Eectromagnetic Nondestructive Evaluation PDF
Eectromagnetic Nondestructive Evaluation PDF
Eectromagnetic Nondestructive Evaluation PDF
Studies in
Applied Electromagnetics and Mechanics
Editors
Volume 23
Previously published in this series:
Volumes 1-6 have been published by Elsevier Science under the series title "Elsevier Studies in Applied
Electromagnetics in Materials".
ISSN: 1383-7281
Electromagnetic
Nondestructive Evaluation
(VI)
Edited by
F. Kojima
Kobe University, Japan
T. Takagi
Tohoku University, Japan
S.S. Udpa
Michigan State University, USA
and
J. Pavo
Budapest University of Technology and Economics, Hungary
/OS
Press
Ohmsha
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 the prior written permission from the publisher.
Publisher
IOS Press
Nieuwe Hemweg 6B
1013 BG Amsterdam
The Netherlands
fax:+31 206203419
e-mail: [email protected]
LEGAL NOTICE
The publisher is not responsible for the use which might be made of the following information.
Organized by:
Japan Society of Applied Electromagnetics and Mechanics
Kobe University
In Cooperation with:
The Japan Society of Mechanical Engineers
The Magnetic Society of Japan
Japan Society for Simulation Technology
Japan Biomagnetism and Biomagnetics Society
The Society of Instrument and Control Engineers
Institute of Systems, Control and Information Engineers
The Japanese Society for Non-Destructive Inspection
The Institute of Electrical Engineers of Japan
The Japan Society of Plasma Science and Nuclear Fusion Research
The Atomic Energy Society of Japan
The Iron and Steel Institute of Japan
Japan Society for Precision Engineering
Cryogenic Association of Japan
Co-sponsors:
Inoue Foundation for Science, Japan
The Kajima Foundation, Japan
Hyogo Science and Technology Association, Japan
Standing Committee:
J.R. Bowler, Iowa State University, U.S.A., Chairman
R. Albanese, Universita Reggio Calabria, Italy
G. Dobmann, Fraunhofer-Institute for NOT, Germany
R. Grimberg, National Institute of R&D for Technical Physics, Romania
H.K. Jung, Seoul National University, South Korea
F. Kojima, Kobe University, Japan
D. Lesselier, DRE-LSS CNRS-SUPELEC, France
V. Lunin, Moscow Power Engineering Institute, Russia
K. Miya, International Institute of Universum, Japan
G.Z. Ni, Zhejing University, China
J. Pavo, Budapest University of Technology and Economics, Hungary
A. Razek, LGEP CNRS-SUPELEC, France
G. Rubinacci, Universita Cassino, Italy
J.N. Sheng, X'ian Jiatong University, China
T. Sollier, CEA Paris, France
T. Takagi, Tohoku University, Japan
H. Takamatsu, KEPCO, Japan
S.S. Udpa, Iowa State University, U.S.A.
V. Vengrinovich, Institute of Applied Physics, Belarus
Secretary:
Futoshi Kobayashi
Graduate School of Science and Technology, Kobe University
1-1 Rokkodai, Nada 657-8501, JAPAN
Tel: +81-78-803-6489, Fax: +81-78-803-6493
E-mail: ende2001 @buna.scitec.kobe-u.ac.jp
E'NDE - Kobe, List of Participants
ALBANESE, Raffaele DOBMANN, Gerd
Univ. Reggio Calabria Fraunhofer-Institut fur Zerstorungsfreie
DIMET, Fac. Ingegneria, Via Graziella, Loc. Feo Priifvergahren
di Vito, 1-89100 Reggio Calabria, ITALY Universitat, Geb. 37, 66123 Saarbrucken.
+39081 7683243 GERMANY
+39081 7683171 +49681 93023855
albanese (ffunirc.it +49681 93025933
dobmann@ izfp.fhg.de
ARA, Katsuyuki
Applied Superconductivity Research Laboratory DOI, Tatsuya
Tokyo Denki University Ashikaga Institute of Technology
2-1200, Gakuendai, Busei, Inzai Ohmae, Ashikaga, Tochigi, 326-8558 JAPAN
Chiba, 270-1382 JAPAN +81-284-62-0605
ara(ff hp.asrl.dcndai.ac.jp +81-284-62-4633
1-doi (ffashitcch.ac.jp
BANKS, H. T.
Center for Research in Scientific Computation, ENDO, Hisashi
North Carolina State University c/o Professor Yoshifuru Saito
Raleigh, NC 27695-8205 USA Graduate School of Engineering. Hosei University
[email protected] 3-7-2 Kajino Koganei,Tokyo 184-8584 JAPAN
+81-42-387-6200
BEUKER. Thomas +81-42-387-6213
H. Rosen Engineering GmbH, cndo@ ysaitoh.k.hosei.ac.jp
Research & Development Dept.
Am Seitenkanal 8, D-49811 Lingen (Ems) ENOKIZONO, Masato
GERMANY Faculty of Engineering, Oita University
+49-591-9136-414 700 Dannoharu, Oita 870-1192 JAPAN
+49-591-9136-121 +81-97-554-7821
+81-97-554-7822
BOWLER, John [email protected]
Electrical & Computer Engineering,
Iowa State University GOTTVALD, Ales
1915 Scholl Road, Ames Iowa, 50011-3042 USA Academy of Sciences of the CR.
5152942093 Insl. of Scientific Instruments
5152947771 Kralovopolska 147, 612 64 Brno
jhowlcrCgPcndc.iastatc.edu CZECH REPUBLIC
+4205415 14224
CHAMONINE, Mikhail +4205415 14402
H. Rosen Engineering GmbH, gottCg'isibrno.c/
Research & Development Dept.
Am Seitenkanal 8, D-49811 Lingen (Ems) HARADA, Yutaka
GERMANY Nuclear Engineering Ltd., Japan
+49-591-9136-414 1-3-7 Tosabori Nishi-ku Osaka 550-0001 JAPAN
+49-591-9136-121 +81-6-6446-9363
MChamonine (g' RosenGcrmany.de +81-6-6446-1746
MUMTAZ, Khalid
Faculty of Engineering, Iwate University
Ueda 4-3-5, Morioka, Iwate 020-8551 JAPAN
+81-19-621-6350
+81-19-621-6373
([email protected]
NAGATA, Shoichiro RUBINACCI, Guglielmo
Faculty of Engineering, Miyazaki University University of Cassino
1-1 Gakuenkibanadai-nishi. DAEIMI, Via di Biasio 43 - 103043 Cassino.
Miyazaki, 889-2195 JAPAN ITALY
nagata@ cc. miya/aki-u. ac.jp +39 0776 299 626
+39 0776 299 628
NAKASONE, Yuji ruhinacci (gunicas.it
Department of Mechanical Engineering
Science University of Tokyo SAKAMOTO, Shintaro
1- 3, Kagurazaka, Shinjyuku-ku, Tokyo, Shinryo Co.
162-8601 JAPAN 41 Wadai Tsukuba Ibaraki JAPAN
+81 -3-3260-4272 (EXT 3356) +81-298-64-6110
+81-3-3260-4291 +81-298-64-6127
nakasone (g-rs.kagu.sut.ac.jp sakamoto.sh @ shinryo.com
Inverse Problems
Direct Problems
Achievement of RFEC Effects in the Nuclear Fuel Rod Inspection by Using Shielded
Encircling Coils, Y.-K. Shin 83
Behavior of the Remote-Field on a Tube with Fan-Shaped Defect,
M. Yasunishi, M. Jsobe and M. Nishikawa 91
Inversion of ECT Signals from Flaws with Tip Variation in Steam Generator Tubes,
S.-J. Song, Y.-H. Kim, C.-H. Kim, E.-L. Kim, Y.-K. Shin, H.-B. Lee, Y.-W. Park and
C.-J. Yim 98
Development of Practical MFES System for Concrete Materials,
5. Nagata, T. Chady, M. Shidouji and M. Enokizono 104
Evaluation of Crack Signals for a Superconductive Pulsed Eddy-Current Probe.
J.R. Bowler, N. Bowler and W. Podney 108
Time-Domain Finite Element Analysis of Remote Field ECT Excited by Pulse
Current, K. Yamazaki ]]6
Materials Characterization
New Sensors
Signal Detection
Abstract.
This paper introduces a computational method for use with eddy current damage
detection techniques. To identify the geometry of a subsurface damage, an optimiza-
tion algorithm is employed which requires solving the forward problem numerous
times. In order for the method to be effective in a practical setting, i.e., in real-time ap-
plications, the forward algorithm must be extremely fast and accurate. Therefore, we
have chosen an approach based on reduced order Proper Orthogonal Decomposition
(POD) methods. This allows one to create a set of basis elements using snapshots with
either numerical simulations or experimental data. The data is organized in an opti-
mal way allowing one to use a reduced number of basis elements, resulting in a fast
algorithm while still obtaining an accurate approximation to the solution. We first de-
rive the model associated with the chosen nondestructive evaluation (NDE) technique
and prove some well-posedness results for the model. We then introduce the proposed
computational methodology and test it on both numerically simulated data as well as
experimental data obtained from a GMR (Giant Magnetoresistive) sensor. The results
demonstrate that the method is extremely efficient and accurate.
An advanced method of damage detection uses a device such as the GMR sensor in the con-
text of eddy current methods [1, 2]. In a standard experimental setting, as depicted in Figure
1, a thin conducting sheet carrying a uniform current is placed above (or below) the sample.
The current within the sheet induces a magnetic field perpendicular to it that in turn produces
H.T. Banks et al. /Electromagnetic Interrogation Techniques for Damage Detection
a current within the sample, called an eddy current. When a flaw is present within the sample,
the flaw disrupts the eddy current flow near the flaw and this disturbance is manifested in the
magnetic flux density detected by the measuring device.
For illustrative purposes, we will assume uniformity along the width (z direction) of the
sample, therefore reducing the three-dimensional problem to a two-dimensional problem.
To test the feasibility of reconstructing the geometry of the damage, we consider the dam-
age (which we shall refer to as a "crack") to be rectangular in shape. In other words, we
assume the crack, located at a certain depth within the sample, has a fixed length and thick-
ness. To further simplify the test problem, we disregard the boundary effects of the materi-
als in the x direction (sample length) by assuming an infinite sample and conducting sheet
in that direction. Because we are considering materials of infinite extent, we will construct
our forward problem by focusing on a small "window", called our computational domain
fi = { ( . i \ t j . ~ ) G R'! : Omm < ./ < 50mm. -35mm < ij < 35m. 771}, centered such that
the left boundary of the computational domain, at location .; = 0, is positioned in the center
of the crack in the .r direction, i.e., the crack is symmetric through the ijz plane at x = 0.
A schematic of the resulting two-dimensional problem is depicted in Figure 2 where it is
assumed that the sample (which is 20mm thick) is composed of aluminum, the conducting
sheet (which is O.lmm thick) is made up of copper and the crack is centered in the y direction
around the center of the sample (i.e., around y = -10mm).
In our computational efforts, we employ the use of the software package Ansoft Maxwell 2D
Field Simulator. Therefore, our equations are formulated to correspond to those used by the
software. (For a full derivation of the resulting equations, see [3].) Although the GMR sensor
measures the magnetic flux density above the sample at a certain location, our equations
are formulated in terms of the magnetic vector potential A, where all the field quantities
are assumed to be phasor quantities [4, 5, 6]. However, one can easily obtain the resulting
magnetic flux density B through the relationship B = V x A.
Using Maxwell's equations in conjunction with Ohm's law and constitutive laws, we
obtain an equation for the magnetic vector potential A given by
where ^ represents the magnetic permeability, n represents the conductivity, ~' is the angular
frequency, and o is a scalar potential.
H. T. Banks et al. / Electromagnetic Interrogation Techniques for Damage Detection 5
Since Eq.( 1 ) contains two unknowns, A and 0, an additional equation is needed to uniquely
determine the solution for A. For this we use an integral constraint given by
between the total current Ics flowing in the conducting sheet (cs) and the total current density
J, within the conducting sheet. This is the second equation used in the software package
Ansoft Maxwell 2D Field Simulator which we use in our computational efforts. However, this
only gives us equations which completely describe the magnetic vector potential and electric
scalar potential in the conducting sheet. Nonetheless, the conducting sheet is the only region
in which a source current of the form ,7, = crV</i is present and hence is the remaining
regions we intuitively assume there is no change in potential, in other words V(/ = 0. This
gives us appropriate equations in which the magnetic vector potential A can be uniquely
determined it appropriate boundary conditions on A are specified.
In Section 1 . 1 we assumed the sample was of infinite extent with the crack being sym-
metric in the ./: direction. In other words on the x boundaries, we assume the fields on both
sides of the boundary oscillate in the same direction. To account for the even symmetry, we
assign Neumann boundary conditions to these boundaries. In a similar manner, we assume
the // boundaries are "sufficiently far" away from the sample and scanning area so as to not
effect the overall measurements. Indeed, as one moves farther away from the sample and
conducting sheet, the magnetic vector potential A tends to zero. Therefore, on the y bound-
aries we assign null Dirichlet boundary conditions. The magnetic vector potential A is thus
determined by
and
H.T. Banks et al. /Electromagnetic Interrogation Techniques for Damage Detection
with
2 Well-Posedness
In this section we consider the existence and uniqueness of a weak solution to the above
boundary value problem on a general domain given by
for which our test problem is a specific example. Let H = L-2(Q) and V = {<.- 6 H l ( Q ) \
<-'( .i-. ymm) = 0 <'(<' fjmaf)} where we use the standard Sobolev space notation, //'(Q)
{(.' G L 2 (fi) : Vr G L"(Q)} and note that we interpret pointwise evaluation of functions
(along the boundary and elsewhere) in terms of a trace operator for which we suppress nota-
tion throughout this paper [7]. We denote by (o. r) = / u ocdu the standard inner product in
H and (o. i-)\- = /^ V6 Vi'Y/a the (//'-equivalent) inner product in V .
We note that in this two-dimensional problem, the term Vo can be proven to be piecewise
constant as done in [3]. In doing this, Vd can be written in terms of the magnetic vector
potential in the conducting sheet by solving for Vo in (4). We can then reduce the system into
an integro-differential equation. Using integration by parts together with natural boundary
conditions and imposed conditions on test functions . 6 V, the variational form is given by
rf(Q- q) = l l q - qll = [(-''I - -''I ) 2 + (.Vi - W i ) " + + (-''-I - ')' + (.V4 - !/4) 2 ] l '" 2 (8)
to be the standard Euclidean norm in R*. We denote by rtfi (see Figure 3) the points in Q
which are either in the damage represented by q or q but which are not in both. In other
words, let il^ represent the points (j. y) in M within the damage given by q and Qq the points
(./. y) in Q within the damage given by q. Then SQ Uq U ilq 12q H i2 q .
Using the terminology above, we have the following theorem. Again, we refer the reader
to [3. 8] for details and proof.
H. T. Banks ct al. /Electromagnetic Interrogation Techniques for Damage Detection
Conducting Sheet
Theorem 2.2. Assume the admissible parameter set Qad "' a compact subset of K8. Then
there exists T = J-([i, ( , 1) such that for source frequencies fs < J-(n, e, J7), q > A(q) is
continuous from Qn<i to V.
3 Computational Method
In this section we discuss the concepts and method of implementation of the POD method in
the context of damage detection. For details on the general POD method, we refer the reader
to [9, 10, 1 1 , 12, 13, 14, 15, 16, 17, 18, 19, 20] and the extensive list of references contained
therein. In the process of detecting a subsurface damage, a device, such as the GMR sensor,
is scanned above (or below) a sample and data is taken. The variation in the field due to
a damage is manifested in the data taken. Therefore, in forming a reduced basis, we want
to incorporate the effects of a damage in the reduced basis. This is accomplished by taking
"snapshots" of the data across various damages. In other words, let q be a vector parameter
characterizing physical properties of the damage such as length, thickness, depth, etc. of the
damage. Then given an ensemble of damages {q,}^|, we obtain corresponding solutions,
{A(q ; )}^ 1 , of the boundary value problem, for magnetic vector potentials which we call
our "snapshots". We then use these snapshots to generate a basis incorporating the properties
of the various damages. (Without loss of generality, we will denote the vector A by its scalar
nonzero component A, i.e., the /l:i component of A.)
The formation of the POD basis can be summarized in a few steps. In order to success-
fully reduce the number of basis elements while maintaining accuracy, a single basis element
must contain aspects of each damage q^. Therefore, as explained in [20, 21], we seek basis
H.T. Banks et al. /Electromagnetic Interrogation Techniques for Damage Detection
(9)
where the coefficients \ ](j) are chosen such that each POD basis element 4>,, / = 1. 2 .Vs,
maximizes
subject to {<,. <bi)L'2(n.C) l l ^ i l l 2 = 1- Forming the basis elements in such a way assures
that a single basis element will contain information from each of the snapshots. From standard
arguments, it can be seen that the coefficients \ ] ( j ) can be found by solving the eigenvalue
problem CV = \\' where C is given by
Hence, the next step in forming the POD basis is to form the covariance matrix C and find
the associated eigenvalues and eigenvectors. Since C is Hermitian positive semi-definite, we
know it possesses a complete set of orthogonal eigenvectors and corresponding nonnegative
eigenvalues. In forming the POD basis, we want to be able to readily decide which basis
elements to use in the reduced basis. To this end, we order the eigenvalues along with their
corresponding eigenvectors such that the eigenvalues are in decreasing order,
Consequently, the /'" POD basis element is defined by Eq. (9) where \ \(j) represents the jth
component of the ith eigenvector of C'.
We now have the full POD basis and need a criterion to decide how many basis elements
are required to accurately portray the data. In other words, we want to choose .V such that
1>7-H"/{<I>''}|1i ~ f>P"n{A(cij)}*.ll. In choosing this number .V, we compute
which represents the percentage of "energy" in ^p<iii{A(q/)}'^l that is captured in span {$_,}*
Then the reduced basis consists of the first .V POD basis elements where ;V is chosen to cap-
ture the desired amount of energy.
We now note that span{$i} -\s, = 1 >7w{--l(q,)}'^ 1 . Indeed, given any .4(q_,), we have
where
fU-(q,) = (.-Hq,).<IV)^a,r) (16)
as { $ , 1 1 are orthonormal in I 2 ( f > . ( T ) .
T
However, to complete the analysis, one needs to be able to calculate A A '(q) where q is a
given parameter not in the set {q^j^j. To this end, we extend the approximation formula to
obtain
where rv A .(q) can be evaluated through different methods. In [3], we explored two different
methods one might choose: a POD/Galerkin method or a POD/Interpolation method. How-
ever, in this paper, we only present results in which we use the POD/Interpolation method
with linear interpolation for the one-parameter simulated results and cubic spline interpola-
tion for all other results. For details on these methods, we refer the reader to [22] and [23].
In [4], we performed several trials in which we assumed we had access to various types of
data, such as the A field or the B field at various points ( x l , y ] ) in Q. We compared and
contrasted the accuracy to which we could estimate the length / of the damage based on
whether the A field or B field was used and whether we considered the field along a single
line, multiple lines or within the entire region (which is not experimentally possible and was
only tested for initial comparisons). From the results presented in [4], we concluded that
extremely accurate results were obtained only when the y component B2 of the magnetic flux
density was used in the cost criterion, i.e., when we used
where 10" is a scaling factor accounting for the low order of magnitude of the field (_B2 is on
the order of 10~*Wb/m 2 ), BN(q) - V x ,4 A '(q) is the reduced order POD approximation in
which .4A (q) is given by (17), and B2 is "data" from a sample we wish to characterize. (In
this section, B-> is obtained from Ansoft finite element simulations to which noise has been
added in the usual manner (see [3, 21]). Furthermore, performing multi-line scans or using
full region data improved the results only marginally and hence did not warrant the extra
effort and time in collecting more extensive data sets. Consequently the results presented in
this section involve only the least squares difference in the B2 field given by (18) along a
single line located 1mm above the conducting sheet.
In the sample results presented, we focus on estimating a single parameter of the crack
assuming all other parameters are fixed or estimating two parameters with one parameter
fixed. The results below summarize the feasibility of determining the length and depth of
a crack. In a specific trial run, ten different data sets (exact data with ten different sets of
added random noise) are used where the relative noise is chosen at a 10% noise level with a
confidence level of 99.7% (3 standard deviations). (Details can be found in [3] and [4].)
In determining the length of the damage alone, we first followed the steps outlined in
Section 3.1 and generated an ensemble of "snapshots". Keeping the crack fixed at a depth of
$111in with a thickness of 2mm, we varied the crack length / from Omm to 4mm in increments
of 0.2mm. Using the snapshots, {^4(/. / )} A j;f 21 , we formed the POD basis and determined that
99.99%; of the energy of the system was captured with a single basis element [4]. However,
in the specific trial runs, we found that using 4 POD basis elements improved the parameter
estimation with data containing no noise, while there was no significant difference between
the use of 5 basis elements over the use of 4 basis elements. Therefore, the results are based
on the use of 4 basis elements in the algorithm.
10 H. T. Banks et al. / Electromagnetic Interrogation Techniques' for Damage Detection
To test the inverse methodology, we first try to identify the length of the damage, /* =
1.3mm. Using data on a single line above the conducting sheet with 10CX relative noise, an
average estimated length of 1.2977mm was obtained with a variance of 0.3237 x 10~ 4 mm 2
across the ten trials.
Proceeding as we did in estimating the length of a damage, to estimate the depth of the
damage, we fixed the thickness of the crack at 0.5mm with length 2mm and varied the depth
of the crack d from 0.25mm to lG.25;;m? in increments of Q.Hntm. Using 5 basis elements,
we were able to accurately estimate a crack depth of Sunn in the presence of 10% relative
noise; an average depth of 8.0G31m/;/ was estimated with a variance 0.1180 x 10~ 3 m/n 2 .
Thus, as in the case of estimating the length of a crack, we can also recapture the depth of a
crack quite accurately and efficiently.
Finally, we estimated both length and depth simultaneously. However, first, in [3, 8],
we discussed in detail a need to modify the assumptions made in the original test problem
to more accurately describe the behavior of experimental data used in the next section. In
short, the computational domain was expanded beyond the edges of the sample and snapshots
were taken of the magnetic flux density data on a single line above the conducting sheet
(instead of the whole region). Recall that for the previous trials, we took snapshots of the
magnetic vector potential for the entire computational domain even though in the inverse
problem we only considered those data points along a single line. Furthermore, we considered
data across the entire length of the sample, instead of just half the sample as done previously.
(For more details , see [3, Chapter 6].) As a result, we also implement these changes in the
two-parameter estimation problem.
We proceed as in the previous estimation problems by first generating an ensemble of
damages. We consider damages with depths ranging from 1mm to l l m m in increments of
2mm in combination with lengths from ().5rm to 3.5rm in increments of Irm (we now
consider longer damages similar to those used in obtaining experimental data). We keep the
thickness fixed at 1mm. A total of 24 snapshots, {B>(d,. / , ) } , / = 1 6, j = 1 4 were
generated using Ansoft. Using 4 basis elements with 10% relative noise added, we estimated
a damage with depth d' = 4mni and length /* = 2cm. We obtained average estimates of d =
4.0855mm with variance 0.551G x 1 0 ~ ' m m - and length / = 1.980rm with variance 0.9G32 x
10~^mm 2 . Thus, even when we estimated two parameters simultaneously, the computational
method proposed produced extremely accurate results.
Since the results using simulated data were extremely promising, we designed an experiment
in which we tried to detect and parameterize a damage within an aluminum sample using
a giant magnetoresistive (GMR) sensor. The sample was constructed of 17 layers of 1mm
thick aluminum plates with a slice cut out of one of the layers to simulate a damage within
the sample (see [3, 8] for graphical representations). The "damaged" piece of aluminum is
moved from one layer to another to simulate damages within the sample at different depths,
and the length of the damage is varied by producing "gaps" of varying size from the aluminum
plate (the thickness of the damage is always fixed at 1mm). As a means of inducing current
within the sample, a thin sheet of copper carrying a uniform current of 3 Amps is placed
above the sample on top of a thin sheet of paper (to avoid direct physical contact between the
sample and the conducting sheet). The GMR sensor measures the amplitude and phase of the
magnetic flux density across a 2m line (along the length of the sample) every O.G35mm. The
data is then filtered through a lock-in amplifier and saved to a file.
The experimental data did not resemble Ansoft simulations from our model: therefore.
H. T. Banks el al. /Electromagnetic Interrogation Techniques for Damage Detection
we chose to snapshot on the data itself in forming the POD basis elements. Furthermore, it
was necessary to filter out the background noise (data for a sample containing no damage)
in order to obtain a representative pattern in the data as a function of the damage within the
sample (see [3] for full details). In filtering out the background noise, we magnified boundary
effects from the edges of the sample. Therefore, in the cost criterion, we only consider data
across the center of the sample. In other words, we use the cost criterion given by
where q is the vector containing the parameters we wish to estimate, B% (q) is the POD
approximation formed using snapshots on the data itself, B\ is GMR data at grid points Xj,
j = 1, ..., n with n total grid points and a and b indicate the indices of the grid points we
consider in our cost criterion.
In estimating a depth of d* = 3mm with corresponding length I* = 1.5cm, we obtained
an estimate of d 2.9759mm and / 1.4G12cm, resulting in a relative error for depth of
R,i = 0.80% and length R/ = 2.59%. Again, even with experimental data, the results were
accurate.
4 Conclusion
Acknowledgments
This research was supported (MLJ) by the NASA Langley Graduate Researcher's Program
under grant NOT-1-52196 and in part (HTB) by the Air Force Office of Scientific Research
under grant AFOSR F49620-01 -1 -0026.
12 H. T. Banks et al. /Electromagnetic Interrogation Techniques for Damage Detection
References
[ I ] Buzz Wincheski and Min Namkung. Development of very low frequency self-nulling probe for inspection
of thick layered aluminum structures. In /99# Review of Progress in Quantitative NDE. Snowbird. Utah.
Aug. 1998.
[2] Buzz Wincheski and Min Namkung. Deep flaw detection with giant magnetoresistive (GMR) based self-
nulling probe. In 7999 Review of Progress in Quantitative NDE. Montreal. Canada. July 1999.
[3] Michele L. Joyner. An Application of a Reduced Order Computational Methodologv for Eddy Current
Based Nondestructive Evaluation Techniques. PhD thesis. North Carolina State University. 2001.
[4] H.T. Banks. M.L. Joyner. B. Wincheski. and W.P. Winfree. Evaluation of material integrity using reduced
order computational methodology. CRSC Tech. Rep. CRSC-TR99-30. North Carolina State University.
1999.
[51 David K. Cheng. Field and Wave Electromagnetics. Addison-Wesley. Reading. MA. second edition. 1992.
[6] Ansoft Corporation. Maxwell 2D Field Simulator - Technical Notes, 1995-1999.
[7] Lawrence C. Evans. Partial Differential Equations. American Mathematical Society. Providence. 1991.
[8] H.T. Banks. M.L. Joyner. B. Wincheski. and W.P. Winfree. Real time computational algorithms for eddy
current bas based damage detection. In progress.
[9] H.T. Banks. R.C. del Rosario, and R.C. Smith. Reduced order model feedback control design: Numerical
implementation in a thin shell model. IEEE Trans. Auto. Control. 45:1312-1324. July 2000.
[10] G. Berkooz. Observations on the proper orthogonal decomposition. In Studies in Turbulence, pages
229-247. Springer-Verlag. New York. 1992.
[ I I ] G. Berkooz, P. Holmes, and J.L. Lumley. The proper orthogonal decomposition in the analysis of turbulent
flows. Annual Review of Fluid Mechanics. 25(5):539-575. 1993.
[ 12] R.C. del Rosario. Computational Methods for Feedback Control in Structural Systems. PhD thesis. North
Carolina State University. 1998.
[13] K. Karhunen. Zur spektral theorie stochasticher prozesse. Ann. Acad. Sci. Fennicae. 37(A1). 1946.
[14] M. Kirby and L. Sirovich. Application ot the Karhunen-Loeve procedure tor the characterization of human
faces. IEEE Transactions on Pattern Anal\sis and Machine Intelligence. 12( I ) : 103-108. 1990.
[15] M. Kirby. J.P. Boris, and L. Sirovich. A proper orthogonal decomposition of a simulated supersonic shear
layer. International Journal for Numerical Methods in Fluids. 10:41 1428. 1990.
[16] K. Kunisch and S. Volkwein. Control of Burgers" equation by a reduced-order approach using proper
orthogonal decomposition. J. Optimization Theory andApplic.. l()2(2):345-371. 1999.
[17] M. Loeve. Functions aleatoire de second ordre. Compte rend. Acad. Sci.. Paris. 1945.
[ 18] J.L. Lumley. The structure of inhomogeneous turbulent flows. Atmospheric Turbulence and Radio Wave
Propagation, pages 166-178, 1967.
[19] J.L. Lumley. Stochastic Tools in Turbulence. Academic Press. New York. 1970.
[20] H.V. Ly and H.T. Tran. Proper orthogonal decomposition for flow calculations and optimal control in a
horizontal cvd reactor. CRSC Tech. Rep. CRSC-TR98-I3. North Carolina State University. 1998. Quart.
Appl. Math, to appear.
[21] H.T. Banks. M.L. Joyner. B. Wincheski. and W.P. Winfree. Nondestructive evaluation using a reduced-
order computational methodology. Inverse Problems. 16. 2000.
[22] Altio Quarteroni. Rjccardo Sacco. and Fausto Saleri. Numerical Mathematics. Springer-Verlag. New
York. 2000.
[23] J. Stoer and R. Burlisch. Introduction to Numerical Analvsis. Springer-Verlag, New York. 2nd edition.
1993.
Electromagnetic Nondestructive Evaluation (VI)
F. Kojima et al. (Eds.)
IOS Press, 2002
1. Introduction
For the sake of simplicity we assume that the resistivity at spatial coordinate r assumes
the value 77, or Tjh, being ;/, and % fixed reference values and 7]t>rih. The forward problem
consists of computing the resistance matrix R when the resistivity rj(r) is ;/, in V and
assumes the value rjb in D\V for a given VcD, D being the region containing the
conductor. The forward problem realizes the mapping VcD > R. The electrodes involved
in the definition of the conductance matrix are located on the boundary of D. Each
electrode is assumed to be to be equi-potential; we also neglect the contact resistance
(shunt model). Let Sk be the part of S=dD in contact with the k-th electrode; we assume
14 G. Rubinacci el al. /Shape Identification in Conductive Materials
also that the current density has no normal component on Sj, the part of S that is not in
contact with any electrodes.
Here we numerically compute the resistance matrix R by means of the complementary
formulations as in [2]:
where R* arises from the scalar potential ( E = -V<p ) based formulation and RA from the
electric vector potential (J = V x T ) based formulation. Specifically, R^ is defined by
PA = 1TRA1 where I is the column vector containing the net currents flowing into the
electrodes and PA is the value of the ohmic power losses computed using the electric
vector potential formulation. Similarly, R* is defined as R * = ( G * j where
PQ =v r G*v n-presents the ohmic power losses and v is the coh mn vector of the
electrode voltages. In particular, (p and T are the numerical solutions of the variational
formulations of the field problem here briefly summarized ([2-3]):
where
The numerical solution of problem (2) and (3) is then obtained by restricting the
minimization to O^, and A'N , finite dimensional subset of & and A1 , respectively, and by
using standard isoparametric shape functions for the scalar potential and edge-elements
based shape functions [2] for the electric vector potential.
The inverse problem consists of identifying the subset V where the resistivity Tj(r) assumes
the value 77,, from the knowledge of the measured (and noisy) resistance matrix R . We
recall that the resistivity in D\V is equal to a known value Tjb.
The key issue for the solution of the inverse problem is the following properties of the
resistance matrices.
Proposition 1. TJ} >TJ2 => R, ^ R 2 where RI is the resistance matrix related to //, , R2 is the
resistance matrix related to 7]2 and R, > R 2 means that R , - R 2 is a positive semi-
definite matrix, i.e. \T (R, - R 2 )x > 0,Vx .
Let T, and T2 be defined as the solution of (3) when the column vector I of the electrode
voltages is given and the resistivities are //, and TJ2 , respectively. Then it follows that
G. Rubinacci et al. /Shape Identification in Conductive Materials 15
where the first and last equality follow from the definition of resistance matrix, the first
inequality follows from the hypothesis and the second inequality follows from (3) when
particularized to ?J2 . The thesis follows from the arbitrariness of I. *
Proposition 2.
where GI (resp. 62) is the conductance matrix due to cr, (resp. <72), Gf (resp. G) is the
conductance matrix due to (7, (resp. <J2 ) but numerically computed by using the scalar
potential formulation, and Rj 4 (resp. R 2 ) is the resistance matrix due to 77, (resp. 7/ 2 ) but
numerically computed by using the electric vector potential. *
hi addition, we notice that (see Appendix) if A and B are two real, symmetric and invertible
matrices, then
Corollary 1.
Corollary 2.
where R 0 ( R, ) is the resistance matrix when TJ -rji in DO (D\) and rj-rjb in Z)\D0 (D\D\).
Property (17) is the key for the inversion procedure. In fact from (17) it follows that:
Therefore, we can establish if >0 cr Z), , or Z), c D0 or (D, CZ! D0 and D0 cz /), ) by simply
analyzing the eigenvalues of the matrix R, -R 0 . A numerical example to validate (18)-
(20) is reported in Figures 1 and 2 showing three different resistivity patterns together with
the eigenvalues for the three different cases (18)-(20).
We now discuss an inversion procedure that takes into account properties (18)-(20) to
retrieve the unknown subset V c D where the resistivity assumes the value rjj.
16 G. Rubinacci et al. /Shape Identification in Conductive Materials
Figure 1. Three different resistivities profile on a gray scale map where white stays for 5Qm and gray for
5kQm for a two-dimensional example. The resistance matrix related to (a) is RA where; > the resistance
matrix related to (b) is RB and the resistance matrix related to (c) is RO
CO 100
Eigenvalue no.
Figure 2. The eigenvalues for the three different cases ( 1 8)-(20) obtained by considering the resistance
matrices RA, RB and RT
The main idea consists of partitioning the domain D in N small non-overlapping parts
QI, ..., QAT. Then by using (18)-(20) we check if Q* is contained in V by computing the
eigenvalues of the matrix R - R A where Rt is the numerically computed resistance
matrix related to a resistivity Tjk defined as:
n h Vre D \ Q t
(21)
Then we set V , the estimate of V , as the union of those Q* such that R - Rt is positive
semi-definite. We notice that, from (18)-(20), it results that KcF . Similarly, it is
possible to build a set V s.t. V c V .
It is worth noting that the computational cost of this procedure grows linearly with N, the
number of Q*'s. Moreover, the matrices RA can be computed and stored thus the inversion
of the data consists of computing only the eigenvalues of R - Rk for k=\, ..., N. Moreover,
the size of the matrix R - R * is modest being a nxn matrix where n+\ is the number of
electrodes that, usually is not greater than few tenth.
The proposed procedure requires a slight modification when the matrix R is affected by the
measurement noise. Specifically, the noise affecting the matrix R may corrupt the
eigenvalues of the matrix R - Rk thus resulting with an undefined sign also when Qt c V .
To manage this situation, we compute a sign index Sk related to Q* defined as:
(22)
G. Rubinacci et al. / Shape Identification in Conductive Materials 17
Figure 3. The reference geometry together with the finite element mesh.
where Ak f is they'-th eigenvalue of the matrix R-R* . The sign index sk ranges from -1
when R - R t is a negative definite matrix (i.e. when F0 c f i ^ ) to +1 when R - R A is a
positive definite matrix (i.e. when Qk c F0 ). For values internal to the interval (-1, +1) the
sign of R - R * is not definite. The effect of noise consists of producing a sign index 5*
different from +1 and -1 also when Q.k c F0 or V0 c Q t . We define a reconstruction with
threshold the union of those Q* with sign greater than e, i.e.:
k\Sk>
(23)
The last problem is the selection of a proper value of the threshold e. To this purpose, we
solve the one parameter minimization problem:
(24)
-R e | 2 (25)
where the matrix R is related to Ve and ||-|| is a suitable matrix norm.
Finally, we notice that the inherit ill-posedness of this inverse problem directly affects the
achievable resolution. In fact, as the dimension of Q/t decreases, the matrix R A tends to the
resistance matrix related to an uniform resistivity equals to Tjh and the noise affecting R
produces a sign index 5* almost independent from k.
4. Numerical examples
Figure 4. The unknown resistivity profile on a gray scale map. White stays for Tjh and gray for 77,
(a) (b)
(c) (d)
Figure 5. The (normalized)/) error functional (a), together with the reconstructions corresponding to
=0.99068 (b) opF0.99075 (c) and ,=0.99079 (d). White stays for 77, and gray for 77,.
Figure 5 reports the reconstruction obtained by the proposed method. Specifically, figure 5a
reports the (normalized) error functional J(e) whereas figures 5b-5c report the reconstructions
corresponding to .<%,pt<+, respectively, being ^,tthe minimizer of/fc) and , e+ the closest
value to opt giving different reconstructions.
G. Rubinacci et al. /Shape Identification in Conductive Materials 19
(a) (b)
(c)
Figure 6. The (normalized) J(e) error functional (a), together with the reconstructions corresponding to
=0.9715 (b) $,pt=0.9716 (c) and &=0.9717 (d). White stays for rfh and gray for 77,.
Figure 6 reports the results related to a case obtained by only changing 77.. Specifically,
77, =7.5Qm is of the same order of magnitude of r)b. In this case the shape of the region of
resistivity 77, has been identified worse than for the previous case since the contrast, i.e. the
ratio 77, lrjh is close to the unity.
5. Conclusions
In this paper we have addressed the inverse problem of resistivity retrieval in Electrical
Resistance Tomography. Specifically, we have shown some properties of the resistance and
conductance matrices and proposed a novel method for solving the inverse problem of
retrieving a resistivity distribution characterized by means of two different values. The key
feature of the proposed method is its low computational cost that increases linearly with the
number of regions used for partitioning the conductive domain under investigation.
Acknowledgements
This work was supported by the Italian MURST and by the Italian Space Agency.
20 G. Rubinacci et al. /Shape Identification in Conductive Materials
Appendix
In this appendix we show that if A and B are two real, symmetric and invertible matrices,
then
A>B A ' < B ' (Ai)
In the following we use this straightforward property:
A > B o C r AC>C r BC (A2)
if C is an invertible matrix.
In fact, thanks to the assumptions made, the matrix B can be factorized as:
B=UAU r
where U is an orthogonal matrix and A is a diagonal matrix with non zero entries on the
main diagonal.
Since A"1/2 VT A U A"1/2 is also a real, symmetric and invertible matrix the canonical problem
becomes to show that
M>I M ' < I (A4)
where M is a real, symmetric and invertible matrix.
In fact, if (A4) holds, then from (A3) it results that
A-1/2UrAUA-'/2>I
A1/2UrA-'UA1/2<I
References
[1] Williams R A and Beck M S (Eds.) 1995, Process Tomography-Principles, Techniques and
Applications, Oxford: Butterworth-Heinemann, 1995
[2] A. Tamburrino, S. Ventre, G. Rubinacci, Electrical resistance tomography: complementarity and
quadratic models Inverse Problems, vol. 16, 1585-1618, 2000.
[3] A. Bossavit, Computational electromagnetism, Variational formulations, Edge elements, Complementarity,
Academic Press, Boston, 1998.
Electromagnetic Nondestructive Evaluation (VI) 21
F. Kojima et at. (Eds.)
IOS Press. 2002
Abstract. A numerical method for finding inclusions based on a formula in the en-
closure method is discussed. We propose a numerical implementation of the formula
by estimating the support function by the slope of the logarithm of the indicator func-
tion, and propose a rule to reject non-regular directions to avoid poor estimation. The
effectiveness of our method is shown by numerical examples.
1 Introduction
where 70. 71 are positive constants with 70 7^ 7i, and \o is the characteristic function of D.
Let u e Hl(tt\D) be a non-constant solution of the equation
Our problem is to extract some information about the location and shape of D from Cauchy
du
data (/, g) = I a rtn, where v denotes the outward unit normal vector to <91 This
problem is closely related to the electrical impedance tomography, and is widely discussed
in mathematical, numerical and practical points of view (e. g. [1, 7, 8] and references cited
therein).
The first author proposed a mathematical formula for finding the convex hull of D for the
cases D consists of polygonal-shaped cavities [2], that is, 71 = 0 in D, and polygonal-shaped
inclusions [3], that is, 71 > 0 in D. In [5], we implemented the formula numerically, and
tested for the case that D is a cavity. However, it is an extreme case, and it is interesting to
consider whether the method works well for more moderate case, that is, D is an inclusion.
In this paper, we proceed the numerical testings of our method for the case that D is an
inclusion. Our main purpose and interest are to test the ability of the numerical method in [5]
for cases that the jump of 7 is small, and there are some inclusions that have different 7 each
other.
22 M. Ikehata and T. Ohe / Finding {he Convex Hull of Inclusions
The contents of this paper are as follows. In section 2, we summarizes the enclosure
method to find the convex hull of inclusions discussed in [3], and also summarize in section 3
a numerical implementation of the enclosure method proposed in [5]. Section 4 gives numer-
ical experiments for some cases to examine the ability of our numerical method for finding
the inclusions. In section 5, we describe some concluding remarks.
In this section, we summarize the enclosure method [3] proposed by the first named author.
LetS"1 = {u> = (u^i, 0*2), |w| = 1} and define
The function ho is called the support function of D. If we know the value of ho on S1, then
the convex hull of D can be obtained by
Therefore, we call the method to determine the support function for finding the convex hull
of D as the enclosure method.
We need three more definitions for the formulation of the reconstruction formula.
Definition (regular direction): The direction u) G S1 is regular with respect to D if the set
{x R2 1 x u> = ho(u)} fl 3D consists of only one point.
where i \/\.
Definition (Indicator function /u>(r. t)): Let (/. g) be Cauchy data on c*O of u that is a
non-constant solution of (I). Define
The reconstruction formula for the support function ///) is following [3]:
Theorem 1. Assume that D is a polygon satisfying diam D < dis(D. 6>Q), and that u is a
non-constant solution of (1). Lei u be regular with respect to D. Then the formula
is valid.
Note that we need no information on the parameter 71 for the reconstruction formula. The
same result has been obtained for the case where D is a cavity (->! = 0), and a numerical
method for this case is discussed in [5]. We also note that the formula (4) is valid for the case
where D ~ DI U >2 U U Dm, 1 < m < oc, each D} is a polygon, and D} n D*. = 0 if
j ^ k (see [4] for this arguments).
M. Ikehata and T. Ohe / Finding the Convex Hull of Inclusions 23
3 Numerical Implementation
Now, let us consider a numerical implementation of the reconstruction formula (4). The im-
plementation method is also the same as that of [5]. However, for readers' convenience, we
explain the details of our method again.
The main difficulty in the numerical implementation is how to estimate asymptotic be-
havior of log |/UJ(T, t}\ for large r. In practical cases, observation points for Cauchy data are
given only on discrete points because of some physical restrictions. Since v(x; r, a;) oscillates
violently for large r and has exponentially growing property, it is very difficult to estimate
I(JJ(T, t) with sufficient accuracy for large r, and is also difficult to estimate ho(uj) using the
direct implementation of the formula (4).
The following lemma is a key point in the proof the reconstruction formula (4), and also
in our numerical implementation:
Lemma 1. Under the same assumption as Theorem 1, there exist A A(CJ, D) > 0 and
A A((jj. D) > 0 such that
Using this estimation, we choose an interval [TI, r2] c R + in which /u;(r, 0) can be estimated
with sufficient accuracy, and estimate ho(uj) by the slope of log |/u;(r. 0)| in [TJ , r^].
The another difficulty is how to distinguish regular directions u;. The reconstruction formula
(4) is valid for regular directions with respect to D. However, it is impossible to give a priori
information on regular directions.
Let u;(#) = (cos 0, sin 9], and assume that u(9] is regular with respect to D in the interval
[9\.9-^\. Since D is a polygon, there exists only one point x () e 3D such that
Thus,
24 M. Ikehata and T. Ohe / Finding the Convex Hull of Inclusions
dhD(u>(9))
is obtained. Equation (6) shows that do not exceed p0. In some practical cases.
dB
we may assume that there exists a known domain D c Q that include the domain D. Then,
the following rejection rule for non-regular u> is applicable:
Rejection rule
dh D (uw(9))
Let R = max|x|, andhD(u(0)) be the estimated value of ho(u(9)). If
xeb d
0 = 00, then reject u(90).
It seems very rare case that one comes across a non-regular direction. However, in the nu-
merical computation, it may be considered that the estimation of hD(u>) becomes worse when
<jj is near to some non-regular directions because of some numerical errors. The effectiveness
of this rejection rule is shown in the next section.
4 Numerical Results
In this section, we show some numerical experiments for our method. We set the domain
0 = {(z. y) | x2 + y2 < 1} throughout this section.
First, we consider the case that the inclusion D consists of one polygon as shown in Figure
1 (a). For parameters 70 and 71, the following three cases are considered:
Case 1-A: 70 < 7i (lo = 1- 7i = 2.0)
Case 1-B: 7o > 7i (7o = 1, 7i = 0.5)
Case 1-C: 70 ~ 7i ho - 1. 7i = 1-001)
The case 1-C is given to test the behavior of our method for the case where the jump of -. is
small.
To give Cauchy data for our inverse problem, the following Dirichlet problem is consid-
ered:
M. Ikehata and T. Ohe / Finding the Convex Hull of Inclusions 25
We solve this direct problem by the finite element method, and give Cauchy data at 32 points
uniformly on <91 For the computation of /U?(T, 0), we apply the trapezoidal rules.
Figure 2 shows the distribution of numerical values of log |/U;(T, 0)| for u> = u>(9) -
(cos 9, sin 6) with 9 = 0, Tr/4, 2?r/4, , ?7r/4 in case 1-A. Distributions of | log/cj(r, 0)|
have corners at T ~ 8. Since the slope value is nearly 1 for r > 8, numerical errors are
dominant for r > 8, and a reliable interval of r for the computation of /U>(T, 0) is [2. 6].
Hence, we compute /u;(r, 0) on [2,6] with step 0.25, and estimate ho(^j) by the slope of
log |/U;(T, 0)| using the least squares linear approximation at u; = t*>(9k) = (cosOk, sin9^)
with $fc = fcyr/16, fc = 0,1,2, , 31. As a priori information for our rejection rule, we
assume that unknown inclusions lie in the disc DR = { ( x , y} \ x2 + y2 < R}, and set R 0.8.
dhD(w(0})
For the estimation of , the forward finite differences are applied.
Figure 3: Results of lines {(x, y) \ (x, y) u}(9^) = /ID(^(^ ))} (left hand side) and estimated convex hull of
inclusions (right hand side) for single inclusion. Broken lines marked by circle are rejected by our rejection rule.
M. Ikehata and T. Ohe / Finding the Convex Hull of Inclusions 27
Figure 4: Results of lines { ( x , y) (x, y) u>($fc) = hD(u(Ok))} (left hand side) and estimated convex hull of
inclusions (right hand side) for two inclusions. Broken lines marked by circle are rejected by our rejection rule.
28 M. Ikehata and T. Ohe / Finding the Convex Hull of Inclusions
uniformly on <9fJ, and the trapezoidal rules are applied for the computation of IU(T. 0). Be-
fore estimation of ho(tjj), we check the distribution of log |/U;(T. 0)|, and decide to compute
IU(T,Q) on [3.9] with step 0.25. The a priori parameter R in DR for our rejection rule is
chosen as 0.8.
Figure 4 shows lines {(x. y) \ (x, y) uj(0k) = h D ( u ) ( 6 k ) ) } and estimated convex hull of
inclusions. The results for case 2-A and case 2-B show that we can obtain good estimations
for the convex hull of two inclusions. However, in case 2-C, we can not find the smaller
domain D? because the effects of the domain D\ may be too large.
5 Conclusions
In this paper, we consider a numerical method for finding the convex hull of polygonal inclu-
sions based on the enclosure method. In the numerical implementation, we have two difficul-
ties. One is caused by characteristics of the function v(x: T. u>), and the other one is because
it is impossible to give a priori information on regular directions. To avoid these difficul-
ties, we propose a method for the estimation of the support function />D(U>) by the slope of
log|/o;(r, 0)|, and a rejection rule for non-regular directions. Our numerical method is exam-
ined by some numerical examples. The numerical results show that our method is effective
for finding the convex hull of polygonal inclusions.
Acknowledgement
This research was partially supported by Grant-in-Aid for Scientific Research (C) (2) (No.
13640152) of Japan Society for the Promotion of Science.
References
[1] M. Bruhl and M. Hanke, Numerical implementation of two non-iterative methods for locating inclusions
by impedance tomography. Inverse Problems 16 (2000) 1029-1042.
[2] M. Ikehata, Enclosing a polygonal cavity in a two-dimensional bounded domain from Cauchy data. Inverse
Problems 15(1999) 1231-1241.
[3] M. Ikehata, On reconstruction in the inverse conductivity problem with one measurement. Inverse Prob-
lems 16 (2000) 785-793.
[4] M. Ikehata, On reconstruction from a partial knowledge of the Neumann-to-Dirichlet operator. Inverse
Problems 17(2001)45-51.
[5] M. Ikehata and T. Ohe. A numerical method for finding the convex hull of polygonal cavities using the
enclosure method, in press in Inverse Problems.
[6] M. Ikehata and S. Siltanen, Numerical method for finding convex hull of an inclusion in conducm its from
boundary measurements. Inverse Problems I6(2(KK)) 1043-1052.
[7] S. Siltanen, J. Mueller and D. Issacson, An implementation of the reconstruction algorithm of A. Nachman
for the 2-D inverse conductivity problem. Inverse Problems 16 (2000) 681-699.
[8] Stavroulakis G E, Inverse and crack identification problems in engineering mechanics. Kluwer. Dordrecht.
2001.
Electromagnetic Nondestructive Evaluation (VI) 29
F. Kojima et a I. (Eds.)
IOS Press, 2002
1 Introduction
Maintenance of aged plants requires non-destructive evaluation (NDE) for predicting initia-
tion of cracks. The NDE based on measurement of passive leakage magnetic fields is one of
the most promising methods so far as ferromagnetic materials are concerned. This method
identifies the magnetization, due to martensite transformation around cracks, from the mag-
netic fields measured outside the material. Since this identification is a severely ill-posed
problem, the system equation must be regularized by an appropriate method. Hitherto this ill-
posed problem has been analyzed with the iterative methods such as an optimization method
with the wavelet technique [1] and neural network method [8].
In the meantime, the authors have shown that the direct method based on the Tikhonov
regularization and L-curve technique can provide successful identification of magnetization
distribution in the permanent magnets used in copy machines and motors [4, 5, 6]. More-
over, this direct method has been applied to the reconstruction of magnetic charge for NDE,
to show that the Tikhonov method with the appropriate regularization term gives accurate
identification of magnetic charges, providing that there is no noise in measurement [7].
In this paper we report the feasibility of the direct method for the noisy measured magnetic
fields. Moreover, an new iterative method based on the conjugate gradient (CG) method with
a variable transformation is also applied to the problem to test its performance.
2 Formulation
Figure 1 shows two planar regions, that is, the charge region fic which contains magnetic
charge, and the measurement region fim on which the magnetic fields generated by magnetic
charges on Q f . are measured.
30 H. Igarashi and T. Honma /A Direct Reconstruction of Magnetic Charge
where q denotes the surface magnetic charge density, which is defined by the divergence
of magnetization. The surface magnetic charges along the fringe of fic are neglected in (1)
assuming that q concentrates near the cracks. In this work, only the ^-components H of the
magnetic field are used for reconstruction because the use of other components is expected
to provide similar results [1].
The charge region S7C is discretized into small triangular elements A*, in which q is linearly
interpolated with the shape function Nt, that is, q = .
The Tikhonov regularization method corresponding to (1) is given by
where || ||, {} and [] represent the 2-norm, column matrix and matrix, respectively, A denotes
the regularization parameter, and [K] is the discrete counterpart of the integral operator in (1)
whose entities are given by
It has been shown that (2) leads to solutions polluted with numerical noise even if there is no
measurement noise [7]. This is due to the fact that the singular vectors of [K] have undesirable
ripples with short wavelength. Note here that the solution to (2) is expressed in the form [2]:
where {ut} and {v,} are the left and right singular vectors of [A'], respectively, a, the ?th
singular value, and fl the filter function defied by /z = a?/(a? + A 2 ).
H. Igarashi and T. Honma /A Direct Reconstruction of Magnetic Charge 31
can also be expanded in the form of (4). The solution to (5) is, therefore, polluted by numerical
noise.
Instead of (2), we consider
where [A/] = Jn {NY{N} dS. Note that the regularization term {<?}f[M]{<?} corresponds to
the discrete counterpart of Jn q2 dS. The problem (6) is equivalent to
where [L] denotes the upper triangular matrix obtained by the Cholesky factorization of [A/]
which satisfies []'[] = [M], ^ the filter function defined by ^ = 7 2 /(7 2 + A 2 ), 7, the
generalized singular value and al 7i/\/l + 72. The generalized singular vectors {xj has
been shown to be smooth in contrast to the singular vectors of [K] [7]. Hence the solution to
(7) is adequately smooth.
The above fact suggests that the iterative solutions to the modified normal equation
where {q} [L}{q], are also smooth because they would be expressed in the form of (8). In
this paper (9) is iteratively solved by the CG method with a controlled stopping criterion for
regularization. In the next section, the effects of noise in measurement on the solutions to (7)
and (9) will be discussed.
3 Numerical Results
The magnetic charge q is reconstructed assuming the original distribution shown in Fig. 2. The
white noise, whose level n is defined by n = ||na||/||{//}||, where na denotes the amplitude
of the noise, is superposed to {H}. The sizes of fic and Qm are set as a = 20 and b 30. The
domain fic is uniformly discretized into triangular elements with 400 unknowns at the nodal
points. The lift-off is 2.0, and the noise level n is set to 0.05 unless notified.
The solutions to (7) for different values of A are shown in Figs. 3-5. The solution seems
adequate when A ~ 0.1. Figure 6 shows the L-curves [3] for different noise levels. The corner
of the L-curve for n = 0.05 was given when A = 0.01 while the error becomes minimum
for A = 0.1, as shown in Fig. 7. Hence the L-curve method yields a bit too small estimate of
the optimal value of A. Moreover there would exist no clear corner of the L-curves when n is
over 0. 1 . Thus the L-curve method cannot be used for such highly noisy cases.
32 H. Igarashi and T. Honma / A Direct Reconstruction of Magnetic Charge
Figure 5: Solution to (7), A = 0.001 Figure 6: L-curves for different noise levels n
Figures 8-10 show the solutions to (9) obtained by the CG method, whose iteration pro-
cesses are stopped when the normalized residue of a linear system | d||/||{//}| is smaller
than the tolerance e. The solutions in Figs. 8 and 10 seem over- and underdamping, respec-
tively, as are in Figs. 3 and 5, whereas that in Fig.9 seems adequate. The CG solution to (5).
in which numerical noise is observed, is shown in Fig.l 1 for reference.
The dependence of the solution errors on e for different noise levels is shown in Fig. 12.
The discrepancy principle states that the appropriate value of e equals to the noise level [2].
For instance, the adequate value of e is 0.05 for n = 0.05 . The error, however, becomes
minimum when e = 0.01 when n = 0.05. The discrepancy principle, therefore, yields a
rough estimate of the optimal value of e.
We conclude that both Tikhonov method and iterative regularization method based on the
CG method in conjunction with the L-curve method and the discrepancy principle, respec-
tively, can be promising methods for the NDE.
H. Igarashi and T. Honma/A Direct Reconstruction of Magnetic Charge 33
Figure 10: Solution to (9), e = 0.001 Figure 11: Solution to (5), e = 0.01
34 H. Igarashi and T. Honma / A Direct Reconstruction of Magnetic Charge
4 Conclusion
In this paper, the Tikhonov method as well as the new regularization method based on the
CG method in which the modified normal equation is solved have been applied to the NDE
problem with noisy measurement data. The L-curve method and discrepancy principle yield
roughly optimal estimate of the regularization parameters. Both regularization methods are
thought to be promising for real NDE.
References
[1] Z. Chen, K.Aoto and S. Kato, Reconstruction of magnetic charges using an optimization method and
wavelet J. JSAEM8(3) (2000) 363-371.
[2] P.C.Hansen, Rank-Deficient and Discrete Ill-posed Problems, SI AM. 1998.
[3] P.C.Hansen and D.P.O'leary, The use of the L-Curve in the Regularization of discretization of discrete
ill-posed problems, SIAM J. Sci. Comput., 14 (1993) 1487-1503.
[4] H.Igarashi, Forward and inverse analyses of cylindrical permanent magnets. J. IEE Japan, vol. 120-A
(2000) 791-797 (in Japanese).
[5] H.Igarashi, T.Honma and A.Kost, Inverse inference of magnetization distribution in cylindrical permanent
magnets, IEEE Trans. Magn., 36 (2000) 1168-1171.
[6] H. Igarashi, On the reconstruction of magnetic source in cylindrical permanent magnets. Inverse Problems
in Engineering Mechanics II (eds. M.TanakaandG.S. Dulikravich), Elsevier, Amsterdam (2000)467-476.
[7] H. Igarashi and T.Honma, Reconstruction of Planar Magnetic Charge Distribution, Proc. ISEM-Tokyo.
621-622, extended version will appear in Int. J. Applied Electromagnetics and Mechanics.
[8] G. Preda, S. Takaya, K. Demachi and K. Miya, Reconstruction of magnetic moments distribution from 2D
scan data using neural networks Proc. Symposium on Electromagnetics and Dynamics (1999) 618-621.
Electromagnetic Nondestructive Evaluation (VI) 35
F. Kojima et al. (Eds.)
IOS Press, 2002
1
MASAHIRO YAMAMOTO AND2 JUN Zou
1
Department of Mathematical Sciences
The University of Tokyo
3-8-1 Komaba Meguro Tokyo 153 Japan
tel: +81-3-5465-8328
fax: +81-3-5465-7017
e-mail : [email protected]
2
Department of Mathematics
The Chinese University of Hong Kong
Shatin, Hong Kong, China
fax: +852-2603-5154
e-mail: [email protected]
1. Introduction.
Let 17 C M3 be a bounded domain whose boundary <90 is of class Cl .
In 1, we consider the quasistatic Maxwell equations:
Key words and phrases, quasistatic Maxwell equations, current source, inverse
problem, uniqueness, stability.
36 M. Yumamoto and J. Zou / Determining a Current Source
More precisely, for this inverse problem, we will discuss the theoretical
subjects:
M. Yamamoto and J. Zou / Determining a Current Source 37
Then there exists a constant C > 0, depending on fJ, T, u;, A, 0, //, such
that
Here we set
and
and
we will explain a role of the uniqueness result for the Tikhonov regu-
larization by the following simple example: For the heat equation:
and
Next let us reconstruct the initial value X]fc=1 sin2/CTrxi from ob-
servation data with errors. For this, the Tikhonov regularization is
widely used, which is formulated as follows: let u(a) = u(a)(xi,t) be
the solution to (2.1) and (2.2) with u(xi,0) = a(x\), 0 < x\ < 1. We
take
and
with given y(x, t) for (x, t) e a; x (0, T) and y(x, 0) for x G 0. Therefore
the argument in Imanuvilov and Yamamoto [4] can complete the proof
of our theorem.
REFERENCES
[1]. H.T. Banks and K. Kunisch, Estimation Techniques for Distributed Parameter
Systems, Birkhauser Verlag, Boston, 1989.
[2]. J. Cheng and M. Yamamoto, One new strategy for a priori choice of regu-
larizing parameters in Tikhonov's regularization, Inverse Problems 16 (2000),
L31-L38.
[3]. C.W. Groetsch, Inverse Problems in the Mathematical Sciences, Vieweg, Braun-
schweig, 1993.
[4]. O. Yu. Imanuvilov and M. Yamamoto, Lipschitz stability in inverse parabolic
problems by the Carleman estimate, Inverse Problems 14 (1998), 1229 1245.
[5]. V. Isakov, Inverse Problems for Partial Differential Equations, Springer- Verlag,
Berlin, 1998.
[6]. J.A. Kong, Electromagnetic Wave Theory, John- Wiley & Sons, New York,
1990.
[7]. P. Neittaanmaki, M. Rudnicki and A. Savini, Inverse Problems and Optimal
Design in Electricity and Magnetism, Clarendon Press, Oxford, 1996.
42 M. Yamamoto and J. Zou /Determining a Current Source
[8]. V.G. Romanov and S.I. Kabanikhin, Inverse Problems for Maxwell's Equa-
tions, VSP, Utrecht, 1994.
[9]. M. Yamamoto, A mathematical aspect of inverse problems for non-stationary
Maxwell's equations, Int. J. of Applied Electromagnetics and Mechanics 8
(1997), 77-98.
[10]. M. Yamamoto, On an inverse problem of determining source terms in Maxwell's
equations with a single measurement, in Inverse Problems, Tomography, and
Image Processing (1998), Plenum Press, New York, 241-256.
[11]. M. Yamamoto and J. Zou, Simultaneous reconstruction of the initial temper-
ature and heat radiative coefficient, to appear in Inverse Problems 17 (2001).
Direct Problems
This page intentionally left blank
Electromagnetic Nondestructive Evaluation (VI) 45
F. Kojima et al. (Eds.)
IOS Press, 2002
Abstract. The use of standard nodal finite elements in the numerical solution of
Maxwell's equations is marred by the occurrence of spurious modes causing severe
stability problems, if no proper gauging is performed. Moreover, in case of cor-
ner or edge singularities there might be solutions that due to a lack of regularity
cannot be approximated by nodal finite elements at all. On the other hand, it is
well-known that curl-conforming edge elements avoid such difficulties, since they
are much closer to the variations! formulation of boundary and initial-boundary
value problems in electromagnetics.
In this contribution, we are concerned with efficient numerical solution techniques
in terms of adaptive multigrid methods based on edge element discretizations with
respect to adaptively generated hierarchies of triangulations of the computational
domain.
In particular, we deal with multigrid algorithms whose basic ingredients are hybrid
or distributive smoothing processes that take care of the nontrivial kernel of the
discrete curl-operator. The characteristic feature is an additional defect correction
on the subspace of irrotational vector fields. Adaptive grid refinement can be per-
formed by means of efficient and reliable a posteriori error estimators. We present
residual estimators based on an appropriate Helmholtz decomposition of the error
into an irrotational and weakly solenoidal part.
The performance of the multigrid algorithm and the a posteriori error estimator is
illustrated by some numerical results.
1 Introduction
Physical processes based on electromagnetic phenomena can be adequately
described by Maxwell's equations which in differential form are given as fol-
lows
where e and p, are the electric and magnetic permeabilities, a is the electric
conductivity, and Jj refers to an intrinsic current density.
For slowly varying fields the time derivative of the dielectric induction is
usually neglected resulting in the so-called eddy current equations that are
of parabolic type
Fl T
of irrotational vector fields and set Ho(curl; /?) := H0(curl; l?)nH(curl; J?).
We shall be concerned with the numerical solution of edge element discretized
variational problems of the form:
Find j HO (curl; I?) such that
where
where ts is the tangential unit vector with respect to the edge E 6 /i(T)
(cf., e.g., [5]).
Specifying the basis fields accordingly, we are thus led to the H(curl)-conforming
edge element spaces
48 R.H. W. Hoppe /Adaptive Multigrid Techniques in Electromagnetic Field Computation
into the subspace Nd5(0(/2;7fc) := (qh Ndi >0 (tf;7fc) | curl qn = 0} and
its I/2 -orthogonal complement Ndf 0 (j?;7fc), a standard smoother like the
Gauss-Seidel iteration, applied to the edge element discretized problem, per-
forms well on the subspace Nd^"0(/2; Tk) of weakly solenoidal discrete vector-
fields, but insufficiently on the subspace Nd 0(17; Tk) of irrotational discrete
vectorfields. Therefore, the idea is to apply a hybrid smoothing process that is
characterized by an additional defect correction step on the subspace of irro-
tational vectorfields. Taking into account Nd 0(.(?;7fc) = grad 5i,o(/?;7jfc),
where Si,o(ft;7fc) is the standard FE-space of continuous, piecewise linear
finite elements, on level t the hybrid smoother works as follows:
Step 1; Gauss-Seidel sweeps (edge element discretized problem)
Given an iterate j^ , perform K > 0 Gauss-Seidel iterations on the edge
element discretized problem (9) resulting in the smoothed iterate j^ .
Step 2; Correctional Gauss-Seidel sweeps (irrotational part)
Compute the residual
2 3 4 5 6
(3 = 0.100 0.15 0.98 0.16 0.98 0.16 0.99 0.16 0.99 0.16 0.99
(3 = 0.500 0.15 0.97 0.16 0.97 0.16 0.98 0.16 0.98 0.16 0.99
0 = 1.000 0.15 0.96 0.16 0.97 0.16 0.97 0.16 0.97 0.16 0.98
0.15 0.96 0.16 0.96 0.16 0.96 0.16 0.97 0.16 0.98
0 = 10.00 0.14 0.88 0.16 0.93 0.16 0.96 0.16 0.97 0.16 0.97
0.10 0.75 0.13 0.91 0.15 0.92 0.16 0.95 0.16 0.96
Table 1: Multigrid convergence rates (effect of hybrid smoothing)
A posteriori error estimation is a prerequisite for grid adaptation techniques.
Here, we follow [2] and sketch a residual type a posteriori error estimator for
edge element approximations of boundary value problems of the form
curl a curl j + /?j = f in Q C R3 (10)
j A n = g on dQ (11)
discretized by means of the lowest order H(curl)-conforming edge elements
of the first family.
The residual based error estimator, which will be given in terms of the energy
norm
|||j||| 2 := (acurlj,curlj) 0 ,f? + (cj,j) 0 ,fl ,
relies on an appropriate evaluation of the residual with respect to an approx-
imation jh Ndi(]7; Th) of (10), (11). Assuming for simplicity jh A n = g on
dQ, it is easy to see that the total error eh := j - jh is in HO (curl; /?) and
satisfies the defect equation
vectorfields.
In the sequel we assume that a decomposition (12) can be found such that
HO" (curl; i?) is continuously embedded in H1(/?) D HO (curl; /?) and
with the local contributions <J^ , 1 < i/ < 2, and <5p given by
Level || 0 1 2 3 4
4
p = io- 4.21 4.70 4.90 4.98 5.01
/3 = 10-' 4.21 4.70 4.90 4.98 5.01
ft = 1.00 4.21 4.70 4.89 4.98 5.01
0 = 10+2 4.22 4.68 4.87 4.96 5.01
(3 = 10+4 4.22 4.66 4.84 4.92 4.96
Acknowledgments. The work has been supported by grants from the Fed-
eral Ministry for Education and Research (BMBF) under Grant No. 03H07AU1-
8 and Grant No. 03HOM3A1.
References
1. R. Beck, P. Deuflhard, R. Hiptmair, R.H.W. Hoppe, and B. Wohlmuth, "Adap-
tive multilevel methods for edge element discretizations of Maxwell's equa-
tions", Surveys of Math, in Industry, Vol. 8, pp. 271-312, (1999)
2. R. Beck, R. Hiptmair, R.H.W. Hoppe, and B. Wohlmuth, "Residual based a
posteriori error estimators for eddy current computation", M 2 AN Math. Mod-
elling and Numer. Anal. 34, 159-182, (2000)
3. P. Grisvard, "Elliptic Problems in Nonsmooth Domains", Pitman, Boston, 1985
4. R. Hiptmair, "Multigrid method for Maxwell's equations", SIAM J. Numer.
Anal. 36, 204-225, (1999)
5. J.-C. Nedelec, "Mixed finite elements in R3", Numer. Math. 35, 315-341, (1980)
52 Electromagnetic Nondestructive Evaluation (VI)
F. Kojima el al. (Eds.)
IOS Press. 2002
1 Introduction
to use the prior information that the sought defects are of prescribed conductivity contrast
with respect to their environment, their 3-D maps ideally consisting of distributions of
black and white voxels (in the exterior and in the interior of the defects) as is the case
already in [5] and as it has been much studied in 2-D geometries since the early approach
[6] of binary-specialized modified gradient methods.
Previous investigations [1,2] have shown that anomalous magnetic fields are modeled
with fair accuracy -in manageable computational time on a standard work-station- by the
aforementioned integral formulation, and we do not come back to this topic here. We mostly
outline below how this hybrid inversion method can be properly developed, and give a few
typical results. Test cases at this still preliminary stage come from synthetic data calculated
on canonical 3-D models of mm-sized parallelepipeded void defects in an Inconel 600 plate
probed in the 150 - 300 kHz frequency range.
2 The algorithm
Let us refer to the configuration sketched in Fig. 1. A possibly multiply-connected 3-D defect
il of constant conductivity a (it will be 0 in the numerical examples) is embedded within
a non-magnetic linear isotropic metal plate of conductivity cr2. Eddy currents are generated
in the plate using a time-harmonic (circular frequency o>, time-dependence exp ( ]ujt}) coil
source carrying electrical currents Jc at fixed location in air (permeability /^0, permittivity
f ( ) ). Perturbation of the eddy current pattern produces an anomalous magnetic field that is
observed in a prescribed planar surface S in air.
Figure 1: Configuration of study: a pancake-type source probe is set at fixed location above the damaged metal
plate, here sketched as a void parallelepiped, the time-harmonic magnetic field being sampled in a plane of
observation at fixed height above the plate.
P = C-J2 (3)
tells us that some contrast source J 2 in T> radiates a certain magnetic field in S whose dis-
crepancy with data is measured by vector residual p"; the state equation
e = 0-2 \ E2 - J 2 (4)
tells us that this contrast source is consistent with a contrast distribution \ (here, purely
real) in Z>, the degree of consistency being appraised by vector residual e, letting E2 =
E2()+,J2.
As for the sought binary aspect of the contrast distribution (valued to \ c or to 0 at any
r in P), and in order to use gradient methods that require us to restore differentiability with
respect to this distribution, it is enforced via the nonlinear transformation
where 0 is a strictly positive, real-valued tuning parameter which is controlling the slope of
the strictly monotonous real function ^ (bounded between 0 and 1 and varying from 0 to
1 when T is increased), the r-dependence being implied. In practice, decreasing 9 yields an
electrical contrast \ closer to \c (resp., to zero) for a given positive (resp., negative) T at same
D. Dos Reis et al. / Eddy-Current Evaluation of 3-D Defects in a Metal Plate 55
The solution itself now relies on the construction of two sequences of rn and J 2 (corre-
spondingly, of E 2 ) that are both function of space r (implied) in T> -once chosen proper first
guesses at n 1- so as a suitable cost functional F is decreased from one iteration to the
next. Using |. 5 and |.| p as norms on L2 (S) et L2 (D) and (later on) corresponding scalar
products {, }s and {, }X), the general expression of F reads as
where weight r/5 H C l l ^ 2 is constant for given data and where weight rj-p = ||cr2 xE 20 ||p
is contrast-dependent.
The determination of the contrast source J 2 is carried out first. At iteration n, one simply
chooses J 2 = J!,'"1 + a n v n , where the complex-valued, r-dependent (3-component) vector
v" is a conjugate-gradient direction of displacement of the Polak-Ribiere type and where the
complex-valued constant parameter a" is a coefficient of displacement. One sets for n > 1
where * denotes the adjoint operation, weight rf^~l and residuals pn~l and en~l being those
associated with the iterates of order n 1 (since one is interested into the gradient of F
at J 2 '~ J ). Accordingly, minimizer an of cost functional F (J 2 -1 + an v n ) for the just deter-
mined set of directions v n is written in closed form as
According to the above, contrast source and field are now updated to new values J 2 and
E2l = E!,7"1 + a "^ 2 2 v "- However, the residual of the updated state equation, written as
e = (7-2 \ E2 J 2 , valued at 'e" when the contrast is the present iterate xn~l Xc ^ (rH~1)'
might not be close enough to zero in order to be satisfied with. So, proper updating of the
contrast comes by letting T" = rn~l+j3ndn, where real-valued r-dependent scalar dn is again
of the Polak-Ribiere type and where real-valued parameter (3n is a displacement coefficient.
For //. > 1
(at /). = 1, d l = yd'1). gd'n is the real-valued gradient (Frechet derivative) at r" l, both
contrast source J 2 and field E2 being now kept fixed, of a weighted-in squared norm Fp of
the above state residual e; in conveniently expanded form, Fp reads as
56 D. Dos Rets et al. /Eddy-Current Evaluation of 3-D Defects in a Metal Plate
where ^ is the derivative of ^ with respect to r, where the upper bar denotes complex
conjugation, where 3? denotes the real part, and where residual e71 at r""1 has been defined
before; here, ( ) represents the usual scalar product of (3-component) vectors at given lo-
cation. (Notice that the weight rj^'1 has been kept constant for simplicity, which means
that one uses it simply as a normalization factor, and that one is differentiating the factor
ika Xc * (T) E^ - J^lp only.) As for /?", the minimizer of Fp (r""1 + 3" d n ) into the set of
directions rf", it is obtained in closed form as
The algorithm now proceeds by alternating an update of contrast source J 2 via equations
(6-9) and an update of contrast r via equations (10-13) inside each voxel of the prescribed
search domain T> until the value reached by the cost functional F is small enough, or until
some plateau is reached. In the latter case, the tuning parameter 0 might be reduced in order
to push contrast values x either to 0 or to \>, and equivalently to re-distribute voxels into
darker ones (at positive r) and lighter ones (at negative r), from which the construction of the
two sequences is then started again.
3 Numerical results
A number of results has been obtained by the above inversion algorithm. Those shown in
figures 2 and 3 consist of gray-level 2-D cross-sectional maps at successive depths (0.1 mm
step) of the retrieved 3-D distributions of the contrast \ (after about 2500 iterations, when the
cost functional has notably decreased) and of evolution curves of the cost functional F as a
function of the number of iterations.
The example is the one of a 2 mm thick slab of conductivity a-2 = 1 MS/m where volumi-
nous void defects are sought. A thick circular coil (0.6 and 1.6 mm inner and outer radii, 0.8
mm height) is centered 0.9 mm above the slab at x = 0.75 mm and y = 0 to avoid undue
symmetries with respect to the search domain > (the latter is centered on the z axis and dis-
cretized into 52 x 52 x 20 cubical voxels of 0.1 mm 3 volume). The frequency of operation is
150 kHz, which is corresponding with a plane-wave skin depth 6 in metal of about 1.3 mm.
One considers in figure 2 and 3 two void parallelepipeds opening in air (the top surface of the
slab), both of 1.1 x 1.1 mm 2 squared horizontal cross-section, the first one 1 mm deep and
centered at .r = 1.1 mm, y = 1.1 mm, the other 0.5 mm deep and symmetrically located
at.;: = 1.1 mm, y = 1.1 mm; one considers a single void opening in air in figure 4, which is
still of 1.1 x 1.1 mm 2 squared horizontal cross-section and 1 mm depth, but its axis is now
inclined by 22.5.
Data consist of the three components of the anomalous magnetic (Figs. 2-4) or of its
single vertical (z) component (Fig. 4) collected at nodes 0.1 mm apart in a planar surface S of
overall area 10 mm 2 and centered 1.55 mm above T>. The inversion is started from a two-level
contrast map valued at x = 0.5 \c for the three shallower layers of voxels of T> and at 0.01 \ r
for the deeper ones, or from a contrast map that evolves with depth as \ = 0.5 \ r exp ( ; M )
D. Dos Reix et al. / Eddy-Current Evaluation of 3-D Defects in a Metal Plate 57
(Fig.4). No refreshment (decrease of 0) has been enforced at this stage of the study, a proper
0 having been equated to 1 by numerical experimentation.
Figure 2: Retrieval of two voids of different depths in a metal slab (see text for details). Cross-sectional gray-
level maps of the conductivity contrast \ at successive 0.1 mm-stepped depths using all three components of the
anomalous magnetic field (a), and the vertical component only (c). Exact maps are given in (b) for comparison.
From those results retrievals appear almost identical (and the cost functionals evolve sim-
ilarly) when all three components of the field are used and when the vertical one is the only
one used. This is rather comforting in terms of application to real measurement configura-
tions, e.g., by displacing a small coil probe throughout a plane of observation parallel with
the probed metal plate. However, only the shallower parts of the voids, down to about 2/3
of a skin depth, are fairly well retrieved, the volume of those shallower parts being increased
with respect to the exact one and those of the deeper parts being decreased. Such shadowing
and blurring due to skin effect are expected phenomena that are faced independently of the
inversion algorithm used.
Nevertheless, the results displayed here show that the proposed hybrid technique is rather
effective. This is illustrated by the retrieval of two close voids (Fig. 2-3) that appear both
fairly located and well discriminated, the inclination of a single one (Fig.4) being well re-
produced. Obviously one needs an appropriate initial guess to do so. An initial contrast that
exponentially decreases as a function of the depth-to-skin-depth ratio from the usual mid
value of \(./2, does not lead to a better retrieval than an initial contrast that is set to Xc/2,
in the three first layers of voxels and to almost 0 below (the cost functional reaches a value
one-third larger in the first case than in the second case).
58 D. Dos Reix et al. /Eddy-Current Evaluation of 3-D Defects in a Metal Plate
Figure 3: Evolution of the cost functional F as a function of the number of iterations for the two voids as studied
in Fig. 2. Dotted line, case (c); solid line, case (a).
Figure 4: Retrieval of an inclined void in a metal slab (see text for details). Cross-sectional gray-level maps of
the conductivity contrast \ at successive 0.1 mm-stepped depths (using all three components of the anomalous
magnetic field). The initial contrast \ is chosen as the two-level one in (a), and is chosen as the one exponentially
decreasing with depth in (c). Exact maps are given in (b) for comparison.
D. Dos Reis et al. / Eddy-Current Evaluation of 3-D Defects in a Metal Plate 59
4 Concluding remarks
So far, data have been assumed to be collected throughout a planar surface (not only along a
probing line within a user-prescribed symmetry plane of a defect as is often assumed); this
appears to be a pre-requisite when mapping geometrically unknown 3-D objects. Two config-
urations, other than this multistatic configuration (fixed source - moving field probe), could
be attacked similarly, e.g., when source and probe devices are moved together (a reference
case would be a SQUID-based magnetometer [7]), or when the variation of the impedance of
the probe itself is observed [8]. Work is presently in progress for such two configurations.
However the difficulty in dealing with such new configurations does not lie in the func-
tional analysis, which is carried out as the one sketched here, but in the numerical burden,
(one has to construct induced sources for each location of the source device) and in the lack of
information (in the second situation, one has to extract pertinent information from a quantity
expected to vary more smoothly than the anomalous field).
In addition it remains to study in further detail the influence of the initial guess and of
refreshment, and the advantages of frequency-hopping (using data at successive frequencies,
from a low one to a high one), whereas the consequence of data noise and model errors
should be appraised closely, the synthetic data here having been modeled using the same
discretization as the one used in the inversion itself.
References
1 1 ] D. Dos Reis et al., Extended Born domain integral models of diffusive fields, Proc. COMPUMAG'Ol
Evian, 2001, IV-74-75.
[2] D. Dos Reis et al., On the modeling of 3-D inclusions in conductive media using extended Born models
in the diffusive regime, Int. J. Appl. Electromagn. Mechan. (2002) in press.
[3] A. Abubakar and P. M. van den Berg, Three-dimensional inverse scattering applied to cross-well induction
sensors, IEEE Trans. Geosci. Remote Sensing 38 (2000) 1669-81.
[4] A. Abubakar et al., A conjugate gradient contrast source technique for 3D profile inversion, IEICE Trans.
Electron. E83-C (2000) 1864-1874.
[5] V. Monebhurrun et al., 3-D inversion of eddy current data for Nondestructive Evaluation of steam gener-
ator tubes, Inverse Problems 14 (1998) 707-24.
16] L. Souriau et al., Modified gradient approach to inverse scattering for binary objects in stratified media,
Inverse Problems 12 (1996) 463-81.
|7] A. Ruosi et al., High Tc SQUIDS and eddy-current NDE: a comprehensive investigation from real data to
modelling, Meas. Sci. Technol. 12 (2000) 1639-48.
(8] T. Takagi et al., ECT research activities in JSAEM Benchmark models of eddy current testing for steam
generator tube, Parts 1 and 2. In: R. Collins et al. (ed.), Nondestructive Testing of Materials. IOS Press,
Amsterdam, 1995, pp. 253-64 and 313-20.
60 Electromagnetic Nondestructive Evaluation (VI)
F. Kojima et al. (Eds.)
IOS Press, 2002
1. Introduction
obtain much more information from the object. Ludwig et al [3] and Thompson [4] studied the
basic principle of the Lorentz type of EMATs, and Hagi et al [5] numerically analysed
transmission of ultrasonic waves by such an EMAT on the basis of FEM.
In this paper, we analyse inspection by an EMAT of the Lorentz type, including reception
process of reflected waves from the defect and the boundary of the object. In the next section
we derive the basic equations describing the transmission process by the EMAT. In section 3,
we lay down the basic equations for wave propagation exited by the Lorenz force and the
reception process by the EMAT. In section 4, we simulate ultrasonic inspection and attempt to
identify an inside crack by scanning the EMAT probe. We employ FEM for analysis of the
electromagnetic fields and the finite difference method for analysis of wave propagation. In
the last section, we summarise the obtained results.
Figure 1 shows the sketch of the cross section of a simple EMAT of the Lorentz type. A
coil of the track shape is located between the specimen and a couple of permanent magnets
with opposite poles. When a high frequency current is supplied to the coil, the induced
dynamic magnetic field generates the eddy current near the surface of the conductive
specimen. The free charges in the eddy current are subjected to the Lorentz force under the
static bias magnetic field by the permanent magnets. The free charges collide to the neutral
moleculars and ions, which generates the local motion of the objective material. The ions and
the free charges in the conductive material are also subjected to the Coulomb force in the
parallel and the opposite directions of the electric field corresponding to the eddy current. It is
known that the local motion due to the Coulomb force to the ions and the free charges are
cancelled in the conductive material. In a result, only the local motion in the direction of the
Lorentz force takes place. It is also known that when the frequency of the driving current is
less than 100 MHz, the wavelength of the exited motion is much lager than the mean free path
of the charged particles in the room temperature, cf. [4]. This fact means that the above
external force by the Lorentz force can be regarded as a continuous field in the space and the
time regions, i.e., it can be identified with the Lorentz force determined by the vector fields of
the current and the magnetic flux density.
where H is the magnetic field, J the current density, E the electric field, and B the magnetic
flux density. Note that the electric displacement and the induced current by it can be neglected
62 T. Mitsuda and E. Matsumoto / Numerical Analysis of Ultrasonic Inspection
in the frequency range of the EMAT. The magnetic and the electric constitutive equations of
an electromagnetic material are given by
where fj, is the magnetic permeability and a the electric conductivity of the material.
Introducing the vector and the scalar potentials A and 0, we can write
If we divide the current density into the forced current density /0 and the eddy current density
Je such that J=jQ+Je, we can derive the differential equations for the potentials around the
EMAT
In the following discussions, we assume that the electromagnetic fields are two-
dimensional, and the boundary conditions are the forced current density in the coil, the
magnetic flux density in the permanent magnets, and the electric conductivities in the coil
and the specimen. We also assume that the material is non-magnetic. For efficiency and
accuracy, we divide the electromagnetic fields into the static fields due to the permanent
magnets and the small dynamic fields due to the driving coil.
We shall show the numerical result for the static and the dynamic electromagnetic fields.
When* andy are the two-dimensional coordinates as shown in Fig.l, the vector potential
and the current take the forms
T. Mitsuda and E. Matsumoto / Numerical Analysis of Ultrasonic Inspection 63
For analysis of the static magnetic field by the permanent magnets, we employ FEM with
first order rectangular elements. The length of the minimum edge of a rectangular element
is 0.5mm. Each permanent magnet has a square cross section with 5mm edges, and two
magnets are located at 15mm distance. The flux density in the permanent magnets is LOT.
We analyze the magnetic field in the 60mmx50mm quadrangle region around the magnets.
The magnetic field is assigned to be normal to the boundary of the analysis region. Since
there is no current and the scalar potential vanishes in this case, from (4) the component A
of the vector potential is found to be governed by the Laplace's equation. The calculated
equipotential lines of A is shown in Fig. 2.
We next consider the dynamic magnetic field generated by the coil of the EMAT. In this
case, the component A and the scalar potential ^ are governed by the system of partial
differential equations (4) and (5). We use similar finite elements to the static analysis in the
region containing the specimen of aluminum with quadrangle cross section SOmmxlOmm.
The driving coil is made of copper wire of the truck shape with quadrangle cross section
Immxlmm and located over the center of the specimen surface with 1.0mm lift-off. The
distance between the centers of two wires is 7mm. The material parameters used in the
calculation are as follows. The permeabilities of the environment (the air), aluminum and
copper are assumed to equal to the one of the vacuum 1.26xlO~6H/m, the conductivities of
aluminum and copper are, respectively, 3.64xl07S/m and 5.99xl07S/m, and the
environment is assumed to be an insulator. We supply one cycle of sinusoidal current to the
coil such that the direction of the current in each wire is opposite to each other. The current
density has the amplitude 1.00xl07A/m2 and the frequency 100kHz. Figure 3 indicates the
distribution of the induced eddy current density in the specimen at the instance of its
maximum. We see that the eddy current is locally distributed near the surface of the
specimen faced to the EMAT coil.
If the magnetic flux density and the eddy current density are obtained, the Lorentz force per
unit volume is given by
F = JexB (7)
64 T. Mitsuda and E. Matsumoto / Numerical Analysis of Ultrasonic Inspection
Since we are concerned with the two-dimensional problem discussed in the previous
section, from (6) the magnetic flux density B and the Lorentz force F take the forms
B-\Bx,By,o), F=(Fx,Fy,Q)=(-JeBy,JeBx,o). (8)
Thus the Lorentz force lies on the x-y plane, and its distribution has symmetry with respect
to the y-axis from the location of the coil and the magnets, refer to Fig.l. That is, since Je
and By are antisymmetric and Bx is symmetric with respect to the y-axis, from (8) Fx and
Fy become, respectively, symmetric and antisymmetric. Figure 4 shows the simulated
components of the Lorentz force near the surface of the specimen. Here the forced current,
the specimen and the coil are the same as in the previous section. Each of the permanent
magnets has 3mmx3mm square cross sections and the flux density LOT, and they are
located just below of each wire of the coil with common center lines. It is found that each
components of the Lorentz force has two peaks corresponding to each pair of wire and
magnet and the maximum strength of the horizontal component Fx is larger than that of the
vertical one Fy. In other words, this type of EMAT mainly excites the transverse wave.
Since the eddy current and the Lorentz force are localized within the skin depth of the
dynamic magnetic field, an elastic wave is excited in the neighborhood of the boundary
faced to the EMAT.
30 60
where double indices in each term are summed from 1 to 3, u is the displacement, p the
material density, a the stress tensor, and A and jU the Lame's elastic constants. The strain
tensor e is defined by
The external body force in (9) is given by the Lorentz force (8). The mechanical fields are
expressed in terms of the material coordinate system, i.e., in the Lagrangian form, so that the
velocity and the acceleration of a material point are denoted by the partial derivatives with
respect to time.
The elastic wave excited near the boundary surface by the Lorentz force is reflected from
other boundary surfaces and the flaws such as cracks and inclusions. When a material point is
moving under the static magnetic field B0 by the permanent magnets, vxJB0 term appears in the
effective electric field. Thus the induced current takes the form
The current (12) plays a similar role to the forced current in the coil for the transmission
process. That is, for the reception process of the EMAT, instead of (4) and (5) the vector and
the scalar potentials are governed by
Then the waveform of the reflected wave is given by the voltage variation in the coil
generated by the electromagnetic induction due to the perturbed magnetic field.
Similarly to the case of the electromagnetic fields, we assume that the mechanical fields are
also two-dimensional. Then the displacement induced by the Lorentz force (8) always lie on
the x-y plane, so that the induced vxU0 term is in the direction of the z-axis. Thus also in the
reception process of the EMAT, the electromagnetic fields are two-dimensional such that the
magnetic field lie on the x-y plane and the electric fields are directed to the z-axis. It is easy to
see that similarly to the transmission process, the EMAT coil is more sensitive to the
horizontal motion of the specimen.
In this section, we first simulate the total process of the transmission and the reception of the
66 T. Mitsuda and E. Matsumoto / Numerical Anal\sis of Ultrasonic Inspection
elastic wave by the EMAT. As we have discussed so far, the transverse and the longitudinal
waves are exited near the boundary surface of the specimen. For the numerical calculation of
the elastic waves, we employ the finite difference method such that a difference formula of
higher order is applied to the time domain. Using the same EMAT as in the previous section,
we transmit the elastic wave into the specimen. In this case we set the amplitude of the
current density in the coil 1.00xl07A/m2 and the frequency 1.0MHz. The specimen is
aluminum with quadrangle cross section 300mmx50mm and a crack with 20mmx2mm
cross section exists at the centre of the specimen. Figure 5 shows the wave fields near the
crack at time 12us. The longitudinal and the transverse components of the wave field are given
by V u and V x u, respectively. We see that the longitudinal wave is reflected from the
crack corners and the transverse wave is reflected from the crack surface and diffracted
beyond the crack. This implies that a mode transformation from the transverse to the
longitudinal waves takes place at the corner of the crack. The received waveform by the coil of
the EMAT is given in Fig.6. The first peak indicates the incident wave and the second one at
16.5 us and the third one at 33.0 us indicate the reflected transverse waves from the crack
surface and the opposite boundary surface, respectively. As shown in Fig.4(a) and Fig.4(b),
the EMAT also excites the longitudinal wave by the component Fy of the Lorentz force.
Although the amplitude of the exited longitudinal wave is not negligibly small compared with
the one of the transverse wave, the signal of the longitudinal wave is not clear in Fig. 6. This
fact comes from that in the reception process the EMAT is not sensitive to the longitudinal
motion of the specimen similarly to the transmission process. This characteristic is convenient
to ultrasonic inspection by means of excluding the effects of other waves with different speeds
bringing errors to the estimation. From Fig.4(a) the source of the transverse motion does not
have a uniform distribution and has two peaks. We see however that the wave front
transmitted by the EMAT is almost straight near the center of the path from Huygens principle,
which means that the wave is nearly a plane wave. This property of the exited waves is also
convenient to the inspection.
T. Mitsuda and E. Matsumoto / Numerical Analysis of Ultrasonic Inspection 67
We next scan the EMAT probe linearly from the original center position in the right
direction by keeping the constant lift-off. Figure 7 shows the distribution of the flying time
of the first reflected wave for each position of the probe. By converting the half of the
flying time into the distance by means of the speed of the transverse wave, we can identify
the location and the width the crack. We see that the estimated position is somewhat far
from the boundary on the EMAT side. The cause may be that we have not taken into
account the transmission time of the incident wave 1 (is and the observed flying time of the
wave was longer. The estimated crack edge is not sharp because the cylindrical waves
including the longitudinal wave are reflected from the crack edge by the collision of the
plane transverse wave, which makes the peak of the pure reflected transverse wave
ambiguous. Although taking into account these points, we can state that the EMAT in
consideration can detect the inside crack in certain accuracy.
5. Conclusions
References
[1] H. Ogi, M. Hirao and H. Fukuoka, Acoustoelastic Stress Measurement on Railroad Rails Using
Electromagnetic Acoustic Transducer, Trans. JSME 60A (1994) 881-887.
[2] H. Ogi and M. Hirao, Line-Focusing of Ultrasonic SV Wave by Electromagnetic Acoustic Transducer, J.
Acoust. Soc. Am. 103 (1998) 2411-2415.
[3] R. Ludwig and X-W. Dai, Numerical Simulation of Electromagnetic Acoustic Transducer in the Time
Domain, J. Appl. Phys. 69 (1991) 89-98.
[4] R. B. Thompson, Physical Acoustics 19. Academic Press, New York, 1988.
[5] H. Ogi, M. Hirao, K. Minoura and H. Fukuoka, Quasi-Nonlinear Analysis of Lorentz-Type EMAT by
Finite Element Method, Trans. JSME 61A (1995) 638-645.
68 Electromagnetic Nondestructive Evaluation < VI)
F. Kojima et al. (Eds.)
1OS Press. 2002
Abstract. It is important to develop an efficient approximation model for the natural crack and
the narrow gap in finite element analysis of ECT problems. In this paper, a finite element model
for approximation of the crack in the ECT problems is proposed. In the proposed model, the
anisotropic conductivity is introduced for the finite element method based on the
edge element. Numerical results and experimental results of the ECT problems are shown, and
the validity of the proposed model is verified by comparing both the numerical results and the
experimental results.
1. Introduction
Various types of numerical methods in eddy current testing (ECT) problems have been
proposed and the validities have been verified. However, the conventional numerical method,
such as finite element method (FEM) and boundary element method, have a problem on
practical use in the analysis of the natural crack and the crack with narrow gaps, because a
large amount of computer resources are necessary in generating the accurate computation
model. The number of unknowns for the final simultaneous equations for the crack model
increases in proportion to the complexity of the crack shape.
On the other hand, the aim of the eddy current testing is to detect the crack from ECT probe
response caused by the change of the eddy current due to the crack. The ECT probe response
is the change of the coil impedance and it is obtained from the magnetic flux linkage of the
probe coil. Accordingly, the response is given by the integral equation taking account of the
vector potential. It is more important to calculate the impedance of the coil than a detailed
analysis of the eddy current distribution in the ECT problem. Therefore, the approximation
methods [1,2] of the crack were proposed in the numerical analysis.
In this paper, a finite element model to approximate the crack in the ECT problems is
proposed. In the proposed model, the anisotropic conductivity is introduced to the finite
element method based on the hexahedral edge element. The proposed finite element model is
applied to an ECT problem with a narrow gap crack. The computational results and the
experimental results of the ECT problem are compared and the validity of the proposed model
is verified. The applicability of the proposed model as the approximation method of the
natural crack is shown by the comparison with the experimental results for a fatigue crack
with several contact conditions.
. Tanaka et al. / Finite Element Model for Crack with Narrow Gap 69
2. ECT problem
In this section, the calculated ECT signals are compared with experimental results to
estimate the conductivity for the finite element model. An ECT problem consists of two
conductor blocks and a coil as shown in Fig. 1. There is a gap between the two blocks in the
model, and the gap is assumed a crack. The pancake-type coil moves parallel to x-axis. The
lift off is O.lmm and the internal diameter l.lmm, the external diameter 2.2mm. The
frequency of the exciting current is 100 kHz. The coil moves from -3.5mm to 3.5mm with
steps of 0.5mm.
The finite element code [3,4] using hexahedral edge elements was applied to the model. The
influence of the conductivity on the ECT signal was investigated. The ECT signals for the gap
of 0.05mm are shown in Fig 2. In the case of 2.5xl07 S/m, the behavior of the computation
results is similar to the experimental results near Omm. Accordingly, the conductivity of the
computation model was assumed to be 2.5 xlO 7 S/m. The computational results of the ECT
signal agree with the experimental results for each gap as shown in Fig. 3.
Fig. 2. Impedance change and coil position, (a) gap:0.5 mm, (b) gap:0.05 mm.
where ^ is the permiablity, a the conductivity, CD the angular frequency, A the magnetic
vector potential, N the interpolation function and Vthe region to be analysed.
The computational resources for the eddy current analysis around the crack increase with
the complexity of the shape of the crack, because many finite elements are necessary to
generate an accurate model of the crack. A special element for the crack is necessary to
complete the eddy current analysis in practical time.
Fig. 3. ECT signals and gap, (a) gap: 0.5 mm, (b) gap: 0.1 mm, (c) gap: 0.05 mm.
Fig. 4 shows eddy current vectors in a finite element containing the crack. The eddy current
depends on the crack shape and the contact conditions of the crack in the element. Each
component of the eddy current vector can be expressed by
where eris the conductivity of the conductor, Jis the eddy current. C is a constant by which
the change of the equivalent conductivity for each direction in the element is expressed. The
constant C for each element is determined in the relation to the crack shape and the contact
conditions. The value changes form 0% to 100%.
When the approximation in (2)-(4) is introduced to (1), the following equations are obtained.
M. Tanaka et al. / Finite Element Model for Crack with Narrow Gap 71
Where <? is the matrix which expresses the anisotropic conductivity due to the crack.
3. Numerical results
The ECT model in Fig. 1 was used to investigate the validity of the proposed model. The
maximum values of the ECT signals are plotted in Fig. 5. The computational results of the
conventional model and the experimental results are same as Fig 3.
The proposed finite element model was applied to the gap in Fig. 1. The thickness of the
finite element was 0.025mm. The labels (0.1%,.., 5%) in Fig. 5 denote the values of C\ in (2).
In these cases, the finite element includes the thinner crack than the thickness of the finite
element. When the Cx decreases, the calculated values move continuously to zero from the
calculated results of 0.05 mm. We can verify the proposed finite element model is able to
simulate the thinner crack than the traditional finite element.
The applicability of the proposed method to approximate the natural crack was investigated
by the comparison between the computational results of the finite element method and the
experimental results of fatigue crack model. Fig. 6 shows the fatigue crack model for the
experiment. Both sides of the crack was pressed or expanded by the mechanical load.
Therefore, the contact condition is changed by the change in the mechanical load.
0.25
0.20
0.15
0.10
0.05
Fig. 7. Computational results of the proposed finite element model and Experimental results of
the fatigue crack model.
Fig. 7 shows the comparison between the computation results of the proposed finite element
model and the experimental results. The maximum values of the impedance change for each
load are plotted in Fig. 7. In this case, the width of the fatigue crack without the load is from 5
M. Tanaka et al. / Finite Element Model for Crack with Narrow Gap 73
jam to 10 fo.m, and the coil was moved along the dotted line. The computation results are same
as Fig. 5. The ECT signals of the experimental results change in proportion to change of the
contact condition of the fatigue crack. The ECT signals of the proposed finite element model
are small compared with the experimental results, and the ECT signals have the similar
tendency to the experimental results and move continuously to zero from the experimental
results. Therefore, it seems that the computation model simulated the contact conditions that
both sides of the crack is strongly pressed than the experiment.
From these results, the possibility of the proposed method as the approximation method of
the fatigue crack was verified, when the parameter o* in (6) can be given by any functions.
4. Conclusions
A finite element model for the crack in the analysis of the ECT problem was proposed. The
proposed finite element model was applied to the ECT problem and the computational results
and the experimental results were compared.
(1) The finite element model which conductivity is approximated by the anisotropic
conductivity was proposed.
(2) The finite element code based on the edge finite element using the proposed finite element
model was applied to the ECT problem with a narrow gap. Both the conventional finite
element model and the proposed finite element model were compared, and it was verified
that the proposed finite element model is able to model the narrower crack than the
conventional finite element model.
(3) The computational results of the proposed finite element model and the experimental
results of the ECT model with the fatigue crack model were compared, and the
applicability of the proposed model as the approximation method of the natural crack was
shown.
References
[1] H.Fukutomi, H.Huang, T.Takagi and J.Tani, Identification of Crack Depths from Eddy Current Testing
Signal, IEEE Transactions on Magnetics, Vol. 34, No.5,pp.2893-2896,1988.
[2] M.Tanaka and H.Tsuboi, Finite Element Model of Natural Crack in Eddy Current Testing Problem,
Proceedings of IEEE CEFC2000, Milwaukee, USA, June 2000, p.209.
[3] M.Tanaka, M.Tsuboi et al, Computational Results of the Benchmark Problem Step 5-3: Conductor Plate
with a Support Plate, Electromagnetic Nondestructive Evaluation (IV), S.SS. Upda et al.(Eds.), IOSS
Press, 2000.
[4] M.Tanaka, K.Ikeda and H.Tsuboi, Fast Simulation Method for Eddy Current Testing, IEEE Transactions
on Magnetics, Vol. 36, No. 4, July 2000.
74 Electromagnetic Nondestructive Evaluation (VI)
F. Kojima el al. {Eds.)
IOS Press. 2002
Abstract. In eddy current analysis for the eddy current testing (ECT), it is essential to
estimate the electromagnetic Fields around cracks accurately because the fields contain
information as to the cracks. In order to get accurate solutions, a method of the eddy
current analysis for ECT is proposed by using the surface integral equation method. The
unknowns of the surface integral equations are the loop electric and surface magnetic
currents. The loop electric current satisfies automatically the condition that the divergence
of the surface electric current is zero. The magnetic surface current satisfies the condition
that the total magnetic charge is zero. In modeling the cracks, a crack element introduced.
In order to check the effectiveness and adequacy of the proposed method, we solve the benchmark
models of ECT.
1. Introduction
Eddy current testing (ECT) is a type of non-destructive methods for detecting a crack in a
conductor. In eddy current analysis for ECT, we have to evaluate the electromagnetic fields
accurately for detecting the crack. In the case of employing the integral equation method for
ECT analysis [1, 2], there exist two methods. One is a direct method. We determine the
electromagnetic fields directly by solving the surface integral equations. The other is an
indirect method. We determine the electromagnetic fields by introducing the dipole current
replacing the crack. In this paper, we discuss how to get the accurate solution by the direct
method formulated by the surface integral equation method.
Applying Green's theorem, we can obtain integral representations of electromagnetic fields from
Maxwell's equation [3]. Obtaining the electromagnetic fields on the surface of the conductor and
enforcing the continuity conditions of the tangential and normal components of the electromagnetic
fields, we get surface integral equations whose unknowns are the surface electric and magnetic
currents, Js and Ks [4]. Here, we call the method employing the surface integral equations with
the surface electric and magnetic currents as the unknowns "conventional SIEM". At the
sharp edge, the normal and tangential directions can not be defined and the continuity
conditions become void. Therefore, the conventional SIEM is employed for solving eddy
current problems when the skin depth is small and the surface of the conductor is flat [5]. In ECT
analysis, we have to solve accurately eddy currents near the crack including sharp edges and comers.
The skin depth for ECT is large for finding the crack in the conductor. In addition, as the crack \\idth
becomes narrower compared with the depth, the surface integral equations become ill conditioned and
so we cannot get accurate computed results. For ECT analysis by the surface integral equations, a
K. Ishibashi / Eddy Current Analysis by Integral Equation Method for ECT 75
crack element for formulating the eddy current with the narrow crack width has been introduced [6].
In order to get accurate solutions even when the skin depth is large, an approach to analyze the
eddy current has been proposed by employing surface integral equations whose unknowns
are loop electric and surface magnetic currents, I? and Ks [7]. By introducing Ie, the
essential electromagnetic condition that the divergence of Js is zero is satisfied
automatically. By employing a constant surface element for Ks, the condition that the sum of
the magnetic charge is zero is satisfied automatically. Here, we call the method employing
the surface integral equations with the loop electric and surface magnetic currents as the
unknowns "improved SIEM". In this paper, we employ the improved SIEM for ECT
analysis. In order to check the effectiveness and adequacy of the proposed method, we solve the
benchmark models of ECT and compare the computed results with the experimental ones [8].
The integral representations of electromagnetic fields are derived from Maxwell's equation by
applying Green's theorem [3]. The magnetic and electric fields, //op and Eop, in space with a
permeability and permittivity, /j. 0 and e 0> and the magnetic and electric field, //jp and Elp, in a
conductor with a permeability and conductivity, /j. \ and a , are given as follows.
where co is the angular frequency //ep and "ep are the exciting magnetic and electric fields at a
calculating point P0 produced by electromagnetic sources, respectively, S is the surface of the
conductor, the subscript/? denotes the value at P0 and j = V^T . Also, we define
J = nxHx, Hn=n- Hx , /if =-nxEx , En=n-Es
with the electric and magnetic fields on the surface of the conductor Es, Hs and the unit normal n
directed from the inside to the outside of the surface,
The surface electric and magnetic currents have to satisfy conditions that the divergence of ./. is zero
76 K. Ishibashi / Eddy Current Analysis b\ Integral Equation Method for ECT
and the sum of the surface magnetic charge is zero. In order to satisfy the former condition
automatically, we use a loop current 1 f circulating along the edge of the surface element. In order to
satisfy the latter condition automatically, we use a constant surface element for As.
Next, we derive the integral equations suitable for If and K^ to be determined.
Obtaining the normal component of the magnetic field on the surface of the conductor from ( 1 ) and
(3) and considering the boundary condition of the continuity of the magnetic flux density B. that is
,u 0 /i-// =/j,/r# , we get
Choosing P0on the surface of the conductor and taking the vector product on (4) with n. we obtain
The unknowns of (5) and (6) are the loop electric and surface magnetic currents. / / and
AS, because Js and Hn are given as functions of A and Ky By dividing the surface of the
conductor to be analyzed into n small elements and introducing I f and Ks. we obtain from (1)
where M is the direction of the loop current, C is the line integral route along the loop
current and A 5 is area of the small element.
We solve a benchmark model of ECT [8]. We analyze eddy currents of a metal plate (140x140x1.25
mm) with a crack (10x0.2x0.75 mm) as shown in Fig. 1.
The ratio of the crack depth to the plate thickness is 0.6 (/D=60 %). Conductivity and relative
permeability of the plate are 10000 S/cm and 1. An exciting coil with 140 turns is placed above the
crack. Inner and outer diameters of the coil are 1.2 mm and 3.2 mm, the height of the coil is 0.8 mm.
The current of the coil 7C is 1/140 A and the frequency is 300 kHz. The gap between the lower surface
of the coil and the upper surface of the plate is 0.5 mm (1=0.5 mm). We call the value L "Lift-off".
In the numerical analysis, the surface of the conductor is divided into equal surface elements
(0.5x0.5 mm). The circulating current I ( flows along the edge of each element. The surface
magnetic current Ks is supposed to be constant on each element.
In solving the surface electric and magnetic current of the crack surface, we use the conventional
SEM and introduce the crack element [7]. The matching points for solving the surface integral
equations are positioned at the center of each surface element.
K. hhibashi / Eddy Current Analysis by Integral Equation Method for ECT 77
In solving the surface integral equations, the iterative sub-domain method is used [9]. The surface
of the plate is divided into two regions: one is a region including the crack (near region)
and the other is a region far from the crack (far region). The surface electric and
magnetic currents, Js and Ks, on the near and far regions are solved alternately until
convergence. First, Js and K& on the near region are determined by considering those
on the far region to be source currents. Next, <^ and K$ on the far region are
determined by considering those on the near region to be source currents. At the first
step, Js and ./^ on the far region are supposed to be zero. The rate of the convergence in
the iterative procedure to obtain Jsg is shown in Fig.2. The convergence is checked by the
equation as
where V denotes the computed value of Jsg and the value with the subscript k denotes the
one obtained at the k-th iteration and that with the subscript k+1 denotes the one obtained at
next iteration. The maximum value of e is in the case that the exciting coil is positioned
at the center of the plate.
(9)
(10)
Coil location ( m m )
Impedance change of the exciting coil is obtained as the coil moves from the middle of the plate
along the crack at every 1 mm. The real component of the impedance change along the crack from the
middle of the plate is shown in Fig. 3 and the imaginary component is shown in Fig. 4. The computed
results are shown by the solid line and experimental values [8] of these components are shown by the
dotted line.
4. Conclusions
We have studied an approach for analyzing eddy currents induced in a conductor with a
crack by employing surface integral equations. The unknowns of the surface integral equations
are loop electric and surface magnetic currents. These unknowns are simple elements but these satisfy
automatically the essential electromagnetic conditions that the divergence of the surface electric
current is zero and the sum of the surface magnetic charge is zero.
As the width of the crack become narrower, the integral equations become ill conditioned.
The crack element for formulating the eddy current at the crack is introduced. In solving the
surface integral equations, the iterative sub-domain method is used. The computed results
were compared with the experimental ones and we find that these results were different a
little bit. We compared also with the results obtained by the conventional SIEM and we find
that the results were almost the same. The unknowns of the improved SIEM are about 3/4
of those of the conventional SIEM.
Acknowledgment
The author of this paper acknowledges the members of the Integration of Eddy Current
Testing Research Committee of the Japan Society of Applied Electromagnetics for their
suggestive discussion about the experiment and analysis.
References
[1] R. Albanese, G. Rubinacci, and F. Villone, "An Integral Computational Model for Crack Simulation and Detection via
Eddy Currents," Journal ofComputational Physics, Vol. 152, pp.736-755, 1999.
[2] J. Bowler and S. Jenkins, "Eddy-current probe impedance due to a volumetric flaw," Journal of Applied Physics, Vol.70,
No.3,pp.ll07-1114, 1991.
[3] J. A. Stratton, Electromagnetic Theory, McGRAW-HILL, pp.464^66, 1941
[4] Johnson J.H. Wang, Generalized Moment Method in Electromagnetics, John Wiley & Sons, 1991
[5] K. Ishibashi, "Eddy Current Analysis by Boundary Element Method utilizing Impedance Boundary Condition," IEEE
Trans. Magn., Vol.31, No.3, pp. 1500-1503, 1995
[6] K. Ishibashi, "Numerical Analysis of Eddy Current Testing by Integral Equation Method," IEEE Trans. Magn., Vol.37,
No.5,pp.3229-3232,2001.
[7] K. Ishibashi, "Eddy Current Analysis by Integral Equation Method utilizing Loop Electric and Surface Magnetic Currents
as Unknowns," IEEE Trans. Magn., Vol.34, No.5, pp.2585-2588,1998.
[8] T. Takagi, M. Hashimoto, H. Fukutomi, M. Kurokawa, K. Miya, H. Tsuboi, J. Tani, T. Serizawa, Y. Harada, E. Okano
and R. Murakami, "Benchmark models of eddy current testing for steam generator tube," International Journal of
Applied Electromagnetics and Mechanics,Vol.5, pp.149-162, 1994
[9j K. Ishibashi, "An Iterative Computational Method for ECT by BEM," ELECTROMAGNETIC NONDESTRUCTIV
EVALUATION. Studies in Applied Electromagnetics and Mechanics 12, Ed. by T. Takagi, J.R. Bowler and Y. Yoshida,
1OS Press, pp.31-36, 1997.
This page intentionally left blank
Eddy Current Testing
This page intentionally left blank
Electromagnetic Nondestructive Evaluation (VI)
F. Kojima et al. (Eds.)
IOS Press, 2002
1. Introduction
The nuclear fuel rod that houses nuclear fuel pellets is the first defense wall against the
catastrophic radioactive contamination. Consequently, special attention has been paid to
maintain the perfect integrity of the rod and even the irradiated spent fuels have been stored
underwater for a long time. Nuclear fuel rods are inspected by the helium leakage testing
during the manufacturing process. However, tiny pinhole type defects are sometimes
discovered while they are in service [1]. Other than the helium leakage testing, the
ultrasonic and eddy current testing are frequently used [2-4]
In this paper, a new inspection method is proposed that uses the remote field eddy current
(RFEC) phenomena [5-7] The RFEC technique has been applied for the inspection of
conducting tubular products such as oil and gas pipes using the inner diameter bobbin probe
This technique is characterized by the double through-wall transmission of the energy
before it reaches the sensor located in the remote field region. Between the two through-
wall transmission areas, potential valley and phase knot had been found by the finite
element modeling [6]. In ferromagnetic pipe inspection, typical spacing between exciter
and sensor coils is about 3 pipe diameters and the operating frequency is usually very low,
from 10 Hz to 400 Hz. The probing signal is the electromotive force (emf) induced in the
sensor coil and both amplitude and phase information can be used as the signal. The phase
signal is in fact the phase difference between the exciter and sensor. These RFEC signals
are closely related to the pipe wall condition, thickness, permeability, and conductivity and
show distinctive characteristics such as 1) they are equally sensitive to inner diameter (ID)
and outer diameter (OD) defects, 2) they are rather insensitive to probe wobble or lift-off,
and 3) the defect signal always shows double defect indications, each occurs as the exciter
and sensor coils pass a defect.
84 Y.-K. Shin /Achievement of RFEC Effects in the Nuclear Fuel Rod Inspection
Recently, the application of this technique has been extended to the inspection of plate
structure by focusing electromagnetic energy into the specimen through the use of
appropriate shielding [8-11] The use of shielding to achieve RFEC effects in the plate
structure inspired the author to design a shielded encircling RFEC probe for the inspection
of nuclear fuel rod. The operating frequency and effects of shielding are studied by using
the finite element analysis In this paper, the shielding for exciter to realize RFEC effects,
the decision on the sensor coil location, and the effects of sensor coil shielding are
presented The simulated signals using the proposed probe have very similar characteristics
to those of conventional ID RFEC signals They show that the probe is equally sensitive to
both ID and OD defects, and the phase signal strength is linearly related to the defect depth
where A , ./ , fj, a are the magnetic vector potential, the source current density,
permeability, and conductivity, respectively Assuming sinusoidal steady state and perfect
axis symmetry of the rod, coils, shielding structures and defects, the governing equation in
the cylindrical coordinate system can be written as follows
where (o is the angular frequency. By applying the Galerkin's weighted residual method
to the above equation, the following elemental matrix equations are obtained
Here, N, and N, are shape functions at each node point in an element These elemental
matrix equations are then summed up to form a global matrix equation and solved for the
magnetic vector potential at every node point
Y.-K. Shin /Achievement of RFEC Effects in the Nuclear Fuel Rod Inspection 85
Since RFEC signal is the cmfof the sensor coil, nodal values over the sensor coil area are
used for em/ calculation as follows:
(7)
Where A and r are the centroidal values of magnetic vector potential and the radial
distance at each element Since emf is expressed in the form of a complex phasor, both
amplitude and phase can be calculated
Since the nuclear fuel rod is accessible only from outside, an encircling exciter coil
covered with laminations of iron is designed to force the electromagnetic energy to
penetrate into the rod. A zircaloy metal rod is used for modeling. Material properties and
dimensions used in the finite element modeling are summarized in table 1.
To achieve RFEC effects in the nonmagnetic zircaloy fuel rod, many different operating
frequencies are tested and 700 kHz is finally chosen in this work. Fig. I shows the
logarithmic scaled equipotential plot and equiphase plot. Each plot has a unique pattern that
was named as the potential valley and phase knot in the numerical modeling study of
ferromagnetic pipe inspection [6]. The presence of these patterns indicates that the RFEC
effects are achieved
TABLE 1. Material properties and dimensions used in the finite clement modeling.
FIGURE 1. a) Log scaled cquipotcnlial plot and b) equiphase plot at the frequency of 700 kH/..
86 Y.-K. Shin /Achievement ofRFEC Effects in the Nuclear Fuel Rod Inspection
outside
FIGURE 2. Comparison of magnetic vector potentials on the surface outside and inside the fuel rod:
a) log scaled magnitude, b) phase.
In this work, the sensor has to be placed outside the rod Therefore, the sensor coil needs
to be placed in the region where the potential inside is higher than outside. Also, such
region is where the double through-wall transmission of energy is completed Fig. 3 shows
the changes occurred in the equiphase line distribution when an OD or ID defect (0 636 mm
width x 0318 mm depth) exists in that region Whether it is an OD defect or ID defect,
changes in ; the equiphase line distribution direct toward the outside, where the sensor coil
should be located. These changes in the plot enable us to predict that the signal would be
sensitive to both OD and ID defects From Fig 2, proper locations for the sensor coil can be
decided approximately depending on the choice of signals Since magnitude signal would
require excessive amplification in the experiment, this paper pays more attention to the
phase information. Fig. 2 displays only the magnetic vector potential, not the sensor coil
emfso that the exact location cannot be decided from the figure In this work, the spacing
between exciter and sensor coils is selected to be 9 54 mm (0 984 OD) Since exciter and
sensor coils move together as one probe, the sensor location is equivalent to the exciter to
sensor spacing. The size of the sensor coil used is 1.272 mm for both width and height
Fig. 4 shows simulated phase and magnitude signals for both OD and ID defects whose
depth is 50 % of the wall thickness. RFEC signals usually show double defect indications
as the sensor and exciter coil pass a defect. As a result, the distance between the two
indications in a signal is proportional to the spacing between exciter and sensor In Fig 4,
the horizontal axis is the axial distance between the centers of exciter coil and defect, and it
Y.-K. Shin /Achievement ofRFEC Effects in the Nuclear Fuel Rod Inspection 87
FIGURE 3. Changes in equiphase lines when a defect exists near the exciter: a) OD defect, b) ID defecl.
FIGURE 4. OD and ID defect signals; (a) Phase signal, (b) Log scaled magnitude signal.
is converted with respect to the OD of rod. Thus, 0 OD on the horizontal axis means that
the exciter coil passes a defect. Since the sensor coil is located in front of the exciter, the
first defect indication should appear at -0.984 OD on the horizontal axis. However, signals
in Fig. 4 do not clearly show the defect indication at that location. OD defect signals show
small indications but they are not strong as much as the indication appeared at 0 OD.
Furthermore, even the indication appeared at 0 OD is affected very much by the shielding
applied outside the exciter coil. When the defect passes the edge of the shield (-0.29 OD),
signals change abruptly. Observing the locations of unexpected signal variations, it is
evident that the external fields from the exciter directly affect the sensor signal.
In the conventional ID RFEC inspection, the sensor coil is located in the remote field
88 Y.-K. Shin /Achievement of RFEC Effects in the Nuclear Fuel Rod Inspection
region to avoid the direct contact with the fields from exciter and to monitor the indirect
fields that have penetrated the pipe wall twice Since the sensor coil in this work is located
close to the exciter, it is necessary to prevent direct contact with the exciter fields For this
reason, the sensor coil is also shielded. The width and height of iron shielding used for the
sensor coil are 2.54 mm and 2.04 mm, respectively
The signals obtained by the shielded sensor are dramatically improved as can be seen in
Fig. 5 The first indication evidently appears and it is even greater than the second
indication whose strength has changed little Also, the influence from the external fields in
the second indication is removed. Fig 5 also compares signals from both OD and ID
defects. Phase signals from OD and ID defects are almost the same Strengths of the
magnitude signal are also similar. Thus, it can be said that the probe designed in this work
is almost equally sensitive to OD and ID defects.
At this point, the choice of operating frequency made in this work is double-checked
Defect signals obtained from the same probe at the frequency of 100 kHz, 400 kHz, and
700 kHz are compared in Fig. 6 As the operating frequency is increased, the strength of
defect indication gets higher However, the middle portion between the two indications
changes very much depending on the operating frequency
FIGURE 5. OD and ID defect signals obtained by shielded sensor; (a) Phase signal, (b) magnitude signal.
FIGURE 6. Frequency dependence of the phase signal from a 50% deep OD defect
Y.-K. Shin /Achievement ofRFEC Effects in the Nuclear Fuel Rod Inspection 89
FIGURE 8. Linear relationship between the phase signal strength and (he defect depth.
To improve the signal behavior in the middle portion, new shielding is added between the
exciter and sensor and at this time the same size of coils and shielding structures are used.
The resulting signals from the same defects are shown in Fig. 7. The two indications in both
phase and magnitude signals become identical. Signal behavior between the two indications
has also improved successfully, especially in the phase signal.
Another finding is that the changes in phase signal strength is linearly related to the
defect depth variation. Fig. 8 shows that the phase signal strength changes to 5 ", 10 ", and
16 as the defect depth varies 25%, 50%, and 75%, which is approximately linear.
6. Discussion
In this work, two kinds of shielding are used. One is used for the exciter to force
90 Y.-K. Shin / Achievement of RFEC Effects in the Nuclear Fuel Rod Inspection
electromagnetic energy to penetrate into the rod and the other is used for the sensor to
prevent the direct contact with the exciter fields The results obtained in this study show
that the same shielding for exciter and sensor results two identical defect indications in a
signal even though their purpose of use is not the same Also learned is that the smaller coil
with lighter shielding produces a stronger signal than the larger coil with heavier shielding
7. Conclusion
Based on numerical modeling results, this paper proposes an encircling RFEC probe for
the inspection of nuclear fuel rods The shielding of exciter coil to achieve RFEC effects in
the rod is shown to be successful and the proper sensor location and effects of its shielding
are discussed. The signals obtained by the proposed probe have very similar characteristics
to those of conventional ID RFEC signals They show that the probe is almost equally
sensitive to OD and ID defects and the phase signal strength increases linearly as the defect
depth is increased These results show promise that the proposed probe could be used
effectively in the inspection of nuclear fuel rods Experimental verification is under way
and will be reported in the near future
Acknowledgements
This paper was accomplished with Research Fund provided by Korea Research
Foundation, Support for Faculty Research Abroad Suggestion on the sensor location by
professor Yushi Sun at Iowa State University is greatly appreciated
References
1 DM Suh. K S. Sim. W.J Kwon. J H Kim. and C H Park../. A.'.V\7' 18(I99X) X5-91
2 D.S. Koo. Y.K. Park, and E.K. Kim. J. KS\T 16 (1996) 29-13.
3. J.D. Allinson. An Eddy Currcnl Tcsl Procedure for Defective End Cap Welds. Zircatcc Precision
Industries. Inc. Preliminary Report. 1992
4 J.B. Hallctt. el al.. .A tatertals Evaluation 42 (19X4) 1276-12X0
5. T.R. Schmidt. Materials Evaluation 42 (19X4) 225-230
6. W. Lord. Y.S. Sun. S.S. Udpa. and S. Nalh. //:/:/: Trans. Magn. 24 ( I 9 X X ) 435-438.
7 J.K. Yi. and D.L. Athcrton, Materials Evaluation 56 (I99X) 771-773
X. Y.S. Sun. J. Si. D Coolcy. H C. Han. S.S. Udpa. and W Lord. IEEE Trans Mapi. 32 (1996) 15X9-1592.
9 Y.S. Sun. W. Lord. L Udpa. S Udpa. S K Lua. K H Ng. and S Nalh. Thick-Walled Aluminum Plate
Inspection using Remote Field Eddy Current Techniques. In Review of I'rogress in OM)E. D O
Thompson and D A Chimcnti(cd ). Plenum 16. New York. 1997. pp 1005-1012
10 Y.S. Sun. L Udpa. S. Udpa. W. Lord. S Nath. S K. Lua. and K.H Ng. Materials Evaluation 56 (I99X)
94-97
II. YS Sun. and T. Ouyang. Detection of Cracks in Multi-Layer Aircraft Structures with Fasteners using
Remote Field Eddy Current Method. In I'mceeiling ofSI'lE3994. A K Mai (cd.). Newport Beach. 20<)0.
Electromagnetic Nondestructive Evaluation (VI) 91
F. Kojima et al. (Eds.)
1OS Press, 2002
Abstract. The numerical simulation for the remote-field eddy current testing has
been applied to non-ferromagnetic tubes with fan-shaped defects. The results
obtained from the simulation code which uses 3D finite-element calculation is
good agreement with the experimental results. As scanning the small detector coil
along the axis of the tube, we obtained the Lissajous figure in which the maximum
magnitude appeared at just below the defect or at the side opposite the defect.
However each Lissajous is inverse in the trajectory on a complex plane. To
investigate these phenomena, the distribution of the magnetic flux density has been
evaluated. The magnetic flux density has the maximum value below the defect
when an exciter coil passes through a fan-shaped defect. Its value gradually
decreases toward to the center of the inner tube, and then increases toward to the
side opposite the defect. On the other hand, the phase gradually shifts from 0
degrees to 150 degrees. Therefore the magnitude and the trajectory of the Lissajous
would be related to the position of the detector coil.
1. Introduction
Fig. 1 shows a schematic diagram of a RFEC probe and energy flow paths. When the
current is driven on the exciter coil in a conducting tube, two components of the magnetic
field are propagated. One is called as the direct-field component coupled by an exciter coil.
The other is called as the indirect-field component coupled by eddy current. The indirect-field
component goes through out the tube wall and return back into the inner tube wall. The
propagated wave shifts 90 degrees in phase through the tube wall. Then the difference of the
phase between the indirect-field component and the direct-field component is about 180
degrees. The indirect-field component dominates in the remote-field region. On the RFECT, a
detector coil is placed on the remote-field region to pick up the indirect-field component.
To verify the validity of a numerical simulation code using edge-based finite elements
and the magnetic vector potential method, it was applied to the case as shown Fig.l. The
sample tubes are fabricated from brass. An exciter coil and a detector coil have the same
dimensions.
Fig.2 (a) and Fig.2 (b) show comparisons between experimental results and calculated
results on amplitude and phase of the voltage induced by the magnetic field in a tube,
respectively. The induced voltages are obtained as the detector coil moved with the exciter
coil fixed. In the case of no defect, the calculated results on the magnetic field correspond
with the experimental results, so that the calculation on magnetic field is noted to be carried
out reasonably.
Fig.2 Comparisons between the experimental results and the calculated results
M. Yasunishi et at. /Behavior of the Remote-Field on a Tube 93
To identify the azimuthal dependence on the defect signal, simulations were carried out
in the case of a dual exciter RFEC probe with a small detector coil moving around the inner
surface of the tube. A small detector coil is set as shown in Fig.3 and Fig.4. The position a,
the position b and the position c are the point at 0 degrees just below the defect, the point at
90 degrees and the point at 180 degrees on the side opposite the defect, respectively. The
circumferential angle of the fan-shaped defect is defined as shown in Fig.5. Table 1
summarizes the dimensions, material properties and test conditions.
Fig.6 shows the Lissajous figure associated with a fan-shaped defect (width: 1.5mm,
depth: 50%, circumferential angle: 90 degrees). It is observed that the magnitude of Lissajous
is the minimum value at the position b, and the maximum value at the position a or the
position c. In Fig.7, the trajectory of the Lissajous figure on the position a is in the direction
opposite to that on the position c.
The behavior of calculated remote-field signals shows good agreement with experimental
results |2'.
Fig.3 Schematic diagram of a dual exciter RFEC probe with a small detector coil
Fig.4 the set positions of a small detector coil Fig.5 Definition of the circumferential angle
94 M. Yasunishi et al. / Behavior of the Remote-Field on a Tube
exciter coils-
outer diameter 19.0mm
inner diameter 7.5 mm
thickness 3.0 mm
number of turns 500
current 57.4 mA
frequency 3400 Hz
interval 35 mm
tube-
outer diameter 25.4 mm
inner diameter 19.2mm
thickness 3.1 mm
material brass
relative permeability 1
conductivity 1.47X10 7 S/m
Fig.7 The trajectory on a complex plane about the Lissajous figure on the position a and the position c
M. Yasunishi et al. / Behavior oj the Remote-Field on a Tube 95
From the calculated results, the phenomena of the defect signal depending on the position
of the small detector coil are as follows. It is observed that the magnitude of the Lissajous is
the minimum value at the position b (the point at 90 degrees) and the maximum value at the
position a (just below the defect) or the position c (the side opposite the defect.). And the
trajectory of the Lissajous figure on the position a is in the direction opposite to that on the
position c. We had predicted that the magnitude of the Lissajous was the maximum value at
only the position a because the point was just below the defect. However, on the position c
also the magnitude of Lissajous appears the maximum value.
To explain these phenomena, the distribution of the axial magnetic flux density was
evaluated, when an exciter coil passed through a fan-shaped defect. Fig.8 shows the
configuration the fan-shaped defect and the exciter coils for calculation. Fig.9 shows the
amplitude of the axial magnetic flux density on the small detector coil and vicinity. Fig. 10
shows the phase of the axial magnetic flux density at the same state with Fig.9.
In Fig.9, the magnetic flux density has the maximum value below the defect when an
exciter coil passed through a fan-shaped defect. Its value gradually decreases toward to the
center of the inner tube, and then increases toward to the side opposite the defect. In Fig. 10,
the phase of the axial magnetic flux density shifts gradually from the position a toward the
position c. The phase shows about 0 degrees on the position a and about 150 degrees on the
position c.
In order to understand the behavior of the magnetic flux density at the small detector coil,
the eddy current in the tube around the exciter coil is evaluated. Fig. 11 shows the distribution
of the eddy current density, when an exciter coil passes through a fan-shaped defect. In Fig. 11,
the density of the eddy current is concentrated around defect. The indirect-field component
induced by the localized eddy current propagates on the outside of the tube. Then, the
indirect-field component on the side of the defect more propagates than that on the side
opposite the defect. Considering the law of conservation of energy, the indirect-field
component on the side opposite the defect decreases as much as that increasing on the side of
the defect. The indirect-field component on the side opposite the defect less propagates than
that in the case of no defect. Hence, at the region of the detector coil, the indirect-field
component appears much strength on the side of the defect and becomes week on the side
opposite the defect in comparison with the case of no defect. Then, the distribution of the
axial magnetic field is found to change gradually from the defect toward to the side opposite
the defect inner tube as shown in Fig.9.
Considering the equation of skin depth, the phase shift on the side of the defect (depth:
50%) becomes half value. The phase of indirect-field component on the side of the defect
appears different value from that on the side opposite the defect. At the region of the detector
coil, the magnetic field consists of summation between the indirect-field component
propagating from the exciter coil near the defect and that propagating from the other exciter
coil. Then, the phase of the magnetic field on the side of the defect appears different value
from that on the side opposite the defect. Hence, the phase of the axial magnetic flux density
affected by the fan-shaped defect shifts gradually from the defect toward to the side opposite
the defect inner tube as shown in Fig. 10.
5. Conclusions
Fig.8 The configuration the fan-shaped defect and the exciter coils for calculation
Fig.9 The distribution of the amplitude of the axial magnetic flux density
Fig. 10 The distribution of the phase of the axial magnetic flux density
M. Yasunishi et al. / Behavior of the Remote-Field on a Tube 97
Fig. 11 The distribution of the eddy-current density in the tube around the exciter coil
the calculated results, the phenomena of the defect signal depending on the position of the
small detector coil are as follows. The magnitude of the Lissajous is the minimum value at
the position b (the point at 90 degrees from the defect) and the maximum value at the
position a (just below the defect) or the position c (the side opposite the defect.). And the
trajectory of the Lissajous figure on the position a is in the direction opposite to that on the
position c. The behavior of calculated remote-field signals shows good agreement with
experimental results.
To explain these phenomena above, the distribution of the axial magnetic flux density
was evaluated, when an exciter coil passed through a fan-shaped defect. The magnetic flux
density has the maximum value below the defect. Its value gradually decreases toward to the
center of the inner tube, and then increases toward to the side opposite the defect. And the
phase of the axial magnetic flux density shifts from 0 degrees to 150 degrees gradually from
the position a toward the position c.
When an exciter coil passes through a fan-shaped defect, the density of eddy current is
concentrated around defect. The indirect-field component induced by the localized eddy
current propagates on the outside of the tube. Then, the indirect-field component on the side
of the defect more propagates than that on the side opposite the defect. Hence, the indirect-
field component on the side opposite the defect less propagates than that in the case of no
defect. Therefore, at the region of the detector coil, the indirect-field component appears
much strength on the side of the defect and becomes week on the side opposite the defect in
comparison with the case of no defect. Then, the distribution of the axial magnetic field is
found to change gradually from the defect toward to the side opposite the defect inner tube.
As has been described, the change of the magnetic flux density is associated with a fan-
shaped defect. If a detector coil becomes smaller and more sensitive, the more detail
evaluation about the feature of a fan-shaped defect can be performed. Also the propagation of
the energy flow of the electromagnetic field can be made clear.
References
[1] T. R. Schmidt, The Remote Field Eddy Current Inspection Technique, Materials Evaluation, 42(2),pp.225-
230,(1984)
[2] H. Higuchi, M. Isobe, M. Nishikawa, Defect Characterization by Using Remote Field Eddy-Current
Testing, Electromagnetic Nondestructive Evaluation, IOS Press, pp 153-160, (1997)
98 Electromagnetic Nondestructive Evaluation (VI)
F. Kojima et al. (Eds.)
IOS Press, 2002
Abstract. This paper reports our recent endeavor to develop automated, systematic inversion tools
by the novel combination of neural networks and finite element modelling for eddy current flaw
characterization in steam generator tubes. Specially, this paper describes 1) the construction of
databases with abundant flaw signals from 2-D axisymmethc flaws with tip variation using finite
element models, 2) the extraction and selection of sensitive features for flaw classification and
sizing, and finally 3) the inversion of ECT signals by use of two neural networks for flaw
classification and sizing. In addition, this paper also presents the performance of proposed
inversion tools for classification and sizing of 2-D axisymmetric flaws 1) having symmetric cross-
sections with the variation in tip width, and 2) having non-symmetric cross-sections with the
variation in tip deviation.
1. Introduction
The finite element ECT models used in this work can describe very carefully the
geometry of 2-D axisymmetric flaws with the variations in the tip width and/or deviation,
since they adopts flexible combination of quadrilateral and triangular elements. Figure 1
shows the configuration of ECT of an Inconnel 600 tube with the differential bobbin probe
for FEM simulation of flaw signals to construct the databases.
Figure 1. Parameters for the simulation of the eddy curent testing. ( unit: mm
In this work, we have constructed two databases. Database I has 432 flaw signals
simulated from 2 types of symmetric flaws with the variation in the tip width. Two flaw
types include "ID" (ID flaws with different tip width) and "OD" (OD flaws with different
tip width) flaws. In addition, by changing the depth and tip width of these flaws (with a
fixed width at the opening) as shown in Figure 2, a total of 108 flaws were generated, and
then from each flaw four signals were obtained with four different testing frequencies of
100,200, 300 and 400 kHz
Database II has 384 flaw signals simulated from 4 types of nonsymmetric flaws
with the variation in the tip deviation. Four flaw types include "I-ID" (I-shaped ID flaws
with different tip deviation), "I-OD" (I-shaped OD flaws with different tip deviation), "V-
ID" (V-shaped ID flaws with different tip deviation) and "V-OD" (V-shaped OD flaws
with different tip deviation). Again, by changing the width (at the openning), depth and tip
width of these flaws as shown in Figure 3, a total of 192 flaws were generated, and then
from each flaw, two ECT signals were obtained with two different testing frequencies of
100 and 400 kHz. In the Figure 2 and 3, w, d, wt and 5 denote flaw width, flaw depth, tip
width and tip deviation, respectively.
100 5.7. Song etal. /Inversion of ECT Signals from Flaws with Tip Variation
Fl. Max Resistance F2. Max Resistance Angle F3. Max Reactance
F4. Max Reactance Angle F5. Max Impedance F6. Max Impedance Angle
F7. Starting Angle F8. Ending Angle
F9. Turning phase angle at the point of maximum impedance of the signal
F10. The length up to the maximum reactance point of the signal / The length from the maximum reactance
point of the signal
Fl 1. Total length of the signal / Magnitude of the impedance at the maximum reactance point
Fl. First Peak Resistance F2. First Peak Resistance Angle F3. Second Peak Resistance
F4. Second Peak Resistance Angle F5. |First Peak Resistance-Second Peak Resistancel
F6. First Peak Reactance F7. First Peak Reactance Angle F8. Second Peak Reactance
F9. First Peak Reactance Angle F10. |First Peak Reactance -Second Peak Reactance]
F11. First Peak Impedance F12. First Peak-Impedance Angle F13. Second Peak Impedance
F14. Second Peak Impedance Angle F15. Starting Angle F16. Ending Angle
F17. First Middle Impedance F18. Second Middle Impedance
F19. |First Middle-Second Middle| F20. Junction Point Magnitude F21. Junction Point Angle
F22. Interception Vector Magnitude F23. Interception Vector Phase Angle
5.7. Song et al. /Inversion of ECT Signals from Flaws with Tip Variation 101
To reduce the possible redundancy that might exist in the set of extracted features,
the feature selections for classification and for sizing were performed separately. For flaw
classification, two selection criteria (the single feature classification performance and the
linear correlation coefficients of the specific feature to other features) were adopted. For
flaw sizing, the sensitivity of individual features to the flaw size parameters (flaw depth,
flaw width, tip width and tip deviation) together with the linear correlation coefficients
between two different features were used as selection criteria.
Based on these criteria, feature selection has been performed for both problems.
The result of the feature selection analysis will be addressed in Section 5 together with the
flaw classification and sizing performances.
102 S.J. Song et al. /Inversion of ECT Signals from Flaws with Tip Variation
4. Neural networks
For the automated flaw characterization in tubes, a hybrid of two neural networks
that has been used in the previous works [ 1,2] was used in this work. In this hybrid system,
a probabilistic neural network (PNN) [4] classifier determines the flaw type, and then the
back propagation neural network (BPNN) [5] estimates the size parameters (the tip width
and deviation) of the flaw. Due to the space limitation, the detailed discussion on these
neural networks can not be given here. However, one can refer the references [ 1 ] and [2]
for the complete discussion.
The characterization of flaws with the variation in tip width was carried out using
flaw signals in the database I, with performing the flaw classification by PNN classifiers
and the flaw sizing by BPNNs. For this purpose, the flaw signals in the database I were
split into two parts for training and test. The training set had 240 flaw signals while the test
set consisted of 192 signals. Based on the training set, feature selection for classification
was carried out to choose only one kind of a feature (Fl 1), while 3 kinds of features (F5,
F9, Fl 1) were selected for sizing.
A PNN classifier and the two BPNN size estimators were trained based on the same
training set, and test samples were fed into the neural networks for the performance
demonstration. The PNN showed 100% correct classification rate, since the discrimination
between OD and ID flaws is a relatively easy task. Table 3 summarizes the BPNN
performances for the estimation of the flaw depth and the tip width in terms of the linear
correlation coefficient (r*) between actual and estimated (by BPNNs) size parameters.
BPNN showed very outstanding performance for estimating of not only flaw depth but also
tip width, with r* higher than 0.97 for all 4 cases.
The characterization of flaws with the variation in tip deviation was carried out
using flaw signals in the database II, with performing the flaw classification by PNN
classifiers and the flaw sizing by BPNNs. For this purpose, the flaw signals in the database
II were split into two parts for training and test. The training set had 288 flaw signals while
the test set consisted of 96 signals. Based on the training set, feature selection for
classification was carried out to choose two kind of features (F15, F16) for the
determination of flaw location (ID or OD) and 3 kinds of features (F15, F16, F19) for the
determination of the cross-sectional shape of flaws (I-shape or V-shape), while 7 kinds of
features (Fl, F5, F10, Fl 1, F19, F20, F22) were selected for sizing.
S.J. Song et al. /Inversion of ECT Signals from Flaws with Tip Variation 103
A PNN classifier and the two BPNN size estimators were trained based on the same
training set, and test samples were fed into the neural networks for the performance
demonstration. The PNN showed 100% correct classification rate for the determination of
flaw location, while produced only 54% correct classification rate for the determination
flaw shape. This implies the discrimination of OD flaws from ID ones is a relatively easy
task while the identification of flaw shape is truly difficult. Table 4 summarizes the BPNN
performances for the estimation of the flaw depth, the tip width and the tip deviation in
terms of the linear correlation coefficient (r*) between actual and estimated (by BPNNs)
size parameters. BPNN showed very outstanding performance for estimating of not only
flaw depth but also tip width, with r* higher than 0.97 for all cases under consideration.
6. Conclusions
In the present work, we have established automated, systematic inversion tools for
eddy current testing signals for flaws with the variation either in tip width or tip deviation.
For the classification and sizing of flaws with tip width variation, the proposed approach
showed very outstanding performance not only for the determination of flaw types but also
for the estimation of tip widths as well as flaw depths. For the classification and sizing of
flaws with variation in tip deviation, the proposed approach also showed very good
performances for the determination of flaw location, and for the estimation of tip
deviations as well as flaw depths and widths. However, when the flaws were non-
sysmmetric and inclined in their cross-sections, the determination of the cross-sectional
shape of flaws was turned out to be very difiicult.
Acknowledgements
The authors are grateful for the support provided by a grant from the Korea
Science and Engineering Foundation (KOSEF) and Safety and Structural Integrity
Research Center at the Sungkyunkwan University.
References
1. S. J. Song, H. J. Park, Y. K. Shin and H. B. Lee, in Rew. Prog, in Quant. NDE, Vol. 18A, eds. D. O.
Thompson and D. E. Chimenti, Kluwer Academic/Plenum Publishers, New York, (1999), pp. 881-889.
2. S. J. Song and Y. K. Shin, NDT&E International, Vol. 33 (2000), pp. 233-243.
3. S. J. Song, C. H. Kim, Y. K. Shin, H. B. Lee, Y. W. Park and C. J. Yim, to appear in Rew. Prog, in
Quant. NDE, Vol. 20, eds. D. O. Thompson and D. E. Chimenti, Kluwer Academic/Plenum Publishers,
New York, (2001).
4. D. F. Specht, Proc. of IEEE Int. Conf. on Neural Networks, Vol. 1(1988), pp. 525-532.
5. D. E. Rumelhart, G. E. Hinton and R. J. Williams, in Parallel Distributed Processing, ed. D. E.
Rumelhart and J. L. McClelland, MIT Press, Cambridge, MA, (1986), pp. 318-363.
104 Electromagnetic Nondestructive Evaluation (VI)
F. Kojima et al. (Eds.)
IOS Press. 2002
Abstract An ECT system configurations for concrete material is presented. The experimental devices
are designed for the practical use at the job sites, and also can be operated under high frequency. The
experimental results show the validity of proposed system. Finally we discuss about suitable
specifications of the practical MFES system.
Introduction
Recently, a lot of serious accidents concerned with concrete structures have happened in
Japan. One of the main causes of these accidents is crack grown in low-quality concrete
material, decrepitudes of concrete, fatigue or so on. In order to avoid the accidents and
prolong the lifetime of structures, cracks on the concrete materials should be detected.
Many conventional NDT methods for concrete materials can be listed, AE, ultra-sonic,
thermograph, microwave, and so on. However, generally speaking, these conventional
methods have following common disadvantages;
(1) Large and heavy equipments,
(2) Need large power supplies,
(3) Need complex data processing at laboratories after measurements.
(4) Very expensive.
NDTs on concrete structures are supposed to be done at the job sites, that is to say, out of
doors. From these disadvantageous reasons, NDTs for concrete materials are not used so
often or widely nowadays.
We apply the eddy current testing (ECT) techniques to the concrete material. The ECT
techniques for metallic materials such as steam generator tubes of the nuclear reactors had
been rapidly developed these ten years. Some techniques are still usable for low
conductivity material. Therefore we investigate the availability of ECT system for concrete
materials.
We employ "multi-frequency exciting and spectrogram" (MFES) method[l][2] for ECT
on concrete materials. A basic idea behind the MFES method is to use a complex signal
containing selected harmonic components as an exciting signal and a spectrogram for
precise crack characterization. The frequency characteristics of concrete materials are
unknown, therefore, this method may be able to determine suitable frequency band for
ECT.
S. Nagata et al. / Development of Practical MFES System 105
The aim of this study is to establish the smaller and less expensive MFES system. The
ECT systems developed in the laboratories tend to grow large and expensive ones. Some
of the authors already presented the successful results about the NOT for concrete
structures [3]. These results were measured by means of superior equipments, for example,
extremely high performance A/D converter (20 Msps, 32 bit resolution). They are suitable
for the fundamental studies of NDT in the laboratory, however, they are not suitable for the
practical equipments at the job sites. Also it is not clear how many resolutions or how high
sampling rates are sufficient. We should clarify the required and sufficient specifications
such as resolutions and sampling frequencies and develop the practical MFES systems.
Concrete mortars have low conductivity because they always contain water. It means
that ECT on concrete material should be excited under high frequency due to significant
output signal of crack. We estimated that high frequency operation above 10kHz up to
1MHz, therefore we designed and produced the equipments including the digital function
generator, the power amplifier and pre-amplifier so that all equipments satisfied the
required frequency characteristics.
A block scheme of the MFES system is shown in Fig. 1. The waveform data generated
by calculation were written into two 8bit EPROMs (27C256), and set inside the function
generator. The function generator supplies an exciting coil of a sensor through a power
amplifier. This digital function generator includes high-speed D/A converter and it can
operate at 10 mega sampling per second. It is the same sampling rate of the A/D converter
(PCI-3163) due to synchronizing. For ease of digital signal processing, especially fast
Fourier transform, we determined 1024 sampling numbers per a cycle at the primary
frequency. Waveform were superimposed by 63 sinusoidal components having frequency
from 9.8 kHz ( = 10 MHz / 1024 ) up to 312.5 kHz each 4.9 kHz. The sensor "CONI"
shown in reference [3] was also used in this system.
Figure 2 shows the details of the digital function generator. The reason why EPROMs
are used for waveform memory is to achieve its sampling rate lOMsps. The access time of
106 5. Nagata et at. / Development of Practical MFES System
the waveform memories should be less than 100ns. In order to generate the waveform
containing the arbitrary wave components, the waveform memories had better been the
RAMs, however, system should be complicated because embeded micro computer would
be needed for modifications of waveform data. Standard logic ICs (LS series TTLs) are
used to perform all procedures of D/A convertion within a system cycle 100 ns.
Results of experiments
In order to verify usability of the proposed sensor and system, two experiments were
carried out.
Fig. 3. Present MFES system can detect a defect on the iron plate.
In the first experiment, iron plate (100x200x5mm) consisting of artificial defect were
investigated. The specimen is show in Fig. 4-(a). The measurements were done by scanning
the probe in X direction in steps of 1mm. The liftoff (the gap between the specimen surface
and the sensor) was measured to be 0.5mm. This one is not for concrete material, but we
can confirm the varidity of proposed system as a conventional NDT method.
Selected result of measurements performed on this specimens are shown in Fig. 4-(b).
The defect exists at the point shown at 21mm. The defect on the iron plate is causing
S. Nagata et al. /Development of Practical MFES System 107
spectrograms with peak point for the lowest frequencies. From the result, we can say that
present system is compact, simple and useful for MFES method.
Another experiment is to search for a steel bars inside concrete mortar. The specimen is
show in Fig. 4-(a). Figure 4-(b) shows the result of measurement. At present, any
remarkable signal of steel bar are not founded.
Conclusions
In this paper, we proposed a compact and inexpensive MFES system. This system
contained the function generator and the A/D converter, and their resolutions are 12 bit and
sampling rate lOMsps. Proposed system successfuly detected the defect on the iron plate.
It can be confirmed that proposed system is valid for ECT equipment.
At present we can not find out any remarkable defect signals for concrete structure with
this system. From the results of steel bar searching, fixed sampling rate and frequency
components should be improved. As shown in Fig. 3-(b), lower excitation frequencies are
supposed to be suitable for iron or steel detections due to serious skin effect. The sampling
rate and excitation frequency components should be optimized in further studies.
Some of the authors had already confirmed the validity of MFES method on concrete
defects at lower frequency components with 32bit A/D converter. More resolution of data
acquisition and some digital filtering techniques are supposed to be required.
References
[1] T. Chady, M. Enokizono, "Eddy Current Testing of Stainless Plates by Using Matrix Sensor",
Proceedings of the 7lh Magnetodynamics Conference, April, 1998, pp.107.
[2] T. Chady, M. Enokizono, R. Sikora, "Crack Detection and Recognition Using an Eddy Current
Differential Probe", IEEE Transactions on Magnetics, Vol. 35, No. 3, May, 1999, pp. 1849-1852.
[31 T. Chady, M. Enokizono, K. Takeuchi, T. Kinoshita, R. Sikora, "Eddy current testing of concrete
structures", JSAEM Studies in Applied Electromagnetics and Mechanics 9, in press.
108 Electromagnetic Nondestructive Evaluation (VI)
F. Kojima el al. (Eds.)
IOS Press. 2002
A new eddy current method for detecting small subsurface cracks in aircraft uses a supercon-
ductive driver coil for exciting the electromagnetic field in the test piece and a differential
pick-up loop for sensing perturbations in the field due to flaws, Figure 1. The driver consists
of approximately one hundred closely spaced strips of niobium deposited on a silicon sub-
strate. Only the central separation between strips is significant on the scale of common flaws,
so in calculations of the electromagnetic field the driver can be approximated as two parallel
conducting strips. The driver carries a sequence of current pulses at regular intervals with a
pulse repetition rate of around 100 Hz.
The differential pick-up coil consists of a single niobium wire formed into two adjacent
semi-circular loops, 6 mm diameter, wound in opposite directions. This is known as a double-
D coil. The pick-up circuit couples inductively to a remote SQUID in a 5 Kelvin refrigerator.
The refrigerator keeps the coils, cryogenic cable and SQUID at approximately 6 Kelvin. If
the windings are perfect, mutual inductance between the source and pick-up coils vanishes.
The configuration and response function of the probe is described in more detail in reference
[1].
The pick-up loop signal has been calculated for a subsurface crack in a stack of aluminum
plates using a boundary element scheme. As a first step, the electric field in an unflawed con-
ductor due to the driver coil is determined. This is then used in a boundary element calculation
J.R. Bowler et at. /Evaluation of Crack Signals 109
Figure 1. Superconductive probe represented as two current strips with a differential double-D pick-
up loop. The center of the pick-up loop is located at x = y = 0 and the length of the current strips (in
the y direction) is Id.
of the discontinuity in the electric field at the crack, as represented by a current dipole layer
varying in time. The differential flux through the pick-up loop is then found from the dipole
density using a reciprocity theorem. Application of the theorem requires knowledge of the
electric field generated by the pick-up loop in the unflawed conductor as if it were used as an
excitation coil. The steps in the calculation are summarized below followed by a description
of typical results.
2 Excitation Current
The current in the driver coil is a smoothed function of time known as the Moulder pulse,
M(t);
In (1), r is the period of the pulse, s is a pulse sharpness parameter, c = arccosh(\/2) and
n = 0,l,2,....
It is useful to approximate the Moulder pulse by a simple function of time that has an
elementary Laplace transform allowing the inverse transformation of the field into the time
domain to be carried out analytically [2]. The elemental function chosen is a linear ramp
function with finite duration. A superposition of linear ramps approximates the pulse as a
piecewise linear function. It is possible to approximate the current rise by a single ramp, by
matching the slope of the ramp with that of M(t) at half height, i.e. at M(t) = 1. However,
a better approximation is to use an odd number of contiguous ramps whose end points lie on
M(t). In Figure 2, the normalized current is compared with an approximation of the current
using 11 ramps.
10 J.R. Bowler el al. / Evaluation of Crack Signals
Figure 2. Comparison between the exciting current variation with time and an 11-ramp fit as used in
calculations.
3 Excitation Field
For the simple case of a probe in air above a half-space conductor with conductivity OQ,
permeability HQ and whose surface is in the plane z = 0, the unperturbed electric field due to
the excitation coil, E^(r, t), is a solution of
with the auxiliary condition that the electric field has zero divergence. The unperturbed field
is transverse electric (TE) with respect to the direction normal to the surface of the conductor
and therefore can be written
where ^'(T, t) is a TE potential. Substituting (4) into (2) and (3) and operating with V (zx)
yields
where ty = V2^' with V2, = V2 d^/dz2 and j V (z x J). Expressing the solution in
terms of a Green function;
J.R. Bowler et al. / Evaluation of Crack Signals I]
and
The spatial arrangement of the current drive coil lends itself to two-dimensional Fourier
transform analysis. Also taking the Laplace transform of (8) with respect to time gives
with
In the above equations, u and v are the Fourier space variables, /c2 = u2+v2, j2 = K?+HQ(TOS
and for 7 the root of K? + faffos with positive real part is taken.
The current density in the source coil takes the form
where F = xFx + yFy and the source coil is located in the plane z = ZQ. Fx is odd in x and
y, and Fy is even in x and y. Writing (1 1) in terms of j and transforming in space and time
yields
Assuming that the return current paths (e.g. at x = c) are narrow filaments, F x(u,v] is
given by
where ip is given in (9). Substituting (10) and (12) into (9) gives
Formally inverting (14), with i(s) for a multiple ramp excitation given in reference [2],
in which Q(i) is the Heaviside step function and x(, z, t) is evaluated in reference [2]. The
remaining two-dimensional inverse Fourier transformation must be performed numerically,
for example by means of a fast-Fourier transform routine [3].
112 J.R. Bowler et al. / Evaluation of Crack Signals
4 Defect Detection
To calculate the response of the double-D pick-up loop to field perturbations generated by a
flaw, the electric field produced by the loop in an unflawed conductor must first be calculated.
Then the net flux through the loop is determined using a reciprocity relationship.
The current in the loop will be written in terms of an equivalent magnetic source, M(r, t),
related to the current by
The choice of M(r, t) is somewhat arbitrary provided that it satisfies (17). Usually a closed
current filament in a plane is represented by a magnetic shell bounded by the current path.
Although the shell need not necessarily be in the plane of the loop, calculations are simplified
if it is confined to the plane.
If the current is transverse to the z -direction one can choose to write M(r, i) = zM'(r, t)
and hence reduce the field problem to the scalar form
The potential $? is evaluated by taking the Laplace transform of (20) and expressing the
transformed Green function as
which can be derived using the approach given in reference [4]. In (21), e^ is the Neumann
_ ~ ~
factor; e0 = 1 and em = 2 (m = 1, 2, 3, . . .). The function/ (/c, z z1, s) = 2 <7 (/c, z z1, s)
for the interior of a half-space conductor, with 9 defined in (10). For the double-D pick-up
loop, define the Laplace transform of M(r, t) as
The reciprocity theorem requires that a unit step current is used to to define m(r, s), hence
I(s) = 1/s. Equations (21) and (22) are substituted into the Laplace transform of (20) and the
integration with respect to <j>', pi and 2? performed to give a summation of Bessel integrals with
respect to K. The inverse Laplace tranform can be performed analytically but the K integrals
must be computed numerically and the summation truncated at a suitable order depending on
the required accuracy of the result.
This approach for modeling the double-D pick-up loop differs in detail from, but is similar
to, an alternative analysis for time-harmonic excitation given in reference [5]. A truncated
series of 5 terms is sufficiently accurate in most cases [5].
J.R. Bowler et al. / Evaluation oj'Crack Signals 1 13
The effect of the crack on the electromagnetic field is modeled as a surface distribution of
current dipoles p(r,) [6]. Applying a dyadic form of Green's theorem and assuming that
the normal component of the current density at the crack is zero leads to a Fredholm integral
equation of the first kind for p(r, t):
Jf (r0, t) = lim ( a^x^- - A vV T / G(r, T', t, t') p(r', t1} dS1 dt'.
r->ro \ Cft OX J JQ J So
So
In (23), r0 denotes field points on the crack surface, Jx (r 0 ,f) is the normal component of
the unperturbed current density at the crack and the dyadic Green function kernel is given in
reference [7]. The crack region is discretized into rectangular elements and p(r, t) is assumed
piece-wise constant in space and piece-wise linear in time. Equation (23) is then solved nu-
merically to obtain p(r, t).
6 Numerical Results
The net flux threading the double-D pick-up loop is proportional to dHz/dy [2]. In Figure 3,
the variation in values ofdHz/dy with both time and position above a cracked conductor are
shown. The field derivative is computed along a line above a crack 5 mm long and 1 mm high
buried 4 mm below the surface of the conductor. The crack is centered at (0, 0, 4.5) mm and
oriented so that the unit vector x is normal to the crack surface.
The single negative peak and smaller, positive peaks are characteristic of the signal ob-
tained as the probe is scanned along a line above the position of the crack. The signal mag-
nitude is larger and the negative peak occurs earlier in time for a crack nearer the surface of
the conductor. Thus, the position of the peak in time conveys information about the depth of
the defect. In Figure 4, a comparison between theoretical and experimental data is shown.
The theoretical data is shifted linearly in time to account for delays in the electronics of the
measurement circuit.
The computed data shown in Figures 3 and 4 assumes excitation of the conductor by
infinitely long current strips (d > oo). It is anticipated that the agreement between experi-
mental and theoretical data will be improved by modeling the strip length more accurately.
7 Conclusion
The transient current rise, two-dimensional current source and double-D pick-up loop of a
pulsed, superconductive probe have been modeled. It has been demonstrated that the super-
conductive, pulsed eddy-current probe has an outstanding capability for detecting deep-lying
flaws in layered structures. There is good agreement between theoretical calculations and ex-
perimental data for cracks concealed in a layered metal structure, several millimeters below
the surface. In the next phase of this work, the effect of a fastener adjacent to the crack will
be included in the model.
14 J.R. Bowler et al. /Evaluation of Crack Signals
Figure 3. Calculated response from a 5 mm x 1 mm crack buried 4 mm below the conductor surface.
Values of dHz/dy 2 mm above the conductor surface are shown.
0.6
Figure 4. Experimental and theoretical data for the peak signal from a 5 mm x 1 mm crack buried 3
mm below the surface of a conductor.
J.R. Bowler et al. / Evaluation of Crack Signals 1 15
8 Acknowledgement
Work sponsored by the Air Force Office of Scientific Research STTR Program, contract
F49620-99-C-0056.
References
Abstract. In this paper, we investigate the electromagnetic field in remote field eddy current testing
excited by pulse current. The time-domain finite element method is applied to analyse the field and
voltages of the detecting coil. It is clarified that the time-variation of the detected voltage shows the
effects of magnetic fields produced by the exciting coil current and the eddy current separately. The
magnetic field produced by the eddy current, which include much information of the cracks of the tube,
can be detected with the appropriate separation length between the exciting and the detecting coils.
1. Introduction
The remote field eddy current testing (RFECT) is an effective method that can detect
both the inner and the outer cracks of the tube [l]-[5]. In this case, the magnetic fields can
be considered as the sum of two components. One is the "near field", the other is the "remote
field". In order to detect the crack of the tube, the separation length between the exciting coil
and detecting coil must be wide enough because the information of the crack is mainly-
included in the remote fields. But it is difficult to determine the appropriate separation length
between the coils. In the case of time harmonic (sinusoidal) excitation, the phase shift of the
detected voltage relative to the exciting coil voltage gives information about the eddy current
(remote) field.
In this paper, we investigate the RFECT with pulse exciting currents. The time-domain
finite element method is applied to analyse the fields and voltages of the detecting coil. In
order to understand the detected voltage waveform, we introduce a new classification of the
magnetic fields on the RFECT, which are "primary field" produced by the exciting coil and
"secondary field" produced by the eddy current in the tube. The magnetic field can be
regarded as the sum of these two components as well as the "near field" and "remote field".
It is clarified that the time-variation of the detected voltage shows the effects of the primary
and the secondary fields separately and that the appropriate separation length between the
coils can be decided from the detected voltage waveforms.
2. Calculation method
The formulation of the time-domain finite element method for the analysis of RFECT in
case of axial-symmetric arrangement is:
1 A'+ * - A1
VxA) = Jc + J* = Jc-a (1)
A/
K. Yamazaki / Time-Domain Finite Element Analysis of Remote Field ECT 17
where A is the magnetic vector potential, Jc is the exciting coil current, Je is the eddy current
and cris the conductivity of the tube.
To obtain the "primary filed" and the "secondary field" separately, we introduce a
procedure shown in Fig. 1, which is the combination of the time-domain and the static finite
element analysis. First, the total electromagnetic field is calculated by (1). Second, The
primary field, which is produced by the exciting coil current, is calculated by the static
analysis with the permeability of the tube and the exciting current at each time step from
following equation.
(2)
where A i is the magnetic vector potential produced by the exciting coil current.
Then the secondary field, which is produced by the eddy current of the tube, can be
obtained as follows.
Ai = A A\ (3)
where Ai is the magnetic vector potential produced by the eddy current.
Fig.2 shows the analysed model. The tube is made of the non-magnetic material whose
conductivity is 1.5xl07 S/m. In the finite element analysis, 13,000 triangular elements are
applied to the discretization of the axisymmetric time-domain formulation. Fig.3 shows the
finite element mesh.
First, we examine the accuracy of the time-domain finite element analysis in the case of
the sinusoidal excitation without the crack of the tube. Fig.4 shows the RMS voltage of the
detecting coil due to the separation length between the coils calculated by the frequency-
domain and time-domain finite element analyses. The experimental result is also shown. It
indicates that the result of the time-domain analysis agrees well with the frequency-domain
analysis when the number of the time step is enough. It also agrees well with the
experimental result. The validity of the time-domain analysis is verified from these results.
Next, let us assume that the exciting coil current is the pulse current shown in Fig.5,
which is the half sinusoidal wave of 1 OkHz. The pulse current flows every 1 ms periodically.
The calculated voltage waveforms of the detecting coil due to separation length between
coils are shown in Fig.6. When the separation is short, the voltage peak that corresponds to
the pulse exciting current is remarkable. This peak must be caused by the primary field
produced by the exciting coil current. On the other hand, when the separation is wide, the
voltage peak decrease and a delayed wave with large time-constant appears. This wave is
caused by the secondary field produced by the eddy current in the tube.
To verify the consideration, let us separate the total magnetic field to the primary and
the secondary field by the proposed procedure shown in Fig. 1. Fig.7 shows the results of the
fields at three time instants. When the exciting current is increasing, the direction of the
secondary field produced by the eddy current is opposite to the primary field produced by the
exciting current. On the other hand, when the exciting current is decreasing, the secondary
field and the primary field are in the same direction due to the Faraday's law. The eddy
currents still remain and spread even when the exciting current becomes zero. In this case,
the magnetic field is only the secondary field. It is clarified that the delayed wave of the
detected voltage is caused only by the secondary field produced by the eddy current in the
tube. It can be considered that the delayed waveform includes much information of the tube.
Fig.8 shows the RMS voltage of the detecting coil due to relative position of the crack
and the coil when the coil separation is 36.8 mm. The RMS voltage of the waves shown in
Fig.6 can be calculated as follows:
VKMS = J f v 2 d f (4)
where v is the instantaneous value of the detected voltage and T is the time-interval (1ms).
It indicates that the detected voltage changes due to the depth of the crack and that we
can detect the depth of both inner and outer crack when the separation length between the
coils is wide enough. On the other hand, Fig.9 shows the results when the coil separation is
3.15mm. hi the case of the inner crack, the crack can be detected but the change of the
voltage due to the depth of the crack is small. Furthermore, in the case of the outer crack, we
cannot detect the crack except for the 100% perfect crack.
Fig.8. RMS voltages of detecting coil due to relative position of crack and coil.
(Separation length between exciting and detecting coils is 36.8mm)
122 K. Yamazaki/Time-Domain Finite Element Analvsis of Remote Field ECT
Fig.9. RMS voltages of detecting coil due to relative position of crack and coil.
(Separation length between exciting and detecting coils is 3.15mm)
K. Yamazaki / Time-Domain Finite Element Analysis of Remote Field ECT 123
4. Conclusions
The detected voltage of the remote field ECT excited by the pulse current is analyzed
using time-domain finite element analysis. The primary field produced by the exciting
current and the secondary field produced by the eddy current are separated and displayed
by the proposed procedure. The voltage waveform of the detecting coil shows the effects of
the primary field and the secondary field separately. Both the inner and the outer cracks can
be detected by the pulse excited remote field ECT with appropriate separation length
between exciting and detecting coils. It can be considered that the appropriate separation
length can be decided from the waveforms of the detected voltage. Further work is required
to understand the relation between the definition of "primary and secondary field" and "near
and remote field"
References
[ 1 ] T. R. Schmidt, The Remote Field Eddy Current Inspection Technique. Materials Evaluation, 42, 2, 225,
1984.
[2] W. Load, Y. S. Sun, S. S. Udpa and S. Nath, A Finite Element Study of the Remote Field Eddy Current
Phenomenon. IEEE Trans, on Magnetics, 24, 1, 435, 1988.
[3] H. Y. Lin and S. Y. Sun, Application of "Zoom-in Technique in 3D Remote Field Eddy Current Effect
Computaion. IEEE Trans, on Magnetics, 26, 2, 881, 1990.
[4] M. Isobe, R. Iwata and M. Nishikawa, High Sensitive Remote Field Eddy Current Testing by Using Dual
Exciting Coils, Nondestructive Testing of Materials, Studies in Applied Electromagnetics and Mechanics, 8,
IDS Press, 145, 1995.
[5] H. Fukutomi, T. Takagi, J. Tani and M. Nishikawa, Three-Dimensional Finite Element Computation of a
Remote Field Eddy Current Technique to Non-Magnetic Tubes, Journal of JSAEM, 6, 4, 343, 1998.
This page intentionally left blank
Materials Characterization
This page intentionally left blank
Electromagnetic Nondestructive Evaluation (VI) 127
F. Kojitna et al. (Eds.)
IOS Press, 2002
Abstract In this paper, the possibility of applying the damage induced mag-
netic fields to the detection of mechanical damages in austenitic stainless steel
is investigated by measuring magnetic fields induced by plastic and fatigue
damages and reconstructing the magnetic charges distribution in the material.
Mechanical damages are introduced to test-pieces of SUS304 stainless steel
at first through a tensile and a zero tension fatigue testing. The magnetic
field over the planar surface of the test-piece was measured with the use of
advanced flux-gate sensors. Finally, the distribution of the magnetic charges,
which is considered to be correlated to the status of mechanical damages, was
reconstructed from the measured magnetic field by applying a deterministic
algorithm. The reconstructed distribution of the magnetic charges agrees with
that observed with the magnetic colloidal method, which gives a strong sup-
port to the inspection method using magnetic field measurement.
I. Introduction
Nowadays, for enhancing the safety of various critical structure systems such as the ma-
chinery of a nuclear power plant, new NonDestructive Evaluations (NDE) techniques are
strongly expected by the related industries [1]. A typical example of such needs is an effi-
cient way for predicting the occurrence of cracks, or in other word, to evaluate the status
of damages or degradations in an aged structural component before macroscopic cracks are
initiated. Recently, it has been found that magnetization appears in the vicinity of a fatigue
crack even without any external magnetic field being applied to an austenitic stainless steel
of SUS304 type [2],[3]. The reason of the phenomenon is considered to be that the damages
occurred in the object alternate the magnetic property and even induce a magnetization
without applying external magnetic field (hereafter, we call it natural magnetic field) [2].
This phenomenon means a possibility to detect the mechanical damages by predicting the
distribution of magnetization in the material.
Practically, the distribution of magnetization in a material is impossible to be measured
directly. To predict the magnetization in the material, a possible strategy is to measure
the natural magnetic field produced by these magnetization outside the material and to
reconstruct the distribution of the magnetization through an inverse analysis.
In view of these backgrounds, the correlation of mechanical damages and the natural
magnetization was investigated in this work by measuring the natural magnetic fiux density
nearby the surface of test-pieces in which damages were introduced through a tensile or a
fatigue test. Advanced film type fiux-gate sensors [4] were used in the measurement [3]. The
magnetic fields were measured during and after the tensile test in order to investigate the cor-
relation of the natural magnetization and the plastic deformation. To evaluate the possibility
of observing the damage status of a key structural component in a practical environment, the
natural magnetic- fiux signal was measured in-situ during the fatigue test and its distribution
was also measured but in an off-situ state after selected numbers of strain cyclic. To estab-
lish a scheme for reconstructing the distribution of the damage-induced magnetization, an
128 Z. Chen et ul. / NDE of Fatigue Damage in Austenitic Stainless Steel
approach based on the least square algorithm and the conjugate gradient method [5].[6].[7]
was introduced. Concerning the ill-poseness of the inverse problem dealing with the direct
reconstruction of magnetization vector, the magnetic charge was taken as the unknown to be
reconstructed.
This paper is arranged in the following order. In next section, the setup, procedures
and results of the experiments are described respectively. In section 3. the method for the
magnetization reconstruction is presented with validations using simulated magnetic- fields
and measured natural magnetic fields. Some discussion and conclusions are presented in the
last section.
I TIN -/II 1 1
I Tn IZILL
The file type FG sensors have a sensitivity over 0.001G and a spatial resolution of about 1
mm. The detected magnetic- signals and the corresponding coordinate information are stored
to a personal computer through a A/D converter. The test data were then processed using
the PC with codes developed specially. In Fig.2. the shape of the test-piece, the arrangement
of the sensors and the scanning routine of sensor are depicted. The scanning range and the
pitch were chosen to be 30 minx30 mm and 1 mm respectively.
Z Chen ct ul. / NDE of Fatigue Damage in Austenitic Stainless Steel 129
c) Experimental procedures
The number of test-pieces used in the experiment and their testing conditions are sum-
marized in table 1. All the test-pieces are demagnetized by using an electromagnetic demag-
netizor before being mounted to the material test machine. Strain control was used for all
the tensile and fatigue tests. In the case of the tension test, the magnetic flux density at the
central part of the gauge area is measured in-situ. The distribution of the natural magnetic
field was then measured on a x-y stage after unloading the test-piece at each tmax listed
in Table 1. In the case of the fatigue test, the magnetic field signal measured in-situ was
picked-np at the same location of the tensile testing for each load cycle. The distributions of
magnetic field are measured when the test-piece was unloaded at selected numbers of strain
cycle. The cyclic' loading was stopped automatically when a crack(s) was initiated.
Fig.3 In-situ measurement result of NDT2 Fig.4 Distribution of natural magnetic field
during a tensile test after tensile test (maximum strain: 3%)
Fig.3 shows an in-situ measurement results of the natural magnetic field for the test-piece
NDT2 during a tensile testing. The test condition is shown in table 1 where 3%/6%/9%
means that the test-piece was unloaded and measured after total strain 3%/6%/9% was
reached. From the figure, it is clear that magnetization has been induced due to the plastic
deformation and the magnetic field did not change much at the elastic region. In addition,
the magnetization increases relatively quicker at a low plastic strain region and saturates with
the increment of the plastic deformation. As shown in Fig.4, a peak value of the magnetic
field appears in the corresponding distribution of natural magnetic field at the bottom side of
the test-piece which may be caused by global plastic sliding (Luders band). An observation
using Scanning Electronic Microscope (SEM) showed that the number of slided grains on
the bottom side (where peak of natural magnetic field occurred) is about 2 times more than
that on the upper side while the ferrite grains do not change much. The sliding lines lead
to a natural magnetic field mainly due to two kinds of principles, i.e., anisotropy due to the
sliding and martensitic transformation. Actually, Transmission Electronic Microscope (TEM)
images of the samples cut from NDT2 show that martensitic transformation may happen at
the cross points of the sliding lines. The results of X-ray diffraction verified the existence of
martensitic phase in these regions.
130 Z Chen el al. / NDE of Fatigue Damage in Austenitic Stainless Steel
A difficulty to detect the plastic damages from the natural flux density change is from the
saturation. Fig.5 depicts the in-situ field change of NDT14 at residual strain of 0.08%. 0.363%.
0.945%. 2.20%, 3.78%. From the figure, we can find that the change of the magnetization
becomes saturated at a high residual strain area. Fig.6 gives the measurement results for
test-piece NDT21 which has a central slit (3mm.0.3mm.through) during the tensile test under
conditions shown in table 1. Natural magnetic field was induced again in this sample.
b) Results of the fatigue testing
Fig.7 shows a distribution of the natural magnetic field in test-piece NDT21 after fatigue
cracks were initiated at the two ends of the central slit. The peak value shows clearly the ex-
istence of concentrated magnetization. The observation of the crack tips by using a magnetic
colloidal method and the observations made on a Magnetic Force Microscope ( M F M ) also
support the results of the natural field measurement[3]. The experimental results of another
test-piece NDT20 in which fatigue cracks were initiated at about the same number of strain
cycle verified the phenomenon occurred in the NDT 21 though the peak value of the NDT20
is smaller than that of the NDT21. This difference may comes from the different residual
strain state of the test-piece. These results clearly show that the measurement of the natural
magnetic field is a possible way to identify the existence of fatigue crack.
Fig.7 Distribution of natural magnetic field Fig.8 Distribution of natural magnetic field
after crack initiated after crack initiated
(test-piece NDT21. residual strain: 2.8%. (test-piece NDT20. residual strain: 0%.
strain range: 0.5%. strain rate: -1) strain range: 0.5%. rate: -1)
be inspected. Moreover, such a treatment can significantly reduce the number of unknowns
in the inverse analysis, i.e. great computer burden can be reduced.
At a field point (coordinate vector r) outside the volume V of the magnetized object, the
magnetic scalar potential induced by the magnetization at a source point (denoting as r') is
equivalent to that induced by a related magnetic dipole. Referring to formulae for a magnetic
dipole. the magnetic scalar potential due to magnetizations can be expressed as[5]
I t\ n' J I
Mr ) V 77dv =
r'|
where. b(r) is the leakage flux density vector at the field point r.
where. j(r'j) is a box function with zero values outside the j th cell, <?j is the charge density
at the j th cell, and n is the total number of the field points.
Substituting the above equation into Eq.(2), some calculations show that the magnetic'
flux density b(r^) at an arbitrary field point r, outside V can be obtained from the discretized
charge parameters with the following formula,
where [K] is the coefficient matrix with element k,}. and {q} is the unknown vector of magnetic-
charges.
Matrix [A^]7^/^] usually has a bad condition number, i.e. the solution of Eq.((i) is not
unique from a rigorous point of view. Therefore, regularization is necessary. Usually, a
deterministic optimization method can gives a reasonable solution for such a problem if an
appropriate initial value can be chosen somehow. Fortunately, such an initial condition is not
too difficult to be selected for present problem, e.g. we at least can choose zero distribution if
the charge at the boundary can be taken as vanished values. Based on these considerations, a
typical algorithm of the optimization method (the steepest descent method or its accelerated
version - the conjugate gradient method) is chosen to solve Eq.(6) in view of their advantage
of superior convergent speed and stability.
Actually, if we let s be the residual function. Eq.(6) can be transformed to an optimization
form as.
)\2). (7)
subject to the known constraint conditions on {q} (such a.s boundary conditions) ) ^ 6 . In
Eq.(7). wt is a weight coefficient. b is an observed data of leakage magnetic flux density and
b , ( { q } m ) is a data of magnetic field from an estimated charge distribution at in-th iteration
step. say. { b ( { q } m ) } = [K}{q}>".
The iteration formula for solving Eq.(7) reads
d~rn-\
(9)
Fig.9 Numerical models for the reconstruction of the distribution of magnetic charges
Z. Chen et al. / NDE of Fatigue Damage in Austenitic Stainless Steel 133
The schemes described in the subsection 3.1 was implemented for 1-D and 2-D problems
on a personal computer. Reconstructions of problems depicted in Fig.9 are performed to in-
vestigate the validity of the developed codes. For the 2-D problem, the charges are assumed
to be located in a square region of size 20 mm x 20 mm. The field data are of either a
horizontal or the normal component of the natural magnetic flux density vector which were
acquired from a 2-D scanning in a 30 mm x 30 mm region located HQ over the plane of
magnetic source. These conditions are corresponding to those used in the measurement of
2-D distribution of test-piece. The natural magnetic fields from the damage-induced mag-
netizations of test-piece NDT21 is taken as the experimental data in the validation of the
proposed method. As the magnetizations can be assumed as a constant distribution in the
thickness direction, the inversion is simplified to a 2-D problem. In addition to the above
measured data, several data sets of simulated magnetic field are also adopted for validating
the proposed method in case of a 2-D problem. Moreover, artificial noises are also added
into the input data for investigating the robustness of the proposed method. In the follows,
numerical results for the least square method will be presented.
3.3 Numerical results
q (Wb/m"3)
Fig. 10 Comparison of the distributions of true magnetic charges and that predicted by
using the least square method
Fig. 10 gives results for the two dimensional problem with the initial profile of the charges
selected as zero. The figure is a comparison of the reconstructed and true distribution of the
charges when noise free simulation signals were applied. The lift-off was taken as 1.0 mm and
the iteration was stopped at the 100 step. The total CPU time used was about 10 minutes
in a PC (Pentium II 350,). From the figure, we can find that the gradient method can give
good reconstruction for noise free data. To investigate the robustness, the reconstruction was
also performed with signals including artificial noise. Fig. 11 is an example of reconstruction
for signal with 20% white noise. Here, the percentage of noise is defined as the ratio of the
maximum values of the noise and the field signal. Comparing with the true distribution
shown in Fig.lO(b). one can find that acceptable results are obtained again.
134 Z. Chen et al. / NDE of Fatigue Damage in Austenitic Stainless Steel
Fig.12 depicts a reconstructed distribution of charges in the gauge region of the test-piece
NDT21. This means that the measured magnetic field data was applied in this case. The
magnetic field calculated from the predicted charges shows a good agreement with the mea-
sured ones (Fig.13). Comparing the reconstructed distribution of charges with the locations
of the fatigue cracks which have initiated at the two ends of the central slit, it can also be
found that the reconstructed distribution of charges is reasonable if one pays attention to
the position of peaks and the crack tips. The results of the magnetic colloidal experiment
[3] also show qualitative agreement with the reconstructed distribution of magnetic charges.
These results support the expectation that the proposed method is suitable for investigating
the correlation of the magnetic and status of damage.
Fig. 13 Comparison of the measured magnetic field and the magnetic- field from the
reconstructed magnetic charges
4. Concluding remarks
Based on the experimental results described in the previous sections, the following con-
cluding remarks are obtained:
1. The damage induced magnetization occurs for the SUS304 steel especially in the vicinity
of the fatigue cracks. The natural field can be considered a.s the results of the plastic
damages for tensile test. Detection of the residual strain from the measured natural
field is difficult because of the saturation and the effect of global sliding.
2. In spite of the local minimum problem of an iteration algorithm, the gradient method
has been proved feasible for magnetic charge reconstruction using a zero initial values.
Moreover, it is also demonstrated that an appropriate selection of the maximum iter-
ation number is an efficient regularization for tackling the ill-poseness of such inverse
problem.
3. The reconstruction results using the measured signals of the passive leakage magnetic
field near a test-piece with fatigue cracks support the validity of the present method at
an experimental environment from the point of view of the reasonable charge distribu-
tion referring to the crack profiles.
References
[1 R.B.Mignogna. and et al.. Passive nondestructive evaluation of Ferromagnetic materials during deforma-
tion using SQUID gradiometers. IEEE Trans. Appl. Supercond.. Vol.3. No.l. p!922-1925. 1993.
[2] Report of research committee on electromagneto-fracture and its applications on nondestructive evaluation
of degradation and damages. JSAEM report. JSAFM-R-9803. JSAEM-R-9902.
[3] M.lesaka and et al..Round-robin test for nondestructive evaluation of steel components in nuclear power
plants, in Studi. Appl. Electroinagn. mech.. Vol.14. p39-48. 1999.
[4] Y.Yamada. and et al.. Application of thin-file flux-gate sensor. New Ceramics. No.2. pl-Vls. 199s.
[5] S.Chikazumi. Magnetics, Kyouritsu Publisher. Lt.d. 1969.
[6] S.J.Norton, and J.Bowler. Theory of eddy current inversion. .1. Appl. Phys.. \ol.73. No.2. ">01-~>12. 1993.
!7i Z.Chen and K.Miya. EOT inversion using a knowledge based forward solver. .1. Noudestr. Eval.. Vol.17.
No.3. p!67-175. 1998.
Electromagnetic Nondestructive Evaluation (VI) 135
F. Kojima et al. {Eds.)
IOS Press, 2002
1. Introduction
More than 50 commercial nuclear power plants are in operation and supply electricity of
about 3 X 108 KWh in Japan, which is the one third of total amount of electric energy. The
lifetime of Japanese nuclear rectors was declared by the government to be extended from
30-40 years to about 60 years in 1999. In the underneath of the decision, there are some
important reasons based on the energy strategy. The first is the fact that many Japanese
nuclear reactors have been used for over thirty years as the major supplier of electric power.
Second, the construction of new nuclear energy plant is rather limited mainly due to the
problem of public acceptance. Therefore, they are expected to have longer lifetime based on
technical innovations and reliability study of the components. Thus, the technical
development to evaluate quantitatively structural material degradation and lifetime becomes
very important. One of the most serious concerns is the degradation evaluation of the reactor
pressure vessel. Since that is made of low alloy steel, called A533B in Japan. We adopt the
low alloy steel, A533B, as the first target in this work.
In this paper, we focus on the magnetic nondestructive evaluation to obtain the
quantitative correlation between the mechanical and magnetic properties of the materials.
Electromagnetic nondestructive evaluation approach is one of the most important topics in the
field of nondestructive inspection and examination, since magnetic properties were greatly
sensitive to micro structural changes in ferromagnetic materials through the microscopic
interaction between magnetic domain walls, moments and microscopic defects. Main
magnetic inspection technologies suggested recently are magnetic hysteresis measurement.
136 T. Sukegawa and M. Uesaka /Magnetic Nondestructive Evaluation of Ferromagnetic Steels
Barkhausen noise (BHN) analysis [1] and SQUID sensor [2]. All the approaches could be
promising for detailed description of degraded materials. However, physical models on the
correlation haven't been sufficiently established for practical poly-crystals yet. In the previous
work [3,4], we performed the measurement of magnetic properties that were fatigue degraded
by the cyclic load using the rod samples whose size is based on Japanese Industrial Standard
[5]. Magnetic hysteresis curves changed depending on the loading conditions. The changes in
the coercive force, residual magnetic flux density and permeability were obtained as a
function of the magnitude of residual strain. In this work, we also performed the measurement
of magnetic properties for low alloy steel that were degraded by the tensile and cyclic loads
using the plate samples so that the surface leakage magnetic filed scanning is available. The
first is to reconfirm the previous results in the correlation between the magnetic and
mechanical measurements. For the second purpose, we performed the sensitivity analysis of
magnetic property changes to the mechanical degradation.
2. Experimental method
The size of the plate A533B sample, the condition of heat treatment and its chemical
components are described in Fig. 1.
The samples were cut off from a single ingot and machined. Its thickness of 2mm 1 is
determined so that our mechanical loading machine can make plastic deformation. . Figure 2
shows the B-H curve measurement setup. The copper wire is wound around the yoke for
excitation, which produces magnetic filed. H. in the sample. To pick up the lime derivation of
T. Sukegawa and M. Uesaka / Magnetic Nondestructive Evaluation of Ferromagnetic Steels 137
magnetic flux density, dB/dt, and the other coil wound directly around the sample. A Hall
sensor is also attached to the sample in order to measure the magnetic flux density, B. The
amplitude of the applied alternating current for measurement rose quickly up to the specific
level to avoid temperature rise. After the measurement was finished, the current decreased
gradually so that the sample was demagnetized. Magnetic hysteresis loops were obtained by a
personal computer. The frequency of the alternating current was relatively low so as to avoid
the effect of eddy current.
Zero-tension sinusoidal cyclic load was applied with the amplitude of 450 to 510 MPa
where the yield is about 520MPa. The minimum amplitude is 30 MPa to avoid unexpected
out-of- plain bending.
3. Experimental result
This load conditions for the tensile test in shown Fig.3 are described in Table 1. We
investigated residual strain from 0 % to 7.5 %. We had carried out the first test with the larger
plate samples (Length: 290mm. Thickness: 5mm1) [3,4]. The previous loading condition is
given in Table 2 as the reference.
The difference of the stress-strain relation between the current and previous tests
attributes to the difference of ingot, namely manufacturing process. We calculated the strain
from the elongation of the 50mm-long central longitudinal line on each samples.
Table 3 shows the loading conditions for the cyclic test. These material properties were
the same as those for the tensile test in shown Fig. 1.
0 0 1.80x10-' 1.0
H
529 0.38 4.42xlO 2.46
H
534 1.20 4.95xlO 2.75
620 2.10 4.85xlOH 2.69
642 3.70 4.78X10' 1 2.65
140 T. Sukegawa and M. Uesaka /Magnetic Nondestructive Evaluation of Ferromagnetic Steels
SlOMPa
dB/dH at M = 0
50 Cy.
dB/dH at H = 500
dB/dH at H = -425
0 3000
Field intensity H (A/m)
Fig. 5 Changes of magnetic hysteresis loops
Fig. 6 The sensitive analysis for 510 MPa
for SlOMPa cyclic load. cyclic load.
The change of magnetic hysistesis loops in the cyclic load is shown in Fig. 5. A-26
sample is for SlOMPa that is close to the yield stress. Changes of magnetic hysteresis loops
are fast with respect to the cycle. When we compare it with the result for the tensile load case
as shown in Fig.3, the same tendency is observed. Namely, the shoulder of the loops declines
and Br decreases for early degradation, while the total loops become oblique and H c increases
for heavy degradation. In order to analyze the reason of the similarity, we put the results of
leakage flux distribution for the same sample measured by the GaAs Hall sensor. It is clear
that many small Luders band are generated for the early degradation, while the large Luders
bands are observed for the tensile load [7J.
Thus, the declination of the shoulder and
SOOMPa
Br decrease can attribute to the generation
of the Luders bands. Figure 6 shows the Virgin
results of the sensitivity analysis of
relative magnetic property changes for
510Mpa. Relative Hc, Br, dB/dH at H= 0,
-425, 500 and -1000 are plotted as a
function of the cycle. It is clear that dB/dH
at H= 0 is the most sensitive value for the
early degradation even in case of the
cyclic load.
We also performed the same test for
the cyclic load with the lower amplitude
(SOOMPa) as shown in Fig. 7. The results 50 TV.
of the dB/dH and sensitivity analysis for
amplitude of SOOMPa are shown in Figs. 8 .KKK) -1500 0 ]5(Kl ?(
and 9. The sensitivity analysis clearly Field Inlensily H ( A m )
indicates again that the dB/dH at H=0 is Fig 7 Changes of magnetic hysteresis loops
the most sensitive index for the early for 500MPa cvclic load.
deizradation due to mechanical loads.
T. Sukegawa and M. Uesaka / Magnetic Nondestructive Evaluation of Ferromagnetic Steels 141
io 5
Figures 10 and 11 also show magnetic hysteresis loops and the sensitivity analysis on the
magnetic properties for 480 MPa. In the figures, we don't observe any clear change of H c and
Br, until H)3 cycles. On the contrary, dB/dH at H=0 is increasing gradually, and turns rapidly
rising from 1()4 cycles. It is clear that many small Luders bands are generated rapidly from
1C)4 cycles. Plotted values of dB/dH at H=0 are the average values over 10 time
measurements. These additional results also support our finding that the dB/dH at H=0 is
more sensitive index than the c-parameter for the early degradation due to mechanical loads.
The physical basis on the changes of the dB/dH and the c-parameter could be the same. That
is, the coefficient of magnetic susceptibility, c, is a parameter that contains the synthetic
information dislocation [9]. The only difference depends on the location on the B-H curve
where the two parameters are evaluated.
4. Future work
The next work is to evaluate changes of magnetic property due to mechanical load for
stainless steels and to obtain again the relation between the magnetic and mechanical
properties. In nuclear reactors, stainless steels are used in many parts of structural components
such as mechanical supports. In the first step, we are going to adopt the 18-8 steel called SUS
304 in Japanese Industrial Standard. 18-8 stainless steel is one of nonmagnetic steels and is
also austenite type.
5. Conclusion
We performed the measurement of the magnetic properties of the low alloy steel samples
used as a pressure vessel of nuclear reactors. Magnetic hysteresis curves change depending on
the load condition. We have got the clear correlation between the magnetic properties and the
magnitude of residual strain. We also carried out the sensitivity analysis of magnetic property
changes for the tensile and cyclic load. Considering only our measurement, the dB/dH at H=0
was more sensitive index above the c-parameter for early degradation due to the mechanical
load.
References
[1] D.G. Hwang and H.C Kim: J. Phys., D: Appl. Phys.,Vol. 21, (1988), pp.1807-1813.
[2] J. Banchet, J. Jouglar, P.L. Voillermoz and H. Weinstock: IEEE Trans.on Appl. Superconductivity, Vol. 5
(1995), pp.2486-2489.
[3] A.Gilanyi, K.Morisita, T.Sukegawa, M.Uesaka and K. Miya: Fusion Engineering and Design 42. (1998),
pp.485-491.
[4] K.Morisita, A.Gilanyi, T.Sukegawa, M.Uesaka and K.Miya: Journal of Nuclear Materials 258-263.
(1998), pp.1946-1952.
[5] Japanese Standards Association, JIS Handbook, Iron and Steel, ISBN-542-12601-3, C3050, (1991).
Z-2201,pp.l21-127.
[6] M. Ueasaka T. Sukegawa, K. Miya, K.Yamada, S. Toyooka, N. Kasai, A.Chiba S. Takahasbi,
J. Echigoya, K. Morishita, K. Ara, N. Ebine, Y. Isoba: Proe. of ENDE' 97, Electromagnetic
Nondestructive Evaluation (II), IDS Press (1998), pp.39-48.
[7] Koji Yamada, Shin-ichi Shoji and Yoshihiro Isobe: International Journal of Applied Electromagnetic and
Mechanics 11, IDS Press (2000) pp.27-38.
[8] N. Kasai, S. Nakayama, Y. Hatsukade, M. Uesaka: Journal of Japan Society of Applied
Electromagnetics and Mechanics, Vol. 8, (2000), pp.16-22.
[9] S. Takahashi, J. Echigoya and /. Motoki: J. of Appl. Phys.. Vol. 87(2). (2000). pp.805-813
Electromagnetic Nondestructive Evaluation (VI) 143
F. Kojima et al. (Eds.)
1OS Press, 2002
1. Introduction
In recent years, non-destructive testing has been playing an important role for safety
and maintenance of any structures. By the conventional non-destructive testing, the cracks
of the structures can be detected by using X-ray, ultrasonic, eddy current and so on. We
have been doing the research on the fatigue evaluation of magnetic materials qualitatively
and quantitatively by measuring Barkhausen noise [1-4]. Barkhausen noise is a signal
generated by the movement of the 180-degree magnetic domain inside the magnetic
materials, and the dislocation influences this magnetic domain movement. Therefore, it
can be possible to evaluate the fatigue level by measuring Barkhausen noise even before a
crack appears. In the previous research, we made clear that the power of Barkhausen
noise from INCONLE plates changed a little after applying the thermal fatigue [1]. It was
very difficult to make clear the relationship between Barkhausen noise and the fatigue
level quantitatively and qualitatively because of the complication of the signals.
Barkhausen noise is a very complicated non-linear signal, so we examined the chaotic
behavior of Barkhausen noise to evaluate the relationships between them because the
chaos theory can treat a non-linear signal. First, Barkhausen noises from non-oriented,
oriented and double-oriented silicon steel sheets were examined to reveal the basic chaotic
behavior of Barkhausen noises [2, 3]. And then we made several sensors suitable for
Barkhausen noise [2, 3]. In these papers, we concluded that Barkhausen noise from
magnetic materials could be classified into the chaos. Moreover, we found that the
constructed attractor made from Barkhausen noise changed depending on the angle of the
scratch on the silicon steel sheet and described that the change of the magnetic domain can
be found by the shape of the chaotic attractor [4]. In this paper, Barkhausen noises from
structural steel plates, SS400 are measured and discussed the possibility of the fatigue
evaluation by chaotic attractors from Barkhausen noise.
144 Y. Tsuchicla et al. / Fatigue Evaluation of Magnetic Materials
150 Inclination
time (s) Frequency (Hz) angle (degree)
Fig. 1 Barkhausen noise under alternating magnetic field. Fig. 2 Spectrum of Barkhausen noise
under alternating magnetic field.
C(m)
--
---; ; :-! : ; 1 : -
-1 -0.5 0.5 I
The Barkhausen noise from the non-oriented silicon steel sheet under the alternating
magnetic field is shown in Fig. 1 [1-4]. The controlled sinusoidal wave of the exciting
coils is also drawn in this figure. The Barkhausen noise under the alternating magnetic
field is mainly generated where the variation of the excitation wave is large as shown in
Fig. 1. We examined if Barkhausen noises had a chaotic behavior by the following four
points [5];
(1) The complexity of Barkhausen noise itself in the time domain. Chaotic signals are
very complicated in the time domain.
(2) The existence of the specific discrete peak of the spectra after Fourier transform.
Chaotic signals don't have any specific peaks in the frequency domain.
(3) The value of the correlation function. The correlation function of chaotic signals
becomes zero immediately with the elapse of the time.
(4) The self-similarity of the constructed attractors from Barkhausen noises. The self-
similarity of the attractor is one of the factors as a chaotic behavior.
Fig. 2 shows the spectra measured from the non-oriented silicon steel sheet under the
alternating magnetic filed. It shows the relationship between the frequencies and the
inclination angles of the alternating magnetic field. The spectra don't have any specific
discrete peaks, so this is one factor whether Barkhausen noise can be categorized into the
chaos. Fig. 3 shows the correlation function of Barkhausen noise under the rotational
magnetic field. By this figure, the correlation function becomes zero immediately, so this
is also one factor that Barkhausen noise can be categorized into the chaos. Fig. 4 shows
Y. Tsuchida et at. / Fatigue Evaluation of Magnetic Materials 145
the constructed attractors from Barkhausen noise under rotational magnetic field. The
embedding method was used to make the attractor from time series Barkhausen noise [2, 3,
5]. It is obvious that the constructed attractor draws the self-similar elliptic orbits. This is
also one of the factors for the chaos. Considering these four factors, we concluded that
Barkhausen noise from magnetic materials could be classified into the chaos [2, 3].
The shape of the steel plate for the fatigue evaluation is shown in Fig. 5. This is the
shape to concentrate the fatigue on the center of the plates. Barkhausen noises are
measured at the one point on the center of the plates as shown in Fig. 5. Fig. 6 shows the
measurement system for the fatigue evaluation. First, the computer generates the signals
to excite the specimens. The Barkhausen noise is measured by the computer from the
pick-up sensor attached on the specimen through the low noise wideband preamplifier and
the digital oscilloscope. The low noise wideband preamplifier has a high pass filter, so
this filter eliminates the frequency components around the excitation wave from the
measured signal. Fig. 7 shows the sensor to pick up the Barkhausen noises. This sensor
was selected after the several sensors were examined to measure Barkhausen noise [2, 3].
Fig. 8 shows the photo of the bending stress machine. The antiplane bending stress was
applied to the specimens with the amplitude, 320 MPa and 30 Hz cycles. The S-N curve
of the used specimen is shown in Fig. 9. It is supposed that the used specimen might be
broken around 60,000 times of applying the bending stress by this figure.
90
unit: mm
Fig. 7 Sensor to measure Barkhausen noise. Fig. 8 Bending stress machine
Figs. 10 show the measured Barkhausen noises from the SS400 plates in the case that
the frequency of the exciting coils is 10 Hz. Fig. 10(a) shows the one before the bending
fatigue stress test and Fig. 10(b) shows the one from the bending stressed plate. Here, the
amplitude of the bending stress is 320 MPa and the number of the stress cycles is 20,000
times. Fig. 10(c) and Fig. 10(d) show the Barkhausen noises measured after 40,000-time
and 60,000-time bending stress. From Fig. 9, the specimen was supposed to be broken
around 60,000 times. Actually, some very small cracks could be recognized around Point-
A and Point-B shown in Fig. 5 by using a magnifying glass. However, the visual crack
could not be found around the measured point. From Fig. 10(a) to (d), it is difficult to
discuss the change of the signals depending on the applied bending stress cycles. Figs. 11
show the constructed attractors from the Barkhausen noises shown in Figs 10. Here, the
parameter of the time delay for the embedding method to draw the attractor is selected as
x=l [2, 3, 5]. The shapes of the attractors are changed depending on the cycle number of
the applied bending stress, but it is difficult to say that each attractor has the self-similar
shape. So, the chaotic behavior could not be recognized in the case that the frequency of
the exciting coils is 10 Hz. Figs. 12 show the measured Barkhausen noise in the case that
the frequency of the exciting coils is 100 Hz. It is also difficult to discuss the change of
the signals depending on the applied bending stress by Figs. 12 (a), (b), (c) and (d).
However, each attractor constructed from Figs. 12 has the self-similarity as shown in Figs.
13. These are the characteristics of the chaotic behavior.
Y. Tsuchida et al. /Fatigue Evaluation of Magnetic Materials
Fig. 10 Barkhausen noise before and after fatigue test (frequency of exciting coils : 10 Hz).
Fig. 11 Constructed attractors of Barkhausen noise from the stressed steel plates,
(frequency of exciting coils : 10 Hz, time delay of embedding method T=l)
148 Y. Tsuchida et al. /Fatigue Evaluation of Magnetic Materials
Fig. 12 Barkhausen noise before and after fatigue test (frequency of exciting coils : 100 Hz).
Fig. 13 Constructed attractors of Barkhausen noise from the stressed steel plates,
(frequency of exciting coils : 10 Hz, time delay of embedding method T=l)
Y. Tsuchida et al. / Fatigue Evaluation of Magnetic Materials 149
Compared Figs. 10 with Figs. 12, the similar Barkhausen noises are measured on the
different frequency of the exciting coils but the shapes of the constructed attractors are
different as shown in Figs. 11 and Figs. 13. It is required to find the proper frequency
range of the exciting to use the chaotic behavior of Barkhausen noise for the fatigue
evaluation. Now, it can be said that there are some relationships between the cycle
numbers of the applied bending stress and the constructed attractors by considering Figs.
13. As mentioned before, any cracks could not be found around the measured point, so
the difference of the shapes of the constructed attractors could be related with the change
of the magnetic domain movement due to the dislocation. The specified chaotic
characteristics change should be found to use it for the fatigue evaluation before a crack
appears.
5. Conclusions
References
[1] M. Enokizono and A. Nishimizu, "Effect of thermal fatigue on Barkhausen noise,"/ of M. M. M., 133,
pp.599-601, 1994.
[2] M. Enokizono, T. Todaka and Y. Yoshitomi, "Chaotic behavior of Barkhausen noise induced in silicon
steel sheets," IEEE Trans., on Magnetics, Vol. 35, No. 5, pp. 5421-5423, September 1999
[3] M. Enokizono, T. Todaka and Y. Yoshitomi, "Chaotic phenomena of rotational Barkhausen noise,"/, of
M. M. M., 215-216, pp. 43-45, 2000.
[4] Y. Tsuchida, T. Ando and M. Enokizono, "Chaotic behavior of rotational Barkhausen noises for fatigue
evaluation," JSAEM Studies in Applied Electromagnetics and Mechanics, 9, pp. 647-648, 2001
[5] P. Berge, Y. Pomeau and Ch. Vidal, "Order within Chaos," John Wiley & Sons, 1987.
150 Electromagnetic Nondestructive Evaluation (VI)
F. Kojima el al. (Eds.)
IOS Press. 2002
Abstract. The well known analytical solutions for standard eddy current
problems have been applied to describe non-destructive testing (ndt)
experiments on surface-hardened cast iron samples by measuring the
incremental permeability. The agreement obtained between the measured
and calculated incremental permeability profile curves allows the conclusion
that measured curves can be reliably simulated and discussed in terms of the
governing influencing parameters.
Following the definition in physics, the measurement rules require a well defined
magnetization of the material of interest. A magnetic field HA is applied which - according
to the hysteresis of the material and supposed the field is producing a homogeneous
magnetization above the direction of HA - defines a magnetic flux density BA- A sinusoidal
alternating magnetic field H(d(t) is superimposed to the steady state value H A :
The time variable is t, to is the angular frequency and AH is the so-called incremental
magnetic field. The value of AH has to be small compared with the coercivity Hc of the
material and by a reasonable approach, AH is chosen to be AH < Hc/2. The reason for this
-4. Yashan and G. Dobmann /Measurements and Modeling of Incremental Permeability 151
limitation is the objective to excite - by use of the superimposed Hm(t) - only additional
reversible magnetization changes, i.e., the macroscopic value BA in the hysteresis has not to
be changed. By thinking in a potential model describing the magnetization of the material
in terms of a magnetic domain structure in which domains are separated by Bloch walls, the
value BA fixes the Bloch walls in potential troughs and by application of the oscillating
magnetic field H w (t) the Bloch walls harmonically oscillate in them. The AH has to be so
small that irreversible Bloch wall jumps, i.e. Barkhausen events, do not occur, and the walls
are trapped in the troughs. The Bloch wall oscillations supply with an additional alternating
magnetization AB which - as long as AH is small enough - is also purely sinusoidal
without any excitation of higher harmonics. The incremental permeability value JIA is
defined by the equation (2):
x
AB = AH, i.e., (2)
According to this definition the incremental permeability can be measured only in special
devices (hystrometers) suitable for hysteresis measurements using special shaped specimen
with an encircling coil for magnetic flux measurement (sphere, ellipse, cylinder), where a
steady state magnetizing field HA can be guaranteed and where the Hw(t) is superimposed
with sufficient accuracy in the same direction as HA (Fig. 1). According to the definition it
is obvious that the function (IA(HA) is measured only at discrete values HA. Fig. 2
documents the so determined incremental permeability function with series of 'inner loops'
within the macroscopic hysteresis.
~ AU/A/
Such a procedure is far from non-destructive testing (ndt). Ndt is normally asking for a
scanning of local inspection regions of large components, and not of small specimen which
can homogeneously magnetized. We have in practice a relative inhomogeneous
magnetization and we have not the time to macroscopically select discrete steady state
magnetic field values.
Our institute has introduced into practice a ndt-approach [3], based on eddy current testing
and the use of an eddy current pick-up coil in addition to a local magnetic yoke
magnetization. However, the accuracy of this approach in comparison with the exact
theoretical model according to the definition neither was scientifically and basically
investigated in detail nor validated by theoretical research.
By using the IZFP ndt approach the component under test is locally magnetized applying an
u-shaped yoke magnetization. The magnetization is controlled by a sinusoidal electric
current driving the excitation coil of the yoke, where the frequency for the magnetic
oscillations is in the range between some Hz and some hundred Hz, in order to allow - by
frequency adjustment - to select a certain field penetration of interest. The superimposed
incremental field is produced by a pick-up eddy current coil with the advantage of the
spatial resolution of that coil (Fig. 3). This, in practice, can be a pick-up air-coil or a ferrite-
core-coil, but also a small u-shaped magnetic yoke transducer. The differences of these type
of magnetizers - observing the definition - is their field inhomogeneity and direction
compared with the macroscopic yoke magnetization. The frequency of the incremental
field, i.e., the eddy current frequency has to be much higher than the hysteresis frequency; a
factor of 100 is appropriate.
A. Yashan and G. Dobmann / Measurements and Modeling of Incremental Permeability 153
a)
b)
In the ndt approach not the inclination AB / AH for each inner loop but the eddy current coil
impedance is measured. Applying a current source as HF-generator for exciting the eddy
current coil, an impressed magnetic field AH is assumed., i.e. an incremental field which is
independent from electrical and magnetic loads. In that case, the impedance meter output is
proportional to AB and therefore proportional to ^A-
The eddy current impedance can continuously be measured during a hysteresis cycle and
incremental permeability profile curves J^A(H) can be documented, as long as the magnetic
field is also measured by use of a Hall-element. Fig. 4 shows such profile curves
(incremental permeability versus H in a x-y-presentation, the x-axis is the H-axis) measured
at laser-hardened surfaces of cast iron specimen with different hardening depths (0.5 mm
and 1.0 mm). For smaller hardening depth values a typical double peak structure was
observed which indicates the layered structure with the hardened martensitic surface layer
covering the non-heat-affected mechanically soft core. Basic investigations document the
fact that the positions of the peaks correlate with the coercivity of the hardened layer and
the coercivity of the soft unhardened core. For a given eddy current frequency - in this case
750 Hz - the soft unhardened material (the two inner peaks of the observed curves) can
reliably be detected only up to a certain hardening depth in between 1 mm and 2 mm. Using
154 A. Yashan and G. Dobmann /Measurements and Modeling of Incremental Permeability
the ratio of the peak-amplitudes (amplitude of the soft peak / amplitude of the hard peak) a
calibration can be obtained for a ndt determination of hardening depth.
In order to understand the structure and amplitude contrast in profile curves like those
shown in Figure 4, semi-analytical modeling based on Maxwell equations was performed.
The solution algorithm was implemented taking into account the well-known approach of
Cheng, Dodd and Deeds [4] developed many years ago, which is a valuable tool to
characterize multi-layered media using eddy current testing with a pick-up coil, so far the
layer properties (thickness, electrical conductivity a, magnetic permeability u) and the coil
parameters are known.
In order to apply the solution [4], the specimen geometry was simplified from the cylinder
(Fig. 3a) to the infinite half space (Fig. 5). The electrical conductivity and magnetic
permeability values are assumed to be linear, isotropic and homogeneous in each layer. The
chosen electrical conductivities are (72 = 4.0 MS/m for the soft core andCTI= 3.2 MS/m for
the hardened layer. The used values of the magnetic permeability of both materials are
functions of H A as shown in Fig. 6, based on the measurements of the incremental
permeability for homogeneous cylinders as indicated in Fig. 1 and 2. The coil inner and
outer diameters are 4 mm and 9 mm respectively.
A. Yashan and G. Dobmann /Measurements and Modeling of Incremental Permeability 155
Figure 5: Coil with rectangular cross section of the winding above the conducting and
magnetic permeable half space - the simplified problem geometry for the
solution of Cheng, Dodd and Deeds [4].
Figure 6: The magnetic permeability curves as functions of the applied field HA used
for the calculation, ul (delta), or UI A , is the measured incremental
permeability of a soft steel with the coercive force Hc = 650 A/m.
u2(delta), or \n2&, is the approximated curve derived from the measured
incremental permeability of a hardened steel with the coercive force Hc =
1700 A/m by shifting along the axis H to have Hc = 3000 A/m.
The results of the modeling for the thickness of the hardened layer 0.5 mm, 1.0 mm, 2.0
mm and 3.0 mm are shown in Fig. 7 (right) and compared with the corresponding
measurement results in Fig. 7 (left). The measured curves show the most significant
projection of the complex coil impedance at 750 Hz vs. magnetizing field H which is
alternating with the frequency 0.8 Hz between -6000 A/m and 6000 A/m. The calculated
curves show the imaginary part of the complex coil impedance at 750 Hz vs. magnetizing
field H which is alternating between -6000 A/m and 6000 A/m and affecting the
incremental permeability according the procedure of Fig. 6.
156 A. Yashan and G. Dohmann / Measurements and Modeling of Incremental Permeability
4. Conclusion
The chosen values of the material characteristics (electrical conductivities CTI, (72 and
magnetic permeability values UIA(H A ), U2A(HA)) result in the qualitatively good agreement
between the measured and calculated incremental permeability profile curves. We can
conclude that the measured impedance curves (correlating with the incremental
permeability curves) can be reliably simulated using the analytical solution for the
simplified eddy current approach and discussed in terms of the governing influencing
parameters.
References
Abstract Guaranteeing the security of the steam generator tubes of the pressurized water
reactor (PWR) in the nuclear power plants is a key factor in their safe exploitation. Accordingly,
the eddy current testing is used for the reason of high accuracy and high speed because the
test must be done in a limited period We developed a multi probe that can detect cracks with
distinction between axial and circumferential cracks. The probe is composed with two kinds
of excitation coils of axial and circumferential windings and 32 pickup coils. The excitation
coil of this probe is able to generate uniform eddy current distribution on a wide area
compared with the detection area of the pickup coils. Therefore the probe can ignore the
influence of the shape of excitation current source on detecting cracks, and numerical
analysis for this probe is handled simply by devising boundary condition. From
measurement results, the probe is confirmed to be able to measure direction of cracks with
high sensitivity and high speed.
1. Introduction
It is one of the most important keys to make sure of the security of the steam
generator tubes in the nuclear power plant of the pressurized water reactor (PWR)
type. Eddy current testing (ECT) is applied for this testing [l]. Cracks in actual tubes
are mostly outer axial or circumferential cracks and the crack shape is shallow and
narrow. It is demanded on ECT to having abilities high accuracy and high speed
because the test must be done in a limited period. Tb ensure the reliance of steam
generator tubes, it is important to judge shape of cracks detected by the testing. We
developed an ECT multi-probe that can distinguish between axial and circumferential
direction cracks and also with high sensitivity and rapidity testing.
We had developed a rotating uniform eddy current ECT probe[2] for aids of numerical analysis.
The probe could distinguish between axial and circumferential direction cracks and also with high
sensitivity. But detecting speed of the probe is late because of rotating type. So we developed a novel
multi type ECT with object of rapidity testing. The probe has fundamentally same characteristicsthe
of the developed rotating uniform eddy current probe. The probe is composed with two kinds of
excitation coils of axial and circumferential windings and 32 pickup coils. The excitation coils are
able to generate uniform eddy current distribution on a wide area compared with the detection area of
the pickup coils. The excitation coil for detecting axial cracks is a coil of solenoidal shape. We call
162 M. Hashimoto et al. /Development of ECTMulti-Probe
the coil by the name of the axial winding excitation coil. The excitation coils for detecting
circumferential cracks are two coils wrapped with a fine wire in axial direction on a cylinder divided
into two domains. We call the coils by the name of the circumferential winding excitation coil The
excitation coils suitable for the uniform eddy current distribution in the neighborhood of the pickup
coils are designed by numerical analysis using the finite element method [1]. Fig.l shows eddy
current distributions in the tube by two kinds of excitation coils. From this figure the eddy current
distribution of axial direction is existed to a wide area by range of db 6mm for axial direction on
center of the pickup coils. And one of circumferential direction is existed to a wide area by range of
22mm for circumferential direction. The two excitation coils for detecting circumferential cracks are
installed in the form that is each sliding 90 degrees in the probe in order to test whole circumference.
The excitation coils of this probe are the ability to generate uniform eddy current distribution in tubes
on large enough area compared with the detection area of the pickup coils. Therefore the probe has
the advantageous feature that specific characteristics detecting cracks is not affected by the relation
between the position of the excitation coils and the cracks, and numerical analysis for this probe is
handled simply by devising boundary condition.
It was confirmed that a magnetic field of perpendicular direction to the surface of the tube
represents the crack shapes smartly from the numerical results. Therefore the pickup coils are
installed on a parallel direction to the tube surface. The pickup coils are installed 32 coils of square
type at one line in a center of the excitation coils. The pickup coils detect disturbances of eddy
current distribution due to cracks. The structure and an appearance of the probe are shown in Fig.2
and 3, respectively.
These pickup coils detect distinguish between axial and circumferential direction of the cracks by
changing the excitation coils. Fig.4 shows a control circuit of this testing. A computer processes
controls of changing signals for the excitation coils and the pickup coils, and collections of detecting
data from the pickup coils. Muhiprexer circuits for changing the exciter coil and the pickup coils and
amplifier circuits are attached in the probe. Ability of sampling data is 48000datas/sec of 16 bits on
the measurement system. When the sampling pitch for axial direction is 1.5mm, the measuring speed
becomes 375mm/sec. This is enough speed on the ECT measurement on two dimensions.
We performed experiment using the specimens of JSAEM model [3,4]. Fig.5 shows measurement
result detected an outer axial crack of 40% depth by a mechanical stage for precision moving. In this
figure two peaks of plus and minus are detected due to the existence of the crack. It is confirmed that
the distance from peak to peak of the detecting signal agrees with the crack length. From
measurement results, the probe is confirmed to be able to measure direction of cracks, length of
cracks from peak to peak of signal and to have abilities high sensitivity.
Because the number of coils are limited, the detecting sensitivity of multi-probe is not constant by
position of crack. Fig.6 shows the characteristic of detecting crack by each pickup coil. From this
figure, if the crack is between pickup coils, the drop of sensitivity detected by the multi probe using
32 pickup coils is 15%. Therefore the drop of the detecting sensitivity has done underestimate for the
detecting cracks.
-3.66E-01 3.92E-01
Fig.6 Detecting characteristic for 40% crack
Fig.5 Measurement result for OD 40% crack by multi of each pick-up coils
probe
On this account we examined an interpolate method for measurement data by the multi probe.
Fig.7 shows signal strength distribution of detected result of an outer 40% axial crack for detailed
detecting by the rotational uniform eddy current probe. Figure (a) shows measuring lines on
measurement result of two dimensions display. Figure (b) shows signal amplitude distributions on
each line shown by Figure (a). Figure (c) shows normalized signal distributions which are calculated
their peak value to unify into 1 for the each line. The normalized signals are in good agreement. We
164 M. Hashimoto el al. /Development of ECTMulti-Probe
confirmed that distributions of signal amplitude on each crack of differential depths furthermore. The
result is shown in Rg. 8. The distributions of the differential cracks are in good agreement, too. This
means that it is an important characteristic of the uniform eddy current probe. Therefore we can
handle easily the interpolation of measuring signals by 32 pickup coils using this characteristic.
Rgure (a), (b) and (c) of Fig.9 show a measurement result by the multi probe for 40% outer crack, an
interpolation result of one and a measurement result by the rotational probe compared with
interpolation result of one, respectively. We can understand that the interpolation method is useful for
improvement for the inferior measurement data by the multi probe from this figure.
Fig. 10 shows measurement results of axial and circumference cracks on high speed scanning over
300mm/sec by manual scanning. In this figure SN ratios are 8 for axial crack and 5 for
circumferential crack. We could confirm that the developed probe has ability with high speed testing.
4. Summary
We have presented a novel multi probe that is able to detect cracks with distinction between axial
and circumferential directions and to generate uniformly distributed eddy current through the
excitation coils in the neighborhood of pickup coils. We showed a special feature of the probe by
experiments. We developed the interpolation method using the probe, and confirmed the
interpolation method was useful. The results showed that the developed probe has detection
capability with distinction between axial and circumferential cracks with high sensitivity and high
speed testing.
References
[1] T. Takagi, M. Uesaka., and K Miya, "Electromagnetic Non-destructive Evaluation", IOS Press,
pp9-16, 1997
[2] Mitsuo Hashimoto and Daigo Kosaka: Development of Rotation ECT Probe Detecting for Axial and
Circumferential Crack using Uniform Eddy Current Exciting Coils "Electromagnetic Non-destructive
EvaluationCV)", IOS Press, pp.242-247, 2001
[3] T. Takagi, M. Hashimoto, H Fukutomi, M. Kurokawa, K Miya, H Tsuboi, .M Tanaka, J. Tani, T.
Serizawa, Y Harada, E. Okanoi and R. Murakami "Benchmark Models of Eddy Current testing for
Steam Generator Tube: Experiment and Numerical Analysis", International Journal of Applied
Electromagnet tics in Materials, 5, pp. 149-162, 1994
[4] Takagi, T. and K. Miya., "ECT Round-robin Test for Steam Generator Tubes," Journal JSAEM, 8, pp. 121-129,2000
166 Electromagnetic Nondestructive Evaluation (VI)
F. Kojima el a I. fEds.)
1OS Press, 2002
Abstract. The purpose of this paper is to present new and improved version of the sensor for
Inconel600 tubes testing. The sensor has a high sensitivity to shallow flaws, and good spatial
resolution. Experiments with various samples and interfering structures have been carried out.
Advantages of the proposed sensor and system are presented and confirmed.
1. Introduction
In this paper a new improved version of a sensor for testing tubes made of INCONEL600 is
presented. Very good spatial resolution of the existing sensor PIPE-UN12 [1] and excellent
sensitivity of the sensor PIPE-MAT inspired the authors to develop a new sensor PipeUniBigl
("PUB1"), which offers both advantages at the same time.
First, sensitivity of the sensor is evaluated using calibration tubes of the JSAEM Round
Robin Test [2]. Next, the enhanced ability of the new sensor and Multi-frequency Excitation
and Spectrogram system to discriminate flaw signals against interferences from fasteners is
confirmed.
2. System description
In all measurements the Multi-frequency Excitation and Spectrogram (MFES) system was
used. A basic idea behind the MFES method is to use a complex signal containing selected
harmonic components as an exciting signal and a spectrogram to precise crack
characterization. The spectrogram is a two-dimensional display of the relative amplitude of the
frequency components of a signal from the search coil versus the sensor position. The system
consists of function synthesizer, power amplifier, instrumentation amplifier, AD converter, X-
Y-Z scanner and computer with interface boards. The detailed description of the system could
be found in [3].
3. Sensor description
The sensor "PUB1" is a significantly improved version of the sensor "PIPE-UNI2" [1].
Construction of the sensor is shown in Fig. 1. The sensor contains four excitation coils and a
single search coil mounted together on a core made of a high performance ferrite (MFH-9B).
The excitation coils could be connected in various ways. In a basic configuration all coils
(Ei, 2, 3, 4) are used and they are connected in series. Such arrangement allows us to detect
all kinds of flaws and achieve the same sensitivity for axially and circumferentially oriented
flaws.
T. Chady et al. /Evaluation of the New ECT Sensor 167
3. Sensor evaluation
Calibration tubes of "JSAEM round robin test" [3] were used to evaluate sensitivity and
resolution of the new sensor "PUB1". The inner diameter and the wall thickness of the test
pieces were 19.7 and 1.27 mm, respectively, the same as those of the steam generator tubes in
nuclear power plants. The measurements were done by scanning the probe axially in steps of
0.5 mm and circumferentially in steps of 1 deg. The liftoff was measured to be 0.3 mm. The
excitation signal consists of sinusoidal components having frequency from 40kHz to 220kHz
(18 components).
Table 1. Results obtained for the flaws having length 5mm (specimens NEL9837, NEL9838).
Axial flaws Circum. flaws
VPP SNR Vpp SNR
[mV] [dB] fmV] [dB]
No flaw 0.15 - 0.15 -
OF10 7.5 34 7.5 34
OF20 18.7 42 20.0 43
OF40 74.1 54 77.1 54
OF60 198.0 62 170 61
168 T. Chady et al. /Evaluation of the New ECT Sensor
Fig. 2. Spectrograms obtained for outer flaws having different depth (axial flaws - specimen
NEL9837, circumferential flaws - specimen NEL9838); flaws length 5mm.
T. Chady et at. /Evaluation of the New ECT Sensor
Fig. 3. Signals obtained for circumferential outer flaws having different depth (specimen NEL9838);
flaws length 5mm; testing frequency for which the spectrogram achieves maximum value.
Fig. 4. Signals obtained for axial outer flaws having different depth (specimen NEL9837);
flaws length 5mm; testing frequency for which the spectrogram achieves maximum value.
Fig. 5. Signals obtained for the masked specimen NEL9825. Two groups consist of seven flaws could be
recognized without using any special inverse algorithm.
170 T. Chady et al. / Evaluation of the New ECT Sensor
X[mm] X[mm]
Y[mm] X[mm]
Fig. 7. Spectrograms and signal obtained for inner axial flaw
(depth 20%, length 5mm, tube with expansion area NELOO-21)
X[mm] X [mm]
Y[mm] X[mm]
Fig. 8. Spectrograms and signal obtained for inner axial flaw (depth 20%, length 5mm, tube with expansion area
NELOO-21) in presence of external interferences (steel sleeve and copper deposit)
T. Chady et al. /Evaluation of the New ECT Sensor 171
-30 -25 -20 -16 -10 -30 -25 -20 -15 -10 10-5 0 5 10
X[mm] X[ram] Y[mm]
Y[mm] X[mm]
Fig. 9. Spectrograms and signal obtained for outer circumferential flaw
(depth 20%, length 5mm, tube NELOO-18)
Fig. 10. Spectrograms and signal obtained for outer circumferential flaw (depth 20%, length 5mm,
tube NELOO-18) in presence of external interferences (steel sleeve and magnetic deposit)
Fig. 11. Spectrograms and signal obtained for unknown stress corrosion flaw
(depth and length unknown, tube NEL93-114)
Fig. 12. Spectrograms and signal obtained for unknown stress corrosion flaw (depth and length unknown, tube
NEL93-114) in presence of external interferences (steel sleeve and magnetic deposit)
172 T. Chady et al. /Evaluation of the New ECT Sensor
Fig. 13. Spectrograms and signal obtained for unknown stress corrosion flaw
(depth and length unknown, tube with expansion area NEL95-205)
Fig. 14. Spectrograms and signal obtained for unknown stress corrosion flaw (depth and length unknown, tube
with expansion area NEL95-205) in presence of external interferences (steel sleeve and copper deposit)
Fig. 15. Spectrograms and signal obtained for limited area around flaw
(area marked by dashed line shown in Fig. 14)
Measurements obtained for the sample NEL93-114 are shown in Fig. 11 and Fig. 12. One
could observe that also in this case the peak value of spectrogram has higher amplitude and is
obtained for a lower frequency than without fastener.
Results achieved for sample NEL95-205 are shown in Fig. 13 ~ Fig. 15. Figure 14 shows
results measured for a tube with fastener (SS400 sleeve) and copper deposit (copper tape). In
this case we can observe two peaks of the signal. The peak having highest amplitude (Fig. 14)
corresponds to "crack" in copper tape (connection of tape). The real flaw causes second peak
(Fig. 15). The "false flaw" and "real flaw" could be easily recognized by using spectrograms
(see Fig. 14 and Fig. 15).
It was found, that the relation between signal amplitude and frequency at the point where the
spectrogram achieves the maximum value, could be approximated by the following function:
= a/V^ (1)
where: u - signal amplitude, /- frequency, a, ft- approximation parameters, which depend
on a flaw depth.
T. Chady et al. /Evaluation of the New ECT Sensor 173
Fig. 16. Cross-section of stress corrosion flaws in sample NEL93-114 (maximum depth of flaw 54%)
and in sample NEL95-205 (maximum depth of flaw 57%).
Estimated Real
Specimen Error
depth depth
NEL 93-114 68% 54% 14%
NEL 95-205 59% 57% 2%
In order to identify depth of the flaws the following procedure was applied. First, parameters
a,j3 are calculated for all of the calibration flaws. Next, the calibration curves were utilized
to estimate depth of unknown flaws in sample NEL93-114 and NEL95-205. Results of
estimation were compared with real data obtained a few weeks later from destructive test of
the samples (Fig. 16). The estimation error was 2% and 14% respectively (Table 2).
6. Conclusions
Presented results of measurements confirmed usability of the new sensor and the MFES
system. The sensor is sensitive to shallow flaws and provides high spatial resolution. MFES
system allows us to distinguish real flaws from interferences caused by external structures with
high probability. The algorithm for flaws' depth identification is proposed and its reliability is
confirmed using samples with quasi-natural flaws (stress-corrosion flaws).
Acknowledgements
This study was supported in part by the Research Committee on Nondestructive Evaluation
Technology by Eddy Current Testing of the Japan Society of Applied Electromagnetics and
Mechanics through a grant from 5PWR utilities and Nuclear Engineering Ltd.
The authors would like to express thanks to Mr. Junri Shimone from Nuclear Engineering
Ltd. for his valuable discussion and providing the samples and photos of natural flaws cross-
sections.
Measurements used in this paper were done by the author during his stay in Oita University.
References
[1] T. Chady, M. Enokizono, T. Todaka, Y. Tsuchida, R. Sikora: "Evaluation of the JSAEM Round Robin
Test Samples Using Multi-frequency Excitation and Spectrogram Method", The Sixth International
Workshop on Electromagnetic Nondestructive Evaluation, August, 2000, Budapest, Hungary.
[2] T. Takagi, K. Miya: "ECT round-robin test for steam generator tubes", Journal of the Japan Society of
Applied Electromagnetics and Mechanics, Vol.8, No.l pp.121-129,1999.
[3] T. Chady, M. Enokizono, T. Todaka, Y. Tsuchida, R. Sikora: "A Family of Matrix Type Sensors for
Detection of Slight Flaws in Conducting Plates", IEEE Transactions on Magnetics, Vol. MAG-35, No.5,
pp. 3655-3657, 1999.
174 Electromagnetic Nondestructive Evaluation (VI)
F. Kojima el al. (Eds.)
IOS Press. 2002
Abstract. The rotational magnetic flux type eddy current testing (RMF-ECT) probe
is applied to detection of a small outer crack on a steam generator tube with sources
of noise such as a pipe holder and sludge. From results of experiments, when a small
outer crack exists on a boundary between a pipe holder and sludge, the output signal
of this RMF-ECT probe is influenced by sources of noise. Even if the noise source
exists, this probe can detect the defect according to the kind of the noise source. In
this paper, crack detection characteristics of our RMF-ECT probe under the noisy
environment are presented.
1. Introduction
An eddy current testing technique is very useful to detect cracks on steam generator
(SG) tubes using in nuclear power plants. In order to prevent a big accident of nuclear
power plants, it is very important to develop a sensitive ECT probe for small outer cracks
on SG tubes [1"2l The authors have already developed rotational magnetic flux (RMF)
sensors to detect small reverse-side cracks on thick steel plates or thick stainless steel plates
[3 4]
" . The RMF-ECT probe, which had been reported at the ENDE-Hungary is one of them
[5]
. It is characterized by the rotational excitation. In that paper the RMF-ECT probe was
applied to detection of a small outer crack on 1.27 mm thick SG tubes without sources of
noise. This RMF-ECT probe shows high detection sensitivity and high signal to noise ratio.
The depth, length and width of it are 10 %, 5 mm, and 0.2 mm respectively. In this paper,
the RMF-ECT probe is applied to detect a small outer crack on a SG tube with sources of
noise. In this case, the output signal of this probe is occasionally influenced by sources of
noise. But, this probe can detect a small outer crack depending on the condition.
Fig. 1 shows a structural drawing of the RMF-ECT probe for a SG tube. The lift-off
that is the distance between the ferrite core and a specimen is about 0.3 mm. First, B-coils
are wound on connective parts of the ferrite core to measure excited maximum magnetic
flux densities (Bexmax and Beymax). Next, excitation coils are wound on them. Each of
exciting coils is made of 0.16 mm diameter copper wire (60 turns). And, each of B-coils is
M. Oka and M. Enokizono /Non-destructive Testing by Using the RMF Type Probe 175
made of 0.16 mm diameter copper wire (6 turns). The rotational magnetic field is generated
in a specimen by two-phase excitation currents with 90-phase difference. Five pickup coils,
which are wound on each end of the core leg, can measure the magnetic flux density of
each part near a specimen. Pickup coils on X- and Y-axis are composed of two pickup coils
(See Fig. 3).
3. Measurement system
Fig. 1 The rotational magnetic flux type ECT probe. Fig. 2 The experimental equipment.
Fig. 3 The wiring diagram of pickup coils. Fig. 4 Axes of a SG tube and assembly
drawing of sludge and a pipe holder.
176 M. Oka and M. Enokizono / Non-destructive Testing by Using the RMF Type Probe
We carried out the experiment under conditions as follows: The excitation frequency
is 60 kHz. Because of this the upper bound of the frequency characteristic of power
amplifiers is about 100 kHz. The maximum excitation magnetic flux density (Bemax) was
0.005 T at connective parts of the ferrite core. Smoothing algorithm such as the method of
moving averages was used for better performance of the RMF-ECT probe. We carried out
the experiment as follows: At first, in order to equalize Bexmax and Beymax, the RMF-ECT
probe was put on a no-crack specimen. A computer adjusted excitation voltages so that
Bexmax and Beymax might become the same. When Bexmax and Beymax become so, we set the
RMF-ECT probe on the first measurement position, and began measurement.
Fig. 5 shows two types of a pipe holder, a support plate and sludge. Fig. 5(a), Fig.
5(b), Fig. 5(c) and Fig. 5(d) show a pipe holder with a 1 mm BEC (Broached Egg Crate)
made from SUS405, a pipe holder with a round hole made from SS400, a support plate
with a round hole made from SS400 and mock sludge made from 45 permalloy
respectively.
Fig. 6 shows assembly drawing of mock sludge such as 45 permalloy and a pipe
holder with the 1 mm BEC. Fig. 7 shows the relationship between the output signal (B^ax)
and the probe position. In this case, the specimen is NEL0019 (without a crack) with
sources of noise such as the pipe holder and mock sludge. From this figure, it can be said
that the signal (Bzmax) from the probe is not influenced from sludge so much though it is
greatly influenced by the pipe holder (a 1 mm BEC). Fig. 8 and Fig. 9 show the relationship
between the output signal (Bzmax) and the probe position. In the case of Fig. 8, the specimen
is NEL0017 (with a 20 % OD axial crack). The specimen is NEL0018 (with a 20 % OD
circumferential crack). In both cases, sources of noise are same as Fig. 6. When Fig. 8 and
Fig. 9 are compared with Fig. 7, the outer crack, which is sure to exist in the circle of each
figure cannot be clearly confirmed from the distribution of Bzmax- When source of noises is
the pipe holder (a 1 mm BEC), this probe cannot detect the 20 % OD crack well.
In our experiments, higher excitation frequency than 60kHz could not be used
because of that the limitation of the frequency characteristic of the power amplifier. If a
higher excitation frequency is used, the skin depth gradually becomes small more than the
thickness of the SG tube wall. Then, it is thought that the influence of the source of noise
becomes small if the higher excitation frequency is used.
(Nocrack,NEL0019)
Fig. 7 Bzmax vs. probe position.
with a 1 mm BEC and permalloy.
B. Crack detection characteristics and Signal-to noise ratio (without sources of noise)
Fig. 10 shows the relationship between B^ax and the probe position in the case of the
specimen with a 20 % outer axial crack (NEL0017). Fig. 11 shows the same relation in the
case of the specimen with a 20 % outer circumferential crack (NEL0018). From Fig. 10 and
Fig. 11, this RMF-ECT probe clearly detected a small outer crack. The values of S/N are
15.6 dB (NEL0017) and 9.0 dB (NEL0018) respectively. The definition of signal-to-noise
ratio followed the method of JSAEM.
Fig. 12 shows assembly drawing of a pipe holder with a round hole or a tube plate.
Fig. 13 shows the relationship between B^ax and the probe position in the case of the
specimen without a crack (NEL0019). In this figure, arrows point a signal from the source
of noise. This part is corresponding to the edge of the pipe holder with a round hole. When
Fig. 14 and Fig. 15 are compared with Fig. 13, the outer crack, which is sure to exist in the
circle of each figure, can be clearly confirmed from the distribution of B^ax.
In Fig. 16, the arrow points a signal from the source of noise such as the tube plate
with a round hole. This part is corresponding to the edge of the tube plate with a round hole.
From these figures (Fig. 14, Fig. 15, Fig. 16 and Fig. 17), it can be said that this RMF-ECT
probe can clearly detect a small outer crack when the outer crack do not exist near the edge
of the source of noise And, this probe can obviously distinguish a direction of a small outer
crack.
Fig. 18 shows assembly drawing of a pipe holder with a round hole and mock sludge
such as 45 permalloy. From this figure, it can be said that the signal from the probe is not
influenced from a pipe holder and sludge so much. The influence of sources of noise
appears only in the boundary of both. The relationship between B^ax and the probe position
is shown in Fig. 20 and Fig. 21. In these figures, a small outer crack is detected. But, this
probe cannot distinguish a direction a small outer crack.
Fig. 12 Assembly drawing of a pipe holder (No crack, support plate, NEL0019)
with a round hole. Fig. 13 Bzmax vs- probe position.
0 - 10 X-posnon (mm)
(Axial crack, support plate, NEL0017) (Circumferential crack, support plate, NEL0018)
Fig. 14 Bzmax vs. probe position. Fig. 15 Bzmax vs. probe position.
M. Oka and M. Enokizono /Non-destructive Testing by Using the RMF Type Probe 179
(Axial crack, tube plate, NEL0017) (Circumferential crack, tube plate, NEL0018)
Fig. 16 Bzmax vs. probe position. Fig. 17 Banax vs. probe position.
Fig. 19 Bzn,ax vs. probe position. Fig. 20 dB^^ vs. probe position.
5. Conclusions
In this paper, we present the rotational magnetic flux type ECT probe for detection of
small outer cracks on SG tubes under the noisy environment. From results of experiments,
the following findings are obtained.
(1) We proposed the RMF- ECT probe to detect effectively small outer cracks on SG tubes
with the source of noise such as a pipe holder and sludge.
(2) If the source of noise is only a pipe holder with a round hole or a tube plate, this
RMF-ECT probe can detect a small outer crack.
(3) But, when the source of noise is a pipe holder with a 1 mm BEC, this RMF-ECT probe
cannot detect a small outer crack.
(3) The signal from this probe is not influenced from sludge (45 parmalloy) so much.
180 M. Oka and M. Enokizono /Non-destructive Testing b\ Using the RMF Type Probe
(4) It is thought that the influence of the source of noise becomes small due to the skin
effect if the higher excitation frequency is used.
References
[ 1 ] M. Kurokawa, R. Miyauchi and K. Enami, "New Eddy Current Probe for NDE of Steam Generator".
Electromagnetic Nondestructive Evaluation (III), IOS Press, Vol. 15, pp. 57-64, 1999
[2] P. Y. Joubert, D. Miller, D. Plack and E. Savin, "Multi-Detector Eddy Current Probe for the
Non-Destructive Evaluation of Steam Generator Tubes, Designed for an Imaging Approach". Electromagnetic
Nondestructive Evaluation (III), IOS Press, Vol. 15, pp. 34-44, 1999
[3] M. Oka and M. Enokizono, "Rotational Magnetic Flux Sensor with Three Axis Search Coil for
Non-Destructive Testing", Nonlinear Electromagnetic Systems, IOS Press, pp. 357-360, 1998
[4] M. Oka and M. Enokizono, "Eddy Current Inspection by Gradio-Magneto Sensor Using MI Device".
Journal of the Magnetics Society of Japan, Vol. 24, No. 4-2, pp. 859-862, 2000
[5] M. Oka and M. Enokizono, "Eddy Current Testing Using Rotational Magnetic Flux Type Probe for SG
Tube", Studies in Applied Electromagnetics and Mechanics 21. ELECTROMAGNETIC
NONDESTRUCTIVE EVALUATION (V), IOS Press, Vol. 21, pp. 234-241, 2001
Electromagnetic Nondestructive Evaluation (VI) ' '
F. Kojima et al. (Eds.)
IOS Press, 2002
Abstract: The possibility of employing a planar mesh type probe for the evaluation of
near-surface material properties as well as to inspect the presence of defects has been
investigated in this paper. The impedance of a planar type coil in proximity of any metal
surface is a complex function of many parameters including near-surface material
properties. A two-dimensional model of a planar mesh type probe has been developed for
the analytical calculation of magnetic vector potential., flux-linkage and impedance. The
impedance of the probe is used for the evaluation of the near-surface material properties.
A simple neural network model has been developed for the post-processing of output
parameters from the measured impedance data.
/. Introduction
This paper is concerned with the research required to assess and inspect near-surface
material properties using a planar type probe having mesh configuration. The impedance of
the coil can be used to detect most kind of flaws and to evaluate degradation of materials.
There are many different causes of degradation including environmental stress, corrosion and
thermal treatment etc. In some circumstances this leads to dangerous situation. For example,
material degradation of aircraft's outer surface may result in a catastrophic accident. In
industries it is a major concern to be able to assess the coating quality, strength, contamination
and hidden corrosion in a simple, versatile, effective and efficient way.
It has been demonstrated that the physical properties of a material undergo deterioration due
to fatigue and aging [1, 2] so it is necessary to be able to assess the actual state of the material.
Planar type meander coils have been used for the evaluation of near-surface properties and are
reported in [3, 4]. The use of meander type probe for the inspection of printed circuit boards
has been described in [5]. The aim is to extend the modeling technique to planar mesh type
probe and to investigate the feasibility of applying it to estimate the near-surface material
properties. A simple neural network model has been developed for the post-processing of
output parameters from the measured impedance data.
182 S.C. Mukhopadhyay et at. /Evaluation of Near-Surface Material Properties
The probe consists of two coils: exciting coil carries a high frequency current and generates
the electromagnetic field. The induced magnetic field in the surrounding material modifies the
resultant magnetic field distribution which is picked up by the sensing coil. Both coils employ
mesh type configuration as shown in Fig. 1. The impedance of a coil in proximity with a
metal surface varies in a complex way with different parameters such as permeability and
conductivity of near-surface of the material, lift-off and pitch of the coil, operating frequency
etc. and is usually used for the inspection of near-surface material properties. It is very
difficult to get a simple expression to determine the near surface material properties from the
measured impedance of the probe. For the derivation of the impedance characteristics of the
coil an analytical model has been developed. In the analytical calculation only one pitch both
along the X and Y-axis are considered. The 2-dimensional representation along the Y-axis is
shown in Fig. 2 and this corresponds to the situation that the probe is used for the inspection
of electroplated material. The conductivities and permeabilities of different sub-regions are
shown in Table 1. Another model is used to model the system along the X-axis in which the
parameters along the Y-axis are replaced by the parameters along the X-axis. The pitch along
the X-axis may or may not be similar in size depending on the configuration of the coil.
Pitch# 1
* X
Periodic
Sensing coil
Exciting coil
Axea#2 Y
Ar.
Area#4 Insulating
Area#S layer
Intermediate
layer
Metal surface
The lift-off is designated by h. Details of the analysis for meander type probes has been
described in [6]. Using the same procedure the modeling for the mesh type probe is done. The
governing equation for area#l and area#2 is given by,
cy oz
The governing equation for area#3, #4 and #5 is given by,
d2 A(y,z) d2A(y,z)
(2)
dy2 dz2
S.C. Mukhopadhyay et al. /Evaluation of Near-Surface Material Properties 183
Area*l z^h O -o
Area#2 O-ez-eh O
-o
CT2
^2
-<>"
^3
The vector potential v4(x,y,z,t) is assumed to be independent along the x-coordinate and is
function of y and z variables only. Assuming the span of the meander coil is periodic and m
being integer, we can write,
a/I = nrn (3)
The method of separation of variables is used for the solution and each part is solved
separately.
The equations of the magnetic vector potential for each and every region are given below.
The equation for magnetic vector potential for area#l is given by,
(4)
(7)
All the harmonic components are separately calculated and used in the solution.
Two boundary conditions VXH = 0 and V. = Oare applied to each intersection of the sub-
regions. The following conditions are obtained.
Ai(0, h) - A2 (0, h)
A 2 (0,0) = A 3 (0,0) (12)
A 3 (0, -dl) = A4(0, -dl)
A4 (0, -d2) = A5 (0, -d2)
and
1 8A,
2=0 M
(13)
Xf & a?
2=-(d\+d2)
Using the above boundary conditions all the equations are solved for the determination of the
unknowns, Ci to Cg.
Assuming the depth of the coil to be unit meter, the expression of flux is given bv
(14)
So the flux ^ is given by,
(15)
We know (16)
So the transfer impedance per unit meter of coil depth along the Y-axis is thus given by,
Similarly taking one coil pitch along the X-axis the expression of transfer impedance per unit
meter of coil depth Z2 can be obtained. So the total transfer impedance of the coil per unit
meter is given by the vector sum of Zi and Z2.
The transfer impedance of the probe has been calculated from the model described in the
last section for different varying parameters. Fig. 4 shows the variation of the resistive part of
the impedance as a function of conductivity of the intermediate layer for different coil pitch at
S.C. Mukhopadhyay et ai / Evaluation of Near-Surface Material Properties 185
an operating frequency of 500 kHz. It is assumed here that total surface area covered by the
probe with different pitches are kept same. It is seen from Fig. 4 that the resistive part
decreases with the increase in conductivity and the value is more for larger coil pitch. Fig. 5
shows the variation of the resistive part of the impedance with the thickness of the
intermediate layer at an operating frequency of 500 kHz. The thickness of top layer is
maintained constant at 0.25 mm. Fig. 6 shows the variation of the resistive part of the
impedance with the thickness of the top layer at an operating frequency of 500 kHz. The
thickness of intermediate layer is maintained constant at 50 urn. If the coil pitch is increased
beyond 2.1 mm the resistive part instead of increasing starts decreasing. So there is an
optimum value of the coil pitch which is to be selected for the best results.
OQT
o.oes
Coll pitch = 2 1
o,oe
CoM pitch = 2 1
Fig. 5 Variation of resistive part of impedance with the thickness of intermediate layer
Fig. 6 Variation of resistive part of impedance with the thickness of top layer
186 S.C. Mukhopadhyay etui. /Evaluation of Near-Surface Material Properties
The experimental set-up is shown in Fig. 7. The probe is placed on the electroplated
material and the transfer impedance (i.e., the ratio of the voltage of the sensing coil to the
current of the exciting coil) is measured with the help of Hewlet Packard make impedance
analyzer HP 4194A.
J Impedance Analyzer j
HP4194A
Base
In order to verify the performance of the probe the electroplated material is simulated in the
laboratory and by placing the probe on that the impedance has been measured. Table 2 shows
the impedance data of the meander probe at 250 kHz and 500 kHz for a base of 5 mm copper
plate and the top layer of 0.25 mm insulating layer. The intermediate layer is 0 05 mm thick
tungsten and molybdenum material.
Table 3 shows the impedance data of meander probe at 250 kHz and 500 kHz for a base of 5
mm copper plate and the top layer of 0.25 mm insulating layer with variable thickness of
tungsten material.
Table 3 Results for variable thickness
The transfer impedance of the probe depends on many parameters including the surface
properties of the material in a complex way and there is no simple method to evaluate them
from the impedance data. In order to make the process fast and simple, a grid system has been
developed by plotting imaginary part of the impedance versus its real part for the parameters
of interest at one particular frequency. A conductivity-thickness measurement grid is shown in
Fig. 8 corresponds to an operating frequency of 1 MHz. The grid can be made for any two
variable parameters for a 2-dimensional grid or any 3-variable parameters for a 3-dimensional
grid. The data used for the generation of the grid system are obtained from the model. Each
node of the grid system in Fig. 8 has fixed conductivity and thickness. The measured
impedance data is plotted on the grid as shown by O in Fig. 8 and the surface properties are
estimated by interpolation technique. The measured impedance data doesn't match very well
with the results from analytical model so a correction factor is introduced while the measured
data is plotted on the grid system.
1 .Tl
f 1 .7
2 SET /
7SE7 conductivity (S/m)
Instead of using the grid system the surface properties can be estimated with the help of
neural network model. A schematic representation of a simple neural network model is shown
in Fig. 9 in which the real and imaginary part of the impedance are used as the inputs. The
conductivity and thickness (or lift-off) are considered as outputs. Matlab's neural network
toolbox has been used for the solution. The data used for the generation of grid system are
used to train the neural network. Depending on the situation the model can have two, three or
even more number of outputs. A simple model has been developed and the results obtained
from it are compared with that obtained utilizing grid system which is reported in [6] is shown
in Table 4 It is seen that the level of error is consistent. The level of the error can be reduced
using more data for training the network.
7. Conclusions
In this paper the possibility of using planar type mesh coil for the estimation of near surface
material properties has been investigated. An analytical modeling has been developed from
which the impedance of the coil is calculated. The impedance of the coil is used for the
estimation of near-surface material properties. The approach of estimation by using a grid
system has been described and an alternative to this approach is to use neural network has
been proposed.
188 S.C. Mukhopadh\a\ el al. /Evaluation of Near-Surface Material Properties
Table 4 Comparison of results obtained using grid system and neural network
References:
[1] Y.Shi and D.C.Jiles, "Finite element analysis of the influence of a fatigue crack on magnetic properties
of steel", Journal of Applied Physics, vol. 83, no. 11, 1 June, 1998, pp. 6353-6355.
[2] N.J.Goldfine and D.Clark, "Near-surface material property profiling for determination of SCC
susceptibility", 4* EPRI balance-of-plant heat exchanger NDE symposium, WY, June 1-2, 19%.
[3] N.J.Goldfine. "Magnetometers for improved material characterization in aerospace application".
material evaluation, March 1993, pp. 396-405.
[4] N.J.Goldfine, D.Clark and T.Lovett, "Material characterization using model based meandering winding
eddy current testing (MW-ET)", EPRI Topical workshop: Electromagnetic NDE applications in the
Electric Power Inductry, Charlotte, NC, Aug 21-23, 1995.
[5] S.Yamada, H.Fujiki, M.Iwahara, SC Mukhopadhyay and F.P.Dawson, "Investigation of Printed
Wiring Board Testing by using Planar Coil Type ECT Probe", IEEE transaction on magnetics, vol. 33.
no. 5, Sep. 1997, pp. 3376-3378.
[6] S.C. Mukhopadhyay, S.Yamada and M.Iwahara, "Investigation of Near-Surface Material Properties
Using Planar Type Meander Coil", presented in JAS, March 16-17, 2000, Australia, will be published
in JSAEM Studies in Applied Electromagnetics and Mechanics.
Electromagnetic Nondestructive Evaluation (VI) 189
F. Kojima et al. (Eds.)
1OS Press, 2002
/. Introduction
2. CFRP Samples
Three 20-mm-thick CFRP planks with hidden slots were prepared as the crack sample. The
width and length of each sample were 250 mm and 200 mm, respectively. Each sample has two
slots. The slots were artificially made on the back surface of the samples. Each slot has a
different depth; the depth means the distance between the sample surface and the slot. The
depths of the slots are 5.0, 7.5, 10, 12.5, 15 and!7.5 mm. The width and length of all the slots
are 40 mm and 1 mm, respectively. The schematic figure of the crack sample is shown in Fig.
l(a).
A 2-mm-thick CFRP plate that included an insulator film inside the plate was prepared as
delamination sample. The width and length of the sample were 100 mm and 200 mm,
respectively. The insulator film made of polyethylene terephthalate (PET) was inserted between
the central layers during the process. The film is in the shape of an ellipse. The long
diameter, short diameter and the thickness of the film is 20 mm, 10 mm and 12 urn,
respectively. The schematic figure of the delamination sample is shown in Fig. l(b).
The carbon-fiber is of the cross texture type, and the packing factor of the carbon-fiber is
0.6 in this study.
3. SQUID-NDE System
The new SQUID-NDE system was constructed in a normal laboratory in National Institute of
Advanced Industrial Science and Technology (AIST). The system is composed of a low
temperature superconductor (LTS-) SQUID, SQUID electronics, a dewar with a 3-mm-thick
bottom, a scanning stage with two stepper-motors, current supplier, lock-in amplifier and PC.
The schematic diagram of the system is shown in Fig. 2. The sample is set on a table of the
scanning stage, and moved under the SQUID. AC current is injected into the sample, and
SQUID measures the magnetic field induced by the current flowing in the sample. The PC
controls the stepper-motors through RS232C interface. The PC also controls the lock-in
amplifier through GPIB interface. The SQUID output was gathered through the lock-in amplifier.
Fig. 1 Schematic figures of the samples, (a) Crack samples, (b) Delamination sample.
Y. Hatsukade et al. /Nondestructive Detection of Inside Cracks and Delamination 191
The amplitude and phase difference of measured magnetic field are stored in the PC. A data
acquisition and stage-control program was developed using LabVIEW. The minimum distance
between measuring plane and sample surface (standoff distance) is 3.5mm.
In this work, we used a small low-rc SQUID loop that formed a first order gradiometeric loop
in order to reduce the ambient magnetic noise [6]. Two washer coils of the SQUID loop are
connected in series and reverse. The microscopic photograph of the SQUID is shown in Fig. 3.
Two washer coils have the same dimension. The outer and the inner dimension of one single
washer coil are 650 x 650 urn2 and 50 x 50 urn2, respectively. The distance between the centers
of each washer coil (base line) is 770 um. The magnetic flux noise of the SQUID measured by a
flux locked loop (FLL) operation in the normal laboratory was about 5 x 10"6 Oo/ATHz (0.2
in the white noise region (5Hz-1200Hz).
4. Method
The direct current injection method was adopted to induce current flow in the samples. For the
crack samples, two electric terminals made of copper were prepared to inject current into the
samples. The electric terminals were stuck on both ends of the sample surface by electrically
conductive adhesive tape. AC current was injected into the sample from the one terminal and
flowed out from the other terminal. The schematic view of the method is shown in Fig. 4(a).
(a)
Fig.4 Injected current method; (a) without and (b) with a return current pad.
192 K Hatsukade et at /Nondestructive Detection of Inside Cracks and Delamination
For the delamination sample, we used a thin copper pad for return-current to extract the
anomalous magnetic signal due to the delamination in the sample [5]. The pad was stuck by
the adhesive tape under the sample inserting a thin insulator sheet between the sample and
the pad at one end of the sample. An electric terminal was also stuck on the sample surface
at the opposite end. The schematic view of the method is shown in Fig. 4(b).
5. Measurements
In the detection of the slots, the crack sample was set on the scanning table so as to make the
slot be parallel to the x-direction. Currents of 300 mA and 300 Hz were injected into the samples
to induce current flow toward the y-direction. The distance between the sensor and the surface
of the sample was set about 5.5 mm in the experiments. The SQUID was set to measure the first
order gradient component in the x-direction of the z-component of the magnetic flux density,
dBz/dx. We measured the dBz/dx along line parallel to the x-direction above the slots. The
sampling space was 2 mm in the x-direction.
In the detection of the delamination, current of 60 mA and 700 Hz was injected into the sample
to induce current flow toward the y-direction. The standoff distance was about 5.5 mm. First,
the sample was scanned along line above the center of the delamination in the x-direction. Next,
the area around the delamination was scanned. The sampling space was 1.5 mm in the x-
direction and 6 mm in the y-direction.
6. Results
All results on the measurements of the crack samples are shown in Fig. 5 together. The
section of the sample is also shown together. Anomalous downward signal peaks due to the slots
appears above the slots. The signal amplitude due to the slot decreases as the depth of the slot
increases. The slots below up to 10mm from the sample surface were successfully detected. We
could not detect signal peaks due to the slot below 12.5mm.
Fig. 5 All results on the hidden slots. The signal peaks due to slots below up to 10mm
were successfully detected. The schematic shape of the slot is represented by dark square.
Y. Hatsukade et al. /Nondestructive Detection of Inside Cracks and Delamination 193
Fig.6 Result on the delamination. Signal peaks due to the delamination appear
above the delamination in both curve (I) and (II). The upward and downward
signals on both sides in the curve (I) and jumps in (II) are due to the edge effect.
Fig. 7 Contour maps of magnetic signals dBz/dx due to the delamination. (a) Amplitude, (b) Phase
difference. The position and size of the PET film in the sample is shown together by perforated line in
both figures.
The amplitude and the phase difference of the dBz/dx measured along the line above the
delamination are shown in Fig. 6 as a curve (I) and a curve (II), respectively. Signal peaks due to
the delamination appeared above the delamination in both curves. The pair of large upward and
downward peaks at the sample edges in (I) and the jumps on both sides in (II) are due to the
edge effects. Figures 7 (a) and 7(b) show the contour maps of the amplitude and the phase
difference measured around the d elamination, respectively. The area w here the edge e ffects
appeared is excluded. The center peaks shown in Fig. 6 appear to form a shape of an ellipse in
both figures. The position and size of these ellipses approximately agree with those of the PET
film in the CFRP sample plate. These results show that the SQUID-NDE method can be applied
to the detection of the delamination in CFRP.
194 Y. Hatsukade et al. /Nondestructive Detection of Inside Cracks and Delamination
7. Discussion
The potential of SQUID-NDE method to detect the inside shallow delamination in CFRP
plate is experimentally clarified. We discuss the detection ability for deep-lying delamination.
The penetration depth and the sensitivity of a SQUID determine the detection ability for deep-
lying defects. We roughly estimated the required frequency and amplitude of injection current
and SQUID parameters to detect a delamination, which located 10mm below from the surface in
a thick CFRP plank. In this case, injection current of lower frequency than 1200 Hz is necessary
because of the skin effect. We experimentally investigated the frequency dependence of the
magnetic field signal due to the delamination by measuring the delamination sample while
changing the frequency of the injection current from 20 Hz to 1000 Hz. The amplitude of the
signal exhibited little dependence on the frequency. However, the phase difference exhibited
dependence as shown in Fig. 8. The phase displacement due to the delamination increases as the
frequency increases. It shows that measuring the phase difference is effective in the frequency
range above about 300 Hz. Then, current of near 1000 Hz is appropriate to detect the
delamination below 10 mm using the phase difference.
The current density at 10 mm from sample surface reduces by a factor of e compared to that
on the surface due to the skin effect while injecting the current of 1200 Hz into CFRP. If it is
assumed that the disturbed current due to the delamination resemble to a current dipole, the
signal peak amplitude of the dBz/dx is in proportion to 1/h3, where h is the distance between the
measuring plane and the delamination. Using the minimum standoff distance 3.5mm, the signal
peak amplitude is reduced by a factor of about 10 compared to that in this study. So, the
injection current of 8 A and 1200 Hz should be sufficient to detect the same size delamination
located below 10 mm with same signal-to-noise ratio as in this study, using the amplitude of the
magnetic field signal. In this estimation, we assumed a CFRP plank of 10 mm thickness and the
same width as used in this study. The amplitude of the injection current is rather large and not so
practical.
There is other way to reduce the amplitude of the current by enlarging the area of the pick-up
coil. In this study, we used a small first order gradient SQUID loop because we only had that.
However, it is better to use larger pick-up coil, for example 10mm x 10mm, to detect deep-lying
delamination. The sensitivity of a SQUID with the large pick-up coil will increase about 40 times
on assumption that the pick-up coil inductance is 5 nH and the input coil inductance is the
Ketchen type and 5 nH [3]. Then the injection current of 200 mA and 1200 Hz is enough to
detect the delamination below 10mm using both the amplitude and phase difference of the
magnetic signal. The condition of the injection current is reasonable in practice. It suggests that
the SQUID-NDE method has a possibility to detect the deep-lying delamination in thick CFRP
materials.
The electric conductivity a of CFRP is about 100 - 1000 times lower than metals such as
copper or aluminium. The current injection method is easier method to induce much current flow
in CFRP than a method with an induction coil. However in the practical NDE of aircraft and
space structures, non-contact method is desired in many cases. We will develop the non-contact
SQUID-NDE method for such cases. We have to resolve some problems with the standoff
distance and imbalance of pick-up coils for the development.
8. Summary
Acknowledgement
The authors would like to thank Dr. K. Chinone of Seiko Instruments Inc. for the low Tc
SQUID.
References
Abstract The influence of plastic deformation on the magnetic properties has been
studied in Ni50Cu50 alloy. The magnetism of this alloy before and after plastic defor-
mation is analyzed by the self-consistent renormalization (SCR) theory of spin fluc-
tuation. The spontaneous magnetization at OK decreases considerably as a result of
plastic deformation. At the same time the Curie temperature also becomes lowes.
This phenomenon shows the existence of atomically ordered state in this alloy. The
ferromagnetic state changes to paramagnetic along the antiphase boundary between
superpartial dislocations even at 4.2K. The paramagnetic state increases in proportion
to the dislocation density. The paramagnetic state extent in the long distance from the
APB. The paramagnetic APB ribbon exerts an influence to the Curie temperature. The
magnetism in the Ni50Cu50 alloy, including the plastically deformed ones, can be
explained by the SCR theory of spin fluctuation. The dislocations are observed by the
electron microscopy.
1 Introduction
Several remarkable magnetic transitions due to plastic deformation were discovered in inter-
metallic compounds and atomically ordered alloys. In the Ni3Al intermetallic compound, for
a example, the introduction of dislocations by plastic deformation causes the susceptibility
as well as the Saturation magnetization to reduce considerably [1]. In Pt3Fe ordered alloy,
which is antiferromagnetic before plastic deformation, the application of plastic deformation
causes it to become quite strongly ferromagnetic even at room temperature [2]. These mag-
netic transitions due to plastic deformation were explained from the viewpoint of the atomic
rearrangement in the antiphase boundary (APB) between superpartial dislocations [3].
The magnetic transition due to plastic deformation includes several interests, which con-
nect the fundamental problem of the ferromagnetic origin. The atomic arrangement in the
APB is an important factor of the magnetic transition. We can get the valuable informa-
tion on the relation between the magnetic ordered state and the atomic arrangement. The
magnetic state of the APB ribbon enhances the neighbouring state and changes to the same
state as the APB ribbon in the long range distance. The atomic arrangement of the APB rib-
bon is paramagnetic in Ni3Al intermetallic compound and the paramagnetic region spreads
around the APB ribbon at least as far as 100 atomic distances. An other example is Fe-Al
intermetallic compounds, which are paramagnetic above 35at % Al concentration. The Fe-
Al compounds change to ferromagnetic by plastic deformation [4]. The ferromagnetic APB
ribbon changes neighbouring magnetism to ferromagnetic in the long distance and the extent
depends on the Al concentration [5]. The magnetic extent suggests that two magnetic states
H. Satoh and S. Takahashi / Change of Magnetic Properties Due to Plastic Deformation 197
exist near the APB ribbon for the same atomic structure in Ni-Cu alloy. The magnetic or-
dered state is explained by the two different models in general, the localized electron model
and the itinerant electron model. It has been metallurgically believed that the Ni-Cu alloys
form a substitutional solid solution with a face-centered cubic structure over the entire range
of concentration, even in the stoichiometric composition. The atomically ordered state has not
been ascertained experimentally in Ni50Cu50 alloy. It is difficult to indicate the atomic order-
disorder change by X-ray and the electron diffraction, since the difference between atomic
scattering factors of Ni and Cu atoms is too small. The existence of the atomically ordered
state is controversial. The decrease of Ms(0) and Curie temperature by plastic deformation
in NisoCuw alloy should be explained by the same model as the other ordered alloys and
intermetallic compounds.
Our group has discovered a decrease in the values of Ms(T) and Tc before and after
plastic deformation in Ni-Cu alloy [6]. This peculiar phenomenon can be explained by the
introduction of superpatial dislocations in Ni-Cu alloy. We confirmed it from the viewpoint
of the existence of APB in order Ni-Cu alloy. In the present study, firstly, our interest is to
examine the validity of the self-consistent renormalization (SCR) theory of spin fluctuation
[7][8] and the change of MS(T), Tc due to plastic deformation. Secondly we aimed to exam-
ine the magnetic transition due to plastic deformation in view of the existence of APB before
and after plastic deformation by magnetic influence.
2 Experimental Procedure
Ni50Cu5o alloy was prepared by cold crucible levitation melting at first and the alloy is
melted by arc melting again to obtain chemical homogeneity on a water cooled copper hearth
in an argon gas atmosphere at a pressure of approximately 93 kPa.
The buttons were homogenized at 1073K for 72 h and annealed at 653K for 120 h to get
the atomically ordered state. We could not obtain higher uniformity by melting the samples
many times in the arc furnace: the different values of magnetic properties were measured in
the different parts of a button. We could prepare rather uniform composition by melting it in
two processes.
The samples dimensions of approximately 2.5 x 2.5 x 5.0mm3 were cut from the but-
ton by Servomet spark cutting machine. The samples were tested in compression at room
temperature by Instron-type machine. Samples for the magnetic measurement dimensions of
2.5 x 2.5 x 2.5mm3 were cut from the spark cutting machine and were chemically polished to
remove surface damage. The steady-field magnetization was measured with a SQUID mag-
netic fluxmeter(Quantum Design) in the temperature range of 4.5K to 100K.
3 Experimental results
The magnetization curves of Ni50Cu50 alloy with different plastic strains were measured in
the temperature range of 4.5 to 100K. Figure 1 shows the isothermal magnetization curves for
the samples with plastic strain =0 %. The Ni50Cu50 alloy is ferromagnetic at low tempera-
tures and the spontaneous magnetization at 4.5K decreases slightly with plastic deformation.
The spontaneous magnetization, M$(0), at OK and the Curie temperature TC have been ob-
tained from the magnetization curves. The analysis of Ms(0) and TC is given as follows:
Figure2 shows that the data fit nicely on straight lines in Ms(T)2 versus H/M in Ni.50Cu5Q
198 H. Sutoh and S. Takahashi / Change of Magnetic Properties Due to Plastic Deformation
Figure 2: Arrotplot
alloy, which are commonly used in TC (Arrot plots). The data for all the plastically deformed
samples were represented well in the form of M(H, T)2 = MS(T) + bH/M(H. T) at all
temperature, except in the region of weak magnetic field. The spontaneous magnetization
at TK, Ms(T], can be obtained by the extrapolation of linear relation to H=0. The Curie
temperature TC shows a slight decrease with plastic deformation. The spontaneous magneti-
zation, MS(T), in the plastically deformed Ni^Cu^o alloy is analyzed was the SCR theory
of spin fluctuation. In the SCR theory the temperature dependence of Ms(T) has the form
of Ms(T)2 = A/(0) 2 - //r 2 at low temperatures and .\fs(T)2 = C(7?/3 - T 4 / 3 ) in a fairly
wide temperature range below TC- Figure3 shows the temperature dependence of the square
of the spontaneous magnetization MS(T}2 against T2 in the A r / 5 oCu 5 o samples with =0~
30 % strain. All the samples have the simple relationship that .U5(0)2 is proportional to T2
at low temperature. The spontaneous magnetization at T=OK, A/s(0), is determined by the
extrapolation of this linear relation of MS(T)2 against T2 to T=0. The values of .Us(0) for
A r ?5oCu 5 o alloy are given in table 1. Figure4 shows the temperature dependence of Ms(0)2
H. Satoh and S. Takuhashi / Change of Magnetic Properties Due to Plastic Deformation 199
on T4/3 in Ni50Cu5Q alloy with =0~30 %. MS(T)2 has a T4/3 dependence in the wide
temperature range below TC for each sample. The value of the Curie temperature is obtained
by the extrapolation of the linear relation to MS(T) = 0. The value of Tc is nearly the same
as that obtained through the Arrot plots. The valued of Tc are shown in table 1. The temper-
ature dependence of Ms(T) of Ni50Cu50 alloy including the plastically deformed ones, is
expressed well in the form of
some scattering exists. The scattering would be caused by the lack of uniformity of the sam-
ples.
4 Discussion
The experimental results indicated that the magnetism changed remarkably, In particuler the
spontaneous magnetization and Curie temperature decrease as the strain increases. The or-
dered structure is Ll 0 -type in Ni-Cu alloy. The change of magnetism was explained accord-
ing to the formation of APB in order-state (Ll 0 -type) by plastic deformation. It changes from
ferromagnetic to paramagnetic to destroy the cycle of atomic arrangement. The experimental
results of temperature dependence of Ms(T) are consistent with the results of the SCR theory
of spin fluctuations in .V/ 5 oCu 5 o alloy with plastic deformation. According to the SCR theory
H. Satoh and S. Takahashi / Change of Magnetic Properties Due to Plastic Deformation 201
Figure 7: is the coefficient of the T^3 term in the Ms (O)2 versus T4/3C plot
MS(T)2 /3
- T4/3) (3)
where F\ and /? are parameters depending on the band structure, and the Curie tempera-
ture is represented as
(4)
where a = I N ( E F ) , i.e. the product of the intra-atomic exchange constant, I, and the
density of states, N ( E F ) at the Fermi level, EF. As both Ms(0) and TC decreases with
plastic deformation, it seems a plausible conclusion that the factor (a - 1) decreases and the
parameter R increases with plastic deformation, keeping [\ nearly constant.( in equation (3)
corresponds to r\/R. ( decreases with increasing plastic strain in tablel. The increase of R
is consistent with the decrease of (. The increase of R due to plastic deformation and the
constant F\ are obtained in Ni^+xAl^-x intermetallic compounds. As the TC decreases
with plastic deformation, a = I N ( E F ) decreases as e increases. R can be represented by
N(EF) and the first and second derivative of N(EF) by.
Plastic deformation would exert some influence on the derivative of N(EF) as well as
N(EF). The decrease of a is consistent with the increase of R.
It gave experimentally a simple relationship between TC and Ms(0) in Ni-j^xA^-x
compounds, i.e.
T4/3( = ATMs(O) 2 (6)
where K is a constant. Equation (6) can be obtained by supposing that (a I)/R =
(a0 - I)/ RQ, where a0 and R0 are the values at T = OK. The relationship of equation
(6) has been explained in the present study. Figure? shows the relationship between Tc'3(,
and A/s(0) 2 . The relationship indicates that the magnetism in the Ni50Cu50 alloy, including
the plastically deformed ones, can be explained by the SCR theory of spin fluctuation. The
values of A/.s-(0) are a function of the number of atomic configurations in the vicinity of APB,
202 H. Satoh and S. Takahashi / Change of Magnetic Properties Due to Plastic Deformation
which increases proportionally to the dislocation density. The change of the spontaneous
magnetization due to plastic deformation can be written as
where N0 and ^Ni are the total number of Ni atoms and their magnetic moment before
plastic deformation, respectively. N\ is the number of Ni atom near the APB. .\\ can be
written as
Ni = CnS2/a2rp (8)
where rpis the size of APB ribbon between superpartill dislocations. The number of these
atomic moment increases as the strain increases. This shows the relationship of size of APB
ribbon. The relationship between spontaneous magnetization and dislocation density is shown
theoretically in comparison with the experimental results. The value of dislocation density
has not been obtained by direct method such as the electron microscope. We can estimate the
dislocation density of the present samples theoretically to be p 10 12 cm~ 2 . The number of
Ni atoms, which carry no moment near the APB, is ten times of that constituting the APB.
The different atomic configuration from the ordered state is restricted in the APB ribbon.
There exist paramagnetic and ferromagnetic states for the same atomic configuration near
the APB ribbon. The two different states for the same atomic structure near the APB were
confirmed in the other alloys and intermetallic compounds.
Thus, there exist paramagnetic state near the APB ribbons in Ar?50Ci/5o alloy. The rela-
tionships between the spontaneous magnetization and dislocation are depending on localized
electron model. The exchange interaction in long distance cannot be explained in Ni-Cu alloy.
Further, the itinerent electron model of Ni-Cu alloy cannot be adapted directly. The results
are consistent with SCR theory.
References
[1] De Boer FR,Shinkel CJ.Biesterbos J and Proost S 1969 J.Appl.Phys.40 1049
[2] Crangle J 1959 J.Phys.Radium 20 435
[3] Takahashi S and Dceda K 1983 Phys.Rev.B 28 5225
[4] Taylor A and Jones RM 1958 J.Phys.Chem.Solids 6 16
[5] Takahashi S.Umakoshi Y 1991 J Phys Condens Matter 3 5805
[6] Takahashi S,Li XG.Chiba A 1996 J Phys Condens Matter 8 5101
[7] Moriya T.Kawabata A 1973 J.Phys.SocJapan 34 639;35 669
[8] Sasakura H,Suzuki K,Masuda Y 1984 J Phys Soc Japan 53 754
Electromagnetic Nondestructive Evaluation (VI) 203
F. Kojima et al. (Eds.)
1OS Press, 2002
1. Introduction
Natural gas transmission pipeline inspection employs the magnetic flux leakage (MFL)
method, where the pipe wall is magnetized axially by permanent magnets and leakage
fields due to defects are detected by an array of Hall sensors. Two major parameters in this
experiment that affect the detection capability are magnetization level and sensor spacing
in the array. A systematic way of quantifying the influence of these factors and optimizing
the tool design is provided by the Taguchi method.
In pipeline inspection, the design quality characteristic of the inspection tool is defined
by the probability of detection (POD) of a critical defect. Thus for the tool to be optimum
the POD should be as large as possible. The optimum condition can be found by
conducting experiments (simulations) and determining the contribution of each
experimental factor to the POD of critical defect. The study presented in this paper follows
the standard approaches described in [1] and [2]. A model based POD evaluation method
[3] is employed in calculating PODs.
2. Taguchi Method
column. Two 2-level factors can be combined in four possible ways, (1,1), (1,2), (2,1), and
(2,2). When any two columns of an array contain these combinations the same number of
times, the columns are said to be orthogonal or balanced. For example, column 1 and
column 2 in Table 1 have all the above combinations once. This also holds for the other
pairs of columns.
Uncontrollable factors (noise factors) are included in a second OA (outer array), which is
used in conjunction with the array of controllable factors (inner array).
The Taguchi approach involves listing the factors and the levels of each factor that need
to be considered, designing the experiment, conducting the experiment, analyzing the
results, and determining the optimum condition and effects of each factor. A 0.5"x0.5"
rectangular defect, of 15% depth and 45 surface angle, is used as the critical defect for
POD calculations, in this Taguchi study.
Four factors with two levels each were considered in this study. Based on prior
experience, magnetization level, sensor spacing, radial liftoff, and circumferential shift
between sensor and defect have obvious influences on the measured signal. The factors can
be classified into controllable or uncontrollable (noise) factors, depending on whether or
not its level can be controlled.
Magnetization level is a controllable factor. In pipeline inspection, permanent magnets
are used to establish magnetic fields in the pipe wall. Magnetization level, denoted by Br, is
a design parameter characterizing the strength of the permanent magnets. Intuitively, the
MFL signal at a high magnetization level is stronger than that obtained at low
magnetization level. The two magnetization levels used in this Taguchi analysis are 0.8T
and 1.2T.
Radial Liftoff is defined as the distance from a sensor to the inner surface of the pipe wall.
It is an uncontrollable factor. When the liftoff distance increases, the MFL signal decreases
monotonically. Radial liftoff is a random variable whose level represents its mean value.
The two levels selected are 0.05" and 0.1".
Sensor spacing in the circumferential sensor array is a controllable factor. Ideally one
would wish to reduce the number of sensors used in the tool without compromising the
POD of a critical flaw. Circumferential shift, identified as the lateral or circumferential
distance between sensor and defect, is an uncontrollable factor. A combination of sensor
spacing and circumferential shift is used to determine the positions of sensors in the
circumferential direction, as shown in Figure 1. In the FEM prdeiction, signal is discrete
corresponding to a sensor spacing So, specified by the element size in the circumferential
direction, and the maximum circumferential shift d equals So/2. Signal magnitudes for
different circumferential shifts can be obtained by interpolating the FEM signal. A
Gaussian distribution of circumferential shifts with an assumed mean (level) chosen as a
fraction of the sensor spacing is used to calculate the POD.
Z. Zeng et al. /Optimization of Pipeline Inspection Tool 205
Levels
Factors
Level 1 Level 2
Magnetization level (A) 0.8T 1.2T
Controllable
Sensor spacing (B) 1" 2"
Radial liftoff 0.05" 0.1"
Uncontrollable
Circumferential shift 0 25%
(N
(N (N
c CN fN
Two levels of sensor spacing, 1" and 2", and two levels of circumferential shifts, 0 and
So/4, are used in the Taguchi study. The factors and levels are summarized in Table 2.
Table 3 shows the inner and outer orthogonal arrays used in the Taguchi study. Note that
the last columns of both the inner array and the outer arrays in Table 3 are not used,
because we have only two controllable and two uncontrollable factors. With the factors and
206 Z. Zeng et al. /Optimization of Pipeline Inspection Tool
Fig 2 2-D signals in the surface perpendicular to the axial direction at the center of the flaw
(a)flr=0.8r (b)B = \.2T
levels chosen as shown in Table 2, the number of experiments is the same as that of full
factorial design. However, the method will be more efficient if more factors and/or more
levels are considered in future research.
To obtain complete data, each trial run of the inner array must be repeated for each of the
four noise combinations (in outer array). That is, the experiments consist of the
computation of POD for each combination of magnetization level, sensor spacing, and
distributions of radial liftoffs and circumferential shifts.
FEM simulations are performed to predict signals. With each simulation of the FEM at a
particular magnetization level, we obtain a signal in the 3-D space (radial, circumferential,
and axial). Since we are interested in the peak value of the signal, we can consider only the
signal at the center (in the axial direction) of the flaw. The 2-D signals perpendicular to the
axial direction with axial coordinate at the center of the flaw obtained with 5r=0.8T and
1.2T are shown in Figures 2 (a) and (b).
For each run of the inspection tool, we have a random radial liftoff and a random
circumferential shift. In order to account for the effects of the uncontrollable factors,
multiple simulations are performed assuming Gaussian distributions for radial liftoffs and
circumferential shifts. Using 2-D interpolation, we can find signal magnitudes at the
corresponding radial liftoffs and circumferential shifts, and thus find the probability density
function of the signal magnitudes.
x 10"3
A A
' POD:098491
<
; \ 1
PFA: 0 05 i
T: 188. 7057'
; i
Without flaw/ With flaw
/
/ ',
I \
^' . ' V
50 160 170 180 190 200 210 22
Peak Flaw Signals
Besides random variations in radial liftoff and circumferential shift, there is also
measurement noise present in the signal, which is independent of any of the four factors
being considered. This common measurement noise should be taken into account when
calculating the 16 PODs. Typically, we assume the noise to be Gaussian with zero mean
and variance 25 Gauss.
Conducting each of the 16 simulations in Table 3, pdfs of measurement outputs with
and without flaw were found. Figure 3 presents the pdfs of measurement outputs when Br
- 1.2T, sensor spacing = 1", mean radial liftoff = 0.1", and mean circumferential shift = 0.
The POD and PFA are calculated as
> 0)
and
PFA=p(y\x0)<fy (2)
where T is the threshold, p(y \ x,) and p(y \ x0) are the pdfs of signal magnitudes with
and without a flaw respectively. The threshold was selected so that the PFA is 0.05. All 16
PODs are shown in Table 3.
5. Analyses
The results of the Taguchi experiments were analyzed using the standard steps. First, the
factorial effects (main effects) were evaluated and the influences of the factors were
determined in qualitative terms. The optimum design parameters and the performance
under this condition were also determined from the factorial effects. In the next phase,
analysis of variance (ANOVA) was performed to identify the relative influence of the
factors.
The signal to noise ratio measures the sensitivity of the quality characteristic investigated
in a controlled manner, with respect to uncontrolled (noise factors). The S/N ratio is
defined as
where MSD is mean squared deviation from the target value of the quality characteristic.
Consistent with its application in engineering and science, the value of S/N is desired to be
large; hence the value of MSD should be small. In our case, we want the quality
characteristic to be as large as possible. In this case, the MSD is defined as
MSD = (l/yf + \/yl + \/y2, + )/n (4)
In (4), y, is the result of the /th repetition, and n is the number of repetitions. The S/N ratio
calculation is based on data from all observations of a trial condition. The set of S/N ratios
can then be considered as trial results without repetitions. The four S/N ratios under each
trial condition (controllable) are shown in Table 3.
208 Z. Zeng et al. / Optimization of Pipeline Inspection Tool
-0.150
-0.918
/ -1.238
j /-2.007
A, ' A 2 B, ~~j~~
Factors and levels
Fig 4 Main effects for S/N analysis
5.3 Main Effects
In Taguchi analysis, the optimum condition is identified by studying the main effects of
each of the factors. The main effects indicate the general trend of the influence of the
factors.
To compute the main effect of each factor at each level, we add results (S/N ratios) for
trials including the factor at this level, and then divide by the number of such trials. For
example, to compute the main effect Ajof Factor A (magnetization level) at level 1 (0.8T),
we look in the column for A in the inner array in Table 3 and find that level 1 occurs in
experiment numbers 1 and 2. Thus the main effect of AI is
The average effects of other factors are computed in a similar manner and listed in Table 4
and plotted in Figure 4.
From these results, we can conclude that in the four trials studied the combination A2Bi
leads to maximum POD, and Factor A (magnetization level) has more influence than
Factor B (sensor spacing) on the POD.
- = 3.589 (6)
N
where (S/N)t is the S/N ratio of the ih trial; N is the number of trials; T = ^(S/N), is the
1=1
summation of all results. The term T2/N in (6) is called the correction factor, C.F. The
sum of squares for a factor (e.g. Factor A) can be calculated from:
Z. Zeng et al. / Optimization of Pipeline Inspection Tool
where L is the number of levels; H* is the number of trials for Factor A at level k; Ak is the
main effect of Factor A at level k and hence (Ak - n k ) is the summation of S/N ratios when
the level of Factor A is k. Subtracting all the factor sums of squares from the total sum of
squares, the residual is referred to as error sum of squares, expressed as
Se=ST-SA-SB (8)
Degree of freedom (DOF) is an important and useful concept that is defined as a measure
of the amount of information that can be uniquely determined from a given set of data.
DOF of a factor equals one less than the number of levels. For a factor A with two levels,
AI can be compared with only A2, not with itself. Thus a two level factor has DOF=1, i.e.,
/4 = 1 . DOF of an experiment (total DOF or fT ) is one less than the number of trials.
DOF of the error term is given by
Variance measures the distribution of the data about the target value. Since the data is
representative of only a part of all possible data, DOF rather than the number of
observations is used in the calculation.
. Sum of Squares S
Variance = --- or V - (10)
Degrees of Freedom /
The variance ratio, commonly called the F statistic, is the ratio of variance due to the
effect of a factor to the variance due to the error term. This ratio is used to measure the
significance of the factor under investigation with respect to the variance of all factors
included in the error term. The F values obtained in this study are /v=87.055, FB=2.5S9,
Fe=\. Assuming a confidence level of 90%, the F values determined from the F-table
(available in most statistical handbooks) for both Factors A and B are /ro.i(l>l)=39.864,
where the two parameters in the parentheses are DOF of Factor A and DOF of error term,
respectively. Since FA > Fo.i(l,l) > FB, we say that Factor A (magnetization level) does
contribute to the sum of squares at the confidence level of 90%, while Factor B (sensor
spacing) does not contribute to the sum of squares at this confidence level.
The pure sum of squares is defined as the sum minus the degrees of freedom times the
error variance:
S'A=SA-fA-Ve (11)
S'e=St+(fA+fB)-Ve (12)
The percent contribution for any factor is obtained by dividing the pure sum of squares
for that factor by ST and multiplying the result by 100.
Results of the above analyses are listed in Table 5 under the part "before pooling", where
f, S, V, F, and P stand for DOF, sum of squares, variance, variance ratio, and percent
contribution, respectively.
When the contribution of a factor is small, the sum of squares for that factor is combined
with the error Se. This process of disregarding the contribution of a selected factor and
subsequently adjusting the contributions of other factors is known as pooling. In Table 5,
we find that the percent contribution of Factor B (sensor spacing) is only 1.753%. Thus
Factor B should be pooled. The new error variance is computed as:
210 Z. Zeng et al. /Optimization of Pipeline Inspection Tool
r = v^ (13)
With a pooled Ve, all S' vilues must be modified to reflect pooling. The new malysis
results with Factor B being pooled are also listed in Table 5. The percent contribution of
Factor A is now 94.061%. The contribution of errors is 5.94%.
6. Conclusions
The Taguchi method is an effective way of optimizing an experiment setup. This study
considered two factors each with two levels and, the optimum combination A2B|, that is,
1.2T magnetization level and 1" sensor spacing was determined to be optimal with
respective to the probability of detecting a critical flaw.
Analyzing the variance ratios, we find that the magnetization level contributes to the
sum of squares within 90% confidence level, while sensor spacing does not.
In the ANOVA table without pooling, the percent contribution of sensor spacing is
considered to be too small and is therefore pooled. The percent contribution of
magnetization level is 94.061% after pooling.
These results were obtained by considering a .5"x.5"xl5% deep critical defect. This
defect was chosen on the basis of the smallest element in the FE mesh. If the critical defect
is smaller, the sensor spacing will have a greater effect on the POD. Also considering more
factors and levels can increase the effectiveness of the study.
References
[ 1 ] R. K.. Roy, "A primer on the Taguchi method", Van Nostrand Reinhold, 1990.
[2] K. Dehnad, "Quality control, robust design, and the Taguchi method", Pacific Grove, Calif, 1989.
[3] Z. Zhang, Y. Zhang, L. Udpa, S. S. Udpa, "Probability of detection model for Pipeline inspection",
QNDE, Vol 16, 1997, pp. 1307-1314.
Electromagnetic Nondestructive Evaluation (VI) 21
F. Kojima et al. (Eds.)
IOS Press, 2002
1 Introduction
crack model and geometrical configuration adopted in the previous study. The
biggest advantage of the scheme is that it can deal w i t h cracks t h a t have d i s t r i b u t e d
conductivity as shown in left-hand side of Fig. 1. Although an artificial neural
network has ability to simulate any kinds of non-linear function, it is very diffi-
cult to simulate exact mapping between two spaces, especially in ca.se of inverse
mapping. Key idea of the scheme is that it is not necessary to take all crack pa-
rameters into consideration for the reason that only size of cracks is i m p o r t a n t in
view of practical inspections. Thus, mapping t h a t the scheme dealt with reduced
into much more simple one. and accurate reconstruction became possible. It was
revealed that the scheme is robust against noises, and profiles of natural ( r a c k s oc-
curred in a steam generator tobe was successfully reconstructed by the scheme[7].
On the other hand, the scheme has significant drawback; no criterion t h a t indi-
cates accuracy of the reconstruction is available unlike reconstruction schemes using
model-based approaches[8). because it is impossible to calculate eddy current signals
from a reconstructed crack.
In the present paper, the authors enhance the scheme by means of a neuro-fuzzy
system in order to overcome the drawback. An overview of the scheme is illustrated
in Fig. 2. The difference between the scheme proposed here and the one in the
previous study is a fuzzy system that processes output of the neural network. The
fuzzy system computes a parameter that will be proposed in the present paper and
gives degree of credibility on the output.
Crack
Pancake Probe
crack profile
Neural network
equally excited, while large vlaue of;; indicates that only one neuron is excited and
other neurons keep calm. Figure 3 also illustrates a scheme to compute the ;;-value.
Since the sixth neuron is excited most and its output value is 0.81. output value
of the third neuron is. for instance, supposed to be (1 + 0.81) exp{;;|5 2|} 1.
Six examples of a neural network's responses to unknown signals and calculated
p-values are listed in Table 2. A glance at Table 2 reveals that the p parameter
well represents degree of ''confidence" that the trained neural network had in its
response.
41 16 34
47 1 1 27 1634
12 42 9 47 43
15 18 16 5 49 48 2638
44 29 3 8 1 10 33 9
40 32 38 27 9 26 9 3O3C
28 15 4 13 16 26 27 44 20 7 33 48 27
21 10 1 15 15 44 19 36 16 26 22 45 35 7 3
Case 2
10 0 mm
2 10
44 3 2 43
12 1 31 18 23 40
39 1 13 2 22 4O 22 14 3 2 27 7
17 41 26 12 16 34 39 41 25 35 49 39| 12 I
46 37 32 14 35 3 41 21 47 31 5 46 34
26 19 28 9 4 1 17 11 28 13 45 34 15 2936
34 17 41 16 2 3 37 6 11 9 40 26 13 22
0.2.0
1.5
1.0
05
Case 4
Figure 4: True and reconstructed profiles of cracks with computed y>-value.
N. Yusa and K. Miya / Profile Reconstruction of Conductive Cracks 217
5 Conclusion
The present study have developed an inversion scheme that reconstructs crack profile
from eddy current signals by means of a neuro-fuzzy system. A neural network
simulated inverse mapping and a fuzzy system gave a criterion that indicated to
what extent the neural network had confidence in its outputs. The p parameter was
proposed in order to overcome the drawback of the original scheme. Cracks with
distributed conductivity were reconstructed from simulated eddy current signals
by means of the inversion scheme. The paper revealed that the parameter gives
criterion of credibility of reconstructed profile. It, is concluded that the scheme can
deal with modeled natural cracks and its reliability is greatly enhanced by the fuzzy
system that calculates the parameter without obtaining eddy current signals from
the reconstructed crack.
References
[1] T. Takagi and K. Miya, ECT round-robin tost for steam generator tubes. Journal of the
Japan Society of Applied Electromagnetics and Mechanics, Vol. S, No. 1, pp. 121-129, 2000.
[2] Y. Tsujikura, T. Nakamura, K. Matsueda and S. Masamori, Current maintenance activities
at nuclear power plants and in other industries, Journal of the Japan Society of Applied
Electromagnetics and Mechanics, Vol. S, No. 2, pp. 159-172, 2000. (in Japanese)
[3] W. R. Randle, B. D. Woody, Caution about Simulated Cracks in Steel for Eddy Current
Testing. Material Evaluation, January 19S9. pp. 44-48.
[4] Z. Badics, Y. Matsumoto, K. Aoki, F. Nakayasu, A. Kurokawa, Finite element models of
stress corrosion cracks (SCC) in 3-D eddy current NDE problems, Nondestructive testing of
materials, IOS Press, Amsterdam, pp. 21-29, 1995.
[5] N. Yusa and K. Miya, Numerical analysis of JSAEM round robin test. Electromagnetic
Nondestructive Evaluation(V), IOS Press, Amsterdam, pp. 325-332, 2001.
[6] F. Matsuoka and A. Kameari, Calculation of three dimensional eddy current by FEM-BEM
coupling method, IEEE trans, on Mag. Vol. 24, No. 1, pp. 182-185, 1988.
[7] N. Yusa, Z. Chen and K. Miya, Quantitative profile evaluation of natural cracks in a steam
generator tube from eddy current signals, International Journal of Applied Electromagnetics
and Mechanics, (to appear).
[8] Z. Chen, K. Miya and M. Kurokawa, Rapid prediction of edrly current testing signals using
A-4> method and database, NDT&E International, Vol. 32, pp. 29-36, 1999.
218 Electromagnetic Nondestructive Evaluation (VI)
F. Kojima et at. (Eds.)
IOS Press. 2002
1. Introduction
In eddy current testing (ECT) of steam generator (SG) tubes, signals are sometimes
rendered noisy by the presence of other structures, such as support plates and tube sheets
outside of test materials. Moreover, copper and magnetic deposits will also adhere to the
outer surface of the tube near the support plates and tube sheets. The support plates and
tube sheets here are sometimes made of ferromagnetic materials. To solve ECT inverse
problems, a fast forward solver is required because the ECT signals must be accurately
computed many times. Some fast pre-computed database approaches [1,2] have been
proposed, but a problem still exists because it is impossible to treat the ferromagnetic
materials. By the extension of these database approaches, a new method [3] is proposed to
solve the ECT problems including ferromagnetic materials as noise sources.
A round-robin test of SG tubes with noise source outsides is introduced by the Japan
Society of Applied Electromagnetics and Mechanics (JSAEM). The noise sources here
include support plates, tube sheets, copper and magnetic deposits, as well as their
combination. Crack shape reconstruction problems with support plate noise have been
solved [4] by a multi-frequency method and inverse analysis. However, the multiple noise
source problems have never been discussed as far as authors' knowledge. Experimental
results of the round-robin test using a transmit-receive type ECT probe are shown. Noisy
signals are processed by a multi-frequency method [5] and other methods as well. Several
artificial cracks are reconstructed and inverse results are shown in this paper.
The round-robin test includes eight kinds of test samples, as shown in Table 1. The
arrangement of tubes and noise sources is shown in Fig. 1. An ECT probe used here is
H. Huang and T. Takagi /Inverse Analysis of Steam Generator Tube ECT Signals 219
Deposits
Crack
(a) Straight tube and a support plate (b) Expansion zone and a tube sheet
Fig. 1 Arrangement of tube and noise sources
side of it is the magnetic deposits noises, and the right side is the support plate noises.
Signal shown here is the Y component of the complex measured signal. As a processing
method of ECT, the support plate noises are rotated to X direction. Much larger noises are
due to the magnetic deposits in Y component graph.
!<%)
Plate
-.100 -200 -100 0 100 200 100
Fig.3 Signal of NEL-0017 with a support plate Fig.4 Support plate and magnetic deposits
and magnetic deposits (400kHz) noises
H. Huang and T. Takagi / Inverse Analysis of Steam Generator Tube ECT Signals 221
Inverse analysis is performed from the signals processed by the base-line subtraction
method. Experimental results of noise-free cases of this probe are also available and the
transform coefficients from measurement data to numerical data are obtained using OD60%
as a standard. B-scan data are used, 21 scan points along the crack direction. Numerical
data and the transformed measurement data are compared in Fig. 6. When signals of
OD60% are used as standards, signals of OD20% also fit each other very well except some
residual noises.
Using the processed signals for the inverse scheme needs more time to compute the
forward problem, but there are no difficulties when using our fast signal estimator.
Combining the signal processing method with normal inverse approaches, the actually
depth of a crack can be predicted. It is a normal approach [4] to solve the inverse problem
by minimizing the least square error between the estimated signals and observation signals.
The following evaluation function is used in our studies:
2 /<pos| ; 2
(3)
where J is an evaluation function, x is the vector characterizing the shape of the cracks,
S'comp(x) are the predicted signals related to the vector x, and S'obs are the observation
signals. Af p o s = 21 is the total observation points and / is the point number. The steepest
descent method is applied to minimize the least square error.
222 H. Huang and T. Takagi / Inverse Analysis of Steam Generator Tube ECT Signals
.1 :
(a) From numerical data (NEL-0017) (b) From experimental data (NEL-0017)
(c) From numerical data (NEL-0021) (d) From experimental data (NEL-0021)
To apply the fast forward ECT signal simulator [3], a "suspect region" which includes
the possible crack region should be defined. For reconstructions of axial cracks, a suspect
region used here is a 9* 1.27*0.2mm3 block, which is divided into 9*5*1 elements.
Parameters are the depths of each 1mm column, along the crack direction. Shape
reconstructions of the sample NEL-0017 and NEL-0021 are shown in Fig. 7. Comparison
of reconstruct results and true results is shown in Table 2. Crack depth reconstructed from
experimental data agrees with the true depth well, and the predicted length of OD20%
(NEL-0017) is 1mm longer than the real value because of the residual noises. Database
used here is for axial cracks only. To reconstruct the circumferential crack shapes of sample
NEL-0018, NEL-0022, NEL-93114, NEL-95205, databases of circumferential cracks are
needed. After preparing a database for circumferential cracks, inverse problems of other
tube samples are also possible to be solved.
H. Huang and T. Takagi /Inverse Analysis of Steam Generator Tube ECT Signals 223
5. Summary
In eddy current testing of steam generator tubes, signals are sometimes rendered noisy
by the presence of other structures, such as deposits, support plates and tube sheets outside
of test materials. Crack shape reconstruction methods from ECT signals with combined
noise sources are discussed in this paper.
1. Round-robin test of SG tubes with noise sources outside is introduced,
arrangement of straight tube, expansion zone of tube, deposits, support plates and
plate sheets are shown.
2. Signal processing methods of the noised experimental signals are discussed.
Multi-frequency method is not able to reduced two or more noises at the same time.
Therefore, base-line subtraction method is used in this study.
3. Inverse analysis method of noised ECT signals is given. The reconstructed results
agreed with real shapes, especially the depths of cracks. Length of OD20% is
reconstructed 1mm longer than true one because of the residual noises.
Acknowledgment
This study was supported in part by the Research Committee on Nondestructive Evaluation Technology
by Eddy Current Testing of the Japan Society of Applied Electromagnetics and Mechanics through a grant
from 5 PWR utilities and Nuclear Engineering Ltd. This study was also supported in part by the Grant-in-Aid
for COE Research (11CE2003) by the Ministry of Education, Culture, Sports, Science and technology.
References
[1] Z. Chen and K. Miya, "ECT inversion using a knowledge-based forward solver," J. Nondestructive
Evaluation, Vol.17, No.3, pp. 167-175, 1998.
[2] T. Takagi, H. Huang, H. Fukutomi and J. Tani, "Numerical evaluation of correlation between crack size
and eddy current testing signal by a very fast simulator," IEEE Trans. Magnetics, Vol.34, No.5,
pp.2581-2584, 1998.
[3] H. Huang, T. Takagi and H. Fukutomi, "Fast signal prediction of noised signals in eddy current testing,"
IEEE Trans. Magnetics, Vol. 36, No. 4, pp. 1719-1723, 2000.
[4] H. Huang, T. Takagi and H. Fukutomi, "Crack reconstruction from noisy signals using a novel ECT
probe," Electromagnetic Nondestructive Evaluation (IV), IOS press, 2000, pp. 143-150.
[5] D. E. Bray and R. K. Stanley, "Nondestructive evaluation: A tool in design, manufacturing and service,"
CRC Press, 1997, pp.418-419.
224 Electromagnetic Nondestructive Evaluation (VI)
F. Kojima el al. (Eds.)
IOS Press, 2002
1 Introduction
The detection of a crack on steam generator tubes is a critical issue for the structural
integrity of the plants. Eddy current testing (ECT) is used for in-service inspection of tubes
in steam generators because of high detectability and rapid scanning process. A transmitter-
receiver pancake type probe coil is scanned along the inner surface of steam generator tubes.
Defects can be detected as a probe impedance or voltage trajectory from the receiver by
exciting the transmitter. The inversion techniques by parameter estimation problems have
been studied computationally and experimentally by many authors (e.g. [1]). However, those
are computationally intensive iterative procedures in which the accurate forward problems
must be solved numerous times. Recently, most efforts in computational savings are directed
to the use of database for an ensemble of the forward solutions. Huang et al [2] and Chen
et al [3] have been developed the current dipole method based on the reciprocity theorem.
In this method, once the forward model can be simulated for the material without flaws, the
ECT signals can be obtained by recalculating the corresponding model only defined on the
suspect region. Those method are very effective if the size of the suspect region become too
small. Banks et al [4] has been proposed the Proper Orthogonal Decomposition methodology
which makes it possible to produce the reduced order forward model. In this paper, a new
approach for parameter estimation is proposed for the sensitivity evaluation of the output
least square method. In order to cut the computational costs drastically, the model output is
directly reconstructed by data sets of the forward solutions. Thus the sensitivities of the cost
can be directly evaluated from the approximated observation space.
N. Okajima et al. / Crack Shape Identification of Steam Generator Tubes 225
The eddy current problems considered here involve a transmitter-receiver pancake type
probe coil as shown in Fig. 1 . The force current density vector is applied to the transmitter, and
then the impedance of the receiver is affected due to the existence of the eddy current excited
by the transmitter coil. Let (t, x) be the time t G [0, oo) and the location x = (xi, x 2 , x3) in
R3. Let VQ be the inspection area of the steam generator tube defined by
V0 = { x = (1)
where di, do and di denote the inner diameter, the outer diameter and the length of the tube,
respectively. The mathematical model of eddy current testing can be described by the mag-
netic vector potential A = (A\,Ai, A3) and the electric scalar potential </>. The mathematical
model can be expressed as
where / is an electric current of the coils, Eu and E? are the electric field with flaw absent
and the difference between the field with the flaw present and with the flaw absent, expressed
by
Eu (7)
Ef (8)
The defect considered here is assumed to be circumferential cracks outside the SG tube and
the profile is represented as shown in Fig. 2 where 7VC denotes the number of division of crack
length and where q = {qi}^ denote the crack profile parameters (the ratio of crack depth to
the nominal thickness of the SG tube). The admissible class of parameter q is assumed to be
given by the set
where tc and qu denote the thickness of the SG tube and the maximum depth of the crack,
respectively. Our aim is to recover defect geometries using probe impedance trajectories.
Parameter estimation problems are applied to our inversion technique. The usual procedure
can be formulated as a minimization of the fit-to-data functional:
(10)
with respect to the geometrical parameters q. Here AZj denotes the measurement impedance
data, while AZ(q) implies the model output corresponding to the defect parameter q. The
model output can be computed by applying the hybrid use of FEM-BEM model (5) combined
with the current dipole method. As is well-known, the minimizer of the fit-to-data functional
EXq) is to implement
= min (12)
r/e[o,oo)
The practical implementations of the algorithm (11) involve the crucial part of computational
volumes. Using the multivariate bilinear spline functions[6], the model output can be approx-
imated by
*pprox(q) = (14)
4 Computational Experiments
In our numerical experiments, the defect was assumed to be the circumferential crack
(width: 0.2 [mm], length: 4.77 [mm]) outside SG tube (inner diameter: d = 19.69 [mm] ,
thickness: tc = 1.27 [mm]) made from INCONEL600 (electric conductivity:CTO= 1.0 x
106 [S/m], magnetic permeability: ^0 = 1-0 x 10~7 [H/m]). Two crack patterns (a stair-step
crack and multiple cracks) were tested for computationbal experiments. The number of finite
elements and boundary elements were set as 1040 and 706, respectively. The frequency of
transmitter coil was 400[fc.//z]. The coils were moved along the crack as shown in Fig. 3. The
scanning positions of coils were set as
where r = 9.145[mm], mp = 20, and 0j were set by a unit of l.75[deg] from 73.41[de<7J to
106.59[de#]. The difference of angle between transmitter and receiver is a 25[de#]. Then,
the solutions were added random noises, and used as 'data' for our inverse algorithm. The
228 N. Okajima et al. / Crack Shape Identification of Steam Generator Tubes
q true
Noise Free 1% Noise 2% Noise 5% Noise 10% Noise
9i 80 80.05 79.68 80.06 79.60 78.25
92 60 60.08 60.18 60.18 61.40 63.04
93 60 60.10 58.94 58.94 54.10 49.86
94 40 40.00 40.55 40.55 43.27 47.39
95 40 40.01 39.99 39.99 40.00 39.86
96 20 19.10 19.76 19.76 18.53 16.75
AaPProx 1.148E-3 2.316E-3 8.816E-2 5.435E-1 2.169
Accuracy 936.3 105.2 55.97 21.33 10.06
q true
Noise Free 1% Noise 2% Noise 5% Noise 10% Noise
91 60 60.31 60.49 59.10 58.96 62.52
92 60 59.97 59.71 60.25 60.71 55.73
93 0 0.27 0.01 0.14 0.31 0.15
94 40 39.85 40.04 40.00 40.00 40.59
95 80 80.00 79.99 79.78 79.44 79.78
96 60 60.03 60.01 60.14 60.37 60.01
Aapprox 7.970E-4 1.112E-2 4.487E-2 2.747E-1 84.11
Accuracy 312.1 240.1 141.7 97.04 27.42
crack shape of initial guess was chosen as no crack. For the optimization routine of Eq.(14),
we used the gradient projection method. Tables 1 and 2 summarize the estimated parameter
results for the data without noise and with 1, 2, 5 and 10% relative noise. The average of
the elapsed time required for estimation procedures was about 84[sec]. In Tables 1 and 2, the
accuracies of the estimated values were evaluated by
tru
Accuracy = |q |q
_ qtrue| (15)
O.003 -0.002 -0.001 0 0.001 0.002 -0003 -0.002 -0001 0 0001 OOO2 0002 -0001 0 0001 0002
5 Concluding Remarks
A fast computational solver was proposed for recovering a crack shape of steam generator
tubes. Computational savings have been achieved by building the approximate observation
space with multivariate B-splines. A very fast recovering scheme was effectively tested with
computational experiments. Current efforts are directed to show the effectiveness of the pro-
posed method to the real experimental data.
6 Acknowledgements
This study was supported in part by the Research Committee on the Nondestructive Evalu-
ation Technology by ECT of the Japan Society of Applied Electromagnetics and Mechanics
through a grant 5 PWR utilities and Nuclear Engineering Ltd.
References
[1] F. Kojima, Computational method for crack shape reconstruction using hybrid FEM-BEM scheme based
on A - <p method, Studies in Applied Electromagnetics and Mechanics, 12 (1997) 279-286.
[2] H. Huang, T. Takagi, and H. Fukutomi, A novel crack reconstruction method for steam generator tube ECT
with sources outside, Review of Progress in Quantitative Nondestructive Evaluation, 20 (2000) 513-520.
[3] Z. Chen and K. Miya, ECT inversion using knowledge based forward solver, J.Nondestructive Evaluation,
17 (1998) 167-175.
[4] H. T. Banks, M. L. Joyner, B. Wincheski, and W. P. Winfree, Evaluation of material integrity using reduced
order computational methodology, Technical Report CRSC, TR99-30, North Carolina State University,
1999.
[5] Z. Badics, Y. Matsumoto, K. Aoki, F. Nakayasu, M. Uesaka, and K. Miya, Accurate Probe-Response
Calculation in Eddy Current NDE by Finite Element Method, Journal of Nondestructive Evaluation, Vol.
14, No. 4, 1995.
[6] C. K. Chui, Multivariate Splines, Vol. CBMS54, SIAM, 1988.
230 Electromagnetic Nondestructive Evaluation (VI)
F. Kojima el al. (Eds.)
1OS Press, 2002
1. Introduction
Eddy-current testing (ECT) is a very useful tool for the detection of defects in metallic
objects. It is applied in the area of inspections of steam generator pipes of atomic power
plants [1], wings of aero plane [2], and printed circuit boards [3]. [4]. and its usefulness has
been reported. But the inspection is not completed without a data processing for the analysis
of features of defects, including existence, position, shape, size, and direction. What is
important here is to carry out the reduction of the undesired components such as noise, offset
signal and its variation which disturb the correct estimation of defect features, simultaneously
with the image analysis. The wavelet-based denoising [6]-[8] is one good choice for this aim.
but generally their bases have nonlinear phase and it might lose correct information on
location of the signal in the wavelet region.
In this paper, we present an image processing method based on wavelet-like subband
technique without this problem. Namely, since bases with complete linear phase (LP) are
adopted, the decomposed signal keeps correct wave shape and phase information in the
transformed domain. The advantages and problems of the proposed method are discussed
through examples.
2. ECT Measurement
The defects on metallic object are examined by using ECT technique as depicted
T. Taniguchi et al. /A Processing Method of Eddy-Current Testing Image 231
in Fig. 1. The ECT probe shown here consists of a meander type coil for the excitation
of eddy-currents in the conductor and figure-of-eight (Fig. 1) or solenoid pick-up coil [4].
The ECT probe usually has a constant output, and detect disturbance of eddy-currents and
generates a peak signal formed by a pair of positive and negative pulses only in the place
where nonuniform distribution of conductor exists. In addition, the output signal contains
measurement noise and offset component (and its variation) which have bad influences to the
feature extraction of defects. Hence, we should remove those components before the analysis
of defect features. As one way to cope with this problem, we present an image processing
method utilizing rnultiresolution technique in the following section.
For the reduction of undesired components in ECT images, many types of filter have
been used. The wavelet-based method also can be interpreted as a kind of octave filter bank
[9], i.e., a set of band pass filters accompanied with sampling rate conversion. But here, it is
desirable that each filter has a linear phase, especially when a manipulation about amplitude
is carried out in the transformed domain. In this case, if the phase distortion results in a
change of waveform, we might have wrong amplitude information.
The system given in Fig.2 has a same structure as that of the discrete wavelet trans-
form (DWT) based method [8], but using exactly LP bases for analyzer and synthesizer, it
can hold correct spatial information also in the transformed domain.
The whole procedure of image processing is given as follows.
(i) The signal is decomposed into J +1 subbands {c.^, , c_ j^, d-j^} through the system
in Fig. 2 (a).
(ii) The subbands which contain undesired components but not defect signal are elimi-
nated (Frequency filtering).
(iii) To reduce the undesired components occupying the same frequency band as the defect
signal, the elements of the subbands not eliminated in (ii) are removed if their absolute value
is smaller than threshold A (Amplitude filtering) [10].
(iv) The signal is reconstructed through the synthesizer in Fig. 2 (b).
Steps (i) to (iv) are applied to each horizontal line of ECT images considering that
232 T. Tuniguchi et al. /A Processing Method of Eddy-Current Testing Image
the measurement is carried out along this direction and they can be dealt as independent
signal in this stage. Each filter is designed through iterative min-max method [11]. [12].
which enables us flexible choice of frequency characteristics.
After reconstruction, the correlation between the theoretical waveform of defect (which
is derived by computer simulation or actual measurement) and ECT image is calculated.
This operation has an effect to emphasize defect-like signal and simplify the image, espe-
cially when the defect signal is small [5]. Then the image is compared with a reference image
obtained from a sample previously known to be without defect. If a peak signal is detected
only in test image, it is considered to be coming from a defect since undesired components
have been already removed.
4. Examples
In this section, the effectiveness and problems of the proposed method are investigated
through examples.
The benchmark data (NEL98-25) shown in Fig. 4 (a) contains defect peaks as shown
inside of the broken ellipses, but due to the existence of the undesired components, the
discrimination of defects is very difficult. Hence the image processing method described in the
previous section is applied to this example. The filters {pk} and {qk} (amplitude responses
are in Fig. 3) are designed with the order of 5 and 11. respectively. Namely. 9 multiplications
are required to calculate one sample when the symmetry of them is considered. First.
T. Taniguchi et al. /A Processing Method of Eddy-Current Testing Image 233
10 20 30 40 50 60 10 20 30 40 50 60
the periodically extended image is decomposed into 4 levels and the signals which belongs
to coefficients {r_i^,c_2,fc,d-4,fc} or have an absolute value smaller than A = 0.0005 are
removed. Then the image is reconstructed, and consequently, the defect peaks become
clearer (Fig.4 (b)) compared to the original. The binarized image shows us the existence
and position of defects.
The results of PCB data processing are also shown in Fig.5. In this case, we should
carry out the comparison between reference (previously known to have no defect) and test
image to discriminate the origin of the peak signal, i.e., they are coming from defect or
distribution of wiring.
In spite of the effectiveness and the phase-distortion-less nature shown above, this
method has a problem as shown in Fig. 6. Both wavelet processing and the proposed method
clearly show the defect peak in undesired components, but residual high frequency compo-
nents are still observed in Fig.6 (c). The low regularity (i.e., low smoothness) of the given
bases is considered as a reason. This is the inferior point of the proposed method to wavelet
processing which is designed with high regularity.
5. Conclusions
In this paper, we have presented an ECT image processing approach using subband
technique for the extraction of the defect. The proposed method consists of two steps: (1) the
234 T. Taniguchi et at. /A Processing Method of Eddy-Current Testing Image
reduction of the undesired signals using multiresolution analysis and (2) actual discrimina-
tion of defect-based signals from others, based on the comparison of two images (Step (2)
is necessary for the case of PCBs). Using LP bases, the correct information on phase is
maintained also in the transformed domain, and the parameters of the filters can be chosen
more flexibly. Through examples, the ability of the proposed method to reduce the undesired
components is demonstrated together with its problem, the poor regularity of the frequency
characteristics.
As a future work, we should develop a design method of LP bases which have a higher
regularity. The examples in this paper concerns only with the exinstence and position of
defects. Hence another theme is the investigation of the usefulness of the proposed method
for the extraction of other defect features.
References
[1] R. Mol, "Evolution of inspection tools for SG tubes," Nuclear Europe Worldscan, vol.3 4,
pp.38-39, 1999.
[2] R. Leclerc. and R. Samson, "Eddy current array probe for corrosion mapping on ageing air-
craft." Review of Progress in Quantitative Nondestructive Evaluation, vol.!9A, pp.489-496,
1999.
[3] S. Yamada, H. Fujiki, M. Iwahara, S. C. Mukhopadhyay and F. P. Dawson, "Investigation of
printed wiring board testing by using planar coil type ECT probe," IEEE Trans, on Magnetics,
vol.33, no.5, pp.3376 3378, Sept. 1997.
[4] D. Kacprzak, T. Miyagoshi, S. Yamada, and M. Iwahara, "Inspection of printed circuit board
by planar ECT probe," IEEE INTERMAG99, GR-04, Aug. 1998.
[5] T. Taniguchi, D. Kacprzak, S. Yamada, M. Iwahara, and T. Miyagoshi, "Defect detection of
printed circuit board by using eddy-current testing technique and image processing," The 5th
International Workshop on Electromagnetic Nondestructive Evaluation, 1999-30, Des Moines.
Iowa, 1999.
236 7". Taniguchi et al. /A Processing Method of Eddy-Current Testing Image
[6] G. Chen, A. Yamaguchi, and K. Miya. "A novel signal processing technique for eddy-current
testing of steam generator tubes." IEEE Trans, on Magn.. vol.34, no.3. pp.642 648. March
1998.
[7] K. Kawata, S. Kumano, and M. Kurokawa "Signal processing of eddy-current signal by 2
dimensional wavelet transform.'' 8th MAGDA Conference. Hiroshima. 1999.
[8] T. Taniguchi. D. Kacprzak. S. Yamada, and M. Iwahara. "ECT image processing for PCB
inspection by using wavelet transform." Journal of JSAEM. vol.9, no.l. March 2001 (in print).
[9] C. K. Chui. "An introduction to wavelets." San Diego. CA:Academic Press. 1992.
[10] D. L. Donoho. "De-noising by soft-thresholding." IEEE Trans, on Information Theory, vol.41.
No.3. pp.613-627. 1995.
[1-1] M. Vetterli. "Filter banks allowing perfect reconstruction." Signal Processing, vol.10, no.3.
pp.219 244. 1986.
[12] M. Ikehara. A. Yamashita, and H. Kuroda. "Design of two-channel perfect reconstruction
QMF." Trans, on IEICE. vol.J75-A, no.8. pp.1333 1340. August. 1992.
Electromagnetic Nondestructive Evaluation (VI) 237
F. Kojirna et al. (Eds.)
IOS Press, 2002
Abstract. In this paper, we have carried out identification of pulse current distribution on a PCB
(printed circuit board) by using the magnetic sensor with high spatial resolutions. Experimental
examinations have shown that the time-varying of pulse current distributions can be identified by
means of our approach, which is expected to clarify electromagnetic phenomena on a PCB.
1. Introduction
Fig. 1 (a) shows the miniature magnetic field sensor having a maximum 250- n m spatial
resolution at a 6 dB degrading point, enabling the measurement of magnetic fields near LSIs
and dense-PCBs. The characteristics and the general specification of the sensor are shown
in table l(a), l(b). As shown in Fig. l(b), at frequencies below 3 GHz, the output is
approximately proportional to the frequency having the gradient of about 20dB/dec.,
permitting magnetic field calibration at less than 3 GHz. If the measurement target is a line
238 N. Masuda and T. Doi / An Approach for Identifying Current Distributions
on PCB, the current-calibration coefficient can be used to convert the measured magnetic
field over the line into current. The design is based on a precise glass ceramic multilayer
board, and the detector is 2 mm wide and 1 mm thick [3,4].
Fig. 2(a) and 2(b) show the simplified PCB, sectional diagram, respectively. Fig. 2(c)
shows the comparison of current waveform caluculated from measured voltage waveform
and that of calculated from measured magnetic field by integrating output voltage waveform
of magnetic field sensor. The voltage was measued at the terminating resistor (50 ohm) by a
differential high impedance probe. The maginetic fields were measured at the night of 20 n
m on the line close to the resistor. It is apparent from Fig. 2(c) that both current values are
agreed well. Thus, it is shown that the current waveform is measured with high accuracy by
means of the magnetic field sensor. Fig. 2(d) shows a magnetic field distribution X
measured by means of the magnetic field sensor. This magnetic field distribution is obtained
at the height of 5mm. Therefore the spatial resolution is degradated. Under such condition,
we consider identifiation of current distribution in the lines on PCB from locally measured
magnetic field.
240 N. Masuda and T. Doi /An Approach for Identifying Current Distributions
H2
(1)
", .
or
t = CX'. (2)
where C is system matrix. And n is a number of measurement points.
Rows of system matrix C corresponds to the spatial characteristics of the sensor.
Consequently, the system matrix C in eq.(2) becomes always a regular matrix. Then,
magnetic field distribution X' to be obtained is obtained uniquely by
x'=c-'x. (3)
Finally, currents in the lines on the PCB can be calculated only to apply the Ampere's
law to the solution X'
Fig.3(a) is spatial characteristics of the sensor, which have practically measured under
same condition in Fig.2. Each element of matrix C is obtained by substituting the value of
each position in Fig. 3(a) to Cy in eq. (1).
j = 1,23 n (4)
Fig. 4(a) shows an example of output voltage waveform of the magnetic field sensor.
The period of pulse is 100 ns (corresponding to 10MHz) and the duration time of the pulse
is 20 ns (corresponding to 20% duty). It is found out that the form of pulse is restored well
which indicates the usefulness of the data-processing procedure. The magnetic fields can be
converted to current by using conversion factor [3], thus waveform shown in Fig. 4(b) is
regarded as current waveform. The dc component of magnetic field waveform was not
close to zero. This dc component is caused by the slight voltage offset of oscilloscope. Fig.
5 shows the time varying magnetic field distributions across the PCB. The time validation of
number 3 line is different from others. Fig. 6 shows a stereoscopic 3-D view of time varying
current distribution. Shades of gray in the figures are correspondence with the value of
currents. There is 10.8 ns delays between the pulse of number 3 line and other lines. The
value of delay time is close to that of shown in table 2.
It is clear that the working current distribution and waveform are ^identified by
applying proposed method in this paper. Futhermore, the current waveforms are identified
as pulse shape.
242 N. Masuda and T. Doi / An Approach for Identifying Current Distributions
Fig.4. Waveform of magnetic field calculated from output voltage waveform of the sensor.
Fig. 5. Magnetic field distrbutions calculated from output voltage waveform of the sensor.
N. Masuda and T. Doi /An Approach for Identifying Current Distributions 243
200
400
Time [x0.2ns] 3 Line
number
600
800
Measurement Termination
position
\
High
frequency
current
\ 2mm \^
"\ l
"" 30mm
- 3 - 2 - 1 0 1 2 3
Measurement position k
distribution on grand plane. In other words, the magnetic fields generate from the currents
distributed on ground plane are included with the measured magnetic fields over a
microstrip line.
However, the currents on ground plane do not have little effect to the calculation,
because the equation (3) performs operation for canceling out the effect. Fig. 8 shows the
time varying current waveform distributed on the test board. It is apparent from Fig. 8 that
the current distribution is identified even if the PCB has microstrip construction.
5. Conclusions
In the present paper, we have proposed the approach for identifying high spatial
resolution current distributions, which utilizes spatial characteristics of a magnetic field
sensor. Identification of current distributions on a simplified printed circuit board have been
carried out by means of our approach. As a result, the experimental examinations show
validity of our approach to identify working current distributions on a PCB. Futhermore, we
have presented that the proposed method is also useful for identification of current
distribution on microstrip lines.
Reference*
(1]T Doi, S. Hayano and Y. Saito, "Magnetic field distribution caused by a notebook computer and its
source searching" , J. Appl. Phys., Vol. 79, No.8, pp.5214-5216, 1996.
[2] T. Doi and N. Masuda, "Leakage magnetic field source searching of micro processing unit on printed
circuit board" , Trans. IEEJ , Vol. 120-A. No. 10, pp 5214-5216, 2000 (in Japanese)
[3] N. Masuda, N. Tamaki, T. Watanabe, K Ishizaka, "A miniature high-performance magnetic-field
probe for measuring high-frequency currents" , NEC R&D, Vol.42, No.2, pp.246-250, 2001.
[4] T. Harada, H. Sasaki, and E. Hankui, "Time-Domain Magnetic Field Waveform Measurement Near
Printed Circuit Boards ", Electrical Engineering in Japan Vol. 125, No. 4, 1998, translated from Denki
Gakkai Ronbunshi, Vol. 117 - A, No. 5, pp. 523-530, May 1997
Electromagnetic Nondestructive Evaluation (VI) 245
F. Kojima et al. (Eds.)
IOS Press, 2002
Abstract. A wavelet basis function (WBF) neural network is proposed to estimate the defect
profile from eddy current inspection data of steam generator tubes. The calibration and phase
gating schemes are first employed. Calibration is used to optimize the frequency and sensitivity
settings, which are used to classifiy the actual defects relative to the reference defects. Phase
gating is employed to improve the visibility of the defect after calibration. A WBF neural
network is subsequently employed to map the signal appropriately to obtain the geometrical
profile of the flaw. Test results showing the effectiveness of the approach are presented. A
calibration curve method is also evaluated for a comparative study of results obtained with the
WBF neural network.
1. Introduction
Inverse methods are used when the parameters governing the behavior of a physical
process are to be estimated from a set of measured signals. However, the inversion
processes are usually ill-posed. In nondestructive evaluation (NDE), the solution to the
inverse problem involves prediction of defect parameters based on probe measurement
signals [1]. One of the earliest and most popular approaches used in industry is the
calibration curve method. An alternative approach involving the use of WBF neural
networks is presented here.
WBF neural networks allow functional approximation using multiresolution wavelets.
In this method, increasing or reducing the number of resolutions can produce multiple
scales of approximation. Several such networks with varying architectures and employing
different wavelets have been studied. Zhang and Beveniste [2] introduced wavelet
networks employing the first derivative of a Gaussian function, which satisfies frames
requirements, as a basis function. The locations of basis functions were determined by
using a density function and then optimized by employing a stochastic gradient algorithm
that is similar to the backpropagation algorithm. Zhang et al. [3] employ an orthogonal
scaling function as a basis function, which provides a unique representation of the function
being approximated. The locations of the basis functions are set using a dyadic scheme
with the number of basis functions at resolution L. This neural network is generally not an
RBF neural network since scaling functions can be radially nonsymmetrical. Lim and
Bakshi et al. [4,5] presented WBF neural networks using both scaling function and wavelet
as basis functions. This study employ the WBF neural network to estimate defect profiles
from eddy current signals obtained in the inspection of steam generator tubes [6].
246 J. Lim ct al. / Wavelet Basis Function Neural Network
The multiresolution nature of the expansion implies that the number of resolutions in the
network architecture can control the level of prediction accuracy and the confidence of the
network prediction can be measured by observing the improvement in outputs.
In this paper, calibration and phase gating schemes are first employed. A calibration
procedure is employed to optimize the frequency and sensitivity settings, which are used to
classifiy the actual defects relative to the reference defects. A phase gating scheme is
employed to improve the visibility of the defect after calibration. Data from sets of sensors
are fused to estimate the defect profile as shown in Figure 1, where WBFNN represents a
wavelet basis function neural network. The data was collected using a pancake coil
rotating probe at four excitation frequencies; 400 kHz, 300 kHz, 200 kHz, and 100
kHz.The four signals are combined using weighted average fusion method into a signal. A
WBF neural network is f nally employed to map the fused signal appropriately to obtain
the geometrical profile of the defect. A calibration method is also evaluated for a
comparative study of results obtained with the WBF neural network.
2. Calibration
Calibration
Curve
Sensor 2 _^j Calibration j-M Phase Gating Defect
Profile
WBFNN
3. Phase Gating
Although the calibration procedure improves the visibility of the defect, additional
artifacts are introduced due to phase wrapping. In order to minimize the artifacts, we
employ a thresholding approach to highlight the defect indication more clearly. The
threshold is computed dynamically based on the statistical properties of the signals within
a swath of data.
The phase gating step includes two thresholding procesures; one for magnitude and the
other for phase. In magnitude thresholding, the signals whose magnitude are less than the
thresdhold level are treated as noise and replaced with zeros both in the horizontal and
vertical components of eddy current signal. In phase thresholding, the signals with phase in
the wrong quadrant (3 and 4) are eliminated. Signals whose phase lie between 5 and 175
are retained.
The weighted average is one of the simplest and most intuitive methods of data fusion.
This technique takes a weighted average of redundant information provided by a group of
sensors [7]. The weighted average of n sensor measurements x, is defined as
n n
x = ^T w,x, , where ^Tw(. =1 and 0 < w, < 1 (4-1 )
1=1 (=1
The weights account for the differences in accuracy between sensor measurements and are
estimated using the variances of the signals as follows:
J_
(4-2)
V
calibration curve
The calibration curve relates the phase of the eddy current signal with the depth of the
defect. We employ linear interpolation techniques to estimate the depth, when a signal
whose phase is intermediate between sample values is to be analyzed. For example, we
first assume that the calibration curve / is piece-wise linear between each of the known
mapping points. The relationship used to find the depth when the signal phase is b is
given by:
(5-1)
c-a
(6-1)
where L is the number of resolutions, A', is the number of translated basis functions at
resolution j, k is the translated point, \p[ is a dilated and translated version of a mother
wavelet, and ck is the center of basis function at translation point k. The architecture of a
J. Lim et al. / Wavelet Basis Function Neural Network 249
typical WBF neural network implementing the above equation is similar to radial basis
function (RBF) neural networks in that both neural networks have a single hidden layer. In
contrast to RBF neural networks, a WBF neural network employs a family of wavelets as
basis functions and has sets of wavelet function nodes depending on the number of
resolutions. In this method, the number of resolutions is either increased or decreased to
control the level of prediction accuracy. The number and location of the basis function
centers C is determined using the well known ISODATA clustering scheme, a is the
standard deviation that is used as a bias of network, R is the length of input eddy current
signal, Q} is the number of training data set, Q2 is the length of defect profile, and W is
the weight matrix relating the input phase and the output defect profile.
Scaling and wavelet functions are employed as basis functions so that the norm error of
the function approximation is minimized using a minimum number of functions [5]. An
intuitive procedure is used to select these functions. Wavelet basis function with a high
degree of smoothness are usually chosen for obtaining good generalization in the case of
multidimensional problems with limited training data. In this paper, the Mexican hat
function is employed as a wavelet.
(6-2)
where a is the standard deviation. This wavelet decays rapidly in the time and frequency
domains and the frame constant is very near unity. A Gaussian radial basis function is used
as a scaling function.
2
The centers of the basis functions can be estimated using the -means clustering
algorithm or one of its several variants. The number of centers at each resolution can be
determined heuristically. Previous studies have shown that the performance of the network
is highly dependent on the location and number of centers. Therefore, an LMS algorithm is
250 J. Lim et al. / Wavelet Basis Function Neural Network
typically employed in order to obtain better performance instead of using the k-means
clustering algorithm not only for determining the optimal locations, but also for
determining the widths of the centers. However, such an algorithm is unsuitable for
training a multiresolution network.
In this paper, we use a dyadic center selection scheme for calculating the location and
number of centers [5]. Since the task at the lowest resolution is to coarsely approximate
the function, the locations of basis functions are chosen using the ISODATA clustering
algorithm. This corresponds to the first term in the approximation given by Equation (6-1).
Subsequently, new centers, which correspond to the second term in Equation (6-1), are
calculated using the dyadic selection scheme to obtain finer approximations. The dyadic
selection scheme is similar to the dyadic dilation scheme in wavelet theory. The centers at
the next resolution are expanded by dividing the Euclidean distance between adjacent
centers at the previous resolution. If too may basis functions are used and overfitting has
occurred, a center pruning algorithm is used in order to eliminate unnecessary centers.
The algorithms were evaluated with eddy current data obtained from defects machined
in tube samples. The data was collected using a three coil rotating probe (two pancake
coils and a plus-point coil) at several excitation frequencies. In this paper, data from
pancake coil rotating probe was used. Figure 5 and Figure 6 show the outputs of the
calibration and phase gating stages respectively. Figure 7 indicates the calibration curve
obtained from calibration data.
Figure 5. Calibration result of vertical component; (a) original, (b) after calibration.
J. Lim et al. / Wavelet Basis Function Neural Network 251
Figure 6. Phase gating result of eddy current signal; (a) before phase gating, (b) after phase gating
Figure 8. Defect characterization result based on calibration curve method and WBFNN.
Figure 8 describes the defect characterization result based on calibration curve and WBF
neural network. The test data was a deep 59% ID axial defect. The result shows that WBF
neural network shows considerable promise in predicting defect profiles. Additional
improvement can be obtained using an adaptive boundary extraction algorithm to locate
the defect area and using data fusion techniques to improve the characterization result.
References
[1] L. Udpa and W. Lord, A Discussion of the Inverse Problem in Electromagnetic NDT, in Review of
Progress in Quantitative Nondestructive Evaluation, vol. T, New York, 1986, pp. 375-382.
[2] Q. Zhang and A. Benveniste, Wavelet networks, in IEEE Transactions on Neural Networks, 3, no. 6,
1992, pp. 889-898.
[3] J. Zhang, G. G. Walter, Y. Miao, and W. N. W. Lee, Wavelet neural networks for function learning, in
IEEE Transactions on signal processing, 43, no. 6, 1995, pp. 1485-1497.
[4] B. R. Bakshi and G. Stephanopoulos, Wavelets as basis function for localized learning in a
multiresolution hierarchy," AICheE J., vol. 39, no. 1, 1992, pp. 57-81.
[5] J. Lim, S. S. Udpa, L. Udpa, and M. Afzal, Multisensor Fusion for 3-D Defect Characterization using
Wavelet Basis Function Neural Network, To appear in Review of Progress in Quantitative
Nondestructive Evaluation, vol. 20, 2001.
[6] J. Stolte, L. Udpa, and W. Lord, Multifrequency Eddy Current Testing of Steam Generator Tubes using
Optimal Affine Transform, in Review of Progress in Quantitative Nondestructive Evaluation, vol. 7A,
New York, 1988, pp. 821-830.
[7] M. A. Abidi and R. C. Gonzalez, Data fusion in robotics and machine intelligence, Academic presss inc.,
San Diego, 1992.
Electromagnetic Nondestructive Evaluation (VI) 253
F. Kojima et a I. (Eds.)
IOS Press, 2002
1. Introduction
NDE of nuclear reactors is the most pressing problem. The generic interests are focus on
the heat transfer tubes and pressure vessels. The currently employed NDE methods are
eddy currents, ultrasonic testing , X-ray measurements etc., but the subject of investigation
is the crack; the discovery of micro-cracks and monitoring of their development. The new
NDE method is introduced recently and the subject is the degradation before the cracks
appear by the use of relationship between magnetism and lattice defects in A533B steel [1]
The mechanisms of magnetic changes due to metal fatigue are different in low carbon steels
such as A533B, and SUS 304 stainless steels; the former is due to the magneto-elastic
coupling energy and strain field of lattice defects and the latter is the magnetic transition
due to martensitic transformation.
Another interest of NDE in stainless steels is detection of stress corrosion cracking
(SCC). The SCC is induced by three factors: the environment, the internal stress and the
chemical in homogeneity. The internal stress is important in SCC as well as metal fatigue.
NDE of the internal stress and the chemical inhomogeneity would be made possible by
using the magnetic methods in connection with the martensitic transformation. The NDE of
internal stress before the appearance of cracks is of our interest and the purpose of the
present study is to investigate the relation between magnetic properties and lattice defects
due to plastic deformation in SUS 304 stainless steel. The magnetic transition due to plastic
deformation has been investigated in SUS 304 stainless steel for a long time. The data are
inconsistent. The reason is that the magnetic transition depends on the experimental
conditions, such as Ni content, test temperature and the method of introducing plastic
254 Y. Okino el al. / Magnetic Transition Due to Tensile Deformation
Table 1 Chemical composition . annealing temperature and grain si/c number of samples
Sample Annealing Grain si/c 1
Chemical composition
number temperature number j
1 C Si Mn P Ni Cr S Fc I323K 6
2 0.06 0.43 1.16 0.33 8.33 18.44 0009 Bal I533K 1 ;
The chemical composition and the grain size of SUS -304 stainless steel samples are
shown in Table 1. Shape and size of sample is shown in Fig 1 In order to change the grain
size, various annealing temperatures (1323K and 1533K) were used after keeping the
samples for 1 hour in the furnace and followed by water quenching Grain sizes of samples
were measured using a optical microscope. The grain size of sample 1 is about 40 urn and is
about 100 urn for sample 2
The samples were plastically deformed using Instron tensile machine at room
temperature with a strain rate of 0.5 mm/min to examine the relationship between
martensitic transformation and tensile deformation The iii;irK-nsitt phast t i a n s f i ' i i : i -
to plastic deformation was detected by the SQUID magnciometei and \-ray d i t l u u .
Coercive force was measured by vibrating sample magnetometry (YSM)
Results of tensile test are shown in fig.2. It shows work hardening in each sample
Sample 2 with coarse grains show striking work hardening at near 400MPa This may be
due to an larger extent of martensitic transformation in sample 2. Yield stress is about 270
MPa for Sample 1 and about 220 MPa for sample 2.
Fig 3 shows the relationship between
magnetization and magnetic field measured by
SQUID magnetometer in sample 1 Magnetization ?
increases with the increase of tensile stress At rupture 5
stress the saturation magnetization reached a J
maximum of about 55emu/g
Fig.4 shows the relationship between magnetization f
and magnetic field for sample 2. In case of sample 2
with coarse grain size very high value of
magnetization ( 1 1 0 emu/g) was obtained at rupture
stress As compared to sample 1 the value of
magnetization is about twice in sample 2.
Fig 2 Stress-strain curves
of samples
Y. Okino el ai /Magnetic Transition Due to Tensile Deformation 255
Fig.5 Amount of martensite phase versus Fig.6 Amount of martensite phase versus
nominal stress of sample 1 nominal stress of sample 2
Figs. 5 and 6 shows the amount of martensite phase (%) versus tensile deformation for
sample 1 and 2. The amount of martensite phase (%) was measured from the saturation
magnetization and by X-ray diffraction methods, and was represented by A and T
triangles, respectively. At rupture stress about 80 % martensitic phase was detected in
sample 2 whereas in sample 1 the martensitic phase was about 40 %.
Magnetic techniques are more sensitive to the amount of ferromagnetic martensitic phase
even at very low tensile stress where the X-ray diffraction method fail to detect the amount
of martensite. For X-ray diffraction pattern, martensite phase is not detected in low strain
area. Martcnsitc phase less lhan 3 7i cannot he measured by X-ray method, because in low
strain area the amount of martensite phase is small. Amount of martensile phase is about
40(7( at rupture stress and increases rapidly alter 6()OMPa in sample 1.
In sample 2 marlensitic phase increases rapidly after tensile stress of a 350 MPa. As
compared with sample 1 martensitic transformation in sample 2 occurred in the early stage
of plastic deformation.
Generally saturation magneli/ation of magnetic material is represented by equation (1)
[2-3].
A/ fl / / / >
(1)
y Inconlinnoiisly movement domain wall are . 10 Pom,mi wall are placed under restraint
placed under restraint bolh ends both end's hy marlensite phase
Y. Okino et al. /Magnetic Transition Due to Tensile Deformation 257
because neighboring atomic restriction is small and lattice fault at near grain boundary
disappeared in grain boundary, therefore, it is hard to for nucleation to occur. Accordingly,
if grain sixe is small, parent phase is stable and amount of martensitic transformation is low.
Fig. 8 shows the relationship between coercive force measured by VSM and the
deformation stress. The direction of magnetic field is parallel to the tensile stress direction.
In each sample, coercive force decreases with the increase of tensile deformation.
Coercive force is given by the following equation:
Y
=2/ s //cosfl ' (4)
/
where 7" is the energy of domain wall at unit volume, r is the radius of curvature.
Accordingly, coercive force at c is given by:
Discontinuous movements of domain wall are placed under restraint at both ends.
marlensite phase and austenile phase exist at constant crystallographic direction.
Accordingly coercive force decreases as tensile deformation increases.
4. Conclusions
Magnetic methods noticed sensitively a very small quantity of martensile phase, which
can't be noticed by X-ray method. Accordingly, the use of magnetic methods are effective
for NDE of martensitic transformation induced by plastic deformation.
In work-induced martensitic transformation, grain boundaries obstruct the growth of
martensite phase. Coarse grain decreases grain boundaries and resulted in martensite phase.
Accordingly, grain boundary makes ms temperature lower.
Finally, the initial condition of SUS 304 stainless steel is important in NDE using
matinetic methods.
KtTcrcnces
[ 1] S.Takahashi, J.Hchigoya and /.Moloki " Magnetixation curves ol plastically deformed l;e inelals and
alloys " JOURNAL OF A1TUKD PHYSICS, Vol.87, No.2. pp 805, J a n u a r y 2000.
|2] S.S.Meeker, M.CLSioul. K.P.Slandhammer and J.Smith: Melall. Trans. A. vol.lM. p619. (1982)
[3] L.K Murr. K.RSl and hammer and S.S. Meeker: Met all. Trans. A. vol.13, p627. (1982)
258 Electromagnetic Nondestructive Evaluation (VI)
F. Kojima el al. (Eds.)
IOS Press, 2002
1. Introduction
SUS 304 stainless steel is commonly used in chemical and nuclear power plant
components because of its excellent tensile strength and resistance to corrosion. However,
the SUS 304 stainless steel undergoes structural degradation in service due to irradiation,
metal fatigue and stress corrosion. Degradation causes transformation of paramagnetic
austenite to ferromagnetic martensite in SUS 304 stainless steel.
The importance of studying NDE techniques is to determine the secular
degradation i.e., the crack initiation before it appears in SUS 304 stainless steel. The
NDE techniques useful for crack detection are magnetic measurements of martensitic
transformation due to internal changes in the material and forecasting [1-2]. By
monitoring the amount and nature of phase changes occurring in stainless steel structure,
one can determine when to take preventive measures and perhaps, intervene at early
stages of the process. Thus the risk of cracking can be minimized while at the same time
increasing reliability of stainless steel components.
The main objective of this research is to get the quantitative relationship between
material damage and martensitic transformation induced by plastic deformation at various
annealing temperatures, and heat treatment sensitization by high precision magnetic
measurement which is a powerful non destructive evaluation technique.
K. Mumtaz el al / The Potential ofNDE by Magnetic Methods 259
2. Experimental Procedure
The material used for the investigation was cold rolled plate of SUS 304 stainless
steel. The chemical composition of the SUS 304 stainless steel is shown in Table 1. All
the specimens used in these experiments are first solution annealed in vacuum at a
temperature of 1323 K for 1 hour and then water quenched.
3. Results
3.1 Effect of annealing temperatures on plastic deformation
The stress-strain curves for SUS 304 stainless steel at room temperature are
shown in Fig. 1. A most remarkable fact in Fig. 1 is the variation of tensile stress, which
decreases with increasing annealing temperatures between 300 and 1453 K. At room
temperature and 853 K with less than 3 % strain a rapid increase in flow stress occurs and
the tensile stress increases quite rapidly. Analogous stress-strain curves are obtained at
annealing temperatures between 1003 K and 1453 K which are characterized by an easy
tensile deformation up to fracture. It can be seen that, in the annealing temperatures
between 1003 K and 1453 K, the strain is high (up to 55 %). The curves are parabolic
type and are smooth up to 55 % plastic strain. The temperature at which the maximum
ductility occurs is between 1203 K and 1453 K. At these annealing temperatures, the
strain-hardening increase continuously with the strain and the martensite is constantly
formed little by little during deformation up to fracture. When the total amount of
martensite formed is small, the strain percentage is relatively large and the total
elongation increased. The elongation does not simply depend on the total amount of
martensite induced during tensile deformation but also on the conditions of martensite
formation. It should be noted that the dislocation density and twins are not sensitive to
crack initiation up to 55 % plastic strain without showing micro cracks above annealing
temperatures of 1103 to 1443 K. In the case of specimen annealed at 1553 K with large
grain size two stages of deformation can be seen in Fig. 1 an early stage followed by
rapid hardening stage. High amount of martensite transformation might cause both easy
260 K. Mumtaz et al. / The Potential ofNDE b\ Magnetic Methods
deformations in the early stage and rapid hardening in the later stage of deformation. The
martensite is formed plentifully in the specimen annealed at this temperature with plastic
strain is about 35 %. The decline in percentage strain at 1553 K and the increase in tensile
stress is found to be essentially dependent on the martensitic transformation. The total
elongation to failure is also decreased. Crack initiation may be induced prematurely at
this annealing temperature, due to the presence of large amount of martensite.
Fig. 1 The stress-strain curves of SUS 304 stainless steel deformed at room temperature
after annealing at various temperature
in the plastically deformed specimens annealed at various temperature originates from a'
martensite transformation.
Fig. 4. The coercive force versus the annealing temperature in SUS 304 stainless steel
grain boundary
Cr deficiency of O.B
grain boundary
(a) (b)
Fig.5 Schematic diagrams (a) showing Cr2:)C6 precipitates along grain boundaries and (b) the chemical
heterogeneity process by HTS near the grain boundary in SUS304 stainless steel
magnetization is 0.09 emu/g before annealing which shows that 6 x I0'2 % ferromagnetic
state in volume is included in the specimen.
Figure 7 shows the saturation magnetization dependence of the annealing
temperature without plastic deformation. The saturation magnetization does not change
by the annealing below the temperature of 923 K. The saturation magnetization increases
by the annealing above 923 K and takes the maximum value of 0.23 emu/g at 1023 K.
The magnetic transition indicates that 0.1% phase changes to martensitic by the annealing
at 1023 K.
4. Discussion
The above results have shown that magnetic methods offered the possibility of
detecting secular degradation i.e., martensitic phase transformation in the SUS 304
stainless steel by an applied magnetic field. The martensite is the phase with superior
magnetic properties. The ferromagnetic susceptibility is larger, about tens of thousands
times in comparison with paramagnetic state. The magnetic susceptibility is very
sensitive in the phase transition. Result shows specifically how the diffusionless
transformation nucleates from austenite, given the free energy difference between the
austenite and martensite phases at the M s temperature. It is evident from the result that a'
martensitic phase occurred only regardless of whether the martensite transformation is
induced by plastic deformation after annealing at various temperatures or heat treatment
sensitization.
Fig.7 The saturation magnetization versus Fig.8 Temperature dependence of the free energies
sensitization temperature in SLJS304 stainless steel. of a' and y phases
All known attempts for calculating the transformation of austenite into martensite
had started from the Gibbs free energies of austenite face centered cubic (y) and
martensite body centered cubic (a1) phases. These free energies are a function of
chemical composition. Figure 8 shows temperature dependence of the free energies of a'
and y phases. Considering the two-phase a1- y field, there is a critical temperature To
where the free energies of austenite and equally composed martensite are identical.
Above TO, a' martensite is of higher energy than y austenite. Beneath To, the energy will
be released when y austenite transforms into a' martensite. At martensite start
temperature MS<TO an energy difference is available as a driving force for the
transformation into a' martensite.
K. Mumtaz et al / The Potential ofNDE by Magnetic Methods 265
The martensitic transformation does not occur by the external factor such as
plastic deformation at above the critical temperature TQ. The martensitic transformation
occurs spontaneously below Ms. The difference between two free energies AGo^"a' at
MS is equivalent to the energy for the martensitic transformation overcoming the
interface energy between the two phases at MS- Between MS and TO the mechanical
energy, AGM/~ a > is necessary for martensitic transformation by plastic deformation,
which is equivalent to the difference AG0Y~a' and the chemical potential AGcY~a'. The
mechanical driving force is caused of AGM^"a at T'. In tensile deformation the driving
force necessary for the initiation of the martensite transformation is attained above the
normal M s temperature. Therefore a change in Ms can be accounted for by simply taking
into consideration the contribution of additional mechanical free energy of the reaction.
We have investigated the martensitic transformation due to plastic deformation
after annealing at various temperatures. By plastic deformation, the martensitic
transformation occurs completely only in the specimen annealed at 1553 K. It is
confirmed by the microscopic observation that the grain size becomes large (100 (am) by
annealing at 1553 K. The present study shows that the larger the grain size is, the more
stable the martensitic state becomes.
Results show that the value of saturation magnetization and coercive force
depends on the grain size. The general aspect of grain boundary is different from the
present result that large grain boundaries make the martensite phase stable. The grain
boundary may exert two influences: 1) dislocation pile up at the grain boundary and 2)
internal stresses. With increasing grain size the dislocations at the grain boundary
accumulates, which in turn increase the amount of martensite transformation. As the
grain size is increased, new martensite is nucleated near the grain boundary. The presence
of dislocation pile ups have the role of increasing dislocation resistance, hence hardening
the matrix, at the same time generating considerable internal stress. It is thought that the
generation of more defects and the increase of local pile up stresses leads to the observed
increase in the martensitic transformation after plastic deformation at 1553 K annealing.
These dislocations pile up may be acted as pinning sites to domain wall moment
and domain wall energy increases at these pinning site. In response to large grain size due
to high annealing temperature, the domains within individual grains must reorient to
minimize their energy, therefore the magneto crystalline anisotropy energy also decreases.
As a result, low value of coercive force He (20 Oe) is obtained at 1553 K.
The reason why large grain size is so much more susceptible to martensitic
transformation may be the changes in slip behavior, dislocations, lattice defects and
tensile properties within this phase have been suggested as possibilities.
Coercive force decreases with increase of saturation magnetization in Fig. 4.
Coercive force gives us information of the shape of martensitic phase. Below 853 K
annealing temperature, the martensitic phase would have the uni-axial form such as the
needle shape or the bi-axial form such as plane shape and they distribute in seclusion.
This martensitic phase has a strong magnetic anisotropy and coercive force is large, when
its volume is small. As the martensitic transformation advances, the uni-axial and bi-axial
phases unite with each other and make a big magnetic cluster. The shape magnetic
anisotropy decreases as the saturation magnetization increases and disappears completely
266 K. Mumtaz et al. / The Potential ofNDE by Magnetic Methods
References
1. S. Takahashi, J. Echigoya and Z. Motoki. J. of applied Physics. 2000. Vol. 87
No. 2, p. 805
2 S. Takahashi, J. Echigoya, T. Ueda, X. Li and H. Hatafuku, J. of Materials
Processing Technology, 2001, Vol. 108. p. 213
3. G. H. Eichermann and F. C. Hall, Tran. Am. S. M. 1953. Vol. 45. p.77
New Sensors
This page intentionally left blank
Electromagnetic Nondestructive Evaluation (VI) 269
F. Kojima et al. (Eds.)
1OS Press, 2002
Development of a Microwave
Focusing Sensor for Nondestructive
Evaluation of Materials
Yang JU1, Kenji SARUTA1, Masumi SAKA1 and Hiroyuki ABE2
'Department of Mechanical Engineering, Tohoku University, Aoba 01,
Aramaki, Aoba-ku, Sendai 980-8579, Japan
2
Tohoku University, 2-1-1 Katahira, Aoba-ku, Sendai 980-8577, Japan
1. Introduction
Microwave nondestructive evaluation (NDE) has become more and more an important
technique with the emergences and applications of new and advanced materials. The primary
advantage of microwave NDE is that the inspection result is not based on the density but the
intrinsic properties of the materials, thus, materials can be inspected sensitively. For example,
in the microwave region, the variation of permittivity for dielectric materials is significantly
larger than the contrast of density. That is why microwave inspection is more sensitive than
the other techniques for testing dielectric materials. Moreover, microwave NDE does not need
any coupling medium because microwave can propagate well in air [1]. In the recent years,
microwave NDE has been used for detecting cracks [2], CT imaging [3], moisture
determination [4], fiber orientation testing [5], thickness measurement of thin films [6], and
so on.
Microwave signal is an electromagnetic wave having a frequency between 300 MHz and
300 GHz (wave length between 1 m and 1 mm). It has a divergent property which usually
decreases the sensitivity and the spatial resolution of the measurement. Therefore, in some cases,
to obtain higher sensitivity and spatial resolution, the standoff distance between the sensor and
sample has to be sufficiently small. For example, in the case of microwave NDE of the
delamination in 1C packages, the standoff distance was required to be 0.5 mm [7]. However,
in some practical environments, such as the detection of fatigue cracks in the piping of power
plants, a larger standoff distance is extremely needed.
In the present paper, an ellipsoidal focusing sensor was developed. A larger standoff
distance was obtained with a higher sensitivity and spatial resolution. By using the focusing
sensor, closed fatigue cracks were detected successfully.
2. Focusing Sensor
The developed focusing sensor consists of a reflector and a feed as shown in Fig. 1. The
geometry of the reflector is shown in Fig. 2. The use of the reflector is based on the concepts
270 K Ju et al. / Development of a Microwave Focusing Sensor
Dimensions in mm Sample
of ray optics [8], according to which the reflector takes the divergent rays from a point source
and converts them into a required beam. The reflector has an ellipsoidal reflecting surface
which is made of metal. A horn antenna is used as the feed, and is placed at one focus, A, of
the ellipsoid. The incident microwave irradiated from the horn antenna is reflected by the
ellipsoidal reflector and is focused at another focus, B, of the ellipsoid. The standoff distance
of the focusing sensor is 11.8 mm. Since the sample is placed far from the sensor aperture, the
higher order modes excited due to the reflector located in the near field of the feed can be
ignored. The higher order modes are evanescent and decay rapidly from the aperture.
3. Experimental Procedure
Dimensions in mm
distinguish the five strips. Here, a is used as a measure of the spatial resolution of the
sensor. The sample was measured by the focusing sensor and the open-ended waveguide
sensor, respectively, with a standoff distance of 11.80 mm, where the scanning pitch in the
x-direction was 0.25 mm.
In addition, to verify the capability of the focusing sensor to detect small defects, fatigue
cracks were inspected. The descriptive geometry of the fatigue cracked specimen is shown
in Fig. 6. Two small fatigue cracks were introduced in different specimens, SI and 52,
respectively. Specimens were prepared as plates having the initial dimensions of
330x35x25 mm, from austenitic stainless steel, AISI 304. To introduce a fatigue crack, an
initial notch was situated in the L-S orientation of the plate, where the direction of crack
growth is considered perpendicular to the longitudinal rolling direction, L, and parallel to
the short transverse direction, S, of the material (ASTM code for crack plane orientation).
272 Y. Ju et al. / Development of a Microwave Focusing Sensor
Maximum
stress intensity Frequency Stress ratio Crack depth
Specimen
factor [, (Hz) R J(mm)
(MPa-m1/2)
SI 22 6 0.1 2.0
S2 22 6 0.1 2.8
The fatigue crack in each specimen was grown from the tip of the initial notch by cyclically
loading the plate in four points bending of tension and tension on the dynamic testing
machine. During the process of fatigue, crack growth was monitored from both sides of the
specimen and the maximum value of the stress intensity factor, A^nux, was determined with
the average value of the crack depths measured on both sides of the specimen. The stress
ratio, R, defined as a ratio of the minimum to the maximum value of the stress intensity
factor, was maintained during the crack growth. After the desired depth of the crack was
reached, the plate was machined and polished to remove the initial notch, leaving a
true fatigue crack in the remaining material. The conditions used for introducing fatigue
cracks in the specimens are listed in Table 1. Since fatigue cracks introduced in the same
material under the same conditions as the present experiment had been evaluated to be
closed by ultrasonic technique [8], the fatigue cracks introduced in SI and 52 are
considered to be closed. The depth of the fatigue cracks, d, indicated in Table 1, is the
average value of the crack depths measured on the both sides of the specimen. In the
experiment, the scanning origin was located at the center of the crack as shown in Fig. 6,
and the scanning pitch was 0.1 mm. The standoff distance was 11.80 mm.
Figure 7 shows the phase measured by using the focusing sensor and the waveguide sensor
for the aluminum strip sample. When a is 1.5 mm, the result obtained by using the focusing
sensor shows five sharp peaks significantly corresponding to the aluminum strips as indicated
by the arrows. However, in the case of the waveguide sensor, the five strips cannot be
Y. Ju et al. / Development of a Microwave Focusing Sensor 273
Fig. 7 Measurement results of the phase for the aluminum strip sample obtained by
using the focusing sensor and the open-ended waveguide sensor
Fig. 8 Measurement results of the amplitude for the aluminum strip sample
obtained by using the focusing sensor and the open-ended waveguide sensor
separated. Therefore, the spatial resolution of the focusing sensor regarding to the phase
measurement was evaluated to be 1.5 mm. Figure 8 shows the amplitude measured by
thefocusing sensor and the waveguide sensor for the aluminum strip sample. Same as the
results of the phase measurement, when a is 1.5 mm, the result shows five sharp peaks
significantly in the case of the focusing sensor, however the five strips cannot be separated in
the case of the waveguide sensor. Therefore, the spatial resolution of the focusing sensor for
the amplitude measurement was also evaluated to be 1.5 mm.
In the measurements of the phase and the amplitude, operating frequency was 109.8 GHz.
At this frequency, a resonance occurred due to the reflection of microwaves from the
aluminum strips and the sensor aperture. By using this resonant frequency, the sensitivity
and the spatial resolution become higher.
Figure 9 shows the measurement results for the fatigue cracks obtained by using the
focusing sensor. An operating frequency was 103.4 GHz. Around x=Q mm, where the fatigue
crack is located, a peak can be observed. It indicates that the focusing sensor can detect fatigue
274 Y. Ju el al. / Development of a Microwave Focusing Sensor
Fig. 9 Measurement results for the fatigue cracked specimens obtained by using the
focusing sensor
cracks on the metal surface effectively. It should be noted that so far the usual microwave
techniques were not able to detect such fatigue cracks. In addition, the magnitude of the peak
as shown in Fig. 9 enlarges corresponding to the increase of the depth of the fatigue crack. It
indicates a possibility to evaluate the depth of the fatigue cracks. In this case, since only the
variation of microwave responses are used, the calibration of the network analyzer is not
needed.
5. Conclusions
To increase the standoff distance between the sensor and sample, a focusing sensor
utilizing an ellipsoidal reflector was developed. The spatial resolution of the focusing
sensor was evaluated to be 1.5 mm when the standoff distance was 11.8 mm. By using the
focusing sensor, fatigue cracks on the metal surface were detected successfully. The
experimental results show a possibility to evaluate the size of fatigue cracks. The present
technique provides a good prospect for the Further applications of microwaves in NDE of
materials.
Acknowledgements
The authors would like to thank Mr. M. Mikami of Tohoku University for preparing the
specimen. This work was partly supported by Japan Society for the Promotion of Science
under Grant-in-Aid for Scientific Research (BX2) 13555023 and 13555192, and the
Ministry of Education, Culture, Sports, Science, and Technology under Grant-in-Aid for
COE Research 11CE2003.
References
[1] D. M. Pozar, Microwave Engineering, John Wiley & Sons, Inc., 1998.
Y. Ju et al. / Development of a Microwave Focusing Sensor 275
[2] N. Qaddoumi, E. Ranu, J. D. McColskey, R. Mirshahi and R. Zoughi, Microwave Detection of Stress-
Induced Fatigue Cracks in Steel and Potential for Crack Opening Determination, Reserch in
Nondestructive Evaluation 12(2) (2000) 87-103.
[3] T. Nakajima, H. Sawada and I. Yamaura, Microwave CT Imaging for a Human Forearm at 3 GHz,
IEICE Transactions on Communications E78-B(6) (1995) 874-876.
[4] Z. Ma and S. Okamura, Analysis and Elimination of the Reflection Influence on Microwave Attenuation
Measurement for Moisture Determination, IEICE Transactions on Electronics E80-C(10) (1997) 1324-
1329.
[5] K. Urabe and S. Yomoda, A Nondestructive Testing Method of Fiber Orientation by Microwave,
Advanced Composite Matrials 1(3) (1991) 193-208.
[6] H. C. Han and E. S. Mansueto, Thickness Measurement for Thin Films and Coatings Using Millimeter
Waves, Research in Nondestructive Evaluation 9(2) (1997) 97-118.
[7] Y. Ju, M. Saka and H. AW, NDI of Delamination in 1C Packages Using Milimeter-Waves, IEEE
Transactions on Instrumentation and Measurement 50(4) (2001) 1019-1023.
[8] S. Silver, Microwave Antenna Theory and Design, Peter Peregrinus Ltd., 1984.
[9] S. R. Ahmed and M. Saka, Quantitative Nondestructive Testing of Small, Tight Cracks Using Ultrasonic
Angle Beam Technique, Materials Evaluation 58(4) (2000) 564-574.
276 Electromagnetic Nondestructive Evaluation (VI)
F. Kojima el al. (Eds.)
IOS Press. 2002
Abstract. The diagnosis of material degradation and the detection of defects in cast duplex
stainless steel have been studied using ultrasonic techniques. This paper describes an
electromagnetic micro sensor consisting of a coil and ferrite core as a transducer for detecting
ultrasonic waves. When this sensor is used in combination with the ultrasonic transmitter, it
can detect the electromotive force caused by the oscillation at the surface of the material.
Results demonstrating the ability of the transducers are presented.
1. Introduction
Due to the good corrosion resistance, high strength and good weld ability of cast
duplex stainless steel, it is widely used in chemical plants, nuclear power plants and other
applications. Cast duplex stainless steel has a double-phase structure, comprised of ferrite
phase and austenite phase. The ferrite content of cast duplex stainless steel ranges from 8
to 30%.
When cast duplex stainless steel is exposed to high temperatures of about 300*C for an
extended period of time of 30,000 hours or longer, the material gets affected by thermal
aging in the ferrite phase. Once thermal aging occurs, in metallurgical terms, spinodal
decomposition progresses in the ferrite phase, finally causing change in chromium
concentration. At the same time, magnetic permeability of the ferrite phase also changes.
As for mechanical changes, fracture toughness decreases, and hardness increases.(l-4)
These mechanical degradations must influence the transmission characteristic of ultrasonic
waves that pass through the material.
In this research, an ultrasonic tool for assessing thermal aging in cast duplex stainless
steel has been studied. As the first step, an electromagnetic micro sensor consisting of coil
and ferrite core was developed.
When this sensor is used in combination with the ultrasonic transmitter, it can detect
the electromotive force due to the oscillation at the surface of the material with the
magnetic bias field. Several kinds of electromagnetic acoustic transducer(EMAT) or
piezo-electric ultrasonic transducer were used to load the oscillation on the surface of the
material. Since the sensor has dimensions that are smaller than 1mm, it can detect the
oscillation in a very small part on the material. Further, it is possible to evaluate the
degradations of the mechanical properties of cast duplex stainless steel due to thermal
aging by using the electromagnetic micro sensor.
Y. Kurozumi and M. Nishikawa /Electromagnetic Micro Sensor 277
Detection Co
0.4mm ^H 0.5mm 2mm
2mm
3. Experimental
3.1 Ultrasonic detection signals by EMS
Figure 4 shows the ultrasonic
detection signals measured with a Distance
forged stainless steel test piece.
20mm
The distance between transmitter
-0.2
and receiver has been varied
30mm
from 20 mm to 40 mm. -0.4
Ultrasonic waves were detected
with sufficient sensitivity. -0.6 40mm
Acoustic velocity can be -0.8V
SOpsec
calculated based on relation
between the flight time of the
first wave and the distance. In Ultrasonic detection signal (Forged Steel)
this case, acoustic velocity is
approximately 3,100 m/sec.
Figure 5 shows the ultrasonic
detection signals measured with a
cast stainless steel test piece. The
-0.5
distance between the transmission
and reception probes had been
-1.0
varied from 11 mm to 20 mm. 19mm
Ultrasonic waves were detected
-1.5V
with sufficient sensitivity even in 40 SOpsec
the case of cast stainless steel.
Acoustic velocity was measured
Figure 5 Typical ultrasonic detection signal (Cast Steel)
by the same method and was
approximately 2,900 m/sec.
These data indicate that the EMS has sufficient sensitivity as an ultrasonic wave
detection sensor.
4. Conclusion
The fundamental characteristic of the electromagnetic micro-sensor were studied and
C-scan images of ultrasonic wave propagation by non-aged and aged cast stainless steel
test pieces were obtained by the sensor. Conclusions made at this point include the
following:
- The electromagnetic micro-sensor has sufficient sensitivity as an ultrasonic wave
detection sensor.
- The resolution of the electromagnetic micro-sensor is 0.5 mm (maximum).
- The distribution of intensity of ultrasonic waves on the surface of cast stainless steel
has been obtained.
This paper describes the application of the electromagnetic micro-sensor to cast duplex
stainless steel. This sensor is expected to be applicable to many other cases.
References
1. O. K. Corpra, H. M. Chung, "Long-Term Aging of Cast Stainless Steels: Mechanisms and Resulting
Properties", NUREG/CP-0091, Vol.2, pp407-428,1987
2. T. Hardin, W. Pavinich, W. Server, "Implication of Thermal Ageing of Cast Stainless Steel and
Weldments", Procof Int Nucl Plant Aging Sympo, pp3 53-362,1988
3. O K Corpra, H M Chung, "Kinetics and mechanism of thermal aging embrittlement of Duplex Stainless
Steels", DOE Report, CONF-870839-6, 32p, 1987
4. O K Corpra, H M Chung, "Long Term Embrittlement of Cast Duplex Stainless Steel in LWR Systems :
Semiannual Report October 1986-March 1987", NUREG/CR-4744, Vol.2, ANL-87-45, 1987
This page intentionally left blank
Signal Detection
This page intentionally left blank
Electromagnetic Nondestructive Evaluation (VI) 285
F. Kojima et al. (Eds.)
IOS Press, 2002
Kenzo MIYA
International Institute of Universality,
1-4-6 Nezu, Bunkyo, 113-0031 Tokyo, Japan
Abstract. A powerful iterative solver for system of equations has been developed. This
solver, Generalized Vector Sampled Pattern Matching (GVSPM), enables us to obtain the
solution of any ill-posed linear system equations, e.g., having rectangular or singular matrix.
The key idea is that the objective function is the angle obtained by an inner product
between the input vector and solution system of equations. This paper reviews the
Generalized Sampled Pattern Matching (GSPM) method, which is the original of proposed
method, and then introduces the GVSPM method along with solution algorithm. The
theoretical background and applications to inverse source problems on magnetic fields are
described. From the applications, the GSPM and GVSPM methods tend to lead the
discontinuous and continuous solutions, respectively. However, the GVSPM has the
advantages of stability and convergence.
1. Introduction
the GSPM is possible to solve for any types of linear equations, the convergence process as
well as mathematical background was unclear at the time. Recently, mathematical
background has been clarified, and the convergence process has been proved [7]. This
leads to improvement of the GSPM.
This paper reviews the GSPM method, and then introduces the improved method,
Generalized Vector Sampled Pattern Matching (GVSPM). The key idea of the GVSPM is
that the objective function is the angle obtained by means of inner product between the
input vector and solution system of equations. The convergence process is described in
analytical way. The applications and comparisons of the GSPM and GVSPM are
demonstrated along with the inverse source problems on magnetic fields. As a result, the
applications reveal that the GSPM and GVSPM methods tend to lead the discontinuous
and continuous solutions, respectively. Moreover, the solution by means of GVSPM
converges with smaller number of iterations for computation.
Y = CX, (1)
where Y and X denote the n-th order input- and w-th order solution/output- vectors,
respectively; C is an n by m rectangular matrix. Even if m=n case, the matrix C is not
always positive definite due to the formulation of inverse problems. (1) can be rewritten by
x=
Y= ,C,, k * *J'C = [C, C2 . Cj. (2)
(2) means that the input vector Y is represented by means of linear combination of column
vectors C/, /=!, 2,..., m, in the system matrix C. The principle of the SPM method is to
search for the patterns representing by the column vectors C,s in order to satisfy the input
vector Yin (1).
Y
Y = * C ' . i = l,2,...,m, (3)
' '
where y is called pattern matching index. If a column vector C, is the most dominant, then
correspoding element of the solution vector X is increased by the quantized discrete value.
From the next calculation steps, the searching patterns are overlapped by the selected
column vector, namely,
,..= v[c.+ct] /=
'' |Y||C,+CJ'
H. Endo et al. / Generalized Vector Sampled Pattern Matching Method 287
where the superscript (k) refers to the k-th iteration. Moreover, C* is the selected pattern in
terms of A:-1-th iteration. (4) means that the amplitude of the solution is approximated by
the concentrating rate of similar solution pattern of C*. The calculation continues until the
pattern matching index y/^ takes 1. This means the obtained pattern (C,+C*) gives the best
pattern matching to the input vector Y. Finally, increment the elements constituting the
output vector X in each of iterations yields the GSPM solution. In this method, calculating
(4) carries out to each of column vectors, therefore, it is essentially required enormous time
and number of iterations for computation.
or Y'=C'X', (5)
where the prime (') denotes the normalized quantities. (5) means that the normalized input
vector Y' is obtained as a linear combination of the weighted solutions JCj|C,j/|Y|, /-I, 2,...,
m, with the normalized column vectors CV|C/|, /- 1, 2,..., m. It should be noted that the
solution vector X could be obtained when an inner product between Y' and C'X' becomes
1. This is the key idea of the GVSPM method.
' * "-t-i*
(6)
CX (
C'X' (*)
Then the solution \(k) is obtained when the function J(X(k)) converges to
f(-\f(li)\ ^ 1 (1\
J\lL )>l. {')
This is the objective function of the GVSPM solution. The objective function of the GSPM
is the angle between the input vector and pattern incrementally composed of column
vectors while those of the GVSPM is the angle between the normalized input vector and
output system of equations. Thereby, evaluation of (7) needs only once an iteration.
Y()
.A. r"T v
- U I , f9\
\o)
/^iV'()
AY"" = Y ' - C X (0) . (9)
C'X'
288 H. Endo et al. / Generalized Vector Sampled Pattern Matching Method
When the deviation AY' becomes zero vector, the objective function (7) is automatically
satisfied. Modification by the deviation vector AY'**"" gives the -th iterative solution
vector X' w , namely,
,7 A y,U-i>
'1 (10)
= C" (
C'X'
S = I-C"C (12)
This means the convergence condition is independent of initial solution in (8). Furthermore,
the solution is always evaluated by means of (6) so that the solution only depends on the
system matrix C.
Let X be characteristic value of the state transition matrix S. Then the determinant of
symmetrical matrix is obtained:
lm - S\ = =0 (13)
It is obvious that the moduli of off-diagonal elements in (13) take less than 1 because of the
normalized column vectors of matrix C\ namely,
, \< 1 (14)
Suppose the modulus characteristic value | A | takes more than 1. Then the column vectors
in (13) become linear independence because of (14). In such a case, the determinant in (13)
is not zero so that the condition | i | < 1 should be satisfied. Therefore, it is proved that the
GVSPM is always carried out on stable iteration.
4. Applications
To visualize the current vector distribution from locally measured magnetic field is one of
the typical inverse problems and fundamental of NDE. Let us consider a unit loop current
model illustrated in Fig. 1[9]. The magnetic field H caused by a loop coil shown in Fig.
l(b) can be analytically obtained as follows:
(15)
where K(K) and E(K) are the first- and second- kind elliptic integrals, respectively. The
system of equations is based on (15) in this example. In this case, the vectors Y and X in
(1) are composed of the measured magnetic field H and current / expressed by (15),
respectively.
(a) Unit loop array (b) An element of the unit loop array
Fig. 2 shows the measured magnetic field distribution generated by 3 excited coils. At first,
each component of the 3-dimensional magnetic field is measured at a surface over the
target, as shown in Figs. 2(b)-(d). Number of measured points is 16 by 16. Therefore, the
input vector Y in (1) becomes 256th order, namely n=256. Second, the system of equations
obtained by (15) is solved in terms of the z component of magnetic field by the GSPM and
GVSPM methods. The estimation points are set to 10 by 10, 16 by 16 and 32 by 32 in this
example. In these cases, numbers of elements in output vectors are 100, 256 and 1024.
Finally, 2-dimensional current vector distribution can be obtained as in Figs. 3 and 4.
2 4 6 8 10 12 14 16 2 4 6 8 10 12 14 16 2 4 6 8 10 12 14 16
The GSPM and GVSPM realize Figs. 3 and 4, respectively. In cases of m<n and m=n, each
of the solutions yields fairly good approximation. On the other hand, in case of m>n, the
GVSPM gives better result than the GSPM in respect of current flowing directions. Since
the GSPM method carries out element-to-element evaluation described as (4), then it needs
a large number of iterations. The results of GSPM and GVSPM have spent 200000 and 100
iterations on computation, respectively. The GVSPM realizes faster computation as well as
good approximation in this case.
Fig. 3 2-dimensional current vector estimation by means of GSPM (/r=256)
Next, searching the current flowing position is carried out using measured magnetic field.
Information coming from the current can only be observed by specific sensor such as
solenoidal coils. In most cases, the current distribution is observed by sensor signal
including some physical space characteristics such as Green's relationship.
Let us consider the one dimensional magnetic field distribution scanned by a solenoidal
sensor coil. Then, the output voltage would be obtained like Fig. 5, and it takes maximum
voltage at the position 100 cm. Using this physical space characteristic of the sensor coil,
the exact current flowing positions are investigated. In this case, the vectors Y and X in (1)
are corresponding to the measured magnetic field obtained by the sensor and the exact
current distribution to be searched, respectively. The system matrix C, shown in Fig. 6, is
composed of the physical space characteristic. This scheme is called deconvolution.
150
Fig. 5 Physical space characteristic of a sensor coil Fig. 6 Elements of system matrix
H. Endo et al. / Generalized Vector Sampled Pattern Matching Method 291
Fig. 7 shows the measured magnetic field caused by two current lines. The measurement is
carried out by a solenoidal coil over the current line surface. Number of measured points is
41 and number of measured points of physical space characteristic is 201. Therefore, Y and
X respectively become the vectors with 41 and 201 orders, namely, m=201, n=41. By
means of the GSPM and GVSPM, Fig. 8 shows the deconvoluted results. The result of
GSPM in Fig. 8(a) reveals the positions of the exact currents has been estimated with much
higher accuracy than that of GVSPM iri Fig. 8(b). This nature is caused by iteration
algorithm. Because the GSPM method carries out element-to-element evaluation, the
solution having discontinuous characteristics is easy to be obtained. On the other hand,
since the GVSPM evaluates with vector type iteration described above, then it tends to be
continuous solutions. The results of GSPM and GVSPM have spent 200000 and 100
iterations on computation, respectively.
5. Conclusions
This paper has introduced the GVSPM as one of the powerful iterative solvers for linear
system of equations. The original version of the GSPM requires an enormous computation
time while the GVSPM algorithm makes it possible to reduce the computation time and the
reliablity of the solution is clearly verified. The distinguished feature is that the objective
function is the angle between the right and left hand terms in linear system of equations.
This makes it possible to solve any types of simultaneous equations, i.e., having row- as
well as column- wide type rectangular system matrices. The convergence process has been
described analytically. As the applications of GSPM and GVSPM, the current searching
problems from locally measured magnetic fields have demonstrated the usefulness of our
solution strategy. Solving for the ill-posed system of equations is inevitable for the inverse
problems. However, the GSPM and GVSPM make it possible to select the physically
existing solution although this solver never uses the matrix inversion. Particularly, the
GSPM and GVSPM tend to be discontinuous and continuous solution, respectively.
292 H. Endo el al. / Generalized Vector Sampled Pattern Matching Method
References
[1] G. Anger, Inverse Problems in Differential Equations, Plenum Press, New York, 1990.
[2] Y. Saito, E. Itagaki, and S. Hayano, A Formulation of the Inverse Problems in Magnetostatic Fields and
Its Application to a Source Position Searching of the Human Eye Fields, J. Appl. Phys., Vol. 67, 1990,
pp.5830-5832.
[3] H. Saotome, T. Doi, S. Hayano, and Y. Saito, Crack Identification in Metallic Materials, IEEE Trans.
Magn., Vol. 29, No. 2, 1993, pp. 1861-1864.
[4] H. Saotome, J. L. Coulomb, Y. Saito, and J. C. Sabonnadiere, Magnetic Core Shape Design by the
Sampled Pattern Matching Method, IEEE Trans. Magn., Vol. 31, 1995, pp. 1976-1979.
[5] Y. Saito and K. Yoda, Method of Generating Energy Distribution, Japanese Patent Application
408315279, 1996.
[6] K. Yoda, Y. Saito, and H. Sakamoto, Dose Optimization of Proton and Heavy Ion Therapy Using
Generalized Sampled Pattern Matching, Phys. Med. Biol., IOP Publishing, Vol. 42, 1997, pp 2411-2420.
[7] D. Sekijima, S. Miyahara, S. Hayano, and Y. Saito, A Study on the Quasi-3D Current Estimation"
Trans. IEE of Japan, Vol. 120-A, No. 10, 2000, pp. 907-912 (in Japanese).
[8] G. Strang, Linear Algebra and its Applications, Academic Press, New York, 1976.
[9] H.Takahashi, S.Hayano, and Y.Saito, Visualization of the currents on the printed circuit boards, IEEE
Vis99 Late Breaking Hot Topics Proceedings, 1999, p37-40.
Electromagnetic Nondestructive Evaluation (VI) 293
F. Kojima etui. (Eds.)
IOS Press. 2002
1. Introduction
The detection and sizing of defects is one of the key technologies for mechanical plant life
management based on fracture mechanics. Improved performance of the available
non-destructive inspection techniques is always required for better detection sensitivity,
better quantitative evaluation and better field applicability. In this work we report on a new
non-destructive inspection (NDI) and evaluation (NDE) technique, based on high frequency
electromagnetic field measurement. In [1], the characterisation of artificial surface defects in
metals was performed by using antennas. The crack depth was evaluated based on the
analysis of data for electromagnetic waves reflected from the metal. However, the angle of
incidence equals the angle of reflection for such waves. Taking into account the fact that the
walls of real cracks are neither flat nor perpendicular to the surface, techniques utilising
reflection are not really very suitable for the characterisation of fatigue cracks. Therefore, a
new technique for the detection and characterisation of surface cracks in metals using a Small
Loop Antenna was developed [2, 3]. When an alternating field is applied to a metallic
material, either directly or indirectly, the existence of defects on the surface of the material
causes changes in the field around them. As a result, the changes in the electromagnetic field
generated by this change in the alternating field can be detected on or near the surface of the
material by using a small antenna. The advantage of this system is that it is possible for the
probe to measure defects without direct contact with the specimen. Additionally, such a
system can detect cracks in ferromagnetic material as well as paramagnetic material [4]. On
the other hand, the measurement time is relatively long in comparison with other non-contact
NDI techniques, because this system requires a conductive wire to be attached to the
specimen surface in order to induce the alternating field.
In this paper we describe the use of a Small Loop Antenna combined with an
induction wire to create a novel non-contact NDI technique. The ability of this sensor to
detect real defects is demonstrated by measurements on an actual metallic specimen with a
fatigue crack of 29mm in length.
294 Y. Sato el al. / Crack Detection by a Small Loop Antenna
A schematic illustration of the principles of the Small Loop Antenna technique is shown in
Fig. 1. When an alternating signal is applied in the conductive wire (induction wire), an
induced signal is generated in the opposite direction beneath the surface of the metal. The
changes in the induced signal flow around a crack lead to a change in the magnetic field in the
vicinity of a crack. This magnetic field has its largest magnitude inside the crack, but part of it
penetrates outside the crack and can therefore be detected by the Small Loop Antenna.
The measurement system consists of a network analyser, an X-Y-Z
sensor-positioning table and a personal computer (Fig. 2). The network analyser measures the
amplitude ratio and phase difference of the transmitted wave compared to the incident wave
[5]. Its operating frequency range is between 30kHz and 6GHz. In this work, the frequency of
the transmission wave was chosen to be IGHz, based on our preliminary experiment [6]. A
computer controlled X-Y-Z positioning table controls the motion of the antennas.
A loop antenna was used to measure changes in the electromagnetic field in the vicinity of the
cracks. The loop antenna contains a coaxial line with a Cu inner conductor and a Cu outer
conductor, as well as a loop that is also closed by a solid Cu conductor. The solid Cu
conductor is soldered at both ends to the inner and outer conductors of the coaxial line. A
schematic illustration of the loop antenna is shown in Fig. 3 (a).
This antenna enables us to measure the potential drop U (t) generated at the output of
the loop antenna, which is expressed as follows [7, 8]:
U(t)=j(oNAB (1)
where j is the imaginary unit, ft; is the angular frequency, TV is the number of coils, A is the
loop area and B is the value of the component of the vector of the magnetic flux density
within the loop, in the direction perpendicular to the loop. Accordingly, the amplitude of U is
proportional to B within the loop antenna [9].
In [2], [3] and [4], it is necessary to fix an induction wire onto the specimen surface in order to
make measurements, which prevents the implementation of simple, high-speed
measurements for a quick inspection process. Accordingly, a new design of antenna sensor
combined with an induction wire has been developed. A schematic illustration of this new
sensor is shown in Fig. 3 (b).
The main design parameters are the length of the induction wire (L), the distance
between the antenna and the induction wire (5), and the loop diameter of the antenna (D).
The main criterion for improving the sensitivity of the new sensor is to increase the
frequency of the applied signal, because a higher frequency yields larger measured values.
However, it is difficult to supply a high frequency signal to the conductive wire without any
devices being connected to it. Accordingly, the resonance phenomenon of the incident and
reflected waves was considered to supply a high frequency alternating signal and to from the
standing wave within the induction wire. If a standing wave is formed within the induction
wire, a high-power alternating signal exists in the induction wire. A IGHz-alternating signal
can form a standing wave in an induction wire 125mm in length.
The measured signals are strongly affected by the distance between the antenna and
the induction wire, 5, because an electromagnetic field is formed by both the supplied and the
induced signals. Furthermore, the electromagnetic field generated by the supplied signal only
has a negative effect on the measured signal, and only the electromagnetic field generated by
the induced signal has to be measured because information about the target material can only
be included on this signal. Six different values of 8 (3.5,4,4.5, 5, 5.5, and 6mm) were tested,
and the maximum signal changes measured above the defect were compared. We found that a
5 value of 5mm was the optimum distance to obtain sufficient signal intensity and to avoid
the effects of the supplied signal.
The antenna diameter was set at 7mm, and diameters of 5mm and 10mm were also
tested. The results of these showed that the smaller diameter antenna gave very small
measurement values, while the larger diameter decreased the sensitivity.
A Cu conductive wire with a rectangular cross section of 2mm x 0.035mm was fixed
296 Y. Sato et al. / Crack Detection by a Small Loop Antenna
at the bottom of the sensor as an induction wire, and the tip of the loop antenna was set at a
distance of 0.2mm from the surface of the specimen.
4. Measurement procedure
Specimens containing two types of defects were prepared in order to evaluate the
applicability of the Small Loop Antenna technique. One of them had artificial defects and the
other contained a fatigue crack.
The design of the specimens with artificial defects, which were made of a
paramagnetic material (Type-316 stainless steel), is shown in Fig. 4 (a). The specimen
contained three artificial defects (Electric Discharge Machining slits, known as EDM slits) of
different depths (called TP-1). These defects all had the same length (10mm) and width
(0.5mm), but had depths of 1mm, 3mm and 5mm respectively.
A fatigue crack was produced by using 4 point bending fatigue crack propagation
test on a plate of paramagnetic material (Type-304 stainless steel), which contained an EDM
slit as a crack initiator, (starting AK was 17.8 MPaVin at the bottom of the initial crack). In
order to make a specimen that contained only a fatigue crack the EDM slit was sliced off after
finishing the fatigue test. The shape and dimensions of the specimen with the fatigue crack
(called TP-2) are shown in Fig. 4 (b). According to surface observations and measurement by
an optical microscope, the length of the fatigue crack was about 29mm.
The origin of the X-Y-Z co-ordinate system was set at the center of the specimen.
The X-axis was defined as being directed along the length of the defects, with the Y-axis
perpendicular to the length of the defects, as shown in Fig. 4 (a). The Z-axis was directed
perpendicular to the surface of the specimen. During the measurements, the loop antenna was
oriented parallel to the Y-axis.
Y. Sato et al. / Crack Detection by a Small Loop Antenna 297
The first set of the measurements with the new probe was performed on specimen TP-1. Fig.
5 (a) shows the distribution of the ratio of the amplitude, R, across the artificial defect. R is
plotted against the position of the centre of the induction wire. R represents the ratio of the
amplitude of the potential drop \U (t)\ of the transmitted wave to that of the incident wave (Fig.
2). The measurement line is parallel to the Y-axis. Measurement was performed when the
induction wire was positioned along the line of the centre of the crack (X=0) and the Small
Loop Antenna was aligned along the crack edge (X=-5). It can be seen from Fig. 5 that as the
sensor approaches the slit, R drops slightly and then rises up again. When the sensor is above
a defect, R reaches a maximum value.
A schematic illustration of the electrical signals and magnetic fields in the vicinity of a crack
is shown in Fig. 6. A change in the induced signal flow around a defect leads to a change in
the magnetic field in the vicinity of the defect. This magnetic field has its largest magnitude
within the defect, but part of it penetrates outside the crack and can be detected by the Loop
Antenna. Consequently, the maximum value for R can be measured above the defect.
The detection sensitivity, AdB, is defined as the maximum change of R above the
crack. Fig. 7 shows the relationship between the detection sensitivity and the slit depth
obtained from Fig. 5. The defects can be detected quantitatively.
298 Y. Sato et al. / Crack Detection hy a Small Loop Antenna
Fig. 8. is a map showing the ratio of amplitudes measured on the specimen with a fatigue
crack by the new sensor. The fatigue crack is located along the line Y=0, from -14.5 to 14.5 in
the X direction. The fatigue crack is clearly detected by the new sensor, although it is difficult
to determine the crack length from the signal distribution. This is due to the asymmetry of the
sensor. Accordingly, in order to improve the signal quality, further development of the sensor
is suggested with special attention to the sensor symmetry.
25
ratio of amplitude , dB
0 -5 -10
position in the Y direction , mm
Fig. 8 Ratio of amplitude maps measured by the new sensor on the specimen with fatigue crack
6. Conclusions
The development of our Small Antenna technique has been summarised. A new type of
sensor, a Small Loop Antenna probe with an induction wire has been developed for
non-contact NDI. The detection performance of this sensor for real defects was demonstrated
by the measurement of a specimen with fatigue crack 29mm in length.
In order to improve the signal quality, further development of the sensor is suggested,
with special attention being given to the sensor symmetry.
300 Y. Sato el al. / Crack Detection by a Small Loop Antenna
Acknowledgements
Partial support for this work by the Grant-in-Aid for COE (Centre of Excellence) Research
(No.llCE2003) from the Japanese Ministry of Education, Culture, Sports, Science and
Technology is gratefully acknowledged.
References
[ 1 ] M. Saka, Materials Evaluation by Using Microwave -Feasibility Proceedings of the Annual Meeting of the
Japanese Society of Mechanical Engineering, Tokyo, JSME, 2000, pp.11-12.
[2] N. Sato, T Shoji, Y. Sato, M. Sato, Development of a novel electromagnetic technique for detection of
cracks by use of micro antenna, Transactions of the Japan. Society of Mechanical Engineers, 63-632, 1999,
pp.925-931
[3] Y. Sato, T. Shoji, Nondestructive Inspection Based on Electromagnetic Technique -Potential Drop and
Micro Antenna techniques-, Proceedings of The Sixth International Conference on Material Issues in
Design, Manufacturing and Operation of Nuclear Power Plants Equipment, 1,2000, pp.271-281
[4] K. Yagi, Y. Sato, T. Shoji, Inspection of Metal Surfaces Containing Cracks by Small Antennas, Review of
Progress in QNDE20, in press
[5] Vector Network Analyser seminar text, Hewlett Packard Co, 1997
[6] K. Yagi, Y. Sato, T. Shoji, Crack Detection by Means of Electromagnetic Field Measurement with Micro
Antennas, Proceedings of APCFS & ATEM01,2001, pp.280-285
[7] R. C. Johnson, H. Jasik, Antenna engineering handbook, second edition, McGraw-Hill, 1984, pp.2-7
[8] L. V. Blake, ANTENNAS, Artech House, 1984, p.332
[9] Nondestructive Testing of Materials, R. Collins et al., IOS Press Amsterdam, 1995
301
Author Index