only) and corresponding averaged intensity
profiles taken along weld line.
160 P. Baniukiewicz and R. Sikora / Automatic Detection and Identication of IQIs in Radiograms
quality of the radiogram needs special algorithms for accurate detection of IQIs.
Exemplary real radiograms and intensity profiles are showed in Figure 2. The intensity
profiles are averaged from dozens of lines taken along the weld line. As one can see,
the signal from the wires is very poorly visible due to the low SNR. On the other hand,
the plastic case form IQI gives a clear signal that is visible in intensity profile. This
gives a possibility to detect position of IQI by means of image recognition algorithms.
The process of detection of IQIs was divided into two steps. In the first stage only the
presence and the position of the indicator are obtained whereas, the second step
provides exact information on the position of particular wires. Thus, it is possible to
evaluate the quality of the radiogram based on the visibility of wires in the wire type
indicator or merging a pair of wires for the double wire indicator. Under fundamental
initial assumptions for both algorithms lies norm [3] that regulates placement of IQI in
the radiogram. In the case of welds, it is expected that IQIs are always located on the
weld line almost perpendicular to it. Thus, the region of the analysis is limited to the
weld line and a small surrounding area.
The plastic case covering the wires in IQI is the best visible part of the whole
indicator in the radiogram as well as in the intensity profile. This fact is used in the
process of detection and localization of IQI. The algorithm works with averaged
intensity profiles taken along the weld line. The position of IQI is obtained from a
fusion of two methods commonly used in the pattern recognition. Those methods are:
1) to measure the dependence between sought pattern p and signal I using Pearson
productmoment correlation coefficient and 2) evaluating phase correlation between
Fourier transform of pattern p and Fourier transform of signal I. The pattern p is
generated artificially as an average calculated from dozens of intensity profiles of IQIs
obtained from various radiograms. The resulting waveform can be assumed as the best
representative of IQI signal holding general features of that signal.The correlation
coefficients were calculated between the pattern p and the window of size p sliding
over the signal I with step of 1:
n Ii p Ii p
ri = (1)
n I ( Ii ) n p ( p )
2 2 2 2
i
where Ii stands for ith position of the window in the signal I, n is the number of
elements of p and Ii. The position of the maximum of function r is correlated with the
position of IQI in the radiogram. In order to make the detection process more reliable
the second criterion based on phase correlation has been proposed. In image processing,
phase correlation is a method of image registration, and uses a fast frequencydomain
approach to estimate the relative translative offset between two similar images. Here,
this approach was applied to onedimensional waveforms, namely intensity profile I
and pattern p:
{
rp = F 1 e
(
i I p )
} (2)
P. Baniukiewicz and R. Sikora / Automatic Detection and Identication of IQIs in Radiograms 161
Figure 3. Plots of Pearson product moment correlation coefficient r and phase correlation rp evaluated
for intensity I (Figure 1b) and searched pattern p. Function R is fusion of r and rp clearly pointing
position x of IQI (x=1182 for radiogram depicted in Figure 1b).
where F1 stands for inverse Fourier transform and I, p are phases of Fourier
transforms of I and p respectively.
Because of the noise and the linear trend that are mostly present in radiographic
images there is no clear maximum in either of the functions visible. Thus, the final
criterion R was proposed as a fusion of correlation function r and phase correlation rp.
All those functions are depicted in Figure 3. The position of the maximum of R stands
for the position of IQI in the radiogram.
In practice, the smallest element readily visible in the area under inspection is used
to determine IQI sensitivity. Thus, the information on the presence or the position of
IQI in the radiogram is not sufficient to draw any conclusions on its quality. One must
know where exactly the wires are located and which of them are visible. For this reason
the second algorithm has been developed. It uses information on location of IQI
obtained from the first method to find positions of wires in the indicator. One of the
well known methods commonly used for detecting straight lines in pictures is the use
of Radon transform or related Hough transform. Using the Cartesian coordinate system
to describe line integrals and projections, the IQI is represented by a twodimensional
function f(x, y) and each line integral by the (, t) parameters. The equation of the line
(being a particular wire in our case) is:
P (t ) =
( , t ) line
f ( x, y ) ds (4)
what yields :
P (t ) =
f ( x, y ) ( x cos + y sin t ) dxdy (5)
where stands for the projection angle, t defines the position of the integral line on the
projection and P(t) is known as Radon transform of function f(x, y) [4].
The most important feature of Radon transform is amplifying straight lines to be
found in pictures invariantly to their rotation angle. This makes it especially useful in
the IQI detection process. Figure 4 shows Radon transform (sinogram) of the
162 P. Baniukiewicz and R. Sikora / Automatic Detection and Identication of IQIs in Radiograms
Figure 4. a) The IQI wires pattern generated artificially according to norms, b) Radon transform of part
of the radiographic image and c) integrated intensity for =6 as function of t. d)Results of detection of
IQI position.
radiogram section that contains IQI. It is clearly visible on the line integral that peaks
from the wires have been amplified in comparison to those original ones depicted in
Figure 2. The detection algorithm of these peaks was realized in the same way as in the
case of the first step described above. The correlation function r (1) is evaluated for all
angles from the sinogram and an artificially generated pattern representing the ideal
IQI signal. The pattern is a sum of six Gaussian functions with increasing amplitudes
(Fig. 4a). But now, unlike the previous case, the wires are under consideration and not
the plastic case. Therefore, the maximum of r is found for two arguments of function
P(t) and it stands for the best matching of the whole artificial IQI pattern to the line
integral for given (Fig. 4c). The angle is the rotation angle of the IQI relating to x
axis whereas t is the radial coordinate of the first wire (beginning of the pattern) in the
indicator. Having and t, the first wire can be expressed as linear function y(x) defined
for analyzed region of interest, which has been found by first algorithm, in the
radiogram:
y ( x ) = tan + ( x x0 t cos ( ) ) + y0 (6)
2
where x0 and y0 stand for the centre of the image. The result of the wires detection is
presented in Figure 4d.
Conclusions
Radiography is the most popular and well known method of nondestructive testing
widely applied in industry. Information on the quality of radiographic picture is
especially important in the case of the method of nondestructive testing. The algorithms
for IQI detection presented in this paper have been applied in the automatic system of
radiogram analysis (ISARView) in the first stage of the whole process of the analysis.
Taking into consideration the fact that even thousands of radiographic pictures can be
taken during manufacturing of e.g. a ship, the methods presented in this paper can
significantly speed up the whole process of nondestructive testing in various branches
of industry.
P. Baniukiewicz and R. Sikora / Automatic Detection and Identication of IQIs in Radiograms 163
Acknowledgments
This work was conducted in a framework of the research project The intelligent
system for radiograms analysis, supported by the Polish Ministry of Science and
Higher Education. Grant no. NNR01003706/2009 (20092012).
References
[1] European Commission sponsored project Development of novel digital radiography technology. To
facilitate the traditionally less research intensive inspection industry sector change from manual film
radiography to automated digital, contract No NMP2 C 2005 515746.
[2] R. Halmshaw, T. Kowol, Image quality indicators in industrial radiography, IE NDT LTD technical
materials
[3] European norm EN 462 part 1:1994
[4] A. G. Rann, and A. I. Katsevich, The Radon Transform and Local Tomography, Boca Raton, FL: CRC
Press, 1996.
164 Electromagnetic Nondestructive Evaluation (XIV)
T. Chady et al. (Eds.)
IOS Press, 2011
2011 The authors and IOS Press. All rights reserved.
doi:10.3233/9781607507505164
Abstract: In this work, a code of finite element method is developed for numerical
simulation of the acoustoelastic effect in pre stressed media. The possibility of
assessing the stress status, by using the EMAT (electromagnetic acoustic
transducer) receiver for precise measurement of LCR (refracted longitudinal) wave,
is investigated. Besides that, the relationship between the polarization of transient
Rayleigh wave and the state of stress is also predicted by simulation.
Introduction
Ultrasonic NDE is one of the most promising methods for quantitative determination of
residual stress in materials. And, most of ultrasonic methods for stress measurement are
based on the effect of acoustoelasticity. The acoustoelastic effect, related to the change
of ultrasonic wave speed and the polarization of Rayleigh wave, is a well established
phenomenon that has been extensively reported in lots of literatures [14]. At present,
the residual stress measurement by using acoustoelastic effect is mainly studied with
experiment method. And simultaneously, the numerical simulation method can provide
an another useful tool for the research of stress measurement with ultrasonic wave. In
this work, the possibility of numerical simulation of acoustoelastic effect related to the
stress status was investigated by using FEM method.
Now in the practical application, the stress measurement using ultrasonic wave is
mainly based on measuring the variation of ultrasonic wave velocity by timeofflight.
As the LCR wave is more sensitive to the inplane stress and easy to yield, it is widely
used in the subsurface stress measurement [9]. In this work, a kind of noncontact
transducer named EMAT is proposed to precisely measure the acoustoelastic effect on
LCR wave.
There are some literatures have demonstrated that the polarization of Rayleigh
wave related to the state of stress is usually more sensitive than the wave speed. The
relationship between Rayleigh wave polarization and state of stress in homogeneous
material was investigated by Junge et al[3]. However, in their analytical model, only a
steady sinusoidal Rayleigh wave propagates in the material. Actually it is very difficult
to generate a pure steady Rayleigh wave in the media, the Rayleigh wave in sinusoidal
mode is usually interfered by the longitudinal and shear wave. In that case, it is
1 Email: peicx2009@gmail.com
C. Pei and K. Demachi / Numerical Simulation of Acoustoelastic Effect in PreStressed Media 165
An ultrasonic wave through a stressed body would give rise to further stresses, and the
traditional theory of acoustoelasticity applied in the unstressed medium cannot be used
here. Instead, based on the deformation process, Duquennoy et al. defined a theory of
three states of a body, shown in Figure 1 [3, 4]. A solid body undergoes a series of
deformations from a stress free state to a static deformation or a dynamic deformation.
The position vector defines the position of a point in the natural state of zero stress
and zero strain. X defines the position of a point in the initial state when the body
undergoes a static deformation due to residual stress during the manufacturing
processes or due to the applied stresses. Similarly x defines the position in the final
states when a dynamic deformation such as the ultrasonic wave through the body. The
displacement of a point from one state to another can be described mathematically as:
u i ( ) X ; u f ( ) x ; u ( ) x X uf ui
Based on the threestate theory, the controlling equation of the acoustic wave in a
solid with initial stress can be written as
( ) X uX
2
i u I 2u I
IK JL + C IJKL
K
+f i
0 (1 NN ) (1)
J L t t 2
where i is the Cauchy stress tensor in the initial state, i the initial strain tensor and
0 the mass density in the natural state, is the acoustic damping coefficient, f is
the external force loading on the body. And for isotropic material the elastic constant
can be expressed as
166 C. Pei and K. Demachi / Numerical Simulation of Acoustoelastic Effect in PreStressed Media
C IJKL IJ KL + ( IK JL + IL JK ) + [( + v1 ) IJ KL
i
+ ( + v2 )( IK JL + IL JK )] NN + 2( + v2 )( IJi KL + KL
i
IJ ) (2)
i i i i
+ 2( + v3 )( JL + JK + IL + IK )
IK IL JK JL
where and are the well known Lame constants and vi (i 1, 2,3) are the
thirdorder elastic constants (TOE constants).
According to FEM, Eq. (1) can be calculated by solving the discrete wave motion Eq.
(3)
[ M ]{U&&} + [C ]{U& } + [ K ]{U } {F } (3)
where [ M ] denotes the mass matrix, [C ] the damping matrix, [ K ] the stiffness
matrix, {U } the nodal displacement and {F } the force vector. Eq. (3) is rearranged
into the following equation by using the explicit integration method in timedomain [7]
{U& }t +t = {U& }t t 2t [ M ]1 [C ]{U& }t 2t [ M ]1 [ K ]{U }t + 2t [ M ]1 {F }t
(4)
{U& }t +t + {U& }t
{U }t +t = {U }t + t
2
f
where B e is the reaction magnetic flux density at a position of the pickup coil
produced by the transient eddy current, N is the number of turns of the pickup coil,
and Si is a surface surrounded by the wire of ith pickup coil. Substituting
B ef A ef into Eq. (5) and applying BioSavat's law, we have,
C. Pei and K. Demachi / Numerical Simulation of Acoustoelastic Effect in PreStressed Media 167
(6)
In order to demonstrate the feasibility of this simulation method, the LCR waves
acoustoelastic effect related to the inplane stress was simulated in a threedimension
model, shown in Figure 2. As the piezoelectric (PZT) ultrasonic transducer is better as
generator; and EMAT is better as detector. A PZT angle transducer was used to
generate the LCR wave, and an EMAT probe was used to receive the time behavior of
the LCR wave signal for different initial stresses x , uniaxially directed along the x
axis. As shown in Table 1[6, 8], two kinds of material have been used here. Where, the
sensitivity constant kc is defined as a relative change in wave speed per unit change in
stress. When the longitudinal wave speed through a wedge is 2720 m/s, based on
Snells law, the incidence angle w should be set at about 26o to generate the LCR
wave in the model.
To give a preliminary judgment of the validity of this method, the ultrasonic pulse was
excited by one damped cycle of a 1 MHz sinusoidal force loading on a wedge with an
angle. The simulation result of the ultrasonic wave field in a cross section at a given
time is shown in Figure 3. The time behavior of the EMAT coil voltage is plotted in
z y
x
Figure 4. Compared to the Ultrasonic wave field, it can be seen that the first pulse in
Figure 4 is the receiving signal of LCR wave.
In Figure 5 and Figure 6, the closeup of the LCR wave signals for two initial
stresses in Al Alloy and rail steel is provided. The acoustoelastic effect with a small
phase delay induced by the stress can be observed. It can be seen that the variation in
LCR wave speed due to acoustoelastic effect in rail steel is much smaller than that in
Al alloy. The relative change in LCR wave speed for different initial stresses is
simulated, shown in Figure 7. It can be seen that the simulation results show a very
good agreement with the theoretical values.
Figure 7. Relative change of LCR waves speed in Al Alloy and rail steel
When Rayleigh wave travels across the surface of isotropic solids, the surface particles
move in ellipses in planes normal to surface and parallel to the direction of the
propagation. The polarization of Rayleigh wave is defined as the ratio between the
maximum inplane, and the maximum outofplane displacement components. As
shown in Eq. (7), the sensitivity constant k p is defined as a relative change in
polarization per unit change in stress
0
k p (7)
0
where, the subscript 0 indicates the parameter at zero stress. In this work, the
polarization of pulse Rayleigh wave excited by the PZT transducer was investigated by
170 C. Pei and K. Demachi / Numerical Simulation of Acoustoelastic Effect in PreStressed Media
FEM. In order to make a comparison with the results based on the theory of steady
Rayleigh wave, three kinds of materials used in Junges paper were chosen here. As
shown in Table 2, k p' is the sensitivity constant in polarization of the Rayleigh wave
in steady mode[6].
Table 2 Material parameters
Material / kg m 3 / Pa / Pa v1 / Pa v2 / Pa v3 / Pa k 'p / MPa
Al Alloy 2719.0 49.1E9 26.0E9 379.0E9 198.0E9 80.0E9 6.673E 5
D54s
Al (99.3%) 2710.0 50.9E9 26.1E9 27.7E9 91.7E9 89.3E9 3.5809E 5
Rail Steel 7800.0 115.8E9 79.9E9 36.0E9 266.0E9 178.5E9 9.011E 6
The model is the same as in Figure 2. In order to yield a pure surface wave, the
o
angel of incidence w is set at about 70 . The simulation result of Rayleigh wave
field in a cross section at a given time is shown in Figure 8. Figure 9 presents the in
plane and outofplane displacement components at point P on the surface of the
unstressed Al alloy model. As can be seen, the two components have a phaseshift
of / 2 . The polarization of Rayleigh wave is presented in Figure 10.
Figure 11. Polarization of Rayleigh wave for different initial stresses in Al alloy model
Figure 11 presents the polarization of Rayleigh wave for different initial stresses in
Al alloy model. It is observed that the polarization increases with increasing tension
and decrease with increasing compression. Figure 12 presents the relative change of
polarization for different initial uniaxial stresses x . It shows that the relative change
in polarization of pulse Rayleigh wave is much larger than the reference values based
on the theory of steady Rayleigh wave.
4. Conclusion
Figure 12. Comparison of acoustoelastic effect of pulse Rayleigh wave with reference values of steady
Rayleigh wave
Acknowledgement
References
[1] Shamachary Sathish, Thomas J. Moarn, Residual stress measurement with focused acoustic wave,
Materials Science and Engineering A 399 (2005), 84 91.
[2] Development of non contact stress measurement system during tensile testing using the
electromagnetic acoustic transducer, NDT&E 39 (2006), 299 303.
[3] Michael Junge, Jianmin Qu, Relationship between Rayleigh wave polarization and state of stress,
Ultrasonics 44 (2006), 233 237.
[4] Dequennoy M. Ouaftouh M., Ultrsonic evaluation of stresses in orthotropic materials using Rayleigh
waves, NDT&E 32 (1999), 189 199.
[5] Shailesh Gokhale, Determination of applied stress in rails using the acoustoelastic effect of ultrasonic
waves, Master's thesis, Texas A&M University, Dec. 2007.
[6] Michael Junge, Measurement of applied stress using polarization of Rayleigh waves. Masters thesis,
School of Civil and Environmental Engineering, Georgia Institute of Technology, August, 2003.
[7] Cuixiang Pei, Zhenmao Chen, Development of simulation method for EMAT signals and application to
TBC inspection, Int. J. Appl. Electromagn. Mech., Vol.33, 2010.
[8] R.T. Smith, R. Stern, and R.W.B Stephens. Third order elastic module of polycrystalline metals from
ultrasonic velocity measurements. Journal of Acoustical Society of American, 40(5):1002 1008, 1966.
[9] Don E. Bray, Wei Tang, Subsurface stress evaluation in steel plates and bars using the LCR ultrasonic
wave, Nuclear Engineering and Design, 207:231 240, 2001.
Electromagnetic Nondestructive Evaluation (XIV) 173
T. Chady et al. (Eds.)
IOS Press, 2011
2011 The authors and IOS Press. All rights reserved.
doi:10.3233/9781607507505173
Abstract. BEMLAB [1] is the open source project implementing Boundary Ele
ment Method (BEM) [2,3] comprehensively. BEMLAB binary packages and the
source code are distributed under GNU LGPL (Lesser General Public License) li
cense terms. The project provides universal library and the reference application,
which is the easiest way for solving problems using BEM. The article presents ex
ample BEMLAB application in modelling comb capacitors used in MicroElectro
Mechanical Systems (MEMS). Finally capacitance will be calculated using BEM
results. Particularly Asymptotic Boundary Conditions (ABC) will be used to model
external space of the capacitor (the geometry of innite halflane) as an internal
BEM problem. This approach will allow to calculate total capacitance with dis
persed capacitance included. Results will be compared with the simplied model.
Keywords. BEMLAB, BEM, Boundary Element Method, MEMS, MicroElectro
Mechanical Systems, Comb capacitors, Comb actuators, Asymptotic Boundary
Conditions, Open source, Objectivity, Multithreading
Introduction
Micro ElectroMechanical Systems (MEMS) are very small devices which dimensions
vary in the range from 0.1m to 0.1mm. They exchange energy either from electri
cal to other type for example mechanical, or mechanical to electrical. The most com
monly known MEMS devices are inkjet printers heads or accelerometers used in airbags.
MEMS design process involves two modelling types electrostatic and mechanical. This
article incorporates electrostatic one and will concentrate on comb capacitors used in
MEMS devices. The comb capacitors design stage problem requires to calculate capac
itance C [4]. One method of calculating capacitance is usage of the simplied model
based on the ideal parallelplate capacitor. However this approach does not incorporate
dispersed capacitances appearing in the corners and thus results are not precise enough.
1 http://bemlab.org/
2 Email: p.wieleba@iem.pw.edu.pl
174 P. Wieleba and J. Sikora / BEMLAB
The most precise modelling method which can be used for associating real physical
properties and processes is the eld method. Field method applied in electromechanics
satises the Laplace equation. The most common numerical method of solving Laplace
equation is the Finite Element Method (FEM) [5,6]. Many open source and commer
cial packages, which can be used to solve problems using FEM exist [7]. However this
method is a domain one and requires discretization of the whole domain. While estimat
ing condenser capacitance it is a worthless overhead. Therefore it is tempting to use the
boundary method. However, there was no universal package which made it possible until
BEMLAB [8,1] came into existence. Now it is possible and fairly easy.
BEMLAB [1] is the open source project implementing Boundary Element Method
(BEM) [2,3] comprehensively. BEMLAB binary packages and the source code are dis
tributed under GNU LGPL (Lesser General Public License) license terms. The project
provides universal library and the reference application, which is the easiest way for
solving problems using BEM. There are also auxiliary programs provided to facilitate
engineers tasks. A very important factor for the end user is the input/output data for
mat, therefore the title library uses format compatible with Matlab Mles, Octave script
les [9] and Scilab sciles [10]. The project is objective and is developed using Unied
Modeling Language. BEMLAB is implemented in C++. BEMLAB uses multithreading
(MT) to speed up calculations on multiprocessor and multicore platforms nowadays
CPUs have at least 2 cores. Threads introduced in C++0x specication are used. As an
example GNU Compiler 4.4 or newer provides C++0x API.
BEM is the numerical method designed for solving physical problems described by Par
tial Differential Equations (PDE) but expressed as Boundary Integral Equation (BIE):
G
c + d = Gd + f Gd (1)
n n
2 (x) = 0 (2)
BIE (1) can be used to solve Laplace equation if proper Green function is applied. Green
function (fundamental solution) for Laplace equation in 2D space is:
P. Wieleba and J. Sikora / BEMLAB 175
1 1
G(R) = ln (3)
2 R
where R is the distance between the observation point for which the BIE equation is be
ing set and the integration point of the element being integrated using the Gauss Quadra
ture.
Comb capacitors may have various shapes, dimensions and types. The article
presents vertical comb capacitors with angular actuation. They are called Angular Verti
cal Comb (AVC) capacitors. Angular vertical comb capacitor nger in a tilted position
was presented in gure 2 on the left side. Whereas on the right side it was presented in the
rest position from the top perspective. Example dimensions of the AVC capacitor used
Table 1. Angular Vertical Comb condenser ngers dimensions.
Description Symbol Dimension [m]
Finger length l 150
Finger width w 4
Finger height h 40
Gap between ngers d 3
Number of ngers in comb capacitor N 12
for driving micro mirror are gathered in table 1 based on [11]. The capacitor is powered
with the voltage of 20V, where potential of 10V is applied to one electrode and 10V to
the second one.
176 P. Wieleba and J. Sikora / BEMLAB
Figure 2. Angular Vertical Comb capacitor nger in a tilted position from the side perspective (left) and in a
rest position from the top perspective (right).
In this section the problem for the AVC capacitor is dened and solved in order to calcu
late capacitance C. Field model is created for the boundary element method (BEM). All
calculations are done using BEMLAB software.
Firstly the geometry have to be dened. Here 2D model is considered. Gap between
electrodes presented in gure 3 is to be modeled. The model was dened in the way so
dispersed capacities are included in calculations. Due to the existence of the open space,
an articial boundary (1t) have to be introduced. The gap is the domain of the problem
P. Wieleba and J. Sikora / BEMLAB 177
marked by (1) . (1) is the boundary around the examined domain (1) . Only nger
halves are modeled because the whole structure of the comb capacitor is repeatable, as
it consists of N nger pairs. Therefore boundary conditions of the second type n = 0
are applied on the left and right side of the presented model. The vertical dashed lines on
the left and right of the gure 3 are placed in the center of particular ngers. The hori
zontal dashed line at the bottom of the gure also cuts particular ngers into two halves.
Therefore zero boundary conditions of the second type (Neumann Boundary Conditions)
n = 0 are also applied. Electrodes are conductors, therefore Dirichlet Boundary Con
ditions (boundary conditions of the rst type) are applied on the boundary of electrodes.
On the left nger (yellow) potential = V is applied and on the right nger (right)
potential = +V is applied.
Almost all boundary conditions of the internal problem were dened. The only prob
lematic boundary is (1t) . The width of the (1t) is a as marked at the gure 3 and is de
ned by d(t) length away from the electrodes. The (1t) boundary placed in the innity:
d(t) (4)
d(t) a (5)
which will increase the domain size and number of boundary elements used. Another one
which will be used in the model, is using of Asymptotic Boundary Conditions (ABC)
[12, chapter 4].
Asymptotic Boundary Conditions [12, pp. 5273] for the halflane are in the following
form:
+ =0 (6)
y a
where a is the width of the halflane. According to the BEMLAB software, the follow
ing notation has been used:
= +0
y a
(7)
= m + n
y
w w
a= +d+ =d+w =3+4=7
2 2
(8)
m = = = 0, 44879
a 7
n=0
When appropriate model denition is created, mesh of the geometry has to be generated.
Figure 4 presented the boundary (1) mesh of the model dened in gure 2. It was
0 5.75 11.5 17.2 23
3.5 6 5
1.75
8 7
1 2
1.75
Z Y
3.5 X 3 4
Figure 4. Boundary mesh of the AVC capacitor air gap between electrodes for the problem geometry dened
in gure 3 with d(t) = 3.
generated using Gmsh software [13]. Corner geometric nodes has been marked with their
numbers from 1 to 8. The modeled problem is 2D, therefore z coordinate equals 0.
Boundary element method requires boundary elements normal vector n to be di
rected outwards the analyzed domain . Therefore special attention has to be brought to
the process of generating boundary mesh and after creating one its correctness has to be
checked. BEMLAB software has the functionality of calculating normal vectors for any
3.5
1.5
x
1.5
3.5
0 20 23
y
dened input mesh. Figure 5 presents normal vectors drawn against the generated mesh.
As can be seen, blue normal vectors n are directed outwards the domain (1) enclosed
P. Wieleba and J. Sikora / BEMLAB 179
by the boundary domain (1) marked with the solid line. The following BEMLAB com
mand was used to calculate normal vectors of mesh dened in the input le dened by
parameter i:
% obem_solve i comb_cap_mesh.m m 7 o normal_vecs.m
Results are written to the output le dened by the command parameter o.
The mesh and boundary conditions described in the previous section are gathered in
one le:
comb_cap_mesh.m
Values of potential and its normal derivative
n are the direct result of BEM. To start
BEM calculations and obtain and n the following command has to be issued:
% obem_solve i comb_cap_mesh.m m 1234 o solution.m
To visualize layout of potential along the domain (1) , values of potential in
internal nodes have to be calculated. This stage is optional and not required to calculate
capacitance C, but is helpful to detect big calculation errors. This stage can be processed
when potential and its normal derivative n are known on the whole boundary . To
proceed calculations the following shell command has to be issued:
% obem_solve i comb_cap_mesh.m S solution.m m 5 \\
I comb_cap_internal_nodes.m o internal_potential.m
where i denes the input le with the input mesh with boundary conditions and
material parameters (as in the previous paragraph), S denes the input le with the
solution on the boundary (potential and its normal derivative n ), I denes the
input le with the internal nodes coordinates, o denes the output le with calculated
internal potential .
Figure 6. Layout of potential inside the AVC capacitors air gap dened in the gure 4.
Figure 6 presents the layout of calculated potential inside the air gap (domain (1) )
calculated using the previous command.
180 P. Wieleba and J. Sikora / BEMLAB
where Q charge gathered on one electrode, U voltage between two electrodes of the
capacitor.
The charge Q can be calculated using the following integral:
Q = d = Dn d = En d = d (10)
n
where the surface charge density, Dn electric displacement eld , En electric
eld, electric permittivity, and Dn = Dn .
Finally the capacitance can be written in the following form:
1
C = d (11)
U n
Normal potential derivative n on the boundary is the direct result of BEM, there
fore it can be directly inserted into equation (11). Figure 7 presents the layout of poten
tial normal derivative
n on the boundary of the right electrode (starting from node no.
1 to node no. 3 of mesh from gure 4). Potential normal derivative n , presented on the
12
10
potential normal derivative
0
0 10 20 30 40 50 60 70
element number
vertical axis, matches the number of the element presented on the horizontal axis.
Firstly the charge Q was calculated for the half of one nger (as presented in the
model):
Q = 144.863 (12)
The integration was done using BEMLAB software by issuing the following command:
P. Wieleba and J. Sikora / BEMLAB 181
% obem_solve i int_electrode.m m 8
integratedValue=[ 144.863 ];
where i denes the input le with geometry to be integrated and values of integrated
function in the geometric nodes.
Then the capacitance per unit length for half of one nger, using the eld model with
ABC applied, was calculated as follows:
Q 144.863
CI(half) = = = 7.24 [F]
U 20 (13)
= 7.24 106 8.854187817 1012 = 64.104 [aF]
The capacitance between two electrodes of one nger (2 halves) of the comb capac
itor with the length l = 150 (the z dimension) and the gap width d = 3 is as follows:
The total capacitance C (total) for the whole comb capacitor with N = 12 ngers is
2N times bigger than capacitance between two electrodes of one nger, as one nger has
two sides (vertical). Therefore the total capacitance can be calculated as follows:
In the previous section calculations using the eld model were proceeded. In this section
the calculation of capacitance using the simplied model will be presented for the com
parative purposes. The model for half of the nger as presented in previous section and
on gure 2, consists of one parallel plate capacitor, which can be calculated as follows:
C h(l d) 40(150 3)
C (nger simple) = = = 106 = 17.354 [fF] (16)
S d 3
9. Summary
Universal, open source, objective and multithreaded BEMLAB software can be suc
cessively used for modelling MEMS devices. Field modelling using boundary element
method includes dispersed capacitance which occurs in comb capacitors, while simpli
ed models not. The difference between calculated capacitance values using eld and
simplied model is 9%:
BEMLAB software also allows three dimensional modelling for more complicated struc
tures or in a tilted position, which cannot be modeled in 2D. BEMLAB input/output text
le format implementation compatible with Matlab Mles, GNU Octave script les and
Scilab .sci les makes cooperation with that software straight forward. Moreover engi
neering tools are provided to simplify cooperation with mesh formats provided by Gmsh
[13] or Netgen [14].
182 P. Wieleba and J. Sikora / BEMLAB
References
[1] P. Wieleba and J. Sikora: Open Source BEM Library, Advances in Engineering Software 40 (8), 564569,
2009.
[2] L. C. Wrobel and M. H. Aliabadi: The boundary element method, vol. 1 and 2, Wiley, 2002.
[3] J. Sikora: Boundary Element Method for Impedance and Optical Tomography, Ocyna Wydawnicza
Politechniki Warszawskiej, 2007.
[4] G. Of, M. Kaltenbacher and O. Steinbach: Fast multipole boundary element method for electrostatic eld
computations, The International Journal for Computation and Mathematics in Electrical and Electronic
Engineering 28 (2), 304319, 2009.
[5] O. C. Zienkiewicz and R. L. Taylor: The nite element method, vol. 1, ButterworthHeinemann, 2000.
[6] R. Sikora: Teoria pola elektromagnetycznego, Wydawnictwa NaukowoTechniczne, 1997. In Polish.
[7] J. Mackerle: Objectoriented programming in FEM and BEM: a bibliography (19902003), Advances
in Engineering Software, 35, 325336, 2004.
[8] BEMLAB homepage, http://bemlab.org/
[9] Octave homepage: http://www.octave.org/
[10] Scilab homepage: http://www.scilab.org/
[11] W. Piyawattanametha, P. R. Patterson, D. Hah, H. Toshiyoshi and Ming C. Wu: A Surface and Bulk
Micromachined Angular Vertical Combdrive for Scanning Micromirrors, IEEE Optical Fiber Commu
nications Conference vol. 1, 251253, 2003.
[12] S. Gratkowski: Asymptotyczne warunki brzegowe dla stacjonarnych zagadnien elektromagnetycznych
w obszarach nieograniczonych algorytmy metody elementw skonczonych, Wydawnictwo Uczelniane
Zachodniopomorskiego Uniwersytetu Technologicznego w Szczecinie, 2009. In Polish.
[13] Gmsh homepage: http://geuz.org/gmsh/
[14] Netgen homepage: http://www.hpfem.jku.at/netgen/
Electromagnetic Nondestructive Evaluation (XIV) 183
T. Chady et al. (Eds.)
IOS Press, 2011
2011 The authors and IOS Press. All rights reserved.
doi:10.3233/9781607507505183
Abstract. The present work concerns the development of pulsed eddy currents
non destructive testing techniques applied to defects analysis. The impulse currents
are produced with capacitor discharge bank [1], [2]. Coupled circuits method
employed is based on the mutual inductances calculation [3] and associated to
Kirchhoff laws yields to a real algebraic equation system which is solved at each
time step. The model developed is used to the defects analysis of a cylindrical
device [4].
Introduction
The eddy currents sensors are used in diverse industrial domains which concern the
evaluation of the physical properties of materials, the load dimensions, defect detection
in a piece etc.
Pulsed eddy current (PEC) sensing is a new emerging technique and has gained
considerable research attention in recent years [5]. Their first applications have been
developed for the measure of material's thickness. The possibility to inject an important
energy in short time with rich spectrum of low frequencies have induced an increase in
use of them in the detection of multidefects in pieces. This technique is used to inspect
the riveted assembly, the slants and the objects of irregular surfaces, it holds the
voltages of becoming the primary means of detecting corrosion in multilayered
structures and to inspect, characterize the coatings [6],[8],[9],[10].
The present work deals with the study of the differential eddy currents sensor in
pulsed diet by using a halfanalytical method coupled with an equivalent electric circuit
[2], [11]. The impulse currents technique is chosen instead of multifrequency study
which induces some difficulties when the number of frequencies exceeds four [7],[10].
In this case the use of impulse currents method present more advantages. The
application of the method concerns the detection of two defects realized in a
conducting cylinder scanned with differential probe. A cylinder with two layers is
investigated too and a comparison with existing data is realized.
1
Corresponding Autor: Hassane MOHELLEBI, aLaboratoire de Gnie Electrique, Universit de Tizi Ouzou,
BP 17 RP, ALGERIA ; e mail :mohellebi@yahoo.fr
184 H. Mohellebi et al. / Use of HalfAnalytical Method for the Detection of Defects in Diet Pulses
The halfanalytical method used is the coupled circuits one based on a discretization of
the inductor and load sections in (r,z) plane. The elementary sections constructed
permit one to define elementary coils with associated elementary parameters (figure 1).
Electrical parameters of the device as mutual and appropriate inductances exploits an
integral calculus.
Ub1 Ub2
Inductor
Rc1 Lc1
Rcm L cm
Shortcircuited load
Ub1, Ub2 are electric voltages of differential probe. Indices n and m are the maximum
numbers of elementary coils of the inductor and shortcircuited load.
This process allows expressing the resistance of every elementary turn, the self
inductance and the mutual inductances between the various turns [12]. The electric
equations of the load and the inductor are given by Eqs (1) and (2) as following:
dI m dI N dI
i + j
0 = Ri I i + L M ij + M ik k (1)
i dt j =1 dt k =1 dt
j i
dI N
dI m dI
k + q (2)
U = Rk I k + L M kq + M ki i
k k dt q =1 dt i =1 dt
qk
H. Mohellebi et al. / Use of HalfAnalytical Method for the Detection of Defects in Diet Pulses 185
Ri, Li and Ii; Rk, Lk and Ik are the resistance, the constant inductance and the current.
Indices i and k are related to load and inductor respectively.
Mij: mutual inductance between both elements i and j of the load.
Mik: mutual inductance between load and inductor elements.
Ij: current in the element j of the load; Ik: current in the element k of the inductor.
Mkq: mutual inductance between both elements k and q of the inductor;
Mki: mutual inductance between the element k of the inductor and the element i of the
load.
Ik: current traversing the element k of the inductor. Iq: current traversing the element q
of the inductor.
Several methods have been proposed for the calculation of the parameters of the
models. We can classify these in two big classes [4], [3]:
 Methods based on the magnetic vector potential
 Methods based on the mutual inductances calculation
2 2 (3)
M = a .b ( k ) K ( k ) E( k )
k k
4. a.b (4)
k2 =
(a + b)2 +h2
c a
M1
b
b
h
8.a (5)
L = a ln 1.75
Rc
Rc : conductor radius
1.1.3. Resistance
The turn resistance of elementary turn coil is given by Eq. (6):
"
R= (6)
S
The relative electric model to the diet pulses is comprised of the equivalent circuit of
the sensorload coupled to the energizing circuit composed of a capacity and a
resistance. The equivalent electric circuit of the system is given by the following
Figure 3 [2]:
ic(t) Rext
Sensor system
Lb
+ Vc(t)
C
Rb
dic (t )
V C (t) (Rext + Rb )ic (t) Lb =0 (7)
dt
dvc (t )
iC (t ) = C (8)
dt
Rb and Lb are the resistance and the inductance of the system loadinductor; Rext:
external resistance; C: capacity.
H. Mohellebi et al. / Use of HalfAnalytical Method for the Detection of Defects in Diet Pulses 187
The resolution of Eq. (7) combined with Eq. (8) yields to the following solution:
1 2 (9)
ic ( t ) = C V0 ( )( e t e t )1 2
1 2
The application considered in the present work is the one studied in [5]. It makes use of
a differential probe used for the tubing inspection generator. The defects realized on the
conducting cylinder are axisymmetric with resistivity W= 106[.m]. The physical and
geometrical characteristics corresponding to differential probe are: coil width: 1.75*10
3
[m], distance between coils: 0.5*103, high in z direction: 0.75*103 [m], internal
radius: 7.5*103 [m], magnetic permeability: 4X*107 [H/m] and the electrical
conductivity: 5.59*10+7 [.m]1(Figure 4).
z
J +J
Inductor turns
Cylindrical object r
+J J
In this section we consider a healthy load (flawless) and loads with different types of
defects (internal, in depth and external defect). The electrical circuit parameters used
are: C = 107 F, Rext = 0.55 Y, Req= 1.6*103 Y, Leq= 1.0054*108 H, V0 = 1V.
188 H. Mohellebi et al. / Use of HalfAnalytical Method for the Detection of Defects in Diet Pulses
0.04 0.045
0.035 0.04
0.035
0.03
0.03
0.025
0.025
z [m ]
z [m ]
0.02
0.02
0.015
0.015
0.01
0.01
0.005 0.005
0 0
75 8 8.5 9 9.5 10 10.5 7.5 8 8.5 9 9.5 10 10.5
r[m] 3 r[m] 3
x 10 x 10
0.045
0.045
0.04 0.04
0.035 0.035
0.03 0.03
0.025 0.025
z [m ]
z [m ]
0.02 0.02
0.015 0.015
0.01 0.01
0.005 0.005
0 0
7.5 8 8.5 9 9.5 10 10.5 7.5 8 8.5 9 9.5 10 10.5
r[m] 3 r[m] 3
x 10 x 10
Figures 6 and Figure 7 represent the variation of the current for different defect types:
0.07 0.1
Defect thickness 20%
Healthy load
0.06 Defect thickness 40%
Load with external defect 0.08 Defect thickness 80%
Load with defect in depth
0.05 Load with internal defect
0.06
0.04
C urrent[A ]
Current[A ]
0.03 0.04
0.02
0.02
0.01
0
0
0.01 0.02
0 0.5 1 1.5 2 2.5 0 0.5 1 1.5 2 2.5
Time[s] 7 Time[s] 7
x 10 x 10
Figure 6. Variation of the current for different Figure 7. Variation of the current with defect
defect types thickness
H. Mohellebi et al. / Use of HalfAnalytical Method for the Detection of Defects in Diet Pulses 189
While examining the results given in Figure 6 we notice that the current increase
every time the sensor is near to the defect. The internal defect with maximum value of
current Im constitute the more detected one because it is most near the sensor, then
comes the defect in depth and finally the external defect.
In Figure 7 there is represented the current response of the differential probe when
considering different defect thickness (20%, 40%, 80%). In this case we notice that the
parameter which varies more is the crest value of the current that increases
progressively as the defect thickness increases.
In this application we consider a load with a combined defect, internal defect (40 %
width) and an external defect (20 % width) (Figure 8). Figure 9 represents the variation
of the differential current i(t) in the case of the load with an internal defect and in the
case of the load with a combined defect, for C=106 F and R=0.2 .
0.045 0.3
Combined defect
0.04 Internal defect
0.25
0.035
0.2
0.03
0.15
C u rre n t[ A ]
0.025
z [m ]
0.02 0.1
0.015
0.05 T0
0.01
0
0.005
0 0.05
7.5 8 8.5 9 9.5 10 10.5 0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1
r[m] 3 Time[s] 6
x 10 x 10
Figure 8. Solving domain with combined defects Figure 9. Differential currents variations
Im[A] T0[s]
1
Internal defect 1.846 10 4.2129 10 7
In Table 1, Im represents the maximum value of the current which results from
difference between two currents turns; T0 is the time corresponding to current when it
vanishes.
According to the results presented in Figure 9, we note a difference between the
differential currents i(t) in the case of the internal defect and the same unknown
obtained in the case of the combined defect. The combined defect supplies amplitude
Im with a time of passage by zero T0 more important. We can thus conclude that in spite
of the internal defect is important; it does not hide the presence of the external defect.
190 H. Mohellebi et al. / Use of HalfAnalytical Method for the Detection of Defects in Diet Pulses
In this application we consider a load with two healthy layers and two layers with
defects.
0.04 0.04
0.035 0.035
0.03 0.03
0.025 0.025
h a u te u r[m ]
h a u t e u r[ m ]
0.02 0.02
0.015 0.015
0.01 0.01
0.005 0.005
0 0
7.5 8 8.5 9 9.5 10 10.5 7.5 8 8.5 9 9.5 10 10.5
rayon[m] 3 rayon[m] 3
x 10 x 10
Figure 10. Load with two healthy layers Figure 11. Load with two layers with defects
The Figure 12 represents the variation of the induced electromotive force (e. m. f) u(t)
in an elementary spire of the inductor for a flawless load, a load with defect, a load
with two layers without defects and a load with two layers with defects.
0.12 1
load w th defect load with defect [1]
0.1 08 healthy load [1]
healthy load
load with defects
load w th two healthy layers
06 healthy load
0.08 load w th two layers with defects
Relative values of e m f
04
0.06
02
0.04
e.m.f [V]
0
0.02
02
0
04
0.02
06
0.04 08
0 01 02 03 04 05 06 07 08 09 1
time[s]
0.06
0 0.2 0.4 0.6 0.8 1 1.2 1.4
Time[s] 7
x 10
Figure 12. Induced e.m.f u(t) variation in an Figure 13. Comparison with results
given in [1]
elementary coil of the inductor
This relative value of induced e.m.f (Figure 13) increases with increase of the
layers number and also with the number of defects. These results with relative value
variation with time allow to the same conclusion than the one given in [1].
4. Conclusion
The halfanalytic model developed in the current study is applied to the non destructive
eddy current testing device. The study is undertaken in pulsed diet and a response of
current is investigated for several defects nature of conducting tube. The case of the
H. Mohellebi et al. / Use of HalfAnalytical Method for the Detection of Defects in Diet Pulses 191
presence of two defects is considered and analysed. A cylinder with two layers is also
studied and the results obtained are in a good agreement according to existing results.
The model developed has shown its capability to analyse multiple defects in a
conducting cylinder or multiple layers one with defect. The advantages of the model
are less time consuming, which is very interesting compared to numerical models
because in this case the air region was note discretized, and its capability to take into
account of the movement of the sensor without changes in domain discretization.
References
Introduction
1
Ammar HAMEL, Electrical Engineering Laboratory, Department of Electrical Engineering,
University of Bjaa, Algeria hamelkane@yahoo.fr.
A. Hamel et al. / ICA Applied to Eddy Current Nondestructive Evaluation 193
1. 2D Electromagnetic equation
! 1 1 2
A j 2
2
F "
r A A J s . A 2 .r.dr.dz (1)
2 r r z 2
R
The energy terms corresponding to the magnetic field, eddy current and source
current are represented by the first, second and third terms, respectively, of the
integrand. The impedance of a filamentary circular loop of radius ri can be calculated
from the magnetic vector potential Ai at ri and the value of the impressed current I s
in the loop as follows:
#
E t j.2 .ri . Ai
Zi (2)
Is Is Is
Hence, the total impedance of the circular coil whose cross section is discretized
into N triangular elements is given by
j.2 .N s N j.2 .J s N
Z
Is
$ ci ci i
r . A %
I s2
$
rci .% i .Aci (3)
i 1 i 1
Where
N s : turn density [turns / m 2 ] ,
% i : area of ith element,
rci , Aci : central values of r , A in ith element.
So the total impedance of a differential probe can be obtained by summing the
impedance of each coil of the differential probe.
194 A. Hamel et al. / ICA Applied to Eddy Current Nondestructive Evaluation
Like other evolutionary techniques, ICA starts with an initial population. Population
individuals called countries are divided into two types: colonies and imperialists that all
together form some empires. Imperialistic competition among these empires forms the
basis of ICA. During this competition, weak empires collapse and powerful ones take
possession of their colonies. Imperialistic competition hopefully converges to a state in
which there is only one empire and its colonies are in the same position and have the
same cost as the imperialist. The pseudo code of Imperialist competitive algorithm is as
follows [7]:
& Select some random points on the function and initialize the empires.
& Move the colonies toward their relevant imperialist (Assimilation).
& Randomly change the position of some colonies (Revolution).
& If there is a colony in an empire which has lower cost than the imperialist,
exchange the positions of that colony and the imperialist.
& Unite the similar empires.
& Compute the total cost of all empires.
& Pick the weakest colony (colonies) from the weakest empires and give it
(them) to one of the empires (Imperialistic competition).
& Eliminate the powerless empires.
& If stop conditions satisfied, stop, if not go to second point.
To start the optimization algorithm we generate the initial population of size N pop .
We select N imp of the most powerful countries to form the empires. The remaining
N col of the population will be the colonies each of which belongs to an empire. Then
we have two types of countries: imperialist and colony. To form the initial empires, we
divide the colonies among imperialists based on their power. That is the initial number
of colonies of an empire should be proportionate to its power. Figure 1 shows an
example of an initial population of each empire.
Imperialist 1
Colony 1
Imperialist 2
Colony 2
Imperialist 3
Colony 3
Imperialist n
Colony n
x ~ U
0, d , ' ~ U
, (4)
Where and are parameters that modify the area that colonies randomly
search around the imperialist. In our implementation and are considered as 2 and
0.5 (Radian) respectively.
Imperialist
New position
of colony
x
' d
Colony
ICA was implemented and then tested with some benchmark functions. The results
which we give hereafter are those found with the function F which is written [7]:
F x sin
4 x 1.1 sin
2 y (5)
Figure 3 shows a 3D plot of the function F . The global minimum of this function
is in the interval 0 ( x ( 10 , 0 ( y ( 10 , is located in
x, y
9.039,8.668 and has the
cost of 18.5547 .
196 A. Hamel et al. / ICA Applied to Eddy Current Nondestructive Evaluation
9 9
8 8
7 7
6 6
5 5
y
y
4 4
3 3
2 2
1 1
0 0
0 2 4 6 8 10 0 2 4 6 8 10
x x
(a) (b)
10 10
9 9
8 8
7 7
6 6
5 5
y
y
4 4
3 3
2 2
1 1
0 0
0 2 4 6 8 10 0 2 4 6 8 10
x
(c) x
(d)
Figure 4. Initial empires (a). Empires at iteration 5 (b). Empires at iteration 10 (c) Empires at iteration 20 (d).
A. Hamel et al. / ICA Applied to Eddy Current Nondestructive Evaluation 197
F unc tion c os t
5
0 15
5
10 17
15
18.5547 18.5547
0 40 80 120 160 200 260 0 20 60 100 140 180 220 260
Iterations Iterations
(a) (b)
Figure 5. Mean and minimum cost versus iterations by PSO (a) and ICA (b).
Results show that with ICA the convergence is reached more quickly than with
PSO. Indeed with PSO convergence is obtained after approximately 240 iterations
whereas with ICA it is obtained after 20 iterations. We noted that the execution time of
iteration is practically the same one for both algorithms. Consequently, ICA finds the
global minimum successfully.
The problem geometry is that studied in [9]. A differential probe is used to scan a
conducting tube (aluminum) having an electrical conductivity c 1M )s / m* . The
geometrical and electrical data of the eddy current differential probe is as follows:
height of a coil according to z: 0.75e3 m, inner radius of a coil: 7.75e3 m, outer radius
of a coil: 8.5e3 m, vertical distance between the coils: 0.5e3 m, number of turns of a
coil: 70. The probe is supplied by a current with intensity of 5mA and a frequency of
100kHz . The impedance measured when the medium of the coil is opposite the lower
edge of the groove is Z m (0.55 j1.45)+ .
The inversion is based on an iterative approach that employs a direct finite element
model to simulate the fundamental physical process, as shown in figure 5. The
inversion algorithm starts with an initial estimate of the groove profile and then
determines the signal by solving a finite element direct problem. The error between the
measured and the calculated signals is minimized iteratively by updating the groove
parameters by keeping the best profile of the previous iteration. When the error is
below a threshold, the profile determined is the desired solution.
198 A. Hamel et al. / ICA Applied to Eddy Current Nondestructive Evaluation
Measured signal
Calculated
Z m
Initial groove signal
profile Forward
Z c Yes Desired
Z m Z c ,  solution
model
Update groove No
profile
Figures 7 (a) and (b) illustrate the evolution of the groove parameters obtained by
PSO and ICA respectively. By PSO, convergence is obtained after 124 iterations. The
value of the height and the depth of the groove are h 3.98mm and p 0.508mm
respectively. On the other hand, by ICA, convergence is practically obtained with 22
iterations and then the value of the height and the depth of the groove are h 3,97 mm
and p 0,509mm respectively. In order to make comparison we used the same number
of particles in PSO as of countries in ICA i.e.30. We noted that the execution time of
iteration is the same one for both algorithms. Consequently, this enables us to
conclude that ICA converges more quickly than PSO algorithm. Nevertheless, with
regard to the precision of the results, PSO proves to be slightly better than ICA.
3 3
x 10 x 10
6 7
Groove depth [m] Groove height
5 Groove height [m] 6 Groove depth
Groove parameters [m]
5
4
4
3
3
2
2
1 1
0 0
0 50 100 150 200 250 300 0 50 100 150 200 250 300
(a) Iterations (b) Iterations
4. Conclusion
The method proposed in this paper, Imperialist Competitive Algorithm (ICA) uses an
evolutionary algorithm in order to inverse eddy current nondestructive evaluation
signals. The aim is to reconstitute the profile of an axisymmetrical groove. The
method uses a finiteelements forward model to simulate the physical process and
Imperialist Competitive Algorithm to solve the inverse problem. This evolutionary
optimization strategy has shown great performance in both convergence rate and global
optima achievement. Indeed, from the experimental results, it can be seen that the ICA
method outperforms the PSO one in terms of speed convergence. However, PSO
remains a method which gives results with a better precision. In continuing this work,
the proposed algorithm could be extended to reconstitute complex profiles.
A. Hamel et al. / ICA Applied to Eddy Current Nondestructive Evaluation 199
References
[1] S. Hoole, S. Subramaniam, R. Saldanha, and J. Coulomb, Inverse Problem Methodology and Finite
Elements in the Identification of Cracks, Sources, Materials, and their Geometry in Inaccessible
Locations, IEEE Trans. Magn., vol. 27 (1991), pp. 3433 3443.
[2] Y. Li, L. Udpa, and S.S. Udpa, Three Dimensional Defect Reconstruction From Eddy Current NDE
Signals Using a Genetic Local Search Algorithm, IEEE Trans. Magn., vol 40, No. 2 (2004), pp. 410
417.
[3] M. Rebican, Z. Chen, N. Yusa, L. Janousek, and K. Miya, Shape Reconstruction of Multiple Cracks
From ECT Signals by Means of a Stochastic Method, IEEE Trans. Magn., vol 42, No. 4, (2006), pp
1079 1082.
[4] M. Cacciola, S. Calcagno, F. C. Morabito, and M. Versaci, Swarm Optimization for Imaging of corrosion
by impedance measurements in Eddy Current Test, IEEE Trans. Magn., vol. 43, No. 4, (2007), pp
1853 1856.
[5] A. Hamel, H. Mohellebi, M. Feliachi and F. Hocini, Particle Swarm Optimization for Reconstitution of
Two Dimensional Groove Profiles in Non Destructive Evaluation, ISEF 2009 XIV International
Symposium on Electromagnetic Fields in Mechatronics, No. 4, September 10 12, Arras, France,
(2009), pp. 1079 1082.
[6] O.C.Zienkiewicz and R.L.Taylor, The Finite Element Method, 4th ed., McGrawHill, 1989.
[7] E. Atashpaz Gargari, C. Lucas, Imperialist Competitive Algorithm : An Algorithm for Optimization
Inspired by Imperialistic Competition, IEEE Congress on Evolutionary Computation, Singapore ,(
2007), pp 4661 4667.
[8] R. C. Eberhart and J. Kennedy, A New Optimizer Using Particle Swarm Theory, proc. Sixth
International Symposium on Micro Machine and Human Science (Nagoya, Japon), IEEE Service
Center, Piscataway, NJ (1995), pp. 39 43.
[9] J.L Thomas, Modlisation du Contrle Non Destructif par Courants de Foucault des Gnrateurs de
Vapeur, Rapport de Stage 3me anne de cycle dingnieur, CEA, France, 1998.
This page intentionally left blank
Transducers and Techniques
This page intentionally left blank
Electromagnetic Nondestructive Evaluation (XIV) 203
T. Chady et al. (Eds.)
IOS Press, 2011
2011 The authors and IOS Press. All rights reserved.
doi:10.3233/9781607507505203
Abstract. In nuclear power plants, there may happen local wallthinning on the
inner surface of a pipe due to the flow of coolant flowing inside the pipe. Pulsed
eddy current testing (pulsed ECT) technology is developed in recent years.
Because of its rich frequency components and applicability of large electric current,
a pulsed ECT method may show promising capability of detecting and evaluating
the defect in the deep region of the material. The aim of this study is to discuss the
feasibility of detection and evaluation of local wallthinning of the bottom surface
in one thick layer and also in the lower layer of twolayer structure pipe using a
pulsed ECT method. Concerning the large thickness of the specimen, a high
sensitive flux gate (FG) sensor has been employed and the corresponding
differential exciting mode has been developed to efficiently apply the FG sensor.
Experimental results show that this FG sensor combined with differential exciting
mode could detect a very small defect that is located in a thick specimen while
which could not be detected by Hall sensor.
Introduction
In nuclear power plants, there may happen local wallthinning on the inner surface of a
pipe due to of the flow of coolant flowing inside the pipe (flow accelerated corrosion
(FAC)). Normally, the thickness of the pipe is very large. In addition, in Japan, there is
a special concern about local wallthinning under an enforcement plate that covers the
outside of a pipe where a branch pipe is connected to a main pipe. Because the
enforcement plate and the pipe wall form two layers of the metal plate, also the
existence of thick insulators outside the pipe, these make it very difficult to inspect the
inside of the pipe by ultrasonic testing method. Pulsed eddy current testing (pulsed
ECT) technology is one of the methods developed in recent years [1, 2]. Because of its
unnecessity of contact between the probe and the inspected specimen, also its rich
frequency components and applicability of large electric current [35], pulsed ECT
method may show promising capability of detecting and evaluating the defect in the
deep region of the material with liftoff. The aim of this study is to discuss the
feasibility of detection and evaluation of local wallthinning of the bottom surface in
1
Corresponding Author: Toshiyuki TAKAGI. Katahiri 211, Aobaku, Sendai 9808577, Japan;
Email: takagi@ifs.tohoku.ac.jp
204 S. Xie et al. / Pulsed ECT Method for Evaluation of Pipe WallThinning of Nuclear Power Plants
one thick layer and also in the lower layer of twolayer structure pipe using a pulsed
ECT method with small liftoff.
When the specimen is rather thick, low frequency is preferred due to the skin effect
phenomenon. It has been pointed out that the sensitivity of pulsed ECT technique can
be greatly enhanced when magnetic sensors are applied instead of pickup coils for
picking up the field. The enhancement can especially be beneficial when deeply buried
defects are of the prime concern [6].
There is a family of magnetic field sensors, such as, SQUID (Superconducting
Quantum Interference Device), FG (flux gate) sensors, AMR (anisotropic magneto
resistive), GMR (giant magnetoresistive) and Hall devices [6]. As we know, SQUID
possesses the highest sensitivity but its use is still limited due to practicality and some
costrelated reasons [7]. Hall sensor is not a bad choice when the specimen is not very
thick and also the defect is not very small, but for the small defect detection in a thick
specimen Hall sensor is not preferred due to its relative big noise level (i.e. poor
resolution, according to the reference [6] and also the experience of the author). Thus
FG, AMR and GMR sensors are good candidates for the small defect detection in
rather thick specimen. In this study, FG sensor was employed for the local wall
thinning detection in large diameter pipes.
Concerning the characteristics of FG sensor, of course, high sensitivity and high
resolution are its obvious merits, but at the same time, small measurement range is its
inherent demerit, normally inside several Gauss. To overcome the trade off of high
sensitivity and small measurement range, in this study, a differential exciting mode has
been developed and the corresponding two identical exciting pancake coils have been
designed.
d N 1 1
u (t ) = + [ sin(n d ) * cos(nt ) + (1 cos(n d )) *sin( n t )] (1)
T n 1 n n
2
An = an 2 + bn 2 = * 1 cos(n d ) (2)
n
2 2
An , = * 1 cos(n d ) = * 1 cos(2n ) (3)
n n
2 2
An ,1 = * 1 cos(n (T d )) = * 1 cos(2 n ) (4)
n n
From the above equation (3) and (4), we can see that when the duty of two square
wave pulse signals is complementary to 1, their spectrum distribution should be
completely the same only except the spectrum of DC component. FFT was applied to
the square wave pulse under condition of different duties (0.1, 0.2,, 0.9), then the
energy (here amplitude was considered as energy) percentage of the fundamental
frequency, which means energy of fundamental frequency divided by the energy of all
the harmonic frequencies (except DC component) has been obtained and shown in
Figure 2.
width and 20mm in thickness. A rectangular slot was arranged along the center line of
the bottom side of the plate in the shorter direction. The size of the slot defect
(widthdepth) is 2010, 1010 or 105 (mm), respectively. The slot defect is to be
detected from the top surface of the plate. The specimen is shown in Figure 3.
Figure 3. One plate specimen with a slot defect simulating local wallthinning in one layer thick pipe
To simulate the local wallthinning on the bottom side of a pipe which is covered
by an enforcement plate, we have prepared two AISI316 austenitic stainless steel flat
plates for our experiments. The size of the two plates is the same, 500mm in length,
300mm in width and 8mm in thickness. One plate is placed on the top of the other. A
rectangular slot has been arranged along the center line of the bottom side of the lower
plate in the shorter direction. The width of the slot is 10mm, and the depth is 1, 3 or
5mm, respectively. The slot is to be detected from the top surface of the upper plate.
The specimen is shown in Figure 4.
Figure 4. Doubleplate specimen with a slot defect simulating local wallthinning in two layer pipes
one time according to the experiment request. Figure 6(b) shows that the output has
good linear property to external magnetic field within the measurement range.
The perturbed field is rather small, thus the small measurement range of FG sensor can
be satisfied.
(2) Eddy current in the specimen decreases very slowly along depth direction and
deeper penetration depth could be expected.
(a) Scanning signal (b) Relationship between value of peak to peak of scanning signal and size of defect
Figure 9. Pickup signal and relationship between signal and defect
(a) Pickup signal using FG sensor and Hall sensor (b) Relationship between peak value and depth of defect
Figure 11. Pick up signal and relationship between signal and defect
4. Conclusion
In the paper, it has been analyzed how to choose a suitable duty of square wave pulse
in pulsed ECT method. A differential exciting mode was developed to efficiently apply
TMFFG sensor in pulsed ECT experiment system. The experiment results state that
local wallthinning in one thick layer pipe (20mm) and double layer pipes (8mm+8mm)
with small liftoff (2mm) could be detected using the above detection mode. In addition,
experiment result shows that FG sensor is much better than Hall sensor for small defect
detection in thicker specimen. This study gives a good foundation for the future
quantitative work.
Acknowledgements
This work was conducted as a part of Nuclear and Industrial Safety Agency (NISA)
project on Enhancement of Ageing Management and Maintenance of Nuclear Power
Plants in Japan and supported by the GrantinAid for the Global COE Program,
"World Centre of Education and Research for TransDisciplinary Flow Dynamics",
from the Ministry of Education, Culture, Sports, Science and Technology (MEXT) of
Japan. The authors would like to thank Mr. Takeshi Sato of Tohoku University, for the
preparation of the specimens and the fabrication of the probe.
References
[1] G.Y. Tian, A. Sophian. Defect classification using a new feature for pulsed eddy current sensors.
NDT&E International 38 (2005) 7782.
[2] J. Kim, G. Yang, L. Udpa, S. Udpa. Classification of pulsed eddy current GMR data on aircraft structures.
NDT&E International 43 (2010) 141144.
[3] M. Fan et al. Analytical modeling for transient probe response in pulsed eddy current testing. NDT&E
International 42 (2009) 376 383.
[4] T. Chen et al. Feature extraction and selection for defect classification of pulsed eddy current NDT.
NDT&E International 41 (2008) 467 476.
[5] M. Morozov, G.Y. Tian and D. Edgar. Comparison of PEC and SFEC NDE Techniques. NDT&E
International 24 (2009) 153 164.
[6] G.Y. Tian and A. Sophian. Study of magnetic sensors for pulsed eddy current techniques. Insight 47
(2005) 277279.
[7] M. Valentino, G. Pepe, A. Ruosi and G. Peluso. Eddycurrent nondestructive measurements with
different HTSSQUID spatial orientations. J. Phys. IV France 8 (1998) 249252.
Electromagnetic Nondestructive Evaluation (XIV) 211
T. Chady et al. (Eds.)
IOS Press, 2011
2011 The authors and IOS Press. All rights reserved.
doi:10.3233/9781607507505211
Abstract. An eddy current testing (ECT) system with a high sensitive AMR sensor
was developed. In our experiments, a specimen of cooper plate with grooves and
slits was used to simulate the cooling grooves of the combustion chamber of liquid
rocket and the defect in it. Three defects, with the width of 0.2 mm, the length of 4
mm and the depth of 0.2 mm, 0.5 mm and 0.8 mm respectively, were made in the
bottom of the grooves. Using a 4.5 mm and 30 turn circular excitation coil, ECT
experiments were done for the AMR sensor with Z, X and Y directions at the
frequency of 2 kHz. The results show that the variance of lift off had big influence
on the ECT result with the Z direction AMR sensor and smaller influence on that
with the X, Y direction AMR sensor. The defects with the depth of 0.8 mm and
0.5mm, 0.2 mm were clearly observed, but the signal of the defect with the depth
of 0.2 mm was very small.
Introduction
For the liquid rocket, liquid Oxygen and liquid Hydrogen are used as the propellant.
The wall of the combustion chamber is made of CuCrZr copper alloy. High
temperature gas over 3000 K is generated in the combustion chamber and liquid
hydrogen flowing in the cooling groove is used for the cooling. Due to the big thermal
gradient and the excessive thermal strain generated in the cooling groove, two kinds of
damages might be generated in the wall of the combustion chamber. One is from the
inner wall side, which is caused by the oxidation/reduction of the oxide layer; another
is the melt damage from the cooling grooves, which is caused by the growth of the
small crack in the grain boundary of the copper alloy. The aim of this research is to
find an effective method for the tiny defect detection and evaluation of the combustion
chamber.
Eddy current testing (ECT) is an effective method to detect defects in conductive
materials. It has the advantage of low price and easy operation. Inductive coil,
anisotropic magneto resistance (AMR), giant magneto resistance (GMR), flux gate,
superconducting quantum interference device (SQUID) have been used as the sensors
in ECT systems [14]. SQUID has the best magnetic field resolution especially at low
frequency. However, SQUID must be cooled by liquid nitrogen or liquid helium, so the
ECT system with SQUID is expensive, complex and is not easy to handle. High
sensitive flux gate is also expensive. The sensitivity of coil become worse at low
212 D. He et al. / Developing ECT System with AMR Sensor for Combustion Chamber
frequency and GMR sensor also has big noise at low frequency. AMR sensor is
cheaper than SQUID and flux gate and has better sensitivity than inductive coil and
GMR sensor at low frequency.
Using the low noise driving circuit developed by us, the magnetic field resolution
of AMR sensor was improved to 12 pT/Hz at the frequency of above 1 kHz [5]. The
driving circuit operated in amplifier mode or feedback mode and the linearity could be
improved for feedback mode, which is important for a moving ECT system in
unshielded environment. ECT experiments were once been done for an aluminum plate
[6]. In this report, we will present our ECT results for the specimen of copper plate
with the X, Y, and Z sensing directions of the AMR sensor.
1. Experimental setup
Figure1 shows the block diagram of the ECT system with AMR sensor. The circular
coil was used to produce the excitation field. Then, eddycurrent was induced in the
specimen. If there was some defect in the specimen, the distribution of the eddycurrent
was changed and the field produced by the eddy current was detected by the AMR
sensor. Through the lockin amplifier, the amplitude signal and the phase signal could
be obtained; then it was sent to a computer for data acquisition and data processing. In
our experiment only the amplitude signal was used. The AMR sensor was fixed with
the XY stage for the scanning.
Figure 1. The block diagram of the ECT system with AMR sensor
1.1. Specimen
To simulate the real combustion chamber of liquid rocket, the specimen was made of
copper alloy plate. Figure 2 shows it. The thickness of the plate was 4 mm. Grooves
with the width of 1 mm and the depth of 3 mm were made on one side of the copper
plate. Three slit defects were made in the bottom of some grooves with the length of 4
mm, the width of 0.2 mm and the depth of 0.2 mm, 0.5 mm, and 0.8 mm respectively.
The surface of the copper was not flat; bending exists in some part of the copper plate,
which caused the variance of the lift off.
D. He et al. / Developing ECT System with AMR Sensor for Combustion Chamber 213
Figure 2. The specimen of copper plate with grooves and slit defect to simulate the combustion chamber of
liquid rocket.
A commercially available AMR sensor of HMC1001 was used [7]. The magnetic field
sensitivity of the AMR sensor was about 3.2 mV/V/Gauss. To improve the field
resolution, higher bias voltage of 24 V and low noise preamplifier were used. To
improve the linearity of the response, a driving circuit operating in feedback mode was
designed [5]. The AMR sensor has flat response till to 5 MHz and the magnetic field
resolution was about 12 pT/Hz at the frequencies of above 1 kHz.
1.3. The excitation coil and the sensing direction of AMR sensor
A 30 turn, 4.5 mm circular coil was used to produce the excitation field. Figure 3
shows the three sensing direction of the AMR sensor. Figure 3 (a) shows the Z
direction AMR sensor. The coil was attached to the bottom of the AMR and the
position was adjusted to get the biggest output for the excitation field. Figure 3 (b)
shows the X direction sensor. The sensing direction of the AMR sensor was along the
direction of the groove. The coil was attached to one side of the AMR sensor and the
position of the coil was adjusted to get the smallest output. Figure 3 (c) shows the Y
direction sensor. The sensing direction of the AMR sensor was perpendicular to the
grooves direction.
D
D E
E F
F
Figure 3. The excitation coil and the sensing direction of AMR. (a). Z direction. (b). X direction. (c). Y
direction
214 D. He et al. / Developing ECT System with AMR Sensor for Combustion Chamber
The conductivity of the copper alloy is about 5.8106 S/m. In our experiments, the
excitation frequency was 2 kHz. From the formula = 1 f , where, is the
penetration depth, f is the excitation frequency, is the permeability of the material and
is the conductivitythe penetration depth is about 1.5 mm for copper alloy at 2 kHz.
The amplitude of the current flow in the excitation coil was about 20 mA. The
amplitude of the magnetic field at the center of the coil was about 2 Gauss. X
direction was parallel to the direction of the grooves, Y direction was perpendicular to
the direction of the grooves, and Z direction was vertical direction perpendicular to the
surface of the specimen. For the XY scanning, first, fixing the X position and line
scanning along Y direction; then changing the X position. Due to the bending of the
specimen, the variance of the lift off was from 0.2 mm to 1mm.
2. Experimental results
We did the ECT experiments for the AMR sensor with Z direction, X direction and Y
direction. For the Z direction sensor, the big excitation field was also detected by the
AMR senor. For the X and Y direction sensor, the excitation field was not detected by
the AMR sensor.
2.1. Z direction
Figure 4 shows ECT results when using the AMR sensor with Z direction. Figure 4 (a)
shows the Line scanning where the defects exist. Figure 4 (b) shows line scanning
where no defects exist. Figure 4 (c) shows the 2D contour map of the XY scanning
results. The defects with the depth of 0.8 mm and 0.5 mm were observed. Due to the
big background field and the influence of the variance of the lift off, the smallest defect
with the depth of 0.2 mm was not observed.
Figure 4. ECT results using AMR with Z direction. (a). Line scanning result where defects exist. (b). Line
scanning result where no defects exist. (c). Contour map of the XY scanning results.
Figure 5 (a) shows the subtraction results of the line scanning Figure 4 (a) and
Figure 4 (b). Figure 5 (b) shows the 2D contour map of the subtraction results. The
D. He et al. / Developing ECT System with AMR Sensor for Combustion Chamber 215
defects of 0.8 mm and 0.5 mm become clearer; A very small signal appeared for the
defect signal with the depth of 0.2 mm.
Figure 5. (a). The subtraction result of Figure 4 (a) and Figure 4 (b). (b). The contour map of subtraction
results.
2.2. X direction
Figure 6 shows ECT results when using the AMR sensor with X direction. Figure 6 (a)
shows the Line scanning where the defects exist. The defect signal was positive at one
end of the defect, and it is negative at another end of the defect. Figure 6 (b) shows the
2D contour map of the XY scanning results. Using the AMR sensor with the X
direction, the influence of the variance of the lift off was reduced. A small signal
appeared for the defect with the depth of 0.2 mm.
Figure 6. ECT results using AMR with X direction. (a). Line scanning result where defects exist. (b). 2D
Contour map of the XY scanning results.
2.3. Y direction
ECT experiments using AMR sensor with Y direction were also done. Figure 7 (a)
shows line scanning results where defects exist. For each defect, a peak and a valley
signal appeared. The defects with the depth of 0.8 mm and 0.5 mm were clearly
216 D. He et al. / Developing ECT System with AMR Sensor for Combustion Chamber
observed. The signal of the defect with the depth of 0.2 mm was small. The influence
of the variance of lift off was also reduced. Figure 7 (b) shows the contour map of the
results.
Figure 7. ECT results using AMR with Y direction. (a). Line scanning result where defects exist. (b). 2D
Contour map of the XY scanning results.
3. Summary
ECT system with AMR sensor was constructed and ECT experiments for the copper
specimen were done using the AMR sensor with the Z, X and Y directions. Variance of
lift off had big influence when using Z direction AMR sensor. The influence of the
variance of lift off was reduced when using X and Y direction AMR sensor. The
defects with the depth of 0.8 mm and 0.5 mm were observed. The signal of the defect
with the depth of 0.2 mm was very small.
Acknowledgment
References
[1] K. Allweins, M. von Kreutzbruck, G. Gierelt, Defect detection in aluminum laser welds using an
anisotropic magnetoresistive sensor array, J. Appl. Phys., 97, 10Q102, (2005)
[2] T. Dogaru, S.T. Smith, Giant magnetoresistance based eddy current sensor, IEEE Trans. Magn., 37,
pp.2790 2793, (2001).
[3] C. Carr, J.C. Macfarlane, The performance of flux gate magnetometers for nondestructive evaluation,
Insight, 41, pp.20 24, (1999).
[4] H. Weinstock, R.A. Welsh, SQUIDs for nondestructive evaluation, Superconductor Industry, 10,
pp.23 28, (1997).
[5] D.F. He, M. Tachiki, H. Itozaki, Highly sensitive anisotropic magnetoresistance magnetometer for
Eddy current nondestructive evaluation, Rev. Sci. Instru , 80, 036102, (2009).
[6] Dongfeng He, Mitsuharu Shiwa, et al., Basic Examination for the Defect Detection of Combustion
Chamber Using ECT, submitted to Journal of JSNDI, (2010).
Electromagnetic Nondestructive Evaluation (XIV) 217
T. Chady et al. (Eds.)
IOS Press, 2011
2011 The authors and IOS Press. All rights reserved.
doi:10.3233/9781607507505217
Abstract. High precision magnetic field sensors are of increasing interest in non
destructive testing (NDT). In particular GMR sensors (giant magneto resistance)
are qualified because of their high sensitivity, high signal to noise ratio and high
spatial resolution. We performed magnetic flux leakage measurements of artificial
cracks with a GMR gradiometer and a 3 axes GMR magnetometer. Cracks of a
depth of 44 m still could be detected with a sufficient high signal to noise ratio.
A semi analytic magnetic dipole model was used for swiftly predicting magnetic
stray fields. The reliable reconstruction based on measurements of artificial
rectangular shaped defects is demonstrated.
Introduction
Magnetic flux leakage testing (MFL) is one of the most popular methods for defect
detection and characterization of magnetic components in NDT. The most prominent
method for a MFL application is the magnetic particle inspection (MP). In this method
magnetic particles accumulate at magnetic stray fields generated by surface cracks in a
magnetized component. By means of MP the number of defects and their length can be
determined precisely. However, the estimation of defect depth and the automation of
MPapplications is rather difficult. To examine the depth and width of inhomogeneities
in more detail it is necessary to measure the stray field near cracks. This can be done by
any magnetic field sensor, such as flux gate sensors, hall probes, SQUID, or XMR
sensors (magneto resistance; X = G (giant), A (anisotropic), etc.) [16].
In this work we used two types of GMR (giant magneto resistance) sensors optimized
for NDT applications: a gradiometer measuring the normal component of the magnetic
flux leakage and a new 3axes magnetometer measuring all three directions in space.
GMR sensors are well suited for NDT applications thanks to their high sensitivity, high
signaltonoise ratio and small sensing area associated with a high spatial resolution. In
contrast to commercial GMR sensors that are usually embedded in a bulky
encapsulation, the sensitive layers of the GMR sensors used in this work are located
1
Corresponding Author: E mail: matthias.pelkner@bam.de. This work is supported by the Federal Ministry
of Education and Research, Germany, project no. 16SV3787.
218 M. Pelkner et al. / Flux Leakage Measurements for Defect Characterization Using NDT
only a few tens of m from the chip edge, resulting in a small sensortosurface
distance. This helps to visualize the defect signature with a very good spatial resolution.
Further, having a fast inverse scheme in mind based on a semianalytic model for MFL
to estimate the defect parameters like width, depth and length, it is helpful to work with
pointlike sensing devices. Therefore the active sensing area should be adapted to the
liftoff and has to be smaller. Otherwise, the field gradients along the active area must
be considered via integration over the whole sensing area, leading to distinct layers
calculation times.
The experimental setup consists of three high precision manipulators (Aerotech) for the
three directions in space, providing a maximum scanning velocity of 0.5 m/s and a
spatial resolution of 1 m. The manipulators are fixed on a 2000 kg bloc of granite
eliminating environmental vibrations.
To record the data of the sensors a fast data acquisition hardware based on a 18 bit
ADconverter with a maximum sampling rate of 600 kHz (National Instruments) was
used. To achieve a high signaltonoise ratio, the sensor signals were preamplified.
(a) (b)
Figure 1. GMR gradiometer measuring the gradient in x direction of the normal field component Hz. (a)
Sketch of the electric circuit of the GMR chip (Sensitec GmbH). The sensing areas (R1 R4, green and red)
are sensitive to Hz. The distance between R2/R4 and R1/R3 is 250 m. (b) Schematic of an artificial crack with
dimensions length l, depth h and width w. The induced polarization on the defect wall is indicated by
plus signs. The white squares represent the sensing elements of the GMR gradiometer above the defect.
The GMR effect was discovered in 1988 by Peter Grnberg [7] and Albert Fert (Nobel
Prize in Physics, 2007). Independently they found a giant magneto resistance (GMR)
in a multilayer structure consisting of ferromagnetic layers divided by nonmagnetic
M. Pelkner et al. / Flux Leakage Measurements for Defect Characterization Using NDT 219
layers which depend on the directions of magnetization in each layer. The effect is
based on quantummechanical scattering of the conduction electrons at the boundary
surface between the ferromagnetic and nonmagnetic layer.
Generally a GMR gradiometer consists of four notshielded elements fabricated as
a Wheatstonebridge as shown in fig. 1(a), resulting in a distinct temperature
independent output signal. The measured quantity is the difference of the magnetic
field (here the normal component Hz) between the two active areas and it is detected by
measuring the bridge voltage. In our case, the sensing areas have a size of only
7060 m and a transfer function of 14 mV/V(kA/m)1 in its linear range between
1kA/m.
The 3axes magnetometer (Sensitec GmbH) consists of three Wheatstone bridges
that allow to measure absolute field values via special flux concentrators. As its
prominent feature the new 3axes magnetometer detects simultaneously all three spatial
components of the magnetic field in the same parallel plane next to the surface under
investigation. Fig. 2 depicts the layout of the sensor. The x and y components are
measured with special flux concentration layers, which are rotated by 90. Both flux
concentrators have a diameter of 140 m. This results in a spatial resolution that is
worse than for the gradiometer described above. The zcomponent is measured by the
field distortion in the vicinity of two flux concentrator strips with dimension about
10160 m and separated by a distance of 30 m. The transfer function for the 3axes
magnetometer is 30 mV/V (kA/m)1 and it has its linear range for fields between
1kA/m. For further details see [8].
Figure 2. Prototype of a 3 axes GMR magnetometer (Sensitec GmbH) measuring the three directions in
space of the magnetic field in the same plane. On the right a SEM picture of the 3 axes magnetometer is
shown [8].
Inversion of MFL signals requires an adequate forward solver for predicting magnetic
stray field signals. To account for defects we employed the magnetic dipole model
introduced by Shcherbinin [9,10]. According to this model MFL signals arise due to
induced magnetic polarizations on the defect walls [11].
This model allows to calculate magnetic stray fields of a rectangular defect with
length l, width w and depth h (see fig. 1(b)). Especially for GMR sensors that can be
positioned very close to the surface of the test piece, it is important to carefully
consider the dimensions of the probe, i.e., the size of their active sensing areas, in the
calculation process. In our calculation we therefore numerically integrated the magnetic
stray field expressions over the active sensor areas of the GMR testing probe (fig. 1(b))
to obtain a fast algorithm.
220 M. Pelkner et al. / Flux Leakage Measurements for Defect Characterization Using NDT
For the inverse calculation of the defect parameters we employed the trusttregion
reflective algorithm to minimize the difference between the predicted( dsyn F(m) )
and measured ( dobs ) GMR sensor data iteratively using an L2 norm approach. Initially,
the forward solution is estimated for an arbitrarily chosen start model m0 (comprising
the defect parameters l, w and h). Subsequently, the soughtafter model parameters are
updated in an iterative fashion by minimizing the objective function
2
dobs j
dsyn with dsyn
j
F(m j )
, (1)
where j denotes the iteration number and F(m j ) the forward solution step as outlined
above. Minimization is continued until convergence is reached yielding an estimate of
the true defect parameters.
2. Results
As test object we used a steel plate with artificial cracks of varying depth (see fig. 3).
The artificial cracks were introduced by electrical discharge machining. The width of
the cracks varies between 100 to 200 m with a length around 5500 m. The depth
ranges between 10 m and 2240 m. The measurements were performed in remanence
(no applied field during measurement), i.e., the test object was magnetized before. The
direction of the applied field was normal to the defect geometry.
Figure 3. Schematic cross section of the steel plate. The field was applied along the x direction.
Using the gradiometer line scans with different liftoffs were recorded. Fig. 4 shows
the gradient of the magnetic stray field as a function of position x. Clear signatures of
the field gradient illustrate the defect positions and the amplitude of H scales with the
depth of the defects [11]. The right panel in fig. 4 illustrates the signals of the defects
with a depth between 44 m and 380 m in more detail. The attenuation of the signal
for an increasing liftoff is in agreement with the theoretical model. Further, it is found
that even the 44 m deep defect can be observed. However, for the 30 m and 10 m
deep cracks no clear signal can be distinguished from the noise. Possible contributions
to the noise level may originate from the roughness of the surface or from spatial
variations of the permeability.
M. Pelkner et al. / Flux Leakage Measurements for Defect Characterization Using NDT 221
Figure 4. Line scan with a gradiometer for three different distances between sensor and surface of the test
object.
2.2. Reconstruction
Magnetic field sources generally create a blurred magnetic field distribution above
the samples surface. This is due to the divergence of the magnetic field and the
position of the sensing device, which in an exterior investigation is naturally always a
certain distance from the source of the field. What starts as a distinct field distribution
inside the sample turns into a broad field distribution when the field sensor is either too
large or positioned at too far a distance above the samples surface. This blurredness
complicates the data interpretation and begs for additional support through signal
processing to evaluate the mechanical state of the sample. Thus, small sensors
providing high spatial resolution in combination with a large signaltonoise ratio are
helpful to gain better insight into the type of defect, its size and its location.
In fig. 5 the measurement and the results of the reconstruction of two defects are
shown. For the reconstruction data in the vicinity of the defect was used to calculate the
defect parameters (approx. 120,000 data points).
The false rendering plots in the upper panel of fig. 5(a) show the experimental and
calculated data of the stray field gradient of a 210 m deep crack. In addition 10 line
scans across the defect, extracted from both data sets, are shown aside.
In the lower panel of fig. 5(a) the results of the reconstruction process are shown,
in which the parameters length, depth and width were reconstructed simultaneously. A
relative fast converging process is observed. The length converges after less then 10
iterations and the depth can be reconstructed after less than 5 iterations. In total this
reconstruction requires only a few seconds on a standard PC.
It is worth mentioning that the crack width does not converge to the real value.
This is in agreement with the sensitivity matrix and is due to the fact that in the
analytical model the crack width has a minor influence to the stray field [11].
Remarkably, we find a good reconstruction quality also for the 44 mdeep defect
although dealing with a significant decrease in SNR, as illustrated in fig. 5(b).
222 M. Pelkner et al. / Flux Leakage Measurements for Defect Characterization Using NDT
Figure 5. False color renderings of two different defects: (a) 210 m and (b) 44 m deep. The top left picture
shows the measurement and the middle one the synthetic generated signal form of the reconstruction
algorithm. Right to them ten lines are exemplary illustrated ( measurement, simulation). In the bottom
diagram the defect parameters and the steps of reconstruction are presented.
Finally, in this section first measurements of the magnetic flux leakage using a new 3
axes magnetometer are shown. The same test plate was used as for the gradiometer.
The prototype 3axes magnetometer measures Hx, Hy and Hz simultaneously at the
same lift off. This provides additional information of the magnetic field which is
required, e.g., for a more voxelbased reconstruction algorithm.
In fig. 6 the results for the Hx and Hz component of the magnetic field are
presented. Panel (a) and (c) show false rendering plots for both magnetic field
components. For each field component three line scans, indicated by the black, red and
green line, are shown in panels (b) and (d). These line scans illustrate the stray field at
the edges (black, green) and the center (red) of the cracks, respectively. As expected
from the analytic model, the stray field signals are most pronounced in the center of a
crack and decline towards the edges. Since a magnetometer was used the defect signals
appear superimposed on magnetic stray fields originating from the magnetic steel plate
itself. The typical signatures of these background fields are a maximum in the center of
the plate for the Hx component and a gradual decline for the Hz component.
The Hy field component (not shown) is characterized by MFL signatures at the
edges of the long sides of the cracks. Since the direction of the magnetization was
along the x direction, only at these end positions distinct signatures of Hy are expected.
M. Pelkner et al. / Flux Leakage Measurements for Defect Characterization Using NDT 223
Figure 6. Stray field distribution represented in a false color renderings. (a) False rendering plot of field
component Hx. (c) Hz component. (b) and (d): Extraction of 3 line scans across the left and right edge of the
crack (black and green) and the center of the crack (red), where field amplitude is maximum.
In fig. 7 the scanning area was adapted to three defects for the Hx component of the
magnetic field. Compared to the gradiometer we observe a somewhat higher noise level.
In addition to the 380 and 210 m deep defects the 44 m deep defect can be imagined.
3. Conclusion
The automation of flux leakage testing is an active research field, where the use of
adapted magnetic field sensors detect stray field distributions of defects above the test
specimen. For this purpose, the sensors require a high spatial resolution and an
excellent field sensitivity. We therefore propose in this work the use of NDTadapted
GMR sensors that enable us to detect several artificial defects with depth down to 44
224 M. Pelkner et al. / Flux Leakage Measurements for Defect Characterization Using NDT
m. Detecting the magnetic field close to the surface offers the advantage to obtain
more detailed information on the geometrical parameters of the defects. Using a few
msized GMR gradiometer, where the sensing MRlayer is positioned close to the
chip edge, we obtained distinct crack signatures with good spatial resolution and
showed that a reconstruction of the length and depth can be reliably carried out. The
width can also be determined, however, less accurate due to the smaller influence of the
width on the stray field. In addition we proposed the use of a 3axes magnetometer
measuring all three field component simultaneously. These results can be regarded as a
further step towards the use of GMR sensors in NDT providing a more complete
picture of the real magnetic field distribution of the samples outside world. This paves
the way for new visualization techniques, precisely defect localization, defect
characterization, and tomographylike mapping techniques.
Electromagnetic testing based on small magneto resistive layers with high spatial
resolution can be provided with relatively high cost effectiveness. This also allows
GMRsensors to follow the miniaturization trend providing an adequate testing method
for quality control of small components. GMRtechnology thus has the potential to
bridge the microgap between the mmsized conventional induction coils for detecting
macroscopic material defects and the scanning magnetic force microscopy for the
detection of field distribution on the nmscale.
References
[1] K. Allweins, M. v. Kreutzbruck, G. Gierelt, Defect Detection in Aluminium Laser Welds Using an AMR
Sensor Array, J.Appl.Phys. 97 (2005), 10Q102
[2] H. Pries, Giant Magnetic Resistance Sensoren in der industriellen ZfP, Report of Annual Meeting of the
German Society of Nondestructive Testing 2008, St. Gallen, Switzerland
[3] Y. Kataoka et al., Application of GMR line sensor to detect the magnetic flux distribution for
nondestructive testing, International Journal of Applied Electromagnetics and Mechanics 15
(2001/2002), 47 52
[4] H. J. Krause, M. v. Kreutzbruck, Recent Development in SQUID NDE, Physica C 368 (2002), 70 79
[5] C. Dolabdjian, G. Wach, L. Perez, Improvement in the detection of subsurface fatigue cracks under
airframe fasteners using improved rotating giant magneto resistance magnetometer head, Paper
presented at the 9th European Conference on NDT, 25 29 September 2006, Berlin
[6] F. Gruhl, M. Mck, M. v. Kreutzbruck, J. Dechert, A scanning superconducting quantum interference
device microscope with high spatial resolution for room temperature samples, Rev. Sci. Instrum. 72
(2001), 2090 2096
[7] P. Grnberg, et al., Layered Magnetic Structures: Evidence for Antiferromagnetic Coupling of Fe Layers
across Cr Interlayers, Phys.Rev.Lett. 57 (1986), 2442
[8] C. Glenske, U.Loreit, New 3D Magnetic Field Sensors with GMR Spin Valve Layers, Paper presented
at the 10th Symposium Magnetoresistive Sensors and Magnetic Systems, 31. March/1. April 2009,
Wetzlar (Germany)
[9] N. Zatsepin, V Shcherbinin, Calculation of the magneto static field of surface defects. I. Field topography
of defect models, Defektoskopija 5 (1966), 50 59
[10] V. Shcherbinin, A. Pashagin, Influence of the extension of a defect on the magnitude its magnetic field,
Defektoskopija 8 (1972), 74 83
[11] F. Frster, New findings in the field of non destructive magnetic leakage field, NDT International 19
(1986), 3 14
Electromagnetic Nondestructive Evaluation (XIV) 225
T. Chady et al. (Eds.)
IOS Press, 2011
2011 The authors and IOS Press. All rights reserved.
doi:10.3233/9781607507505225
Introduction
In these years, technology for imaging aws in structural materials using nondestructive
testing methods has become important for the safety of society. The recent progress
of high speed computational ability of PCs has enabled the widespread use of aws
imaging technology.[1][2][3] Conventionally, piezoelectric probes have been employed
for detecting and imaging aws in structural materials. Piezoelectric probes usually are
in contact with samples directly or are indirectly coupled through water and therefore
cannot easily scan samples quickly. Piezoelectric probes are not available for detecting
aws in large samples.
An electromagnetic acoustic transducer (EMAT) is a major nondestructive (NDT)
device that detects aws in a sample using an ultrasound signal induced by electromag
netic force.[4] EMAT can address the problem of slowspeed scanning because EMATs
can scan without contacting the samples. An EMAT measures a time series of data, that
is waveforms similar to those produced by ordinary piezoelectric transducers. Two kinds
of EMAT are availables, one using Lorenz force and another using magnetostrictive ef
1 Corresponding Author: Yoshihiro Nishimura, AIST, 1 2 1 Namiki Tsukuba Ibaraki, Japan; E mail:
nishimura.yoshihiro@aist.go.jp.
226 Y. Nishimura et al. / Evaluation of EMAT Signals Using Magnetostrictives for Imaging
fects. The latter can be used for easily detecting aws in a sample with ferromagnetic
oxide scale. However, the data generated by EMAT using magnetostrictive effects are
very noisy due to the Barkhausen effect, compared to data produced by ordinary piezo
electric transducers. It is important to decrease such noise to reconstruct images of aws
in samples because the noise from Barkhausen effects is sometimes too large to ignore.
Of course, some noises come from electoronic amplier circuit also. Such noise can be
decreased by staying at one position and averaging over a time series of data derived
repeatedly. However, if the probe scans a large surface area on a sample in a limited time
for imaging , the probe cannot remain long at one position for averaging. The frequency
of these noise is much higher than that of the signicant signal data. A lowpass lter,
that can remove high frequency noise, is one way to address this problem but introduce
another kind of noise. Some other method for reducing noise when imaging aws should
be explored.
1. EMAT Principles
An EMAT is used in an NDT method similar to the conventional ultrasound testing (UT),
except that its probe output signal has more noise and that it is a contactfree method.
The EMAT probe has a coil (or coils) and a strong internal magnet (Figure.1). Pulsing
current induces eddy current on the surface of the inspected material as much as the
eddy current testing (ECT) probe does. The induced eddy current generates an ultrasound
signal by Lorenz force or by magnetostrictive effect( Figure.1). The ultrasound signal
propagates to and reects from the edges of the sample or the aws (Figure.2). The
probe output is observed as a time series of echo signal intensities on the EMAT screen(
Figure.3).
A typical time series of data derived by an EMAT from reference specimen is de
picted in Figure.3. The initial peak induced by the rst current pulse and many subse
quent peaks from multiple reections at the edge of reference sample can be observed.
The initial peak is too large compared to later peaks and so is saturated in the full input
scale range. This measurement was made at 2GHz.
Y. Nishimura et al. / Evaluation of EMAT Signals Using Magnetostrictives for Imaging 227
A sample with some internal aws was prepared. Its geometry and dimensions are de
picted in Figure.4. Its oxide scale was generated in an environment similar to those of
actual use. The 3D scanner depicted in Figure.5 was developed and applied to the
prepared sample. A FA computer controlled the scanning movement of the EMAT probe
and the trigger timing of the pulser/receiver. An AD converter collected the time series
of data from the receiver and the FA computer stored them in its data strorage. A block
diagram of the 3Dscanner is presented in Figure.6. Scanning was done as illustrated in
Figure.7. The specication of the 3D scanner are listed in Table.1.
The measurement was made under the conguration listed in Table 2. The observed
time series of data is very noisy, as illustrated in Figure. 8(a),(b), but averaging mea
sured data(timebasedaveraging) can decrease such noise (Figure.9(a),(b)). These noisy
signals were observed all over the surface and can be decreased by averaging.
228 Y. Nishimura et al. / Evaluation of EMAT Signals Using Magnetostrictives for Imaging
Figure 5. 3D scanner
If the probe scans a large surface area on a sample in a limited time, the probe cannot
stay long at one position for averaging. The timebasedaveraging method thus cannot be
Y. Nishimura et al. / Evaluation of EMAT Signals Using Magnetostrictives for Imaging 229
Figure 8. (a)Raw signal derived from the sample in Figure.4 and (b) Magnied image of reection signal from
bottom edge.
230 Y. Nishimura et al. / Evaluation of EMAT Signals Using Magnetostrictives for Imaging
Figure 9. (a)Signal averaged over 400 raw signals derived from the sample(in Figure.4) and (b)Magnied
image of reection signal in (a)
used for scanning a large area. The power spectra of Figures.8(a) and 9(a) are plotted in
Figures.10(a) and (b). Most of the noise can be minimized by averaging.
A lowpass lter that passes frequencies below 5.5MHz was applied to the time se
ries data in Figure.8(a). The time series data in Figure.8(a) was converted to the one
shown in Figure.11(a),(b). Most of the noise in Figure.8(a),(b) could be removed success
fully. However, the lowpassltered data reveals noises generated by the lter function
at the initial peak in the left half of Figure.11(a).
We tried another noise reduction method. In this measurement, the data was sampled
in the X and Y directions at a pitch of 100m by 3Dscanner. However, the sampling
interval in the Z direction is equivalent to a sampling pitch of 7.5m because the velocity
of the transverse wave in steel is 3000m/s and the sampling frequency is 400MHz in
Y. Nishimura et al. / Evaluation of EMAT Signals Using Magnetostrictives for Imaging 231
Figure 11. (a)Signal ltered from Figure.8 by a low pass lter (in Figure.11) and (b)Magnied image of
reection signal in (a)
Figure.8(a). Compared to the sampling pitch in the X and Y directions, the sampling
pitch in the Z direction is thought to be too small. We therefore tried 1Dpositionbased
averaging. First, we derived a time series data by sampling at 2GHz. Second, we averaged
every 50 samples to derive a time series data equivalent to the one sampled at 40MHz.
We tried averaging every 10 samples, every 20 samples , every 50 samples and every 100
samples and decided to employ evey50samplesaveraging. Sampling the time series
data at 40MHz is though to be equivalent to sampling pitch of 75m in the Z direction.
AD converters supporting under 100MHz or 400MHz are used for common uses of UT
equipments. AD converters supporting 2GHz or more are not popular. But a 2GHz AD
converters was found to be particularly useful for real time noisereduction and imaging
of aws by such noisy EMAT using magnetostrictive effect.
Figure 12. (a)Signal derived by averaging raw data sampled at 2GHz at every 50 samples raw(in Figure.8)
and (b)Magnied image of reection signal in (a)
In this way, the noise depicted in Figure.8 was decreased successfully. The distribu
tions of reection intensities in a sample have to be derived to reconstruct a 3Dimage
of aws in a sample. By applying Hilbert transformation to the time series data in Fig
232 Y. Nishimura et al. / Evaluation of EMAT Signals Using Magnetostrictives for Imaging
ure.12, a curve illustrating their envelope could be derived in Figure.13(a). Hilbert trans
formation provides an analytical expression of the signal and its absolute value provides
an envelope curve of the signal. A sectional image of the sample can be reconstructed by
putting the time Taxis in the vertical direction and the Xaxis in the horizontal direction
and allocating the the reection from the aws at a position (X,T) on the XT plane. An
example of the sectional image of sample is depicted in Figure.13(b). The vertical axis
denotes time and the horizontal axis denotes the position in the X direction. The top edge
of Figure.13(b) means the surface. The reection from the aw and the one from bottom
edge can be identied easily and their depth 33mm can be quite precisely measured in
Figure.13(b) because the velocity of the transverse wave in steel id 3000m/s.
Figure 13. (a)Envelope curve derived by Hilbert transformation and (b) Reconstructed sectional image of
reected signal
4. Conclusion
References
[1] K.Nakahata, S.Hirose, The study of the qualities of the reconstructed image by array probe , JSNDI
preceedings, (2006),39 42
[2] R.G.Pratt:Seismic waveform inversion in the frequency domain I Theory and verication in a physical
scale model, Geographics (1999), 888 901
[3] Toshiki Watanabe:Full wave inversion, JSNDI Jornal vol 53 (2005), 274 279
[4] Nishimura, Y,. Sasamoto, A,. Suzuki, T,. Study of 3D Image reconstruction using EMAT, Int.J. of Ap
plied Electromagnetics and Mechanics, 28(1), p171 176, 2008
Electromagnetic Nondestructive Evaluation (XIV) 233
T. Chady et al. (Eds.)
IOS Press, 2011
2011 The authors and IOS Press. All rights reserved.
doi:10.3233/9781607507505233
8
 9
,,
5
.
' 6 (7 $
. Q6#' %7 %
98#D7 "9I6"%#D
'HSDUWPHQWRI(OHFWURPDJQHWLFDQG%LRPHGLFDO(QJLQHHULQJ)DFXOW\RI(OHFWULFDO
(QJLQHHULQJ8QLYHUVLW\RI=LOLQD8QLYHU]LWQD=LOLQD6ORYDN5HSXEOLF
$EVWUDFW
)
,

)
. .
5
.
3
6

1
)
.
,
5
5
3 8
5
5
)
.H\ZRUGV 8

7
5
7 <

3
,QWURGXFWLRQ
. ,
,
 ,
,

3 6)7
5 .
. .
/6 0 )

1
2
5 5
4 1
5
5
1 ,

53
 5
1
,
1 ,
55
.
3

/ 90
 5
6 5
.
3 "
1 5
,
5 57 15 5
5
)
3 +
.

,

7

7 5
<1
,1 ) ,1 ,

I
,
1
5
3
5
.
,
,. ,1
5
9
,
5
N
7
,
,7
)

,1
,.
 1
,. ,
.
3
8

.
/8 90
.
6 3 T 5
.
7
 5
,
5

)
.
N
,
3
5 8 9 . 1
5 .

(
9,
>
'
7 ,
5
T
7
 5
7 .
 5 !
7 .2
(7 A(A :P !
7 '
.N +,1
L
>
B5
323N
234 M. Smetana et al. / Pulsed Eddy Current Testing Application to Defect Evaluation
1 )

5
9 ,

7 5
) 1
/ 03
5
,,
. )
7 : ?
. 1

8 9 ,1
3

5 8 9 1
5
 5
5
C
)
5 / 0 $,
5 /$07 E(F3  55
. 1
5
7 1
. 5
)N
1
)
8 9
,
3
,,
,
 5 N ,
7 ,1
,
)
.
,
5
,
.
.) 5
,
3
3XOVHGHGG\FXUUHQWHYDOXDWLRQ
8
/
0 
, ).5
).
<

,
3 T ,
,
5 ). 5 1 1 5
547
,
54
,5
55
,
3
.
,

)54
,
1
,
15 , 5
)
7 E:F3
N
,,
)
1C
.
5 
,
3 
7
8 9
<,
1 5
5
. , ,
.
9
)
<
3 1 ,
1
5
5
5

).5
)1 54
3
)
, 5 8 9
7
5
3 T
N)
1
.
N
.
5
,3 &
5 5
1C
1
7
5
, 8 9
1 7 E?F7 ESF3
)LJXUH 8,
5

7 E:F3
M. Smetana et al. / Pulsed Eddy Current Testing Application to Defect Evaluation 235
&
, 5
,

, 7
).5{
,
,
5
5
5
553
,
,
,
5
N

7
5
).5
.
N )
1
5
5
7 3 (3
8

55 .
.
,
54 ,
1 {
543 1 55

,
.
54
1)
7
)
,
54 
 ,
, 5
,5
543
,
54 5  ,
5
)
) 1 .
1
5 

3
,
54
7
,
54
1
. )
1 55
5 55
7 4 .
. ,
<
7 E:F3
3(&QXPHULFDOLQYHVWLJDWLRQV
6
1
5
#8 + 5
)
. 1
3
. ,7 ) 3 S7 .
,
5
'' ?(P$ ,
3
.
 5 K (3?H 'O
. ,1
 5
K (3 ,
. 5
3 8
5
5
7 33
,
.
3:3 ,, 5
>
(3
.
5
5
 5
5
,
7
:3
.
. 5
5
<

1
,
)
5
5
,
3
)LJXUH 5
5
7 5
3 S3
5
5
,
/ 0
3
236 M. Smetana et al. / Pulsed Eddy Current Testing Application to Defect Evaluation
1XPHULFDOVLPXODWLRQUHVXOWV
W :
\ W DWH E
7 /(0
) D E
3
< .
5
.
, K ( ) 3 ?3
)LJXUH 8 % .
5
<
3
55
.
, 1
5
8 %
2
2 5
5
5 
,13
&RQILJXUDWLRQ
5 1
, ) 3 S )
1 K PAA 1 5
. 
,
3
.
 5

, 5 )
5
55 5 ( 3
M. Smetana et al. / Pulsed Eddy Current Testing Application to Defect Evaluation 237
)LJXUH', )
5
3
5
)
) 3 PJR ,
, 5
55
,
 5
5
3 5
5
) 3 :
,
 3 S3 "
1
5

,
5
5
7
,
7 /, 63(0
1
)
5
,
,,<
 S 3 .
, 5
1
1
1 ,
 .
5
,3
)LJXUH55
,
7 , 63 (7 5
(3
)LJXUH55
,
7 , 63 :7 5
(
238 M. Smetana et al. / Pulsed Eddy Current Testing Application to Defect Evaluation
)LJXUH55
,
7 , 63 ?7 5
(3
'
, 63: / 3 :7 S07 )

5
5
 55
,. 3 5
,
5 H
55
,
)
5
,,<
 :H
3 & 1
,
5
5
.
7 , 63 ? / 3 :7 S07
55
,
5
5
 ,
5
,
5
,
5 R 3
5
)
,
,
,
5
3 #5 7
5
,
5 ,
1
5
5
,
63 ?3
5
)
,
1
) ) )
,
&RQILJXUDWLRQ
5
) 3 M
. 
3
)
,
3
5
,

, 5 )
5
55 ( 3
)LJXUH55
,
7 , 63 (7 5
:3
6

55
,
) 3 R3 "
1
5
1
1
)
5 , 5
1 1
)
5
: ,
)
,N ,17 3 H3
95
?
, 5
<
,N,
5
5
?
5
:7 1
<
5 5
1
SH}7 3 (A3
)LJXUH95
? 5
<

3
)LJXUH55
,
7 , 63 (7 5
?3
240 M. Smetana et al. / Pulsed Eddy Current Testing Application to Defect Evaluation
6

55
,
5
5
)
3 ((3 "
1 1.
.
)
5
:7 3 R3 7
5
<
5
.
5
,
3
&RQFOXVLRQ
,
,

4
5 ,
5
.
3 6
.
. 1
.
5
5 . ,
5
.
7 5
,1 5
N
55
,
3 '
,
,,3 T
,5
1
,
5
5

,
>
, 5
.
/5

, 5
07 .
5
5
J55 ,
/5
.
5
03 5
,
5 55
N 1 . 1 ,, 5
5
,13 T
,  5
,
 5
)
,

5 5
)N3
$FNQRZOHGJPHQW
5HIHUHQFHV
5 ,

.
3 6~
"
S( /:AAM07 , S@@ SM?3
E:F &1, )))34

37 Q :H7 :A(A7 E#
F3
E?F 17 37
N.7 U37 97 !37 ,7 $3 >
. .
/"07 "''6
(?M? @:M(7 "#' ,7 D
3 ?(7 :AAM3
ESF Q3 &3 &
7 *3U3 > 3XOVHG HOHFWURPDJQHWLF PDWKRGV IRU GHWHFW GHWHFWLRQ DQG FKDUDFWHULVDWLRQ7
6~ 7 .
3 SA7
. $
37 :AA@7 E#
F3
Electromagnetic Nondestructive Evaluation (XIV) 241
T. Chady et al. (Eds.)
IOS Press, 2011
2011 The authors and IOS Press. All rights reserved.
doi:10.3233/9781607507505241
Abstract. Currently, local wall thinning on the inner surface of a pipe is evaluated
mostly by ultrasonic thickness measurement from the outer surface of the pipe in
nuclear power plants in Japan. However, it has been pointed out that there are two
major issues of this evaluation method. Firstly, this method is not able to evaluate
local wall thinning from the top of a reinforcing plate that covers the outer surface
of a pipe because this two layer structure hinders the propagation of ultrasound.
Secondly, because ultrasonic thickness measurement of a pipe wall is conducted at
certain intervals on the outer surface of the pipe, it may overlook a hole like defect
if the defect size is small compared to this interval. We propose an eddy current
testing (ECT) method using widely spaced excitation and pick up coils to alleviate
these problems. Our experimental results show this method can be applied to in
spection of a double plate that consists of two 8 mm thick austenitic stainless steel
plates, and a defect indication is observed in a much larger area than the defect.
1. Introduction
In nuclear power plants, local wall thinning on the inner surface of a pipe is evaluated
mostly by ultrasonic thickness measurement from the outer surface of the pipe. How
ever, some improvements are still required for thorough and efficient inspection of
local wall thinning to maintain aging nuclear power plants. In Japan, it is considered a
problem that ultrasonic thickness measurement is not able to evaluate local wall thin
ning from the top of a reinforcing plate that covers the outer surface of a main pipe
around a branch pipe perpendicularly connected to the main pipe because this twolayer
structure hinders the propagation of ultrasound (Figure 1). Also, it is of special concern
that ultrasonic thickness measurement may overlook a small hole caused by liquid
droplet impingement (LDI) because ultrasonic thickness measurement of a pipe wall is
conducted at certain intervals on the outer surface of the pipe and the defect size is
sometimes smaller than this interval. Decreasing the interval of course reduces the
incidence of this problem, but piping maintenance in nuclear power plants already
takes an enormous amount of time, and inspection time is not easily allowed to be
extended any longer. The aim of this study is to propose an inspection method for two
1
Corresponding Author: Professor, Institute of Fluid Science, Tohoku University, 2 1 1 Katahira, Aoba ku,
Sendai, Miyagi, Japan; E mail: takagi@ifs.tohoku.ac.jp
242 T. Yamamoto et al. / An ECT Probe with Widely Spaced Coils for Local Wall Thinning
2. Experiments
The ECT probe we used in this study has one excitation coil and one pickup coil
shown in Figure 2. The excitation and pickup coils are 10 mm in outer diameter, 5 mm
in inner diameter and 5 mm in height. The number of turns of the excitation and pick
up coils is 301 and 3,465 respectively. The wire diameter of the excitation and pickup
coils is 0.2 mm and 0.05 mm respectively. The centertocenter distance between these
coils was set to be 50 mm. The equipment used in our experiments consists of a func
tion synthesizer (WF1945B, NF Corporation), a bipolar amplifier (HSA4011, NF Cor
poration) and a lockin amplifier (LI5640, NF Corporation). In the experiments, the
detection signals were measured while the pair of excitation and pickup coils was
moved on the surface of a plate as shown in Figure 3. The excitation frequency was set
to be 1 kHz.
2.2. Specimens
Tables 1 and 2 show the information of the specimens used for evaluation of the per
T. Yamamoto et al. / An ECT Probe with Widely Spaced Coils for Local Wall Thinning 243
(a) Plate with a rectangular slit (b) Plate with a square hole
Figure 5. Specimens
formance of our ECT technique. The material of these plates is 316 austenitic stainless
steel. A reinforcing plate part of piping was simulated by two 8 mmthick plates. An
intact plate was placed on the top of the other plate that has a rectangular slit on the
bottom side as shown in Figure 4. The rectangular slit was cut along a half of the center
line on the bottom side in the shorter direction (Figure 5 (a)). A defect caused by LDI
was simulated by a square hole. The square hole was made at the center of a 12 mm
thick plate (Figure 5 (b)). For each of the rectangular slit and the square hole, three
plates were prepared so as to vary the depth of these simulated defects (1 mm, 3 mm
and 5 mm).
The specimens described above were inspected from the opposite side of the defect to
be detected. Figure 6 shows how we scanned the surface of the specimens. For a double
plate with a rectangular slit, five equallyspaced lines along the longer direction of the
plate were selected as the scanning lines. Each of the scanning lines has a length of 150
244 T. Yamamoto et al. / An ECT Probe with Widely Spaced Coils for Local Wall Thinning
mm at 1 mm intervals. The left three lines run above the rectangular slit while the other
two do not. For a square hole, a 150 mm60 mm area at the center of the surface of the
plate was scanned. The intervals of scanning were 1 mm in the longer direction and 2
mm in the shorter direction of the plate.
Because the specimens used in this evaluation have a widearea surface (500 mm300
mm) and also have a slight warp, the liftoff distance between the probe and the surface
T. Yamamoto et al. / An ECT Probe with Widely Spaced Coils for Local Wall Thinning 245
of the specimens varies with the position of the probe, which causes considerable lift
off noise. To cope with this problem, we processed raw signals obtained from the lock
in amplifier.
Figure 7 (a) shows the Vx and Vy components of ECT signals obtained from one of
the scanning lines going over a rectangular slit in Figure 6 (a). It can be observed that
the rectangular slit induces a big change in the signals around the center position. Al
though the end point of the scanning line is far enough away from the slit, the signal
values do not return to zero at the end point. To make the signal values zero at both
ends of the scanning line, the voltage value on the line that connects both ends of a
signal curve in Figure 7 (a) is subtracted from the voltage value on that signal curve for
each point. Figure 7 (b) is obtained in consequence of this processing.
For a signal distribution obtained as described in Figure 6 (b), the above process
ing is applied in the X direction as well as in the Y direction so as to make the values
on the boundary of the scanning area zero. Figure 8 shows an example of this signal
processing. The removal of gradual changes in the signal distribution highlights a de
fect indication due to a square hole.
2.5. Results
Figure 9 shows the signal distributions obtained from the specimens with a rectangular
slit. To simulate a reinforcing plate part, an intact plate was placed on the top of a plate
with a rectangular slit as shown in Figure 4. The results indicate that our ECT tech
nique can detect a 1 mmdeep defect on the bottom side of two 8 mmthick plates.
Figure 10 shows the signal distributions obtained from the specimens with a square
hole. A 1 mmdeep square hole on the bottom side of a 12 mmthick plate is clearly
recognized in the signal distributions. Defect signals are observed in a much larger area
than the defect (The defect area of a square hole is 10 mm10 mm. The area where a
defect indication appears due to a square hole is about 20 mm40 mm). This property
helps prevent overlooking a small defect.
Figure 11 shows the relationship between the maximum signal amplitude and the
defect depth obtained from Figures 9 and 10. There is an almost linear relationship
between them. It implies that the defect depth can be estimated from the maximum
amplitude of a defect indication.
In above experiments, an LDI defect is simulated by a square hole, but a real LDI
defect tends to have a cone shape. Then, a specimen with a coneshaped hole shown in
Figure 12 was also used for evaluation of the detectability of an LDI defect with our
ECT technique. The specimen is 300 mm in length, 100 mm in width and 7.1 mm in
thickness. The base diameter and the height of the coneshaped hole are 10.2 mm and
3.6 mm respectively. Because the surface of this specimen is not large enough, the
scanning area was reduced from the area described in Figure 6 (b) by decreasing the
width in the X direction from 60 mm to 40 mm. Figure 13 (a) shows the signal ampli
tude distribution obtained by this scanning. The appearance of the defect indication is
almost the same as Figure 10, which was obtained with square holes. Figure 13 (b)
provides the profiles on the lines X=0 and Y=0 in the amplitude distribution. Whereas
the defect is located from 5 mm to 5 mm in both the X and Y directions, a big varia
tion in signal amplitude is still observed at 10 mm in the X direction and at 20 mm in
the Y direction. This indicates that a defect can be found even if the scanning line is 5
mm away from the edge of the defect with the coils aligned parallel to the scanning line.
When the coils are aligned perpendicular to the scanning line, the defect can be found
even if the scanning line is 15 mm away from the edge of the defect (Figure 14).
(a) Parallel to the scanning line (b) Perpendicular to the scanning line
3. Summary
The aim of this study is to propose an inspection method to evaluate local wall thinning
on a reinforcing plate part of piping and detect a defect located out of the scanning line.
We proposed an ECT technique using widely spaced excitation and pickup coils to
detect a defect on the bottom side of a thickwalled plate. The experimental results
show this probe can detect a 1 mmdeep rectangular slit on the bottom side of two 8
mmthick austenitic stainless steel plates and a 1 mmdeep square hole on the bottom
side of a 12 mmthick austenitic stainless steel plate. There is an almost linear relation
ship between the maximum signal amplitude and the defect depth. This implies that the
defect depth can be estimated from the maximum signal amplitude due to a defect.
These defect signals are observed in a much larger area than the defect size. When the
coils are aligned perpendicular to the scanning line, this method can detect a defect
located within 15 mm from the scanning line.
The proposed method can be applied to twolayer structure and detects a defect
even if the defect is not on the scanning line. These properties are expected to alleviate
concern over ultrasonic thickness measurement conducted for evaluation of local wall
thinning of piping in nuclear power plants.
248 T. Yamamoto et al. / An ECT Probe with Widely Spaced Coils for Local Wall Thinning
Acknowledgment
This study was conducted as part of Nuclear and Industrial Safety Agency (NISA)
project on Enhancement of Ageing Management and Maintenance of Nuclear Power
Plants in Japan.
References
[1] T. Yamamoto, T. Takagi and T. Uchimoto, Remote field ECT for evaluation of local wall thinning using
pancake coils, The 14th International Workshop on Electromagnetic Nondestructive Evaluation (ENDE
2009), 130 132.
[2] T. R. Schmidt, The remote field eddy current inspection technique, Materials Evaluation, 42 (2) (1984),
225 230.
[3] T. R. Schmidt, History of the remote field eddy current inspection technique, Materials Evaluation, 47
(1) (1989), 14 22.
[4] Y.S. Sun, S. Udpa, W. Lord and D. Cooley, A remote field eddy current NDT probe for the inspection of
metallic plates, Materials Evaluation, 54 (4) (1996), 510 512.
[5] Y.S. Sun, T. Ouang and S. Udpa, Remote field eddy current testing: one of the potential solutions for
detecting deeply embedded discontinuities in thick and multiplayer metallic structures, Materials Evalua
tion, 59 (5) (2000), 632 637.
Electromagnetic Nondestructive Evaluation (XIV) 249
T. Chady et al. (Eds.)
IOS Press, 2011
2011 The authors and IOS Press. All rights reserved.
doi:10.3233/9781607507505249
Introduction
Titanium because of its mechanical properties (high corrosion resistance, high strength
toweight ratio) is an important material in many leading industries e.g. aerospace. It is
commonly used to manufacture new components such as turbine disc. The disc is
machined from a titanium billet which is a cylindrical solid bar of up to 400mm of a
diameter. Inservice failures of such components can often lead to catastrophic
consequences with significant economic impact. NDT techniques that are commonly
utilized to examine traditional aerospace materials are not always sufficient to detect
subsurface defects created during the manufacturing process in titanium alloys [1]. In
the case of components having large diameter a reasonable technique to apply is
Ultrasound UT. However its effectiveness is limited to shallow regions close to the
components surface. This results in a strong need for new advanced NDT technologies
for the inspection of titanium components during the manufacturing process. We
introduce a novel automated quality control system for the inspection of titanium billets
in response to this need in this paper. The system combines two subsystems: eddy
current testing (ECT) for evaluation of the material not deeper than 5mm beneath the
surface and phased array (PA) ultrasonic for deeper inspection. In this paper the ECT
1
Corresponding Author: West Pomeranian University of Technology in Szczecin, Department of
Electrical and Computer Engineering, ul. Sikorskiego 37, 70 313 Szczecin, Poland; E mail:
tchady@zut.edu.pl.
250 T. Chady et al. / PLL Based Eddy Current Measuring System for Inspection of Outer Flaw
a) b)
The block diagram of the measuring system is shown in Figure 2. A high speed 16bit
multifunction data acquisition converter (DAQ) with D/A and A/D modules (NI USB
6251) are used in the system for both generating and acquiring signals. The DAQ
device in connected via USB interface to the PC class computer which accumulates,
computes data and controls the system. A generated in the D/A DAQ module
sinusoidal signal is gained utilizing a high speed power amplifier (TDA 7294) before
driving transducer's excitation coils. TDA 7294 is very low distortion and noise power
amplifier with DMOS power stage. The driving current was controlled through LEM
Components electronic transducer with no galvanic connection to the main circuit and
was cut off if its value exceeds the maximum calculated for a measuring transducer.
T. Chady et al. / PLL Based Eddy Current Measuring System for Inspection of Outer Flaw 251
DIGITAL
REF IN
PC
TRANSDUCER
OUT
USB
COMPUTER
C/A
TDA 7294 SWICH
POWER AMPLIFIER
Figure 2. Block diagram of measuring system.
The proposed system works with specialized software operating in the National
Instruments' LabView environment. The software interface is shown in Figure 3. The
application is used to drive the DAQ converter in order to control excitation
instruments, configure PLL operating parameters (such as sensitivity and time
constant) and acquire measuring signals (a real and imaginary part of measuring
signal). All parameters of lockin amplifier, DAQ and excitation signal can be
introduced to the system through a configuration file. The application enables external
and internal triggering of acquisition events. During the measurements, real time results
of acquired signal, its absolute value, real and imaginary part, phase shift and finally
plot of real versus imaginary part are presented to an operator. In the system there were
three processing algorithms implemented utilizing statistical analysis, gradient and
lowpass filtering. They allows to eliminate trends in the signal caused by the lift off
and surface roughness. The algorithms compute correction values for a signal basing on
value measured at present step and past ones and can be applied insitu or after the
measurements.
The transducer used in the system is presented in Figure 4. In order to make possible
defect detection located 5 mm beneath the sample surface its dimensions (Figure 4b)
and excitation frequency were optimized using FEM (Finite Element Method) and
Comsol software.
The transducer is built using an Etype ferrite core with excitation coils wound on
exterior columns of the ferrite core and a pickup coil on a central one. The excitation
coils generate contrary directed fluxes in the pickup coil. Therefore, the flux flowing
through pickup coil is close to zero in equilibrium state. If a flaw appearing close to
one of the excitation coils distorts the generated flux flow, the signal different from
zero is induced in the pickup coil. Such configuration of the transducer allows us for
the optimal usage of the dynamic range of the A/D converter module. In order to
achieve the greatest equilibrium the excitation coils of the transducers were driven by
separate power amplifier.
a) 1 b) c) 4
44
3
1 5
3
25
2 2
Figure 4. Photo and view with dimensions of the transducer: a) photo of the transducer, b) 3D view of the
transducer, c) photo of the transducers head; all dimensions are in [mm]; 1 ferrite core, 2 excitation coils,
3 pick up coil, 4 transducer head, 5 sample
T. Chady et al. / PLL Based Eddy Current Measuring System for Inspection of Outer Flaw 253
Figure 5. Photo and view of the titanium plate: a) photo, b) drawing of the notch, c) cross section drawing of
the notch; x axis is the scanning axis
The system is designated for the detection of minor outer flaws OF (casing
discontinuity in the surface at the opposite site of the plate to the scanning side) in thick
titanium specimens. In order to verify the performance of the proposed system several
experiments were carried out.
A titanium alloy plate having artificial EDM notches of different depth d was used
(Figure 5.). The depth was ranging from 10 to 100% of the plate thickness (6.9 mm).
All measurements were carried out for outer flaws.
Results presented in the paper were obtained for 40% notch. First, the Ascan
measurements using different exciting frequencies were carried out in order to find the
optimal one. Figure 6 shows the results of real and imaginary component as well as
amplitude and phase of the signal acquired during the experiments versus excitation
frequency.
From the set of achieved spectrograms one can notice that the frequency range
between 8 and 16 kHz gives the greatest chance to detect the flaw. Considering result
obtained for the real component of the signal (Figure 6a) it is clearly visible that
frequency equal to 11.2 kHz is the optimal one.
In the second stage of the experiments Cscan measurements were done. The two
dimensional distribution of the signals amplitude measured in both scanning direction
over selected area of 40% OF was presented in Figure 7a. In order to better visualize
changes of signal caused by the flaw, plot of amplitude and phase shift as well as real
and imaginary component of a signal measured in selected y line were shown in
Figure 7b and Figure 8. The presented results allow easily to detect the flaw.
254 T. Chady et al. / PLL Based Eddy Current Measuring System for Inspection of Outer Flaw
a)
f [kHz]
b)
f [kHz]
x [mm] x [mm]
c) d)
f [kHz]
f [kHz]
x [mm] x [mm]
Figure 6. Multi frequency results of measurements obtained for 40% OF: a) real part of measured signal
Re(Usig), b) imaginary part of measured signal Im(Usig), c) amplitude of measured signal Usig, d) phase of
measured signal arctg(Im(Usig)/Re(Usig)).
a) b)
Usig [V]
y [mm]
x [mm] x [mm]
Figure 7. Single frequency results obtained for 40% OF: a) amplitude of signal Usig measured in selected
area in both scanning direction, b) amplitude of signal Usig measured in selected y line.
3. Conclusions
The newly developed ECT system for observation of minor defects in titanium alloy
was presented. The system uses lockin technique which allows to detect signals having
the amplitude lower than the noise level by taking into consideration only that
component of a signal which frequency is equal to the reference one. In consequence it
T. Chady et al. / PLL Based Eddy Current Measuring System for Inspection of Outer Flaw 255
a) b)
Re(U) [V]
Im(U) [V]
x [mm] x [mm]
c) d)
leads to increase the possible depth penetration. The results of experiments with the
titanium alloy plate having artificial notches confirm the possibility of using this
system in industrial application. In the future, further experiments will be carried out
for titanium discs with artificial flaws.
Acknowledgement
This work was supported in part by Polish Ministry of Science and Higher Education
and by European Commission sponsored project QualiTi which is a collaboration
between the following organizations: I.S.O.TEST Engineering s.r.l, West Pomeranian
University of Technology (ZUT), Tecnitest Ingenieros S.L., TIMET UK Ltd, TWI Ltd
and Vermon SA. The project is coordinated and managed by TWI Ltd and is partly
funded by the EC under the Research for the Benefit of Specific Groups Project (ref:
FP7SME20071GA222476.)
References
[1] National Transportation Safety Board Aircraft Accident Report NTSB AAR 98/01
[2] http://www.femto.de/datasheet/FEMTO_Product_Overview.pdf
256 Electromagnetic Nondestructive Evaluation (XIV)
T. Chady et al. (Eds.)
IOS Press, 2011
2011 The authors and IOS Press. All rights reserved.
doi:10.3233/9781607507505256
Abstract. Giant magneto resistive (GMR) sensor based magnetic flux leakage
(MFL) technique is proposed for detection of defects in 64 mm diameter track
ropes. Helmholtz coil is used for magnetization and tangential component of
leakage flux from defects is measured using a GMR sensor. This technique is able
to detect local flaws and loss of metallic cross sectional area type defects in the
track rope.
Keywords. GMR sensor, magnetic flux leakage, track rope, finite element model
Introduction
Steel wire ropes are used for material handling in mines and hauling of men in skilift
operations [1]. In wire ropes, corrosion and wear are the main causes for damage such
as formation of local flaws (LF) and loss of metallic crosssectional area (LMA) [24].
LFs are external and internal discontinuities such as broken wires, cracks and corrosion
pitting. Reasons for wire breaking can be due to fatigue, interstrand nicking or
martensitic embrittlement. LMAs are distributed defects such as missing of wires
caused by corrosion, abrasion and wear resulting in loss of crosssectional area.
Periodic inspection of wire ropes is important to assess the structural integrity and to
take corrective actions. Nondestructive inspection of wire ropes is challenging due to
the heterogeneous structure of wire ropes, multiplicity, uncertainty of broken wires and
hostile working environment. Visual and magnetic flux leakage (MFL) techniques are
widely used for nondestructive inspection of wire ropes [4]. A variety of procedures
that use different types of sensors and magnetizing devices have been employed for
detection of defects in wire ropes by MFL technique [610]. Jomdecha et al. [7] used
printed circuitshaped coils connected in series as field sensors and solenoid as
magnetization unit for inspection of wire ropes and reported detection of 2 mm deep
surface defects in a 38 mm diameter wire rope. Kalwa et al. [9] developed an MFL
system comprising of magnetic concentrators, Hall sensors and sensing coils and
suggested that measurement of tangential component is more versatile than normal
component for detection of multiple defects in wire ropes.
1
Corresponding Author: Dr. B.P.C. Rao, Head, EMSI Section, NDE Division, Indira Gandhi Centre for
Atomic Research, Kalpakkam, TN 603102, India, E mail: bpcrao@igcar.gov.in.
W.S. Singh et al. / GMR Sensor Based MFL Technique for Inspection of Track Ropes 257
In this paper, we propose MFL technique for detection of defects on the outer
surface of 64 mm diameter track rope used for transportation of coal. The track ropes
are stationary and are rigidly supported at periodic intervals. They are operated for
about 10 hours every day transporting nearly 3000 tons of coal with the help of 256
numbers of stationary buckets each carrying nearly 1.6 tons of coal. The schematic of
crosssection of the track rope is shown in Figure 1. The track rope has 8 layers of
stranded wires of different diameters as detailed in Table 1. The width of the outer
surface of the Z wire is 6.45 mm and the gap width between two outer Z wires is 0.76
mm. The carriage wheels of the bucket are in contact with the track rope and thus
causing damage to the outer surface of the Zwire. Service induced surface flaws are
the major causes of failure in these ropes. When more than two Z wires of the outer
layer are broken, they will be separated from the adjacent layers.
Recently, GMR sensors are being widely used in MFL testing for detection of
defects in carbon steel plates [10], pipelines [11] in view of their high sensitivity for
low magnetic fields, good signaltonoise ratio (SNR) and high spatial resolution. GMR
sensors are attractive for measurement of feeble magnetic fields from shallow surface
defects and deeply located subsurface defects. They are also useful for detection of
stress and fatigue damage [12]. This paper discusses MFL technique proposed using
Helmholtz coil and GMR sensors for detection of defects in the track ropes. The results
of the developmental studies carried out on artificial LFs and LMA defect in track rope
are discussed. In order to interpret the measured MFL signal from LMA defect in the
track rope, 3D finite element modeling is performed and results are discussed.
used for detection of tangential component (along the scan direction) of leakage flux
from defects and is kept at the middle of the Helmholtz coil.
Measurements are made by moving the GMR sensor and Helmholtz coil together
as a single unit over the track rope in steps of 1 mm and at each axial location GMR
sensor is scanned circumferentially. A constant lift off of 0.3 mm is maintained
between the GMR sensor and the track rope to avoid physical damage to the sensor. In
order to enhance the sensitivity of GMR sensor, its output is first amplified using a
lownoise amplifier consisting of a differential amplifier, notch rejection filter at 50 Hz,
100 kHz lowpass filter and a singleended variable gain amplifier with DC
suppression. The variable gain amplifier is set such that it amplifies the sensor output
10 times. The amplified GMR sensor output is digitized using a 2channel data
acquisition system (16 bit) and stored in the computer for analysis.
&RPSXWHU
'&SRZHUVXSSO\ %LDVVXSSO\$PSOLILHU
PP
6FDQGLUHFWLRQ
*05
6HQVRU
:LUH5RSH PP
+HOPKROW]FRLO
GMR sensor works on giant magnetoresistive effect in which there is a large drop in
electrical resistance of the multilayer for an incident magnetic field due to spin
dependent scattering of electrons at the interface between two ferromagnetic layers
(few nm thick) separated by a nonmagnetic layer. The GMR bridge sensors (AA003
02) manufactured by M/s NVE Associates are used. These sensors measure differential
output voltage and ensure high stability with low noise (sensitivity ~260 VT 1 at 5V
bias voltage). They exhibit linear response in 0.21.3 mT range followed by saturation.
PP
$ % & ' ( ) * +
PP PP
PP
Figure 3. Schematic of track rope having axial and circumferential machined artificial notches.
The magnetizing current in the Helmholtz coil is optimized such that detectable leakage
field comes from all the notches. Current is optimized by studying the MFL signal peak
amplitudes for shallow axial notches (A and C) and shallow circumferential notches (E
and G). The signal amplitude is found to increase up to 6 A, beyond that it is decreased
due to the penetration of significant amount of magnetic flux into the inner layer of the
rope, as shown in Figure 4. Hence, a magnetizing current of 6 A is considered optimum
and is used in this study.
2.0
1.5
1.0
0.5
0.0
0 2 4 6 8 10
Magnetizing Current, A
2.1. LF defects
GMR signals (tangential component of the leakage flux) of axial notches A, B, C and D
are shown in Figure 5. The GMR sensor output for the shallowest axial notch (2.05 mm
deep) is approximately 2.0 times the background noise that comes mainly from 0.76
mm wide stranded structure of the track rope. As can be observed all the four axial
notches are detected by the technique with two distinct peaks correlated to the edges of
axial notches. To access the detection and sizing capability of the technique, peak
amplitude and Full Width at Half Maximum (FWHM) are determined after Gaussian
fitting and interpolation and the results are shown in Table 2. The error in the
260 W.S. Singh et al. / GMR Sensor Based MFL Technique for Inspection of Track Ropes
determination of FWHM is 0.1. As can be noted the amplitude of the MFL signals is
found to increase with increase in notch depth. The FWHM is found to be nearly the
same for all the axial notches, because of the constant notch length (5.5 mm).
MFL signals of circumferential notches viz. E, F, G, and H are shown in Figure 6.
The GMR response of circumferential notches is seen sharp with a single peak and the
amplitude of 1.94 mm deep notch is 3.2 times the background noise. As can be seen
from Figure 6, all the four notches are detected by this technique. The signal amplitude
increases up to a depth of 5.90 mm and then decreases while the FWHM remains
nearly constant (refer Table 2). It is observed from Table 2 that the amplitude of the
circumferential notches is 2 to 3 times higher than that of axial notches of similar depth.
However, the FWHM of circumferential notches is found to be nearly 2 to 3 times
lower than that of the axial notches. These observations are due to the axial
magnetization of the track rope causing circumferential notches to leak out higher flux
in comparison to the extended axial notches. Using a simple threshold FWHM, it is
possible to identify axial and circumferential defects in the track rope. Alternately, a set
of array sensors covering the circumference can be used.
1 .2
A x ia l N o tc h e s C D
1 .0
0 .8 B
GMR Output, V
0 .6
A
0 .4
0 .2
0 .0
0 .2
0 .4
0 50 100 150 200 250 300
S c a n D is ta n c e , m m
2.5
Circumferential Notches H
G
F
2.0 E
GMR Output, V
1.5
1.0
0.5
0.0
0.5
Table 2. The MFL signal amplitude and FWHM of axial and circumferential notches.
Notch Notch Depth, mm Amplitude, V FWHM, mm
A (axial) 2.05 0.58 4.8
B (axial) 4.11 0.91 5.5
C (axial) 5.86 1.22 4.8
D (axial) 7.91 1.25 4.6
E (circumferential) 1.94 1.86 1.8
F (circumferential) 3.88 2.33 1.6
G (circumferential) 5.90 2.56 2.0
H (circumferential) 8.24 2.50 1.7
Photograph of LMA defect and its GMR sensor response are shown in Figure 9. The
depth distribution (profile) of the LMA defect is measured using optical method. As
can be observed, the GMR sensor has detected the LMA with good SNR.
1.2
1.0
0.8
GMR Output, V
0.6
0.4
0.2
0.0
0.2
2 4 6 8 10 12
Scan Distance, mm
Figure 7. Photograph of a 2 mm deep saw cut in track rope and the measured GMR sensor output.
262 W.S. Singh et al. / GMR Sensor Based MFL Technique for Inspection of Track Ropes
2.0
C1 C2 C3 C4 C5
1.5
1.0
GMR Output, V
0.5
0.0
0.5
1.0
0 10 20 30 40 50
Scan Distance, mm
Figure 8. Photograph of series of 2mm deep saw cuts machined in track rope and the corresponding
measured GMR sensor output.
2.2
GMR Output, V
1
2.0
1.8 2
1.6
3
1.4
1.2 4
0 10 20 30 40 50 60 70 80
Scan Distance, mm
Figure 9. Photograph of LMA (42.0 mm length, 9.2 mm width and 3.0 mm depth) machined in the track rope,
GMR sensor response scanned along dotted line direction and optically measured depth profile.
W.S. Singh et al. / GMR Sensor Based MFL Technique for Inspection of Track Ropes 263
3. Conclusion
GMR sensor based MFL technique has been proposed for detection of LF and LMA
type of defects on the outer surface of 64 mm diameter track ropes. Performance of this
technique has been validated by detecting a few sawcuts and EDM notches simulating
cracks in the track rope. For LMA type defects, a decrease in signal amplitude has been
observed and confirmed by 3D finite element model. This proposed MFL technique
can be used for detection of both LF and LMA types of defects in track ropes. In order
to further enhance the detection sensitivity and to implement the technique in field,
work on the use of split type coils and array GMR sensors is underway.
4. Acknowledgements
References
[1] S.S. Udpa and M.O. Patrick, Nondestructive Testing Handbook  Electromagnetic Testing, ASNT, 3rd
Edition, Vol. 5, 2004, p. 437.
[2] D. Basak, S. Pal and D. Chandra Patranabis, In situ non destructive assessment of a haulage rope in a
monocable zigback passenger ropeway, Insight, Vol. 50, No. 3, March 2008, pp. 136 137.
[3] Herbert R. Weischedel and R. P. Ramsey, Electromagnetic testing, a reliable method for the inspection
of wire ropes in service, NDT International, June 1989, pp. 155 161.
[4] Herbert R. Weischedel, The Inspection of Wire Ropes in Service: A Critical Review, Materials
Evaluation, Vol. 43, December 1985, pp. 1592 1605.
[5] N. Sumyong, A. Prateepasen and P. Kaewtrakulpong, Influence of Scanning Velocity and Gap Distance
on Magnetic Flux Leakage Measurement, ECTI Transactions on Electrical Engineering, Electronics
and Communications, Vol. 5, No. 1, February 2007.
[6] E. Kalwa and K. Piekarski, Design of Hall effect sensors for magnetic testing of steel ropes, NDT
International, Vol. 22, No. 5, October 1987, pp. 295 301.
[7] C. Jomdecha and A. Prateepasen, Design of modified electromagnetic main flux for steel wire rope
inspection, NDT&E International, Vol. 42, 2009, pp. 77 83.
[8] Wang Hong yao, Hua Gang and Tian Jie, Research on Detection Device for Broken Wires of Coal
Mine Hoist Cable, Journal of China University Mining & Technology, Vol. 17, No. 3, pp. 376 381.
[9] E. Kalwa and K. Piekarski, Design of inductive sensors for magnetic testing of steel ropes, NDT
International, Vol. 20, No. 6, December 1987, pp. 347 353.
[10] W. Sharatchandra Singh, B. P. C. Rao, S. Vaidyanathan, T. Jayakumar and Baldev Raj, Detection of
Leakage Magnetic Flux from near side and far side Defects in Carbon Steel Plates using Giant
magnetoresistive Sensor, Measurement Science and Technology, Vol. 19, 2008 015702 (8pp).
[11] L. Chen, P.W. Que and T. Jin, A Giant Magnetoresistance Sensor for Magnetic Flux Leakage
Nondestructive Testing of a Pipeline, Russian Journal of NDT, Vol. 41, No.7, 2005, pp. 462 465.
[12] T. Chady and G. Psuj, Data fusion from multidirectional remanent flux leakage transducers for NDT
of stress and fatigue loaded steel samples, IEEE Trans. on Magn., Vol. 44, Nov. 2008, pp. 3285 3288.
This page intentionally left blank
Application of Electromagnetic
Nondestructive Techniques
This page intentionally left blank
Electromagnetic Nondestructive Evaluation (XIV) 267
T. Chady et al. (Eds.)
IOS Press, 2011
2011 The authors and IOS Press. All rights reserved.
doi:10.3233/9781607507505267
Introduction
Phenomena of pipe wall thinning are critical issues on ageing management of nuclear
power plants [17]. Pipe wall thinning management in nuclear power plants is aimed at
providing a life management process ensuring replacement or repair prior to inservice
failure. The main objective of pipe wall inspection is to identify the location of
maximum thinning, to ascertain the extent and depth of the thinning, and to evaluate
the wear rate. Currently the conventional technique is ultrasonic testing (UT) which
provides the accurate resolution for the pipe wall thickness measurement. However, in
UT, we must remove insulation from the piping area at each inservice inspection.
There also exist so many inaccessible locations for the insitu inspection by
surveyors in practical plants. Therefore, continuous surveillance during operation is
much more important than the periodical inspection by UT. Recently, we propose a
monitoring technique of pipe wall thinning using ElectroMagnetic Acoustic
Transducer (EMAT) [8]. EMAT is a noncontacting inspection device that generates an
ultrasonic pulse in the sample inspected [913]. The advantage of our method is that
measurements can be remotely implemented and the exciting test signals can be
automatically reproduced by electrical circuits. EMAT consists of a magnet and a coil
of wire and relies on electromagnetic acoustic interaction for elastic wave generation.
1
Corresponding Author: Fumio KOJIMA, Organization of Advanced Science and Technology, Kobe
University, 11 Rokkodaicho, Nadaku,Kobe 6578501; Japan, Email: kojima@koala kobeu.ac.jp
268 D. Kosaka et al. / Pipe Wall Thickness Measurement by EMAT Using Band Exciting Method
The pulsar signal was taken by two pulses with 2MHz. The reflective waveforms were
recorded by a computer coupled to an oscilloscope. Using PEC, thickness calculation
can be represented by
vt
T= (1)
2
where T , v and t denote thickness, sound velocity and time of flight (TOF). Sound
velocity of SS400 is 3.216103m/s. Detecting waveforms are shown in Figure 3. Figure
3(a) is a waveform of a plate sample specimen with 8mm thickness. Noting that the
time of flight (TOF) t was 4.9s and from Eq. (1), the estimated thickness was taken
as 7.9mm of which error became 0.1mm. Figure 3(b) is a waveform of a plate sample
specimen with 6mm thickness. Since TOF was t = 3.6s, the estimated thickness was
5.8mm where the error was 0.2mm. Taking into account that TOF t becomes
proportional to the thickness, it is more difficult to recognize TOF t for a smaller
thickness.
v
T= (2)
2f
where f denotes a peak interval with respect to the frequency. The detecting
waveforms of 4.7mm and 4.9mm thickness are shown in Figure 4. The f is very clear
that, if T is smaller, then the t of pulseecho method becomes smaller interval while
the f becomes larger interval. The f of 4.7mm thickness and the f of 4.9mm
thickness are clearly distinguishable. This implies that EMAR method is more accurate
measurement technique than PEC method. The f of the detecting waveform of 4.7mm
and 4.9mm thickness are 0.343MHz and 0.330MHz. Detecting thickness of the sample
specimens is 4.69mm and 4.87mm.
Figure 5 depicts the total resolution of the method presented here. In Figure 5, we
admit the detecting accuracy was up to 0.1mm for SS400. On the other hand,
measurement becomes timeconsuming because EMAR requires sweeping exciting
frequencies with wide range. In our experiment, each measurement spent about 2
minutes for obtaining sizing estimate.
In this section, the band exciting resonance method is applied to evaluate a mock FAC
sample specimen. As shown in Figure 8, a 6BSS400 pipe was tested and the corrosion
shape was fabricated to simulate flow accelerated corrosion (FAC). The extent of FAC
was taken as 200 mm with the maximum thinning depth b = 1.0mm. Figure 9 shows
estimating sizing results of the sample specimen by our measurement method. For the
purpose of the comparative discussions, the conventional UT techniques were also
performed.
Figure 9. Estimated sizing results for mock FlowAccelerated Corrosion sample specimen.
4. Conclusions
Two measurement methods for pipe wall thickness measurements were examined
using calibration specimen of SS400. It was shown that EMAR is superior to PEC
for its sizing accuracy. It was shown that the measurement strategy using the
exciting voltage with a wide band has achieved the rapid processing for thickness
measurements. The feasibility studies of the method are summarized for evaluating
mock FlowAccelerated Corrosion (FAC) test samples. Detecting results of the
method agreed well with detecting results of ultrasonic testing.
Acknowledgments
This work has been performed as a part of the National Research Project for
Enhancement of Measures against the Ageing of Nuclear Power Plants sponsored by
the Nuclear and Industrial Safety Agency (NISA). The authors also gratefully
acknowledge the Institute of Nuclear Safety System Inc. (INSS).
References
[1] C. Schefski, J. Pietralik, T. Dyke, M. Lewis, A physical model to predict wear sites engendered by flow
assisted corrosion, US DOE Rep, pp. 149154, 1995.
[2] B. Chexal et al., Flowaccelerated corrosion in power plants, EPRI report, TR106611, 1996.
[3] R. B. Dooley, V. K. Chexal, Flowaccelerated corrosion of pressure vessels in fossil plants, International
Journal of Pressure Vessels and Piping, Volume 77, Issues 23, pp. 8590, 2000.
[4] Flowaccelerated corrosion in nuclear power plants: Application of CHECWORKS at Darlington,
[5] T.R. Allen, P J. King, and L. Nelson: Flow Accelerated Corrosion and Cracking of Carbon Steel Piping
in Primary Water Operating Experience at the Point Lepreau Generating Station, Proceedings of the 12th
International Conference, pp. 773784, 2005.
[6] K. D. Efird, Flow accelerated corrosion testing basics, Pap Corros, p. 16, 2006.
[7] S. Uchida, M. Naitoh, Y. Uehara, H. Okada, N. Hiranuma, W. Sugino, S. Kishizuka and D. H. Lster:
Evaluation of Wall Thinning Rate with the Coupled Model of Static Electrochemical Analysis and
Dynamic Double Oxide Layer Analysis, Journal of Nuclear Science and Technology Vol. 46, No. 1, pp.
3140, 2009.
[8] D. Kosaka, F. Kojima, H. Yamaguchi and K. Umetani, Monitoring System for Pipe Wall Thinning
Management using Electromagnetic Acoustic Transducer, Vol.2, No. 1, pp.3442, 2010.
274 D. Kosaka et al. / Pipe Wall Thickness Measurement by EMAT Using Band Exciting Method
[9] H.M. Fros, Electromagneticultrasonic transducer: Principles, practice and applications, in: Physical
Acoustics, Vol. 14, W.P. Mason and R.N. Thurston eds., Academic Press, New York, pp. 179275, 1979.
[10] B.W. Maxfield and C M. Fortunko, The design and use of electromagnetic acoustic wave transducers,
in: Material Evaluation, Vol.41, pp. 13991408, 1983.
[11] R.B. Thompson, Physical principles of measurements with EMAT transducers, in: Physical Acoustics,
Vol.19, Academic Press, New York, pp. 157200, 1990.
[12] K. Mirkhani et al. Optimal design of EMAT transmitters, in: NDT&E International, Vol. 37, pp. 181
193, 2004.
[13] D. MacLauchlan, S. Clark, B. Cox, T. Doyle, and B. Grimmett, Recent advancements in the application
of EMATs to NDE, in: Proc. of the 16th WCNDT 2004, Montreal, Canada.
[14] K. Kawashima, Very high frequency EMAT for resonant measurement, in: Proc IEEE Ultrasonic
Symposium, No. 2, pp.11111119, 1994.
[15] H. Ogi, Field dependence of coupling efficiency between electromagnetic field and ultrasonic bulk
waves, J. Appl. Phys. 82 (8), 15, pp. 39403949, 1997.
Electromagnetic Nondestructive Evaluation (XIV) 275
T. Chady et al. (Eds.)
IOS Press, 2011
2011 The authors and IOS Press. All rights reserved.
doi:10.3233/9781607507505275
Keywords. Wall thinning, magnetic flux leakage (MFL), ancillary yoke pole,
reinforcing plate, dual layer
1. Introduction
An ageing management for nuclear power plants is quite important issue due to its
longterm operation in Japan. One of ageing problem is wall thinning on pipe. Wall
thinning occurs at orifices, elbows and under reinforcing plates, etc., through Flow
Accelerate Corrosion and Liquid Droplet Impingment erosion. Although several
methods like ultrasonic, eddy current are available tools, the magnetic flux leakage
(MFL) [14] also must be an effective nondestructive evaluation (NDE) technique for
the wall thinning, because this method has already been applied to industry field such
as gas pipeline. It is difficult to assess wall thinning under a reinforcing plate, where
steels are superposed for reinforcements, using ultrasonic and eddy current, because of
the reflection of signal and the wall thickness. Therefore MFL method is expected to
apply for an evaluation of wall thinning under the reinforcing plate.
The fundamental experiment of MFL using a singleyoke was examined and it
was clarified that sizing of a rectangular slit fabricated on steel plates is possible [5].
However, a development of a probe feasible for piping and a feasibility study of MFL
for application in a reinforcing plate are required for practical uses. Therefore, in this
study, an applicability of MFL method for wall thinning monitoring in nuclear power
plants was investigated, and a possibility to evaluate wall thinning under reinforcing
1
Corresponding Author: Hiroaki Kikuchi, NDE & Science Research Center, Faculty of Engineering,
Iwate University, 4 3 5 Ueda, Morioka 020 8551, Japan; Phone:+81 19 621 6890; Email: hkiku@iwate
u.ac.jp
276 H. Kikuchi et al. / Applicability of Magnetic Flux Leakage Method for Wall Thinning Monitoring
plates using MFL was examined. The ancillary yoke poles were utilized to solve the
problem in contact between the magneticyoke and the pipe surface and to achieve
higher sensitivity for estimation of slit depth on piping. Additionally, steels superposed
were prepared and a slit modeled wall thinning was formed in underlayer, then an
evaluation of the slit was performed using MFL method.
Figure 4(a) shows the dimension of the magneticyoke used for experimental
investigating the utilization of ancillary yoke poles. The geometry and the size of the
specimen are shown in Figure 5. The diameter is 60.5 mm and thickness is 5.2 mm, and
carbon steel STPT is used for their material. A slit with 10 mm width is fabricated at
13 Bz 13 Bz
Unit: mm Bx Unit: mm Bx
x x
0 0
9 S15C Pipe 9 S15C Pipe
Slit Slit
(a) with air gap (b) using ancillary yokes
Figure 1. 2D FEM analysis model.
H. Kikuchi et al. / Applicability of Magnetic Flux Leakage Method for Wall Thinning Monitoring 277
0 800
without slit
Leakage flux density Bx (Gauss)
800 0
1000 200
1200 400
without slit
1400 with slit 600
1600 800
16 12 8 4 0 4 8 12 16 16 12 8 4 0 4 8 12 16
Position x (mm) Position x (mm)
(a) x component, Bx (b) z component, Bz
Figure 2. Leakage magnetic flux density distribution with air gap between yoke and pipe.
0 100
without slit without slit
Leakage flux density Bx (Gauss)
80 0
100
120 50
140
160 100
16 12 8 4 0 4 8 12 16 16 12 8 4 0 4 8 12 16
Position x (mm) Position x (mm)
(a) x component, Bx (b) z component, Bz
Figure 3. Leakage magnetic flux density distribution with ancillary yokes between yoke and pipe.
the center part of the pipes, and its depths are 0.5, 1, 4, 4.5 mm, respectively. The
coordinate for measurement system is defined as shown in Figure 4(a). The center part
of slit is the origin of xdirection (horizontal), and the vertical direction to horizontal is
zdirection (the surface of specimen is z = 0).
Figure 4 (b) shows the geometry and the size of ancillary yoke pole adjusted
for the pipe shown in Figure 5. The material for this pole is pure iron. To obtain soft
magnetic properties, the pole was annealed at 700 C for 1 hour in the atmosphere after
its machining. After annealing, the poles were polished to remove oxide layer.
2.3 Results
Figure 6 plots x and z component of magnetic flux density, Bx, Bz against position x
(Here, center of the slit is x = 0). On the measurement, the applied current to the
excitation coil was 3 A, and the magneticyoke was fixed as the center of the yoke
corresponding with the center of the slit, and a gauss meter (Lakeshore 460) scanned
the pipe surface in xdirection at z = 0.5 mm. The pipes with slit depth of 4 and 4.5 mm
were used here. The results obtained with ancillary yoke poles are shown as compared
with measurement without ancillary yoke poles. It is clarified that the higher sensitivity
was obtained when the ancillary yoke poles were used. As for the pipe with slit depth
of 4 mm, the intensity of Bx has no change and Bz has no peak at the edge of slit in the
case without ancillary yoke poles, while the profile using ancillary yoke poles shows
large increment in Bx profile and has a peak in Bz at the edge of slit (x = 5 mm). For the
278 H. Kikuchi et al. / Applicability of Magnetic Flux Leakage Method for Wall Thinning Monitoring
Figure 4. Measuring system for MFL and dimensions of magnetic yoke and ancillary yoke pole.
# 2345mm
w 10 mm
60.5mm
t = 5.2mm 60mm
Figure 5. Dimension of pipe with slit.
20 200
Magnetic flux density Bx (Gauss)
40
100
60
80 0
100 100
120 4mm NAY 4mm NAY
4 5mm NAY 200 4 5mm NAY
140 4mm AY 4mm AY
4 5mm AY 4 5mm AY
160 300
2 0 2 4 6 8 10 12 14 2 0 2 4 6 8 10 12 14
Position x (mm) Position x (mm)
(a) x component, Bx (b) z component, Bz
Figure 6. Experimental leakage magnetic flux density distribution with and without ancillary yoke poles
between yoke and pipe (AY: with ancillary yoke, NAY: without ancillary yoke).
pipe with slit depth of 4.5 mm, the changes in Bx and the peak intensity in Bz at x = 5
mm were enhanced using ancillary yoke poles.
Figure 7 plots x and z component of magnetic flux density, Bx, Bz against
position x, when the probe includes the magneticyoke, ancillary yoke poles and the
gauss meter scanned the pipe surface in xdirection (In this case, gauss meter and
magneticyoke scanned together). The depth of slit changes from 0.5 to 4.5 mm. The x
component Bx of flux has rapid increase at x = 5 mm, and its rate of change increases
with increasing slit depth. On the other hand, the zcomponent Bz of flux has a peak at x
= 5 mm, and its intensity increases with increasing depth of slit. The position x = 5 mm
is consistent with the edge of slit.
Pipe specimens with modeled wall thinning being similar to the practical cases were
H. Kikuchi et al. / Applicability of Magnetic Flux Leakage Method for Wall Thinning Monitoring 279
20 300
Magnetic flux density Bx (Gauss)
60
14 20
57
63.4 Unit: mm
Figure 8. Dimension of pipe with wall thinning. a is Figure 9. Dimension of ancillary yoke
maximum depth, b width of slit and ' central angle of wall pole.
thinning part.
prepared, and MFL method using ancillary yoke poles was examined to assess wall
thinning of those specimens, here.
The configuration and the size of pipe are shown in Figure 8. Its diameter is
114.3 mm and its thickness is 8.6 mm, respectively. The material of the pipe is carbon
steel STPG. We prepared two specimens, A and B. The value of a, b and ' shown in
Figure 8 are 4.4 mm, 100 mm, 130 deg., respectively for specimen A, and those values
for specimen B are 2.15 mm, 50 mm, 90 deg., respectively. The magneticyoke used
here was the same as mentioned in Figure 1 (a), while another ancillary yoke poles
adjusted for the pipe shown in Figure 8 were prepared. Figure 9 shows the geometry
and the size of ancillary yoke pole. The gauss meter was fixed at the center part of the
magneticyoke, and the magneticyoke with the ancillary yoke poles scanned the pipe
surface in xdirection. The x component of magnetic flux density, Bx was measured
with application of 3 A current to the excitation coil. The scans were performed along
with the line passing through the deepest wall thinning point in both A and B
specimens and were also performed along with the line passing through out of wall
thinning area; this means scanning excludes wall thinning part.
Figure 10 shows the distribution of x component of magnetic flux density Bx
against position x. For both specimen A and B, magnetic flux density is constant on the
line passing through out of wall thinning area, while the distributions on the line
passing through deepest wall thinning point reflect the depth profiles of wall thinning
in each specimen (the cases of A: a = 4.4 mm and B: b = 2.15 mm). The magnetic flux
density at deepest wall thinning point in specimen A is larger than that of specimen B.
At deeper wall thinning point, larger magnetic flux density is obtained.
280 H. Kikuchi et al. / Applicability of Magnetic Flux Leakage Method for Wall Thinning Monitoring
40
30
Lebih dari sekadar dokumen.
Temukan segala yang ditawarkan Scribd, termasuk buku dan buku audio dari penerbitpenerbit terkemuka.
Batalkan kapan saja.