Computational Electromagnetics and Model-Based Inversion - A Modern Paradigm For Eddy-Current Nondestructive Evaluation (PDFDrive)
Computational Electromagnetics and Model-Based Inversion - A Modern Paradigm For Eddy-Current Nondestructive Evaluation (PDFDrive)
Harold A. Sabbagh
R. Kim Murphy
Elias H. Sabbagh
John C. Aldrin
Jeremy S. Knopp
Computational
Electromagnetics
and Model-Based
Inversion
A Modern Paradigm for Eddy-Current
Nondestructive Evaluation
Computational Electromagnetics and Model-Based
Inversion
Scientific Computation
Editorial Board
J.-J. Chattot, Davis, CA, USA
P. Colella, Berkeley, CA, USA
R. Glowinski, Houston, TX, USA
Y. Hussaini, Tallahassee, FL, USA
P. Joly, Le Chesnay, France
D.I. Meiron, Pasadena, CA, USA
O. Pironneau, Paris, France
A. Quarteroni, Lausanne, Switzerland
and Politecnico of Milan, Milan, Italy
J. Rappaz, Lausanne, Switzerland
R. Rosner, Chicago, IL, USA
P. Sagaut, Paris, France
J.H. Seinfeld, Pasadena, CA, USA
A. Szepessy, Stockholm, Sweden
M.F. Wheeler, Austin, TX, USA
Computational
Electromagnetics and
Model-Based Inversion
A Modern Paradigm for Eddy-Current
Nondestructive Evaluation
123
Harold A. Sabbagh R. Kim Murphy
Victor Technologies, LLC Victor Technologies, LLC
Bloomington, IN, USA Bloomington, IN, USA
Jeremy S. Knopp
Air Force Research Laboratory
(AFRL/RXLP)
Wright-Patterson AFB, OH, USA
ISSN 1434-8322
ISBN 978-1-4419-8428-9 ISBN 978-1-4419-8429-6 (eBook)
DOI 10.1007/978-1-4419-8429-6
Springer New York Heidelberg Dordrecht London
Library of Congress Control Number: 2013937723
A few years ago, a colleague wrote a note in which he reviewed the state of the
art in eddy-current probes that were to address new problems in the nation’s steam
generators. With respect to calibration standards, he stated
The performance of these probes will not be realized unless calibration standards that
accurately simulate the range of expected defects are used. This means a series of axial
and/or circumferential notches, on both the tube od and id must be used to accurately
calibrate these probes. The ASME Section XI standard with flat bottomed holes is a very
poor representation of cracks, particularly for directional probes. In addition, the cable
between the instrument and the probe should be as short as reasonable and should be low
capacitance, low noise and low loss.
The purpose of this book is to address the entirety of issues raised in this quote,
and beyond, and to effectively resolve them favorably through the use of com-
putational electromagnetics and model-based inversion methods as a replacement
for expensive and unreliable standards. Indeed, we hope to demonstrate that our
assertion, that the computer will soon be the most important instrument in eddy-
current nondestructive evaluation (NDE), is only a mild stretch.
The book, like Gaul, is divided into three parts whose intention is to show that
computational electromagnetics is more than simply solving Maxwell’s equations
for various configurations of “anomalies” in hosts. Rather, it becomes a part of
system science of engineering, which is where we hope to elevate the notion
of ‘nondestructive evaluation’. Maxwell, himself, might not have envisioned the
ramifications of his theory, but surely he would have admired the results. (We are
reminded that a well-known aerospace NDE engineer once told us that had he
realized that you could make money solving Maxwell’s equations, he would have
paid more attention in his undergraduate E&M course.)
In the first part, we compute the Green’s dyad by working with the field equations
directly, without the intervening use of potential functions (Heaviside would be
proud!). Indeed, it isn’t until the derivation of the “vector form” of the volume-
integral equations in Chap. 3 that we first see potential functions explicitly stated.
We applied this approach originally to model electromagnetic responses in plane-
layered anisotropic bodies, such as advanced composites made of carbon-fiber
v
vi Preface
reinforced polymers, which will be an important part of the second volume in this
series. This approach works well in cylindrical coordinates, as is demonstrated in
Chap. 9.
The discretization of the volume-integral equation via Galerkin’s method on a
regular grid is certainly well known in the method of moments, and much of the
remainder of the book attempts to show the advantage of this approach in solving
large problems with reasonable computer resources.
Starting in Part I, and continuing throughout the remainder of the book, we have
sought to impress upon the reader the value of using equivalent electric circuits
or networks to interpret the physical response of the volume-integral equation.
This makes eddy-current NDE a subset of electrical engineering and should allow
those familiar with the basic concepts of electrical engineering lead into the further
development of eddy-current NDE.
The development of advanced probe models in Chap. 6 is original, especially in
treating coils, of whatever shape or orientation, as generalized magnetic dipoles
comprising solenoidal currents flowing in closed loops. This has allowed us to
model rectangular coils, or D-shaped coils, in a consistent numerical manner.
We have validated several of these models against benchmark data, as shown in
Chap. 6. What was especially pleasant, however, was the realization that we could
take the dual of the magnetic-dipole approach and model planar spiral-coil probes
assuming that the source is an electric dipole (Chap. 7). This allows us to compute
capacitive effects, as well as the usual inductive effects, thereby enabling a more
efficient design process for high-frequency applications. Problems of this type—
for example, spiral antennas—are usually treated by boundary-integral equations,
assuming perhaps that the metallic traces are perfect conductors. We have applied
the volume-integral approach, treating the spiral traces as an “anomaly” in free space
that is excited by an electric dipole, rather than an anomaly in a host that is excited by
a magnetic dipole, and are able to compute the capacitive and inductive reactances
and resistance of the probe as a function of frequency. Clearly, this approach allows
a single code to solve a greater variety of problems than originally assumed. We
can now treat resonance phenomena rigorously on the computer, without relying on
trial-and-error laboratory mockups.
Because of the increasing use of transmit–receive (T/R) arrays in eddy-current
NDE, we have included a discussion of N-port analysis of T/R arrays in Chap. 8. We
tie N-port theory of microwave networks with chain matrices to derive equivalent
networks for these arrays. This should help in designing and understanding the
behavior of these networks, especially at higher frequencies, where resonance
effects become important.
Part II is really where the fun begins, and computational electromagnetics
becomes a part of NDE system theory. We believe that the application of sophis-
ticated signal-processing and inversion algorithms lies at the heart of the future
development of NDE into a solid engineering discipline that is capable of handling
the challenging problems that new structures and materials introduce. Indeed, it will
lead NDE into the digital age, in which the computer replaces the oscilloscope as
the instrument of choice.
Preface vii
graduate theses and academic papers but is a commercially viable tool for solving
those problems that the industry needs to solve. VIC-3D R
is our contribution to the
list of codes that solve Maxwell’s equations for profit.
This book is not an introductory text; it will require a good background in electric
circuit theory, especially in understanding the concepts of phasors, impedance and
admittance, magnetic fields, magnetic induction and inductances, and magnetically
coupled circuits, as well as electromagnetic field theory, including Maxwell’s
equations. Material on electric circuits is usually covered in the sophomore year in
electrical engineering courses, whereas the required background in electromagnetic
fields is covered in upper-level undergraduate courses.
In teaching courses on electric circuits, we have found Circuit Analysis, by Elias
M. Sabbagh, Ronald Press, 1961, to be excellent preparation. It covers all aspects
listed above. Although long out of print, it is available on the internet at very low
prices. Of more recent vintage is Linear Circuit Analysis, by Raymond A. DeCarlo
and Pen-Min Lin, Oxford University Press, 2001.
When it comes to senior-level undergraduate texts on electromagnetic theory, we
can do no better than to recommend the classic Fields and Waves in Communication
Electronics, by Simon Ramo, John R. Whinnery, and Theodore Van Duzer, John
Wiley & Sons, New York, 1965. It not only gives a precise development of
electromagnetic theory, including Maxwell’s equations and their applications, but
also derives circuit concepts that are consistent with Maxwell’s equations.
If the reader wishes to develop his/her own code for Green’s functions, he/she
will need to become familiar quite quickly with the notion of “special functions.”
The classic reference on this subject, Handbook of Mathematical Functions, edited
by Milton Abramowitz and Irene A. Stegun, National Bureau of Standards, 1970,
has been superseded by NIST Handbook of Mathematical Functions, edited by
F.W.J. Olver, D.W. Lozier, R.F. Boisvert, and C.W. Clark, Cambridge University
Press, 2010. We have also found Table of Integrals, Series, and Products, by I.S.
Gradshteyn and I.M. Ryzhik, Academic Press, 1980, to be useful. Finally, to tie all
of this together in meaningful computer codes, we recommend Numerical Recipes:
The Art of Scientific Computing, by W.H. Press, B.P. Flannery, S.A. Teukolsky, and
W.T. Vetterling, Cambridge University Press, 1986.
We hope that you, the reader, will find this book useful and that you will agree
that it brings us closer to the goal of making eddy-current NDE a systematic branch
of engineering science, resulting in a more reliable system for making products and
materials safer.
ix
x Acknowledgments
Engines and Sue Vernon of the Naval Surface Warfare Center, White Oak Labs.
Both were early supporters of model-based eddy-current NDE, as well as its
manifestation in VIC-3D R
. Walt introduced us to the Split-D probe and informed
us of its importance in inspecting aircraft engines. He encouraged its inclusion
into VIC-3D R
, and we followed his advice (see Sect. 6.6). It continues to be an
important part of our current research. One of our first papers on modeling eddy-
current probes with ferrite cores using volume-integral equations was coauthored
with Sue. She supplied the data, and we supplied the model calculations. The results
were encouraging, so we thought that we were on to something. Sue introduced
us to carbon-fiber advanced composites and asked if we could do anything with
eddy currents to characterize them. This led us to some very interesting problems in
the electromagnetic modeling of anisotropic media and validating the models with
experimental data. Again, the results were positive and will be a significant part of
our second book.
Our men at Springer, Chris Coughlin and Ho Ying Fan, were a joy to work with.
Erik Wallace, a doctoral student in mathematics at Indiana University, is going to
investigate the “POCS-Iterative” issue that we described above (look for an answer
in our next book), but of more immediate import is the fact that he is a LaTeX guru
and assembled all the document preparation files for Ho Ying to pass to his people
at Springer.
Contents
xi
xii Contents
References .. .. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 435
Index . . . . . . . . .. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 443
Part I
Computational Electromagnetics
Background
Chapter 1
Overview of Methods of Computational
Electromagnetics
1.1 Introduction
Following Burrows, work at the Oak Ridge National Laboratory included efforts
in applying rigorous electromagnetic theory and developing computer programs
to solve problems in eddy-current NDE, thereby bringing the discipline into the
computer age. The primary authors of many of these papers were C.V. Dodd and
W.E. Deeds. See [4–8].
We will consider plane-parallel bodies of infinite extent in the (x, y) plane, which are
made up of layers of homogeneous, anisotropic material. To be specific, we consider
magnetic host materials that are characterized by the following biaxial generalized
electrical permittivity matrix:
⎡ ⎤
εx εxy 0
ε h = ⎣ εyx εy 0 ⎦, (2.1)
0 0 εz
where Jm and Je are anomalous magnetic and electric currents that account for the
presence of flaws, or anomalies, in the otherwise-uniform host material. From here
on we drop the subscript h on the generalized host permittivity and permeability.
Because of the material anisotropy, it is convenient to work with a matrix
formulation of these equations that has been useful in crystal optics, plasmas, and
microwave devices [13–21]. If the body is homogeneous with respect to (x, y), then
Maxwell’s equations can be Fourier transformed with respect to (x, y) and written
as the following four-vector matrix differential equation in the spectral domain:
∂z ẽ = S · ẽ + U · J̃, (2.3)
ky kx j ˜
Ẽz = H̃x − H̃y + Jez , (2.4)
εz ω εz ω εz ω
−ky kx j ˜
H̃z = Ẽx + Ẽy − Jmz , (2.5)
μω μω μω
where the tilde denotes a function defined in the transform domain (kx , ky ), and
⎡
⎤
J˜ex
⎡ ⎤ ⎢ ˜ ⎥
Ẽx ⎢ Jey ⎥
⎢ ⎥ ⎢ ⎥
⎢ Ẽy ⎥ ⎢ J˜ ⎥
⎢ ez ⎥
ẽ = ⎢ ⎥
⎢ H̃ ⎥; J̃ = ⎢ ⎥. (2.6)
⎣ x⎦ ⎢ J˜mx ⎥
⎢ ⎥
⎢ ˜ ⎥
H̃y ⎣ Jmy ⎦
J˜mz
which is called the equation of discontinuity. The superscript (+) denotes the limit
z approaches z from above and the superscript (−) denotes the limit from below.
Equation (2.9) will be used in the next section to compute the Green’s dyad for a
layered workpiece.
Starting with these equations, Roberts [24] has developed a fairly complete
theory of normal modes of biaxial anisotropic media. This work is based on, and
extends, earlier work performed at Sabbagh Associates [22, 23, 25–29]. From here
on we specialize the theory developed in [24] to the case to be considered here,
in which the media involved are transversely isotropic to the z-coordinate. The
dielectric permittivity tensor, in its principal-axis coordinate system, then takes the
form
⎡ ⎤
εt 0 0
ε = ⎣ 0 εt 0 ⎦ . (2.10)
0 0 εz
The entries in S now become
j j
a= kx ky ; α = (−kx ky )
ωεz ωμ
j j
b= (μεz ω 2 − kx2 ); β = (− μεt ω 2 + kx2 )
ωεz ωμ
j j
c= (− μεz ω 2 + ky2 ); γ = (μεt ω 2 − ky2 )
ωεz ωμ
j j
d=− kx ky ; δ = (kx ky ). (2.11)
ωεz ωμ
Let’s introduce some notation: kx2 + ky2 = kt2 , ω 2 μεt = Ωt2 , ω 2 μεz = Ωz2 , ε =
εt /εz . Then the eigenvalues of S are
√
λ1 = kt2 − Ωt2 λ2 = −λ1 λ3 = ε kt2 − Ωz2 λ4 = −λ3 . (2.12)
⎢ 0 ⎥ ⎢ 0 ⎥ ⎢ μ0 ⎥ ⎢ μ0 ⎥
⎢ ⎥ ⎢ ⎥ ⎢ ⎥ ⎢ ⎥
v1 = ⎢ ⎥ v2 = ⎢ 0 ⎥ v3 = ⎢ εt ⎥ v4 = ⎢ − εt ⎥. (2.14)
⎢ 0
⎥ ⎢
⎥ ⎢ ⎥ ⎢ ⎥
⎣ εt ⎦ ⎣ εt ⎦ ⎣ 1 ⎦ ⎣ 1 ⎦
−
μ0 μ0 0 0
When we substitute v1 , v2 of (2.13) into (2.4), with the source currents set to
zero, we find that Ẽz = 0; hence, v1 , v2 are transverse electric (TE) modes, with
respect to z. Similarly, v3 , v4 are transverse magnetic (TM) modes. Note that the
TE modes are orthogonal to the TM modes. This will facilitate the computation
of the Green’s dyadic. v1 and v3 are downward-traveling waves in the z-direction;
i.e., they represent waves that travel in the negative z-direction. v2 , v4 are upward-
traveling waves (in the positive z-direction). We see from (2.4) and (2.5) that all
modes are TEM (transverse electric and magnetic) with respect to z for kx = ky = 0.
Though the theory of eigenmodes that was developed in the preceding section
is applicable to the class of transverse-isotropic media, we will apply it only to
isotropic media. Hence, εt = εz = ε in (2.10), which means that λ1 = λ3 in (2.13).
The flaw, or other anomaly, in a workpiece produces an anomalous current which
is to be determined as the solution of a volume-integral equation. The kernel of this
equation is a Green dyad for a plane-parallel layered workpiece, and we turn our
attention to determining this dyad.
Consider the system shown in Fig. 2.1. The point-source of electric or magnetic
current is in region 0, and we want to compute the fields in this region, or in any
other region. This will give us the Green dyad. If the source is an electric current, and
the field electric, then the resulting dyad is called electric–electric. If the source is
a magnetic current, and the field electric, then the dyad is called electric–magnetic,
and so-on.
We write for the system of four-vectors in the ith region
⎡ ⎤ ⎡ ⎤ ⎡ ⎤ ⎡ ⎤
− j ω μi ky − j ω μi ky λ i kx λ i kx
⎢ ⎥ ⎢ ⎥ ⎢ ⎥ ⎢ ⎥
⎢ j ω μi kx ⎥ ⎢ j ω μi kx ⎥ ⎢ λ i ky ⎥ ⎢ λ i ky ⎥
v1i = ⎢⎢ ⎥ ⎢ ⎥ ⎢ ⎥ ⎢ ⎥
λ ⎥ v2i = ⎢ −λ k ⎥ v3i = ⎢ jω ε̂ k ⎥ v4i = ⎢ − jω ε̂ k ⎥ .
⎣ k
i x ⎦ ⎣ i x ⎦ ⎣ i y ⎦ ⎣ i y⎦
λ i ky − λ i ky − jω ε̂i kx jω ε̂i kx
(2.15)
2.2 Green Dyad for Plane-Parallel Layered Media 11
Z
(N)
ε ,μ
Ν Ν z
N−1
z
2
ε ,μ (2)
2 2 z
1
ε ,μ (1)
1 1 z
0
ε ,μ z (0)
0 0
z
−1
ε ,μ (−1)
−1 −1
z
−2
z
−M+1
ε ,μ (−M+1)
−Μ+1 −Μ+1
z
−M
^
ε ,μ
−M −M (−M)
First of all, we will compute the infinite-space Green dyad, which is the dyad
associated with an infinite, uniform, unlayered medium. This corresponds to the
situation wherein region 0 of Fig. 2.1 extends to z = ±∞; i.e., z0 = ∞, and z−1 = −∞.
Figure 2.2 shows a source located at z = z in region 0, together with the
appropriate field eigenvectors on each side of z . This choice of the eigenvectors is
consistent with the fact that v1 and v3 travel in the negative z-direction, and v2 and
v4 are positively traveling waves. The superscript, (inc), denotes “incident” fields
due to the source.
If the source is an electric current vector, then the right-hand side of (2.9) consists
of the three electric excitation vectors:
⎡ ⎤ ⎡ ⎤ ⎡ ˜ ⎤
0 0 kx Jez /ω ε̂0
⎢ 0 ⎥ ⎢ ⎥ ⎢ ⎥
(e) ⎢ ⎥ (e) ⎢ 0 ⎥ (e) ⎢ ky J˜ez /ω ε̂0 ⎥
fx = ⎢ ⎥ fy = ⎢ ⎥ fz = ⎢ ⎥, (2.16)
⎣ 0 ⎦ ⎣ J˜ey ⎦ ⎣ 0 ⎦
−J˜ex 0 0
12 2 Green’s Dyad for Plane-Layered Media
(+)
Z
(inc) - λ 0z (inc) - λ0 z
b V20 e + d V e
40
Region 0
z
λ z λ z
V 10 e 0 + c e 0
(inc) (inc)
a V
30
(-)
Fig. 2.2 Showing the eigenvectors used to calculate the infinite-space Green dyad. The source is
located at z = z
and if the source is a magnetic current, then the right-hand side consists of the
magnetic excitation vectors:
⎡ ⎤ ⎡ ˜ ⎤ ⎡ ⎤
0 Jmy 0
⎢ −J˜ ⎥ ⎢ ⎥ ⎢ ⎥
(m) ⎢ mx ⎥ (m) ⎢ 0 ⎥ (m) ⎢ 0 ⎥
fx = ⎢ ⎥ fy = ⎢ ⎥ fz = ⎢ ⎥. (2.17)
⎣ 0 ⎦ ⎣ 0 ⎦ ⎣ −kx J˜mz /ω μ0 ⎦
0 0 −ky J˜mz /ω μ0
where f denotes one of the six excitation vectors in (2.16) or (2.17). We will return
to this shortly.
Take the dot product of (2.18) with respect to the TE-mode vectors, v10 , v20 ,
and get
− A(i)α11 + B(i)α12 = F1
−A(i) α12 + B(i)α22 = F2 , (2.19)
where we have used the orthogonality of the TE and TM modal vectors to eliminate
c(inc) and d (inc) . Here
A(i) = a(inc) eλ0 z
B(i) = b(inc) e−λ0 z
F1 = v10 · f
2.2 Green Dyad for Plane-Parallel Layered Media 13
F2 = v20 · f
α11 = v10 · v10 = kt2 (λ02 − ω 2 μ02 )
α12 = v10 · v20 = −kt2 (λ02 + ω 2 μ02 )
α22 = v20 · v20 = kt2 (λ02 − ω 2 μ02 ). (2.20)
Upon solving (2.19) for A(i) , B(i) , and then using (2.20) we get
F1 α22 − F2α12
v10 eλ0 (z−z ) , z ≤ z . (2.22)
−α11 α22 + α12
2
To compute the TM-mode contribution, take the dot product of (2.18) with
respect to v30 , v40 . Proceeding as before, we get for the TM-mode contribution
F4 β33 − F3 β34
v40 e−λ0 (z−z ) , z ≥ z
β33 β44 − β34
2
F3 β44 − F4β34
− v30 eλ0 (z−z ) , z ≤ z , (2.23)
β33 β44 − β34
2
where
F3 = v30 · f
F4 = v40 · f
Therefore, upon combining (2.22) and (2.23) we get for the composite infinite-
field Green dyad:
−F2 α11 + F1 α12 F4 β33 − F3 β34
v20 + v40 e−λ0 (z−z ) , z ≥ z
−α11 α22 + α122 β33 β44 − β34
2
F1 α22 − F2α12 F3 β44 − F4β34
v −
2 10
v30 eλ0 (z−z ) , z ≤ z . (2.25)
−α11 α22 + α12 β33 β44 − β34
2
The Fs that appear in (2.25) are defined in (2.20) and (2.24). They are computed
by taking the dot product of the eigenvectors of (2.15) (with i = 0) with the
excitation vectors in (2.16) and (2.17). For example, the value of F1 that is associated
(ex) (e)
with electric excitation in the x-direction is given by F1 = v10 · fx = −λ0 ky J˜ex .
Continuing in this manner, we find:
(ex) (mx)
F1 = −λ0 ky J˜ex , F1 = − jω μ0 kx J˜mx
(ex) (mx)
F2 = λ0 ky J˜ex , F2 = − jω μ0 kx J˜mx
(ex) (mx)
F3 = jω ε̂0 kx J˜ex , F3 = −λ0 ky J˜mx
(ex) (mx)
F4 = − jω ε̂0 kx J˜ex , F4 = −λ0 ky J˜mx
(ey) (my)
F1 = λ0 kx J˜ey , F1 = − jω μ0 ky J˜my
(ey) (my)
F2 = −λ0 kx J˜ey , F2 = − jω μ0 ky J˜my
(ey) (my)
F3 = jω ε̂0 ky J˜ey , F3 = λ0 kx J˜my
(ey) (my)
F4 = − jω ε̂0 ky J˜ey , F4 = λ0 kx J˜my
When these are combined with the α ’s and β ’s of (2.20) and (2.24), we get for
the infinite-space Green dyad
2.2 Green Dyad for Plane-Parallel Layered Media 15
[J2 v20 + J4 v40 ] e−λ0 (z−z ) , z ≥ z
[J1 v10 + J3 v30 ] eλ0 (z−z ) , z ≤ z , (2.27)
where
J1 J2 J3 J4
J˜ez J˜ez
(ez) : 0 0 −
2λ0ω ε̂0 2λ0 ω ε̂0
J˜mz J˜mz
(mz) : 0 0. (2.28)
2 λ 0 ω μ0 2 λ 0 ω μ0
When we introduce the vectors of (2.15) into (2.27), and then use (2.28), we
obtain an expression for the x- and y-components of the field in terms of the
source components. The z-component of the fields are obtained from (2.4) and
(2.5). For instance, we will compute the x-component of the electric field, due to
the x-component of electric current:
⎧
⎪
⎪ − jω μ0 ky2 J˜ex jkx2 λ0 J˜ex −λ (z−z )
⎪
⎪ + e 0 , z ≥ z
⎪
⎪ 2 λ k 2 2 ω ε̂ k 2
⎨ 0 t 0 t
⎪
⎪
⎪ − jω μ0 ky2 J˜ex jkx2 λ0 J˜ex λ (z−z )
⎪
⎪
⎪
⎩ + e 0 , z ≤ z
2λ0 kt2 2ω ε̂0 kt2
k2 e−λ0 |z−z | ˜
= − jω μ0 1 − x2 Jex . (2.29)
k0 2λ0
16 2 Green’s Dyad for Plane-Layered Media
δ (z − z )
+j az az , (2.30)
ω ε̂0
δ (z − z )
where (+) sign goes with z > z and (−) with z < z . The term j az az
ω ε̂0
is called the “depolarizing” term [34] and follows from the last term in (2.4).
A similar analysis holds for the magnetic–magnetic Green dyad and produces the
same expression as (2.30), except that μ0 is everywhere replaced by −ε̂0 .
The spectral-domain, infinite-space, magnetic–electric Green dyad is given by
⎡ ⎤
0 ±λ0 − jky
(em) e−λ0 |z−z |
G̃(0) (kx , ky ; z, z ) = ⎣ ∓λ0 0 jkx ⎦ . (2.31)
2λ0
jky − jkx 0
The infinite-space dyad that we computed in the previous section serves as the
“incident” field in the layered medium of Fig. 2.1. We are interested in computing
the fields produced when the incident dyad is scattered from the layers.
As before, we focus our attention on region 0, but this time assume that it has a
finite upper-boundary, z0 , and lower-boundary, z−1 , as shown in Fig. 2.3. Region 1,
which lies above region 0, is a half-space, as is region -1, which lies below region 0.
The infinite-field Green dyad is shown in Fig. 2.3. Because of the singularity
of this dyad, we split it into that part, [J2 v20 + J4 v40 ]e−λ0 (z−z ) , which is valid for
z ≥ z , and [J1 v10 + J3 v30 ]eλ0 (z−z ) , which is valid for z ≤ z . The scattered field,
(a0 v10 + c0 v30 )eλ0 z + (b0 v20 + d0v40 )e−λ0 z , is continuous throughout region 0.
2.2 Green Dyad for Plane-Parallel Layered Media 17
(1)
- λ1 z - λ1 z
b V e + d V e
1 21 1 41
z
0
- λ0 (z-z ) - λ0 (z-z )
JV e + JV e
2 20 4 40
-λ z
a V e λ0 + b V e - λ0 + c V e λ0 + d V e 0
z z z (0)
z
0 10 0 20 0 30 0 40
λ (z-z )
J V e λ0 +JV e 0
(z-z )
1 10 3 30
z
-1
λ z
+ c V3,-1 e λ−1
z
a V e −1
-1 1,-1 -1
(-1)
Fig. 2.3 Showing the source, at z , in region 0, together with the incident and scattered fields.
Regions 1 and −1 are half-spaces
where
b1 , d1 , a0 , c0 are scattered fields at z0 and b0 , d0 , J2 , and J4 are incident upon z0 .
Continuity of the fields at the lower boundary, z = z−1 , requires that
a−1 v1,−1 + c−1 v3,−1 − b0v20 − d0 v40 = a0 v10 + c0 v30 + J1 e−λ0 z v10 + J3 e−λ0 z v30,
(2.35)
where now
a−1 = a−1 eλ−1 z−1 , c−1 = c−1 eλ−1 z−1 , a0 = a0 eλ0 z−1 , b0 = b0 e−λ0 z−1
. (2.36)
c0 = c0 eλ0 z−1 , d0 = d0 e−λ0 z−1 , J1 = J1 eλ0 z−1 , J3 = J3 eλ0 z−1
In this case, a−1 , c−1 , b0 , d0 are scattered fields at z−1 and c0 , a0 , J1 , and J3 are
incident upon z−1 .
18 2 Green’s Dyad for Plane-Layered Media
Equations (2.33) and (2.35) include TE- and TM-modes. They are easier to solve
if we separate these modes. Take the dot product of (2.33) with respect to the
TE-modes, v21 and v10 and get
b1 v21 · v21 − a0v10 · v21 = b0 v20 · v21 + J2 eλ0 z v20 · v21
b1 v21 · v10 − a0v10 · v10 = b0 v20 · v10 + J2 eλ0 z v20 · v10, (2.37)
where
Next, compute the TM-modes by taking the dot product of (2.33) with v41
and v30 :
d1 v41 · v41 − c0 v30 · v41 = d0 v40 · v41 + J4 eλ0 z v40 · v41
d1 v41 · v30 − c0 v30 · v30 = d0 v40 · v30 + J4 eλ0 z v40 · v30 , (2.40)
where
where
where
The final expressions for the mode coefficients are given in terms of these scattering
parameters:
(E) (E) (E)
R1 R−1 e−λ0 (2T +z ) J1 + R1 e−λ0 (2z0 −z ) J2
a0 = (E) (E)
1 − R1 R−1 e−2λ0 T
δ (z − z ) ˜
Ẽz = jkt2 (c0 eλ0 z − d0e−λ0 z ) − Jez
jω ε̂0
δ (z − z ) ˜
H̃z = jkt2 (a0 eλ0 z + b0 e−λ0 z ) + Jmz . (2.51)
j ω μ0
From here on we will drop the delta-function terms in Ẽz and H̃z , because they will
be associated with the infinite-field Green dyad.
The coefficients a0 , b0 , c0 , d0 that appear in (2.51) have been previously
expressed in terms of J1 , J2 , J3 , J4 in (2.50). When these results are substituted
into (2.51) we get:
(E) (E) (M) (M)
−R R j ω μ k J R R λ k J
Ẽx = e−λ0 (2T +(z −z)) −1 −1
1 0 y 1 1 0 x 3
(E) (E)
+ (M) (M)
1 − R1 R−1 e−2λ0 T 1 − R1 R−1 e−2λ0 T
(E) (E) (M) (M)
−λ0 (2T −(z −z)) −R1 R−1 jω μ0 ky J2 R1 R−1 λ0 kx J4
+e (E) (E)
+ (M) (M)
1 − R1 R−1 e−2λ0 T 1 − R1 R−1 e−2λ0 T
22 2 Green’s Dyad for Plane-Layered Media
(E) (M)
−λ0 (2z0 −(z +z)) −R1 jω μ0 ky J2 R1 λ0 kx J4
+e (E) (E)
+ (M) (M)
1 − R1 R−1 e−2λ0 T 1 − R1 R−1 e−2λ0 T
(E) (M)
λ0 (2z−1 −(z +z)) −R−1 jω μ0 ky J1 R−1 λ0 kx J3
+e (E) (E)
+ (M) (M)
1 − R1 R−1 e−2λ0 T 1 − R1 R−1 e−2λ0 T
(E) (E) (M) (M)
−λ0 (2T +(z −z)) R1 R−1 jω μ0 kx J1 R1 R−1 λ0 ky J3
Ẽy = e (E) (E)
+ (M) (M)
1 − R1 R−1 e−2λ0 T 1 − R1 R−1 e−2λ0 T
(E) (E) (M) (M)
−λ0 (2T −(z −z)) R1 R−1 jω μ0 kx J2 R1 R−1 λ0 ky J4
+e (E) (E)
+ (M) (M)
1 − R1 R−1 e−2λ0 T 1 − R1 R−1 e−2λ0 T
(E) (M)
−λ0 (2z0 −(z +z)) R1 jω μ0 kx J2 R1 λ0 ky J4
+e (E) (E)
+ (M) (M)
1 − R1 R−1 e−2λ0 T 1 − R1 R−1 e−2λ0 T
(E) (M)
−(z +z)) R−1 jω μ0 kx J1 R−1 λ0 ky J3
+ eλ0 (2z−1 (E) (E)
+ (M) (M)
1 − R1 R−1 e−2λ0 T 1 − R1 R−1 e−2λ0 T
(M) (M)
R1 R−1 jkt2 J3
Ẽz = e−λ0 (2T +(z −z)) (M) (M)
1 − R1 R−1 e−2λ0 T
(M) (M)
−R1 R−1 jkt2 J4
+ e−λ0 (2T −(z −z)) (M) (M)
1 − R1 R−1 e−2λ0 T
(M)
R1 jkt2 J4
+ e−λ0 (2z0 −(z +z)) (M) (M)
1 − R1 R−1 e−2λ0 T
(M)
−R−1 jkt2 J3
+ eλ0 (2z−1 −(z +z)) (M) (M)
1 − R1 R−1 e−2λ0 T
(E) (E) (M) (M)
(2T +(z −z)) R1 R−1 λ0 kx J1 R1 R−1 jω ε̂0 ky J3
H̃x = e−λ0 (E) (E)
+ (M) (M)
1 − R1 R−1 e−2λ0 T 1 − R1 R−1 e−2λ0 T
(E) (E) (M) (M)
−λ0 (2T −(z −z)) −R1 R−1 λ0 kx J2 −R1 R−1 jω ε̂0 ky J4
+e (E) (E)
+ (M) (M)
1 − R1 R−1 e−2λ0 T 1 − R1 R−1 e−2λ0 T
2.2 Green Dyad for Plane-Parallel Layered Media 23
(E) (M)
−λ0 (2z0 −(z +z)) R1 λ0 kx J2 R1 jω ε̂0 ky J4
+e (E) (E)
+ (M) (M)
1 − R1 R−1 e−2λ0 T 1 − R1 R−1 e−2λ0 T
(E) (M)
λ0 (2z−1 −(z +z)) −R−1 λ0 kx J1 −R−1 jω ε̂0 ky J3
+e (E) (E)
+ (M) (M)
1 − R1 R−1 e−2λ0 T 1 − R1 R−1 e−2λ0 T
(E) (E) (M) (M)
−λ0 (2T +(z −z)) R1 R−1 λ0 ky J1 −R1 R−1 jω ε̂0 kx J3
H̃y = e (E) (E)
+ (M) (M)
1 − R1 R−1 e−2λ0 T 1 − R1 R−1 e−2λ0 T
(E) (E) (M) (M)
−λ0 (2T −(z −z)) −R1 R−1 λ0 ky J2 R1 R−1 jω ε̂0 kx J4
+e (E) (E)
+ (M) (M)
1 − R1 R−1 e−2λ0 T 1 − R1 R−1 e−2λ0 T
(E) (M)
−λ0 (2z0 −(z +z)) R1 λ0 ky J2 −R1 jω ε̂0 kx J4
+e (E) (E)
+ (M) (M)
1 − R1 R−1 e−2λ0 T 1 − R1 R−1 e−2λ0 T
(E) (M)
−(z +z)) −R−1 λ0 ky J1 R−1 jω ε̂0 kx J3
+ eλ0 (2z−1 (E) (E)
+ (M) (M)
1 − R1 R−1 e−2λ0 T 1 − R1 R−1 e−2λ0 T
(E) (E)
R1 R−1 jkt2 J1
H̃z = e−λ0 (2T +(z −z)) (E) (E)
1 − R1 R−1 e−2λ0 T
(E) (E)
R1 R−1 jkt2 J2
+ e−λ0 (2T −(z −z)) (E) (E)
1 − R1 R−1 e−2λ0 T
(E)
R1 jkt2 J2
+ e−λ0 (2z0 −(z +z)) (E) (E)
1 − R1 R−1 e−2λ0 T
(E)
R−1 jkt2 J1
+ eλ0 (2z−1 −(z +z)) (E) (E)
. (2.52)
1 − R1 R−1 e−2λ0 T
(me)
H̃(kx , ky ; z, z ) = G̃(s) (kx , ky ; z, z ) · J̃e (kx , ky )
(mm)
H̃(kx , ky ; z, z ) = G̃(s) (kx , ky ; z, z ) · J̃m (kx , ky ) . (2.53)
(ee)(s) s(z −z) −2λ0 T (M) 1 eλ0 (z+z ) −2λ0 z0 (M) 1
G̃yz = − jω μ0 (− jky λ0 ) − e G1,−1 2 − e G1 2
2λ0 k0 2λ0 k0
e−λ0 (z+z ) 2λ0 z−1 (M) 1
+ e G−1 2
2λ0 k0
where c(z − z) = eλ0 (z −z) + e−λ0 (z −z) and s(z − z) = eλ0 (z −z) − e−λ0 (z −z) . The
superscript (a) denotes terms that are convolutional (“Töplitz”) in z and z , i.e.,
depend upon z − z , whereas (b) denotes terms that are correlational (“Hankel”) in z
and z , i.e., depend upon z + z . The Gs are defined in terms of the TE and TM-mode
reflection coefficients:
(E) (E) (E) (E)
(E) R1 R−1 (E) R1 (E) R−1
G1,−1 = (E) (E) −2λ T
, G1 = (E) (E) −2λ T
, G−1 = (E) (E) −2λ T
1−R1 R−1 e 0 1−R1 R−1 e 0 1−R1 R−1 e 0
(M) (M) (M) (M)
(M) R1 R−1 (M) R1 (M) R−1
G1,−1 = (M) (M)
, G1 = (M) (M)
, G−1 = (M) (M)
.
1−R1 R−1 e−2λ0 T 1−R1 R−1 e−2λ0 T 1−R1 R−1 e−2λ0 T
(2.55)
The physical origin of the Töplitz and Hankel terms follows from the definitions
in (2.55), and a graphical illustration is shown in Fig. 2.4.
The Töplitz structure, as shown in the top of Fig. 2.4, arises when the path
between the source point, z , and field point, z, includes reflections from both
boundaries. The total z−directed path length between z and z is z − z + 2T for
path A and z − z + 2T for path B. In each case, the length includes the difference
between the z-coordinate of the source and field points.
The Hankel structure, as shown in the bottom of Fig. 2.4, arises when the path
between the source and field points includes reflections from only one of the
boundaries. The total path length between z and z is 2Z0 − (z + z ) for path A and
z + z − 2Z−1 for path B. In each case, the length includes the sum of the source and
field z-coordinates.
26 2 Green’s Dyad for Plane-Layered Media
Z0
B A A
z’
B
B A
T
B
z
A A B Z −1
Toeplitz
Z0
A A
z’
A
B
z
B B
Z−1‘
Hankel
Fig. 2.4 Illustrating the difference between Töplitz (top) and Hankel (bottom) Green’s functions
(mm)(s) s(z − z) −2λ0 T (E) 1 eλ0 (z+z ) −2λ0 z0 (E) 1
G̃xz = jω ε̂0 (− jkx λ0 ) − e G1,−1 2 + e G1 2
2λ0 k0 2λ0 k0
e−λ0 (z+z ) 2λ0 z−1 (E) 1
− e G−1 2
2λ0 k0
(mm)(s) (mm)(s)
G̃yx = G̃xy
2
(mm)(s) c(z − z) −2λ0 T (M) (M) 1 (E) λ0
G̃yy = jω ε̂0 e G1,−1 − ky G1,−1 2 + G1,−1 2 2
2
2λ0 kt k0 kt
eλ0 (z+z ) −2λ0 z0 (M) (M) 1 (E) λ
2
− e G1 − ky2 G1 2 + G1 2 0 2
2λ0 kt k0 kt
2
e−λ0 (z+z ) 2λ0 z−1 (M) (M) 1 (E) λ
− e G−1 − ky2 G−1 2 + G−1 2 0 2
2λ0 kt k0 kt
(mm)(s) s(z − z) −2λ0 T (E) 1 eλ0 (z+z ) −2λ0 z0 (E) 1
G̃yz = jω ε̂0 (− jky λ0 ) − e G1,−1 2 + e G1 2
2λ0 k0 2λ0 k0
e−λ0 (z+z ) 2λ0 z−1 (E) 1
− e G−1 2
2λ0 k0
Finally, we list the mixed dyadic functions, starting with the electric–magnetic:
⎧ ⎡ ⎤ (E)
⎨ e−2λ0 T G
(E)
−G
(M)
eλ0 (z +z) −2λ0 z0 G1 −G1
(M)
(em)(s) ⎣ 1,−1 1,−1 ⎦
G̃xx = −kx ky −s(z −z) − e
⎩ 2 kt 2 2 kt 2
(E) ⎫
e−λ0 (z +z) 2λ0 z−1 G−1 − G−1 ⎬
(M)
+ e
2 kt 2 ⎭
28 2 Green’s Dyad for Plane-Layered Media
(em)(s) s(z −z) (M) eλ0 (z +z) (M) 2λ0 z−1 e−λ0 (z +z) (M)
G̃xy = e−2λ0 T G1,−1 +e−2λ0 z0 G1 −e G−1
2 2 2
⎧ ⎡ ⎤ (E)
⎨ s(z −z) G
(E)
−G1,−1
(M)
eλ0 (z +z) G −G
(M)
−2λ0 T ⎣ 1,−1 ⎦− −2λ
−ky −
2
e e 00 z 1 1
⎩ 2 kt 2 2 kt 2
(E) ⎫
e−λ0 (z +z) 2λ0 z−1 G−1 −G−1 ⎬
(M)
+ e
2 kt 2 ⎭
(em)(s) −2λ0 T c(z −z) (E) eλ0 (z +z) (E) 2λ0 z−1 e−λ0 (z +z) (E)
G̃xz = − jky e G1,−1 +e−2λ0 z0 G1 +e G−1
2λ0 2λ0 2λ0
(em)(s) s(z − z) (M) eλ0 (z +z) (M) e−λ0 (z +z) (M)
G̃yx = −e−2λ0 T G1,−1 − e−2λ0 z0 G1 + e2λ0 z−1 G−1
2 2 2
⎧ ⎡ ⎤ (E)
⎨ s(z − z) (E) (M) (M)
2
G
−2λ0 T ⎣ 1,−1
− G1,−1
⎦ eλ0 (z +z) −2λ0 z0 G1 − G1
−kx e + e
⎩ 2 kt 2 2 kt 2
(E) ⎫
(M)
e−λ0 (z +z) 2λ0 z−1 G−1 −G−1 ⎬
− e
2 kt 2 ⎭
(em)(s) (em)(s)
G̃yy = −G̃xx
(em)(s) −2λ0 T c(z −z) (E) eλ0 (z +z) (E) e−λ0 (z +z) (E)
G̃yz = − jkx −e G1,−1 −e−2λ0 z0 G1 − e2λ0 z−1 G−1
2λ0 2λ0 2λ0
(em)(s) −2λ0 T c(z −z) (M) eλ0 (z +z) (M) 2λ0 z−1 e−λ0 (z +z) (M)
G̃zx = − jky −e G1,−1 +e−2λ0 z0 G1 +e G−1
2λ0 2λ0 2λ0
(em)(s) −2λ0 T c(z −z) (M) eλ0 (z +z) (M) 2λ0 z−1 e−λ0 (z +z) (M)
G̃zy = − jkx e G1,−1 −e−2λ0 z0 G1 −e G−1
2λ0 2λ0 2λ0
(em)(s)
G̃zz = 0, (2.57)
⎧ ⎡ ⎤ (E)
⎨ −2λ0 T (E)
−G
(M) λ (z +z) (M)
−2λ0 z0 G1 −G1
(me)(s) e G 1,−1 ⎦ e 0
G̃xx
= −kx ky s(z −z) ⎣ 1,−1
− e
⎩ 2 kt 2 2 kt 2
(E) ⎫
e−λ0 (z +z) 2λ0 z−1 G−1 −G−1 ⎬
(M)
+ e
2 kt 2 ⎭
(me)(s) s(z −z) (E) eλ0 (z +z) (E) 2λ0 z−1 e−λ0 (z +z) (E)
G̃xy = e−2λ0 T G1,−1 −e−2λ0 z0 G1 +e G−1
2 2 2
⎧ ⎡ ⎤ (E)
⎨ s(z −z) G
(E) (M)
−G1,−1
eλ0 (z +z) −2λ0 z0 G1 −G1
(M)
−2λ0 T ⎣ 1,−1 ⎦
− ky 2
e − e
⎩ 2 kt 2 2 kt 2
(E) ⎫
e−λ0 (z +z) 2λ0 z−1 G−1 −G−1 ⎬
(M)
+ e
2 kt 2 ⎭
(me)(s) −2λ0 T c(z −z) (M) eλ0 (z +z) (M) e−λ0 (z +z) (M)
G̃xz = − jky e G1,−1 −e−2λ0 z0 G1 − e2λ0 z−1 G−1
2λ0 2λ0 2λ0
(me)(s) s(z −z) (E) eλ0 (z +z) (E) 2λ0 z−1 e−λ0 (z +z) (E)
G̃yx = −e−2λ0 T G1,−1 +e−2λ0 z0 G1 −e G−1
2 2 2
⎧ ⎡ ⎤ (E)
⎨ s(z −z) (E)
−
(M) λ0 (z +z) (M)
−2λ0 z0 G1 −G1
G G
−kx 2 − e−2λ0 T ⎣
1,−1 1,−1 ⎦+e e
⎩ 2 kt 2 2 kt 2
(E) ⎫
e−λ0 (z +z) 2λ0 z−1 G−1 −G−1 ⎬
(M)
− e
2 kt 2 ⎭
(me)(s) (me)(s)
G̃yy = −G̃xx
(me)(s) −2λ0 T c(z −z) (M) eλ0 (z +z) (M) 2λ0 z−1 e−λ0 (z +z) (M)
G̃yz = − jkx −e G1,−1 +e−2λ0 z0 G1 +e G−1
2λ0 2λ0 2λ0
(me)(s) c(z −z) (E) eλ0 (z +z) (E) 2λ0 z−1 e−λ0 (z +z) (E)
G̃zx = − jky −e−2λ0 T G1,−1 −e−2λ0 z0 G1 −e G−1
2λ0 2λ0 2λ0
(me)(s) −2λ0 T c(z −z) (E) eλ0 (z +z) (E) e−λ0 (z +z) (E)
G̃zy = − jkx e G1,−1 +e−2λ0 z0 G1 + e2λ0 z−1 G−1
2λ0 2λ0 2λ0
(me)(s)
G̃zz = 0. (2.58)
30 2 Green’s Dyad for Plane-Layered Media
(i+1)
η
i+1
T R
ξ i+1(-) i+1(-)
i+1
z
i
(+) (+)
λi R
i(+) η ,ξ
η = T i i
i j ωμ i i(+)
λi (i)
ξ =
i j ωε i
(−) (−)
R η ,ξ
i(-) i i
z
i-1
η
i-1
(i-1)
ξ
i-1
Fig. 2.5 Definition of scattering parameters and intrinsic wave-immittances in layered structures
The reflection and transmission coefficients of (2.49) are those of a slab surrounded
by a half-space above and below it. We will derive a recursion relation that will
(E,M)
allow the computation of the reflection coefficients, R±1 , that are used in (2.55)
when there are an arbitrary number of layers above and below the source slab [35].
Of course, there will ultimately be half-spaces that terminate the system at ±∞.
Let region i be bounded above by zi and below by zi−1 . Region i + 1 lies
immediately above region i and region i − 1 lies immediately below. The intrinsic
wave-admittance, ηi , and wave-impedance, ξi , for the TE and TM modes in region
i are
λi λi
(TE) ηi = , ξi = (TM). (2.59)
j ω μi jω ε̂i
(E,M)
We define Ri± to be the reflection coefficient in layer i at the interface with layer
(E,M)
i ± 1, and Ti± to be the transmission coefficient, as shown in Fig. 2.5. Then, as
shown in (2.49), for a slab sandwiched between two half-spaces
2.2 Green Dyad for Plane-Parallel Layered Media 31
(E) μi 2ηi
Ti± =
μi±1 ηi + ηi±1
(M) ε̂i 2 ξi
Ti± = . (2.60)
ε̂i±1 ξi + ξi±1
When we have layers above and below region i, we must replace the intrinsic wave-
parameters, ξi±1 , ηi±1 , by equivalent “load parameters,” ηi± , ξi± , which are defined
to be the surface admittance and impedance in layer i at the interface with layer
i ± 1. The surface admittance is defined to be H̃y /Ẽx at the appropriate interface, and
the surface impedance is defined to be Ẽx /H̃y . These ratios are the same regardless
of which side of the interface they are evaluated at, because Ẽx and H̃y are each
continuous at the interface. Hence, we replace (2.60) by
(E)
eλi Ti − Ri+ e−λi Ti
= ηi (E)
eλi Ti + Ri+ e−λi Ti
(+)
ηi − ηi
eλi Ti − (+)
e−λi Ti
ηi + ηi
= ηi (+)
ηi − ηi
eλi Ti + (+)
e−λi Ti
ηi + ηi
(+)
ηi + ηi tanh(λi Ti )
= ηi (+)
. (2.64)
ηi + ηi tanh(λi Ti )
When we use v3i and v4i , we find that the components of the TM-field at z = zi−1
are given by
(M)
Ẽx = di Ri+ λi kx e−λi Ti + λikx eλi Ti
(M)
H̃y = di Ri+ (− jω ε̂i kx )e−λi Ti + jω ε̂i kx eλi Ti . (2.65)
where we treat the negatively-traveling wave, ai v1i eλi (z−zi−1 ) , as being incident on
the surface z = zi−1 . Therefore, we have
(E)
Ẽx = ai − jω μi ky eλi Ti + Ri− (− jω μi ky )e−λi Ti
(E)
H̃y = ai λi ky eλi Ti + Ri− (−λi ky )e−λi Ti , (2.68)
2.2 Green Dyad for Plane-Parallel Layered Media 33
for the fields at the upper-boundary, zi , of the ith region. Hence, the load admittance
at z = zi is
Equations (2.64), (2.69), (2.66), and (2.70) are the iterations that produce the
immittances that go into the expressions for the reflection coefficients, (2.61).
The iterations are started at the interface of the last slab with an infinite half-space,
for which
+
ηN−1 = ηN+ = ηN , ξN−1
+
= ξN+ = ξN
− − − −
η−(M−1) = η−M = η−M , ξ−(M−1) = ξ−M = ξ−M . (2.71)
Chapter 3
The Volume-Integral Equations
for Plane-Layered Media
λ02 e−λ0 |z−z | δ (z − z ) ˜
Ẽz (kx , ky ; z) = − jω μ0 dz 1+ 2 − Jez (kx , ky ; z )
k0 2λ0 k02
1 d 2 e−λ0 |z−z | ˜
= − j ω μ0 dz 1 + 2 2 Jez (kx , ky ; z )
k0 dz 2λ0
1 d2 e−λ0 |z−z | ˜
= − j ω μ0 1 + 2 2 dz Jez (kx , ky ; z ) . (3.1)
k0 dz 2λ0
Hence, the spatial-domain field is obtained by taking the inverse Fourier trans-
form of (3.1)
∞ −λ0 |z−z |
1 ∂2 1 e
Ez (r) = − jω μ0 1 + 2 2 dz 2 J˜ez (kx , ky ; z )e− j[kx x+ky y] dkx dky
k0 ∂ z 4π −∞ 2λ0
1 ∂2
= 1+ 2 2 dr Φ (e) (r − r )Jez (r ) , (3.2)
k0 ∂ z
where
∞ −λ0 |z−z |
− j ω μ0 e
Φ (e) (r − r) = e− j[kx (x−x )+ky (y−y )] dkx dky . (3.3)
4π 2 −∞ 2λ0
The transform
function in (3.3) is spherically symmetric, because it is a function
of λ0 = kt2 − k02 . Hence, we can transform the two-dimensional integrals into one-
dimensional integrals in the following way: transform to cylindrical coordinates in
both physical and k-space
x − x = r cos φ ; kx = l cos α
y − y = r sin φ ; ky = l sin α . (3.4)
(We are replacing the variable, kt , by l.) Then the integral in (3.3) becomes
2π ∞ −λ0 |z−z |
− j ω μ0 e
Φ (e)
(r − r ) = dα e− jlr cos(α −φ ) ldl . (3.5)
4π 2 0 0 2λ0
When (3.6) is substituted into (3.5), only the J0 term survives the integration over
2π radians, so that
∞ −λ0 |z−z |
(e) − j ω μ0 e
Φ (r − r ) = J0 (rl)ldl
2π 0 2λ0
e− jk0 |r−r |
= − j ω μ0 . (3.7)
4π |r − r|
We decompose the total Green dyad into the “infinite” part, which is the field
produced by the source in infinite space, and the “layered” part, which is due to the
presence of the various layers of the workpiece. If we consider only nonmagnetic
problems (μ = μ0 ) right now, and let J(r ) be the unknown anomalous electric
current produced by the flaw, then the infinite part of the dyad produces the “infinite-
space” contribution to the total electric field:
3.2 The Electric Differential Volume-Integral Equation 37
and
e− jk0 |r−r |
Φ (e) (r − r) = − jω μ0
4π |r − r|
∞ −λ0 |z−z |
− j ω μ0 e
= J0 (rl)ldl . (3.10)
2π 0 2λ0
(s) (1) 1 ∂2 1 ∂2 1 ∂2
Ex (r) = Fx + Fx + Fy + Fz
k02 ∂ x2 k02 ∂ x∂ y k02 ∂ x∂ z
(s) 1 ∂2 (1) 1 ∂2 1 ∂2
Ey (r) = Fx + Fy + 2 2 Fy + 2 Fz
k0 ∂ y∂ x
2 k0 ∂ y k0 ∂ y∂ z
(s) 1 ∂2 1 ∂2 1 ∂2
Ez (r) = Fzx + 2 Fzy + Fz + 2 2 Fz , (3.11)
k0 ∂ z∂ x
2 k0 ∂ z∂ y k0 ∂ z
where
(1) (1)
Fx (r) = dr Gxx (x − x , y − y ; z, z )Jx (r )
(2)
Fx (r) = dr Gxx (x − x , y − y ; z, z )Jx (r )
(2)
Fy (r) = dr Gxx (x − x , y − y ; z, z )Jy (r )
Fz (r) = dr Gxz (x − x , y − y ; z, z )Jz (r )
(1) (1)
Fy (r) = dr Gxx (x − x , y − y ; z, z )Jy (r )
38 3 The Volume-Integral Equations for Plane-Layered Media
(a) (a)
Fzx (r) = dr Gxz (x − x , y − y ; z, z )Jx (r )
(b) (b)
Fzx (r) = − dr Gxz (x − x , y − y ; z, z )Jx (r )
(a) (a)
Fzy (r) = dr Gxz (x − x , y − y ; z, z )Jy (r )
(b) (b)
Fzy (r) = − dr Gxz (x − x , y − y ; z, z )Jy (r ) . (3.12)
∞
(1) − j ω μ0 c(z − z) −2λ0 T (E) eλ0 (z+z ) −2λ0 z0 (E)
Gxx = e G1,−1 + e G1
2π 0 2λ0 2λ0
e−λ0 (z+z ) 2λ0 z−1 (E)
+ e G−1 J0 (rl)ldl
2λ0
2
(2) − jω μ0 ∞ c(z − z) −2λ0 T (E) k0
2
(M) λ0
Gxx = e G1,−1 2 + G1,−1 2
2π 0 2λ0 l l
2
eλ0 (z+z ) −2λ0 z0 (E) k0
2
(M) λ0
+ e G1 2 + G1 2
2λ0 l l
2
e−λ0(z+z ) 2λ0 z−1 (E) k0
2
(M) λ0
+ e G−1 2 + G−1 2 J0 (rl)ldl
2λ0 l l
− jω μ0 ∞ c(z − z) −2λ0 T (M) eλ0 (z+z ) −2λ0 z0 (M)
Gxz = e G1,−1 − e G1
2π 0 2λ0 2λ0
e−λ0 (z+z ) 2λ0 z−1 (M)
− e G−1 J0 (rl)ldl. (3.13)
2λ0
As before, r = [(x−x )2 +(y−y )2 ]1/2 in the Bessel transform. The results of (3.11)–
(3.13) follow from (2.54), upon replacing multiplication by − jkx , − jky , and λ0 by
∂ /∂ x, ∂ /∂ y, ∂ /∂ z, respectively. In the expression for G̃zz we use kt2 = λ02 + k02 .
The integro-differential equation to which we will apply the method of moments
is simply gotten by equating the total electric field, J(r)/σa , to the sum of the
incident field, due to the coil and the infinite-space and layered-space scattered
fields:
The equations of the model, (3.8)–(3.14), can be rewritten to bring out the vector
nature more clearly. In deriving the new equations we repeatedly make use of
integration-by-parts and then use the fact that the current distribution is limited in
(e)
space. For example, consider the term ∂ Ax (r)/∂ x:
(e)
∂ Ax (r) ∂ (e)
= Φ (e) (r − r )Jx (r )dr
∂x ∂x
∂ (e) (e)
= Φ (r − r)Jx (r )dr
∂x
∂ (e)
= − Φ (e) (r − r )Jx (r )dr
∂x
(e)
∂ Jx (r )
= Φ (e) (r − r ) dr . (3.15)
∂ x
Following this pattern we derive two other important results for z-derivatives:
∂ (a) (a) ∂ Jz (r )
dr Gxz (x − x , y − y, z − z )Jz (r ) = dr Gxz (x − x , y − y , z − z )
∂z ∂ z
∂ (b) (b) ∂ Jz (r )
dr Gxz (x − x , y − y, z + z )Jz (r ) = − dr Gxz (x − x , y − y , z + z ) .
∂z ∂ z
(3.16)
∂ ∂ ∂ Jx ∂ Jy
where ∇t = ax + ay is the transverse-gradient operator; ∇t · J = + is
∂x ∂y ∂x ∂y
the transverse divergence; Jt = ax Jx + ay Jy , and
⎡ ⎤
∞ − z) (E) (M)
G1,−1 − G1,−1
(2)(a) − jω μ0 ⎣ c(z λ ⎦ J0 (rl)ldl
Gxx (x − x , y − y , z − z ) = −2
e 0 T
2π 0 2λ0 l2
∞ λ0 (z+z ) (E) (M)
(2)(b) − jω μ0 e G − G1
Gxx (x − x , y − y , z + z ) = e−2λ0 z0 1
2π 0 2λ0 l2
(E) (M)
e−λ0 (z+z ) 2λ0 z−1 G−1 − G−1
+ e J0 (rl)ldl.
2λ0 l2
(3.19)
The development of the Green’s dyad in Chap. 2 assumed that the unknowns were
the dual currents, Jm and Je . By “dual” we mean that one appears in the Faraday–
Maxwell law (the first equation in (2.2)) and the other in the Ampere–Maxwell
law (the second equation in (2.2)). This works well in presenting the theory in a
coherent and consistent manner, but it also means that there will be “dual” Green’s
dyads for electric–electric and magnetic–magnetic interactions, as shown in Chap. 2.
It is advantageous from a computational perspective to be able to use as much
code as possible when computing the various dyadic components, and in order to
accomplish this we find that writing the anomalous magnetic currents as “Amperian
currents” in Ampere’s law is beneficial.
We start with Maxwell’s equations
∇ × E = − jω B
∇ × H = jω D + J(e) . (3.20)
Now H = B/μ (r) = B/ μh + B/μ (r) − B/μh = B/μh − Ma , where μh is the host
permeability and Ma is the anomalous magnetization vector. Thus the second of
Maxwell’s equations may be written
∇ × B/μh = jω D + J(e) + ∇ × Ma , (3.21)
which makes clear that the Amperian current, ∇ × Ma , is an equivalent anomalous
electric current that arises because of the departures of the magnetic permeability of
the workpiece from the host permeability, μh . J(e) , on the other hand, is an electric
current that includes the anomalous current that arises due to differences in electrical
conductivity; J(e) = σh E + (σ (r) − σh)E = σh E + Ja.
Even though the Amperian current is electrical, because it appears as a source
term in the second Maxwell equation (Ampere’s law), we will refer to it as J(m) , to
remind us that it is of magnetic origin, and to distinguish it from J(e) (which now
stands for the anomalous electric current, Ja ). The important point, however, is that
because the Amperian current behaves as an electrical current, we need only use
electric–electric Green functions in the new formulation of the problem, as we now
show.
The coupled volume-integral equations for J(e) and J(m) are
J(e) (r)
E(i) (r) = − E(0) (r) J(e) − E(s) (r) J(e) − E(0) (r) J(m) − E(s) (r) J(m)
σa (r)
μ (r)μh 1 1
B(i) (r) = Ma + ∇ × E(0) (r) J(e) + ∇ × E(s)(r) J(e)
μ (r) − μh jω jω
1 1
+ ∇ × E(0)(r) J(m) + ∇ × E(s) (r) J(m) . (3.22)
jω jω
42 3 The Volume-Integral Equations for Plane-Layered Media
In arriving at the second equation, we have used the fact that B = −(1/ jω )∇ × E,
and Ma = ((μ (r) − μh )/ μ (r)μh )B.
The first part of the first equation in (3.22) is the usual electric–electric model,
(3.14). The remaining parts must be determined, and in order to do this we will use
the form of the equations given in Sect. 3.3. Because J(m) has zero divergence, we
can write
E(0) (r)[J(m) ] = Φ (e) (r − r )J(m) (r )dr
(1)(a) (m)
E(s) (r)[J(m) ] = dr Gxx (x − x , y − y , z − z )Jt (r )
(a) (m)
+az dr Gxz (x − x , y − y , z − z )Jz (r )
(1)(b) (m)
+ dr Gxx (x − x , y − y, z + z )Jt (r )
(b) (m)
+az dr Gxz (x − x , y − y , z + z )Jz (r )
(2)(a) (m)
+∇t dr Gxx (x − x , y − y, z − z )∇t · Jt (r )
(2)(b) (m)
+∇t dr Gxx (x − x , y − y, z + z )∇t · Jt (r ) , (3.23)
∂ ∂ ∂ Jx ∂ Jy
where ∇t = ax + ay ; ∇t · J = + ; Jt = ax Jx + ay Jy , and
∂x ∂y ∂x ∂y
∞
(1)(a) − jω μ0 c(z − z) −2λ0 T (E)
Gxx (x − x , y − y , z − z ) = e G1,−1 J0 (rl)ldl
2π 0 2λ0
(1)(b) − jω μ0 ∞ eλ0 (z+z ) −2λ0 z0 (E)
Gxx (x − x , y − y , z + z ) = e G1
2π 0 2λ0
e−λ0 (z+z ) 2λ0 z−1 (E)
+ e G−1 J0 (rl)ldl
2λ0
⎡ ⎤
∞ − z) (E)
−
(M)
(2)(a) − j ω μ c(z G G
Gxx (x − x , y − y , z − z ) = ⎣ e−2λ0 T ⎦ J0 (rl)ldl
0 1,−1 1,−1
2π 0 2λ0 l2
∞ λ0 (z+z ) (E) (M)
(2)(b) − jω μ0 e G − G1
Gxx (x − x , y − y , z + z ) = e−2λ0 z0 1
2π 0 2λ0 l2
(E) (M)
e−λ0 (z+z ) 2λ0 z−1 G−1 − G−1
+ e J0 (rl)ldl
2λ0 l2
3.4 The Volume-Integral Equations in Terms of Amperian Currents 43
∞
(a) − jω μ0 c(z − z) −2λ0 T (M)
Gxz (x − x , y − y , z − z ) = e G1,−1 J0 (rl)ldl
2π 0 2λ0
∞
(b) − jω μ0 eλ0 (z+z ) −2λ0 z0 (M)
Gxz (x − x , y − y , z + z ) = − e G1
2π 0 2λ0
e−λ0 (z+z ) 2λ0 z−1 (M)
− e G−1 J0 (rl)ldl . (3.24)
2λ0
Furthermore,
∇ × E(0) (r) J(e) = ∇ × Φ (e) (r − r)J(e) (r )dr
(1)(a) (e)
∇ × E(s) (r) J(e) = ∇ × dr Gxx (x − x , y − y, z − z )Jt (r )
(a) (e)
+∇ × az dr Gxz (x − x , y − y, z − z )Jz (r )
(1)(b) (e)
+∇ × dr Gxx (x − x , y − y , z + z )Jt (r )
(b) (e)
+∇ × az dr Gxz (x − x , y − y, z + z )Jz (r )
(2)(a) (e)
+∇ × ∇t dr Gxx (x − x , y − y , z − z )∇t · Jt (r )
(2)(b) (e)
+∇×∇t dr Gxx (x−x , y−y , z+z )∇t · Jt (r ) , (3.25)
with the same expressions holding for ∇ × E(0)(r) J(m) , ∇ × E(s) (r) J(m) .
Chapter 4
Discretization via the Galerkin Method
of Moments
We will discretize (3.22) by employing Galerkin’s method, which uses the same
vector functions for expansion and testing. The spatial derivatives that cause
problems will be removed by the testing process. In order to test these derivatives,
we introduce special vector expansion functions, called “facet elements” and “edge
elements,” that comprise products of pulse and tent functions.
Facet elements have been called “volumetric rooftop” functions by Catedra
et al. [36]. Volumetric rooftop functions have also been used in [37, 38]. These
functions are a generalization of two-dimensional rooftop functions that have been
used in problems involving scattering from two-dimensional structures and three-
dimensional surfaces [30, 32, 33].
Facet elements and edge elements are a subset of a more general class of spline-
generated basis-functions that are based upon higher-order convolutions of the unit
pulse
1, if 0 ≤ x < 1
π (x) = (4.1)
0, otherwise.
The reader is invited to study [31, 39] for a more complete development of the
subject.
We introduce facet elements oriented in the x-, y-, and z-directions, such that the
derivative with respect to x, y, and z is bounded. This ensures that the divergence of
the vector field is bounded, and for this reason facet elements are often referred to
Y
l δy (l+1) δy
Z
m δz (m+1) δz
(q)(e)
as “divergence-conforming.” We write Tklm (x, y, z) for the facet element oriented
(q)(e)
in the qth direction. The expressions for Tklm are:
(x)(e)
Tklm (r) = π2k (x/δ x)π1l (y/δ y)π1m (z/δ z)
(y)(e)
Tklm (r) = π1k (x/δ x)π2l (y/δ y)π1m (z/δ z)
(z)(e)
Tklm (r) = π1k (x/δ x)π1l (y/δ y)π2m (z/δ z) (k, l, m) = (0, 0, 0), . . . , (Nx , Ny , Nz ) ,
(4.2)
where π1m (y/δ y) is the mth unit pulse function and π2k (x/δ x) is the kth tent
function, which is the convolution of π1k (x/δ x) with itself. The T (q)(e) (r)klm are
called “facet elements,” because the qth element is constant over the qth facet of the
klmth cell.
Figure 4.1 shows the position of the tent and pulse functions for the facet element
(x)(e)
Tklm (x, y, z).
(x)(e)
The support of Tklm (x, y, z) is shown in Fig. 4.2. We assume that the conductiv-
ity is constant, with the value
(k+2,l+1,m)
(k+1,l+1,m)
(k,l+1,m)
(klm) δy
(k+1,l,m+1)
δz
δx
(x)(e)
Fig. 4.2 Support of Tklm (x, y, z). The conductivity is assumed to be constant within each cell of
dimension δ x × δ y × δ z
within each cell of dimension δ x × δ y × δ z. σmax and σmin are, respectively, the
maximum and minimum conductivities in the problem and Vc is the conductivity
volume-fraction.
We expand the anomalous electric current vector in terms of the facet elements as
and will then use these same basis-functions for testing the integral equations.
Because J(m) (r) = ∇ × Ma (r), we expand Ma (r) in “curl-conforming” edge
elements, which have the required differentiability of the curl operation
∑ MKLM TKLM
(x) (x)(m)
Mx (r) = (r)
KLM
∑ MKLM TKLM
(y) (y)(m)
My (r) = (r)
KLM
∑ MKLM TKLM
(z) (z)(m)
Mz (r) = (r), (4.5)
KLM
48 4 Discretization via the Galerkin Method of Moments
(l,m+1) (l+1,m+1)
δz
(l,m) (l+1,m)
Y
δy
(x)(m)
Fig. 4.3 Support of Tklm (x, y, z) in (y, z)-space. The support extends one cell in the x-direction,
normal to the page. The permeability is assumed to be constant within each cell of dimension
δx×δy×δz
where
(x)(m)
TKLM (r) = π1K (x)π2L (y)π2M (z)
(y)(m)
TKLM (r) = π2K (x)π1L (y)π2M (z)
(z)(m)
TKLM (r) = π2K (x)π2L (y)π1M (z). (4.6)
These functions are called edge elements because the expansion coefficient,
(x)
MKLM , is the (constant) value of Mx along the x-directed edge, (y = (L + 1)δ y, z =
(y) (z)
(M + 1)δ z). There are similar interpretations for MKLM and MKLM .
(x)(m)
The support of Tklm (x, y, z) is shown in Fig. 4.3. We assume that the magnetic
permeability is constant, with the value
(m) ∂ Mz ∂ My
Jx = −
∂y ∂z
= ∑ MKLM π2K (x)π2L
(z) (y)
(y)π1M (z) − MKLM π2K (x)π1L (y)π2M (z)
KLM
4.2 Testing the Integral Equations 49
(m) ∂ Mx ∂ Mz
Jy = −
∂z ∂x
= ∑ MKLM π1K (x)π2L (y)π2M
(x) (z)
(z) − MKLM π2K (x)π2L (y)π1M (z)
KLM
(m) ∂ My ∂ Mx
Jz = −
∂x ∂y
= ∑ MKLM π2K
(y) (x)
(x)π1L (y)π2M (z) − MKLM π1K (x)π2L (y)π2M (z)
KLM
(m) ∂ 2 Mx ∂ 2 My
∇t · Jt = −
∂ y∂ z ∂ x∂ z
= ∑ MKLM π1K (x)π2L
(x) (y)
(y)π2M (z) − MKLM π2K (x)π1L (y)π2M (z) , (4.8)
KLM
where the last term is the transverse divergence of the Amperian current.
The first step in discretizing the integral equations, (3.22), is to substitute the
expansions, (4.4) and (4.8), for the anomalous currents into the equations and then
to “test” the resulting equation. By “testing” we mean taking moments, which is
done by multiplying a functional equation by a test function and then integrating
over space. When the test function is the same as the expansion function for the
unknowns, the method is called the Galerkin variant of the method of moments. We
will test each component of the electric equations of (3.22) by the corresponding
(q)(e)
facet function, Tklm (r), and each component of the magnetic equations by the
(q)(m)
corresponding edge function, Tklm (r). The procedure is straightforward, but quite
lengthy, so we will show only the results. The terms of (3.22) that involve spatial
derivatives are “mollified” by using the two vector identities that are defined in
Appendix A.1.
The discretized electric equation is:
⎡ ⎤ ⎡ (x) ⎤(ee) ⎡ (x) ⎤
E(ix) Q 0 0 J
⎣ E(iy) ⎦ = ⎣ 0 Q(y) 0 ⎦ ⎣ J(y) ⎦
E(iz) 0 0 Q(z) J(z)
⎡ (xx) (xy) (xz) ⎤(ee)
G G(0) G(0) ⎡ (x) ⎤
⎢ (0) ⎥ J
⎢ (yx) (yy) (yz) ⎥ ⎣
− ⎣ G(0) G(0) G(0) ⎦ J(y) ⎦
(zx) (zy)
G(0) G(0) G(0)
(zz) J(z)
50 4 Discretization via the Galerkin Method of Moments
⎡ ⎤(ee)
(xx) (xy) (xz) ⎡ ⎤
G(a) G(a) G(a)
⎢ (yx) (yy) (yz) ⎥ J(x)
−⎢
⎣ G(a) G(a) G(a) ⎦
⎥ ⎣ J(y) ⎦
(zx) (zy) (zz)
G(a) G(a) G(a) J(z)
⎡ ⎤(ee)
(xx) (xy) (xz) ⎡ ⎤
G(b) G(b) G(b)
⎢ (yx) (yy) (yz) ⎥ J(x)
−⎢
⎣ G(b) G(b) G(b) ⎦
⎥ ⎣ J(y) ⎦
(zx) (zy) (zz)
G(b) G(b) G(b) J(z)
⎡ (x) ⎤
(em) M
− G(0) G(a) G(b) ⎣ M(y) ⎦ , (4.9)
M(z)
where the Q’s are tri-diagonal matrices, the G(0) ’s the infinite-space matrices, the
G(a) ’s the convolutional layered-space matrices, and the G(b) ’s the correlational
layered-space matrices. The infinite-space matrices are convolutional, also. The
superscript (ee) denotes electric–electric matrices and (em) denotes electric–
magnetic matrices. The J’s are the unknown electric currents and the M’s are
the unknown magnetic polarization vectors. The last block in (4.9) is simply a
shorthand representation of the three blocks above it, except that it represents
electric–magnetic interactions.
The discretized magnetic equation is similar to (4.9) and is given by
⎡ ⎤ ⎡ (x) ⎤(mm) ⎡ (x) ⎤
B(ix) Q 0 0 M
⎣ B(iy) ⎦ = ⎣ 0 Q(y) 0 ⎦ ⎣ M(y) ⎦
B(iz) 0 0 Q(z) M(z)
⎡ (xx) (xy) (xz) ⎤(mm)
G G(0) G(0) ⎡ (x) ⎤
⎢ (0) ⎥ M
(yx) (yy) (yz) ⎥
+⎢ G G
⎣ (0) (0) (0) ⎦ G ⎣ M(y) ⎦
(zx) (zy)
G(0) G(0) G(0)
(zz) M(z)
⎡ ⎤(mm)
(xx) (xy) (xz) ⎡ ⎤
G(a) G(a) G(a)
⎢ (yx) (yy) (yz) ⎥ M(x)
+⎢
⎣ G(a) G(a) G(a) ⎦
⎥ ⎣ M(y) ⎦
(zx) (zy) (zz)
G(a) G(a) G(a) M(z)
⎡ ⎤(mm)
(xx) (xy) (xz) ⎡ ⎤
G(b) G(b) G(b)
⎢ (yx) (yy) (yz) ⎥ M(x)
+⎢
⎣ G(b) G(b) G(b) ⎦
⎥ ⎣ M(y) ⎦
(zx) (zy) (zz)
G(b) G(b) G(b) M(z)
⎡ (x) ⎤
(me) J
+ G(0) G(a) G(b) ⎣ J(y) ⎦ , (4.10)
J(z)
4.3 Solution via Conjugate Gradients 51
where B is the incident magnetic flux density due to the coil, the superscript (mm)
stands for magnetic–magnetic interactions, and (me) stands for magnetic–electric
interactions. The magnetic–magnetic Q matrices are a little more complicated than
the electric–electric ones.
The system of equations, (4.9) and (4.10), that is produced by the method of
moments contains a dense matrix, as opposed to the sparse matrices that are
produced by finite-element or finite-difference techniques. Hence, it is necessary
to develop efficient means of solving the system if the volume-integral method is
to be viable in solving realistic three-dimensional problems. When the number of
unknowns is 3,000, we can use a direct matrix-decomposition solver, such as
the LU-decomposition [41], but for large problems we use the conjugate-gradient
algorithm [42].
The conjugate-gradient algorithm, being an iterative method, requires many
matrix-vector multiplications. In order for this method to be useful, therefore, there
must be an efficient method for computing these products. Fortunately, because we
have formulated the volume-integral equations on a regular grid, there exists a very
efficient numerical scheme for computing vector-matrix products.
The triple sums that appear in (4.9) and (4.10) consist of three-dimensional
convolutions, or two-dimensional convolutions and one-dimensional correlations,
which means that we can use three-dimensional FFTs to compute them.
The appropriate theorems (in one dimension) that relate discrete Fourier
transforms and convolutions and correlations are ( ⇐⇒ denotes a discrete transform-
pair):
If g( j) ⇐⇒ G(n)
h( j) ⇐⇒ H(n)
k=0 g( j + k)h(k) = N ∑k=0 g(k)h(k − j)
Then N1 ∑N−1 1 N−1
⇐⇒ G(n)H(−n)
= G(n)H(N − n)
∑
1 N−1
N k=0 g(k)h( j + k) = N1 ∑N−1
k=0 g(k − j)h(k)
⇐⇒ G(−n)H(n)
= G(N − n)H(n)
∑
1 N−1
N k=0 g(k)h( j − k) = N1 ∑N−1
k=0 g( j − k)h(k)
⇐⇒ G(n)H(n), (4.11)
52 4 Discretization via the Galerkin Method of Moments
⎡ ⎤ ⎡ ⎤⎡ ⎤
y0 m0 m−1 m−2 m−3 x0
⎢ y 1 ⎥ ⎢ m1 m0 m−1 m−2 ⎥ ⎢ x1 ⎥
⎢ ⎥=⎢ ⎥⎢ ⎥. (4.12)
⎣ y 2 ⎦ ⎣ m2 m1 m0 m−1 ⎦ ⎣ x2 ⎦
y3 m3 m2 m1 m0 x3
Rewrite this in the expanded form (padding with zeros to get a power of two) in
order to achieve a circulant-matrix:
⎡ ⎤ ⎡ ⎤⎡ ⎤
y0 m0 m−1 m−2 m−3 0 m3 m2 m1 x0
⎢ y 1 ⎥ ⎢ m1 m0 m−1 m−2 m−3 0 m3 ⎥ ⎢
m2 ⎥ ⎢ x 1 ⎥
⎢ ⎥ ⎢ ⎥
⎢y ⎥ ⎢m m3 ⎥ ⎢ ⎥
⎢ 2⎥ ⎢ 2 m1 m0 m−1 m−2 m−3 0 ⎥ ⎢ x2 ⎥
⎢ ⎥ ⎢ ⎥⎢ ⎥
⎢ y 3 ⎥ ⎢ m3 m2 m1 m0 m−1 m−2 m−3 0 ⎥ ⎢ x3 ⎥
⎢ ⎥=⎢ ⎥⎢ ⎥, (4.13)
⎢ ∗ ⎥ ⎢0 m3 m2 m1 m0 m−1 m−2 m−3 ⎥ ⎢ 0 ⎥
⎢ ⎥ ⎢ ⎥⎢ ⎥
⎢ ∗ ⎥ ⎢ m−3 0 m3 m2 m1 m0 m−1 m−2 ⎥ ⎢ ⎥
⎢ ⎥ ⎢ ⎥⎢ 0 ⎥
⎣ ∗ ⎦ ⎣ m−2 m−3 0 m3 m2 m1 m0 m−1 ⎣ 0 ⎦
⎦
∗ m−1 m−2 m−3 0 m3 m2 m1 m0 0
where the ∗ denotes a discarded entry. Hence, the sequences to be FFT’d are:
(m0 , m1 , m2 , m3 , 0, m−3 , m−2 , m−1 ) and (x0 , x1 , x2 , x3 , 0, 0, 0, 0), and the output se-
quence is (y0 , y1 , y2 , y3 , ∗, ∗, ∗, ∗). The order of the entries in the sequences is very
important.
Now, for correlations:
⎡ ⎤ ⎡ ⎤⎡ ⎤
y0 m0 m1 m2 m3 x0
⎢ y 1 ⎥ ⎢ m1 m2 m3 m4 ⎥ ⎢ x 1 ⎥
⎥ ⎢
⎢ ⎥=⎢ ⎥. (4.14)
⎣ y 2 ⎦ ⎣ m2 m3 m4 m5 ⎦ ⎣ x 2 ⎦
y3 m3 m4 m5 m6 x3
Rewrite this in the expanded form (padding with zeros to get a power of two) in
order to achieve a circulant-matrix:
4.3 Solution via Conjugate Gradients 53
⎡ ⎤ ⎡ ⎤⎡ ⎤
y0 m0 m1 m2 m3 m4 m5 m6 0 x0
⎢ y 1 ⎥ ⎢ m1 m2 m3 m4 m5 m6 0 ⎥ ⎢
m0 ⎥ ⎢ x 1 ⎥
⎢ ⎥ ⎢ ⎥
⎢y ⎥ ⎢m m1 ⎥ ⎢ ⎥
⎢ 2⎥ ⎢ 2 m3 m4 m5 m6 0 m0 ⎥ ⎢ x2 ⎥
⎢ ⎥ ⎢ ⎥⎢ ⎥
⎢ y 3 ⎥ ⎢ m3 m4 m5 m6 0 m0 m1 m2 ⎥ ⎢ x 3 ⎥
⎢ ⎥=⎢ ⎥⎢ ⎥, (4.15)
⎢ ∗ ⎥ ⎢ m4 m5 m6 0 m0 m1 m2 m3 ⎥ ⎢ 0 ⎥
⎢ ⎥ ⎢ ⎥⎢ ⎥
⎢ ∗ ⎥ ⎢ m5 m6 0 m0 m1 m2 m3 m4 ⎥ ⎢ ⎥
⎢ ⎥ ⎢ ⎥⎢ 0 ⎥
⎣ ∗ ⎦ ⎣ m6 0 m0 m1 m2 m3 m4 m5 ⎣ 0 ⎦
⎦
∗ 0 m0 m1 m2 m3 m4 m5 m6 0
where the ∗ denotes a discarded entry. Hence, the sequences to be FFT’d are:
(m0 , m1 , m2 , m3 , m4 , m5 , m6 , 0) and (x0 , x1 , x2 , x3 , 0, 0, 0, 0), and the output sequence
is (y0 , y1 , y2 , y3 , ∗, ∗, ∗, ∗). The order of the entries in the sequences is very important,
and also don’t forget to negate the frequencies in the transform of the x-sequence.
To summarize: we expand the original data, padding with zeros as necessary to
get a circulant matrix, and then take FFTs.
Y = A X, (4.16)
where Y denotes the known left-hand side and X denotes the vector of unknown
currents. The conjugate gradient (CG) algorithm starts with an initial guess, X0 ,
from which we compute R0 = Y − A X0 , P1 = Q0 = A ∗ R0 , where A ∗ is the adjoint
operator that corresponds to the conjugate-transpose of the matrix blocks in (4.9)
and (4.10). In addition, we have a convergence parameter, ε . Then for k = 1, . . ., if
Test = Rk /Y < ε , stop; Xk is the optimal solution of (4.16). Otherwise, update
Xk by the following steps:
Sk = A Pk
Qk−1 2
ak =
Sk 2
Xk = Xk−1 + ak Pk
Rk = Rk−1 − ak Sk
Qk = A ∗ R k
Qk 2
bk =
Qk−1 2
Pk+1 = Qk + bk Pk . (4.17)
54 4 Discretization via the Galerkin Method of Moments
The convolution and correlation operations that are a part of A and A ∗ are
evaluated by using the FFT, as described in the preceding section. This, together
with the fact that the storage requirements are reasonably modest, is the reason why
the conjugate gradient algorithm becomes attractive for solving potentially large
problems in our model.
Appendix
Let the testing vector function, B(r), have a finite support, then the curl-operator
can be transferred from one vector function to another within an integral:
B · ∇ × Adr = A · ∇ × Bdr + ∇ · (A × B)dr
= A · ∇ × Bdr + (A × B) · dS
= A · ∇ × Bdr, (4.18)
VIC-3D R
computes impedances from field calculations by using the reaction
concept. We will follow Harrington [47, pp. 116–120], in developing this concept.
Harrington defines reaction as
[a, b] = E(a) · J(b) dV , (5.1)
where E(a) is the field due to source a and J(b) is source b. The reciprocity theorem
states that
that is, the reaction of field a on source b is equal to the reaction of field b on
source a.
Harrington also shows that the driving-point and transfer impedances of a linear
network can be defined in terms of reactions by
[ j, i] [i, j]
zi j = − =− . (5.3)
Ii I j Ii I j
In Fig. 5.1 we show the classical bistatic arrangement. This arrangement subsumes
the usual “driver-pickup” and “remote-field” configurations. The transmitter in
Fig. 5.1 is excited, and the receiver drives an infinite-impedance load. Hence, the
equivalent circuit is as shown in the lower part of Fig. 5.1.
The circuit equations are
V1 = jω L1 I1 + jω M13 I3
0 = jω M13 I1 + (R3 + jω L3 )I3 . (5.4)
I1 is the actual current flowing in the exciting (transmitter) coil and I3 is the
distributed anomalous current due to the flaw. I1 is a circuit current and I3 a
distributed current.
From the second equation in (5.4) we have
Transmit Receive
Flaw
I1
V1 L1 L2 V2
Transmit Receive
L3
Flaw I3
jω M13
I3 = − I1 . (5.5)
R3 + jω L3
V0 = jω M12 I1 + jω M23 I3
[2, 3]
= jω M12 I1 − , (5.6)
I2
where we have once again used (5.3) in writing the second term.
The transfer impedance, Z01 , is defined to be
V0
Z01 =
I1
[2, 3]
= jω M12 −
I1 I2
[2, 3]
= z12 − . (5.7)
I1 I2
Now, z12 represents the direct coupling between the transmitter and receiver; as
such it does not take into account the presence of the flaw, which is represented by
the distributed current, I3 , but, rather, represents the background against which the
flaw must be detected. In remote-field inspection, the transmitter and receiver are
sufficiently far apart (remote) that z12 ≈ 0; that is, the direct coupling is practically
zero.
We can look at this another way, as well. Because z12 has nothing to do with
the flaw, we can subtract its effect initially, even if the bistatic arrangement is not
remote-field. Hence, we get for the impedance change that is due solely to the flaw:
[2, 3]
ΔZ = − , (5.8)
I1 I2
where
[2, 3] = E(2) · J(3)dV . (5.9)
flaw
Note that I1 is the current that actually drives the transmitter, but I2 is an apparent
circuit current that drives the receiver, when the receiver is treated as a transmitter,
as well. (It is a feature of reciprocity that it transforms transmitters into receivers
and receivers into transmitters.)
Because I1 and I2 are circuit currents, they are scalars that can be set to unity.
If we normalize I1 and I2 to unity, then E(2) in (5.9) is the incident field within the
60 5 Computing Network Immittance Functions from Field Calculations
flaw, due to one ampere in the receiver coil, when the receiver acts like a transmitter.
Note that J(3) is independent of E(2) , because there is only one source for J(3) ,
namely the one-ampere current in the true transmitter, coil 1.
VIC-3D R
computes the reaction [2, 3] by means of a scalar product of two
discrete vectors. This can be seen by substituting the current expansion, (4.4),
into (5.9):
[2, 3] = E(2) · J(3)dV
flaw
∑
(3)(x)
= JKLM E (2)(x) (r)T (x)(e) (r)dV
KLM flaw
(3)(y)
+ JKLM E (2)(y) (r)T (y)(e) (r)dV
flaw
(3)(z) (2)(z) (z)(e)
+ JKLM E (r)T (r)dV
flaw
∑
(3)(x) (2)(x) (3)(y) (2)(y) (3)(z) (2)(z)
= JKLM EKLM + JKLM EKLM + JKLM EKLM , (5.10)
KLM
(2)
where EKLM is the KLMth moment of E(2) . This expansion is extremely important
and will appear throughout this text.
The equivalent circuit for a differential-probe system is shown in Fig. 5.2. The
exciting coil is driven by the ac source, while the two sense coils are connected
Sense
I1 Coils +
L2 V2
L1
V1
Op-Amp
Exciting
Coil L3 V3
−
L4
Flaw I4
V1 = jω L1 I1 + jω M14 I4
0 = jω M14 I1 + (R4 + jω L4 )I4 , (5.11)
where I1 is the actual current flowing in the exciting coil and I4 represents the effects
of the distributed anomalous electric current that flows within the flaw. As before,
I1 is a circuit-current and I4 is a distributed current. The second equation in (5.11)
shows that the distributed current is due only to the exciting current:
jω M14
I4 = − I1 . (5.12)
R4 + jω L4
Now, because of the symmetrical placement of the two sense-coils with respect
to the exciting coil, we have M12 = M13 . This holds regardless of the presence of
the flaw, because the flaw effects are included in the second term in (5.13). Hence,
(5.13) becomes
V0 = jω (M24 − M34)I4
[3, 4] [2, 4]
= − . (5.14)
I3 I2
We call the ratio, V0 /I1 = Z01 , the open-circuit transfer impedance of the linear
network consisting of the coupled circuits of Fig. 5.2. In terms of the reactions, then,
we have
[3, 4] [2, 4]
Z01 = − . (5.15)
I1 I3 I1 I2
I1 is the actual current in the exciting-coil, whereas I2 and I3 are fictitious currents
in the two sense-coils, when these coils are treated as transmitters when applying
reciprocity theory. If I1 = I2 = I3 = 1, then (5.15) is expressed solely in terms of
reactions
We calculate the driving-point impedance first. This is the impedance seen at the
terminals of the probe in the absence of the flaw; that is, I4 = 0 in (5.11). This
impedance consists of two parts, the first being the contribution of the coil and
the second the contribution of the ferrite core. We will compute each of these
contributions by using the reaction principle.
(1)
The reaction of field E(1) on source Je is
(1)
[1, 1] = Je · E(1) dV . (5.17)
The source with superscript 1 is the primary source due to the exciting coil. If Ic is
the current in the exciting coil, then the driving-point impedance (or self-impedance)
seen by the coil is
[1, 1]
Z=− . (5.18)
Ic2
If we normalize the excitation to be Ic = 1, then
Z=− Je · E(i) dV
=− (Jc + ∇ × MC ) · E(i) dV
=− Jc · E(i) dV − MC · ∇ × E(i) dV
=− Jc · E(i) dV + MC · jω B(i) dV , (5.19)
where we have replaced the superscript 1 by (i) to denote incident fields. The
transference of the curl operator in going from the second to the third equation is
valid for M with finite support (see Chap. 4, Appendix A.1). Jc is the current density
in the coil and MC is the magnetization of the core.
Upon substituting the expansions for the magnetic solution vectors, (4.5), into
(5.19), we get
Z=− Jc · E(i)
+ jω ∑ MKLM BKLM + MKLM BKLM + MKLM BKLM ,
(x) (ix) (y) (iy) (z) (iz)
(5.20)
KLM
(i)
where BKLM is the KLMth moment of B(i) . The first term is the contribution of the
coil to the driving-point impedance and is computed in Appendix A.1; the second is
the contribution of the core. The scalar product in this term is reminiscent of (5.10).
5.4 Computation of Impedance Changes due to the Presence of a Flaw 63
Here, the crucial thing that must be computed is the reaction between the anomalous
current produced by the flaw and the incident field of a coil. In the case of the
differential ferrite-core probe, we want to compute [2, 4] and [3, 4] of (5.16).
(1)
The reaction of field, E(2) , on source, Je , is
(1)
[1, 2] = Je · E(2) dV . (5.21)
The source with superscript 1 is the primary source due to the exciting coil and
superscript 2 denotes scattered fields (and their sources) due to the flaw. If Ic is the
current in the exciting coil, then the change in impedance due to the flaw, as seen by
the coil is
[1, 2] [2, 1]
ΔZ = − =− 2 , (5.22)
Ic2 Ic
where we have dropped the superscript 2 and replaced the superscript 1 by (i) to
denote incident fields. The transference of the curl operator in going from the second
to the third equation is valid for M with finite support (see Chap. 4, Appendix A.1).
Upon substituting the expansion for the magnetic solution vectors, (4.5), and the
corresponding one for the electric current, (4.4), into (5.23), we get
∑
(x) (ix) (y) (iy) (z) (iz)
ΔZ = − JKLM EKLM + JKLM EKLM + JKLM EKLM
KLM
(x) (ix) (y) (iy) (z) (iz)
− jω MKLM BKLM + MKLM BKLM + MKLM BKLM . (5.24)
This is a sum of scalar products of the electric current and magnetic polarization
solution vectors with the incident electric field and magnetic flux-density moment
vectors.
64 5 Computing Network Immittance Functions from Field Calculations
Appendix
where
∞ −λ0 |z−z |
1 e + R0 e−λ0 (z+z )
G00 (r, z; r , z ) = aφ J1 (rl)J1 (r l)ldl . (5.26)
2π 0 2λ0
R0 is the reflection coefficient from the top surface of the workpiece and is calculated
using the ideas presented in Chap. 2.
We assume that the coil has a rectangular cross-section and carries a current,
Ic = 1A, that is uniformly distributed over this cross-section with Nc turns per square
meter (Jc = Nc aφ ). The integrals with respect to (r , z ) over the coil are evaluated
first:
G00 (r, z; r , z ) · Jc (r , z )r dr dz =
coil
∞
Nc I (r1 l, r2 l) F2 (z1 , z2 , z) + R0 e−λ0 z F1 (z1 , z2 )
J1 (rl)dl , (5.27)
2π 0 l 2λ0
where (r1 , r2 ) are the inner and outer radii of the coil and (z1 , z2 ) the bottom and top
coordinates of the coil.
Then
Zc = − Jc · E(i) dV
= −2π Nc E (i) (r, z)rdrdz
= jω μ0 (2π )2 Nc rdrdz G00 (r, z; r , z ) · Jc (r , z )r dr dz
coil coil
I (r1 l, r2 l)∞
= jω μ0 2π Nc2 rdrdz
coil 0 l
F2 (z1 , z2 , z) + R0 e−λ0 z F1 (z1 , z2 )
× J1 (rl)dl
2λ0
∞ 2
I (r1 l, r2 l) F3 (z1 , z2 ; z1 , z2 ) + R0F12 (z1 , z2 )
= jω μ0 2π Nc 2
dl .(5.28)
0 l3 2λ0
A.1 Calculation of Circular Coil Impedance 65
The functions F1 (z1 , z2 ), F2 (z1 , z2 , z), F3 (z1 , z2 ; z1 , z2 ) and I (r1 l, r2 l) are defined
by [27]
b
F1 (a, b) = e−λ0 z dz
a
e− λ 0 a − e− λ 0 b
=
λ0
b
F2 (a, b, c) = e−λ0 |c−z| dz
a
= I1 + I2 + I3 + I4
(+)
I1 = −2x1 /λ0 − e−λ0 x1 /λ02 , x1 > 0, x1 = zi − z2
= −eλ0 x1 /λ02 , x1 < 0
(−)
I2 = −2x2 /λ0 − e−λ0 x2 /λ02 , x2 > 0, x2 = zi − z1
= −eλ0 x2 /λ02 , x2 < 0
(−)
I3 = 2x3 /λ0 + e−λ0x3 /λ02 , x3 > 0, x3 = zi − z2
= eλ0 x3 /λ02 , x3 < 0,
(+)
I4 = 2x4 /λ0 + e−λ0x4 /λ02 , x4 > 0, x4 = zi − z1
= eλ0 x4 /λ02 , x4 < 0
b
I (a, b) = zJ1 (z)dz , (5.29)
a
Maxwell’s second equation defines the three current systems that we are inter-
ested in:
∇ × H = jω D + Jc + σ (r)E
= jω D + Jc + σh E + (σ (r) − σh)E
= j ω D + Jc + Jh + Ja , (5.30)
where jω D is the displacement current, Jc the coil current, Jh the host current, and
Ja the anomalous current due to the anomalous conductivity, σ (r) − σh .
Setting aside consideration of the displacement current, the three remaining
current systems are going to be modeled by the three coupled circuits of Fig. 5.3 in
order to get a simpler interpretation of the volume-integral equation. Mch = Mhc is
the mutual inductance between the coil and host, Mca = Mac is the mutual inductance
between the coil and anomaly, and Mha = Mah is the mutual inductance between the
host and anomaly.
The loop equations (Kirchoff’s Voltage Law) for this system are
The circuit elements, R and L, should not be taken literally. They represent such
things as energy loss and storage, which occur in both lumped circuits (such as this
one) and distributed fields.
Ic Ih
Rc
+
Mch
Coil V0 Lc Lh Rh Host
−
Mca Mha
Ia
La
ω 2 Mca
2 (R + j ω L ) − j ω 3 M M M
h h ch ha ca
+
(Rh + jω Lh )(Ra + jω La ) + ω 2 Mha
2
ω 2 Mch
2
= Rc + jω Lc +
Rh + jω Lh
ω 2 Mca
2 (R + j ω L )2 − j2ω 3 M M M (R + j ω L ) − ω 4 M 2 M 2
h h ch ca ha h h
+ ! " ch ha .
(Rh + jω Lh ) (Rh + jω Lh )(Ra + jω La ) + ω 2Mha
2
(5.34)
The first two terms in (5.34) are the free-space coil driving-point impedance
and the third is the additional term due to the host in the absence of an anomaly
(the “unflawed host”). This term vanishes when the coil is well removed from the
workpiece, so that Mch → 0.
68 5 Computing Network Immittance Functions from Field Calculations
The final term is the change caused by the presence of an anomaly. It vanishes
under two distinct conditions: Ra → ∞, which, as we have seen before, corresponds
to the absence of the flaw, and Mca , Mha → 0. This condition corresponds to the
vanishing of Einc and the integral operator in (5.33), thereby forcing the anomalous
current, Ia , to vanish. This occurs when the coil is well removed from the anomaly.
Let’s take another look at (5.34). Each of the last two terms can be interpreted
as an “Einc · J” with an appropriate Einc and J. If Ia = 0, then the solution of the
second equation in (5.31) for Ih is Ih = − jω Mch Ic /(Rh + jω Lh ), and when this is
multiplied by Einc = − jω Mch Ic we get the second term in (5.34), after negation and
setting Ic = 1. Hence, we interpret Ih as being the anomalous or “scattering” current
when Ia = 0, and − jω Mch Ic as the incident field acting in free space.
The more usual case in the hierarchy is Ia = 0, which means that (5.33) applies.
If we solve this equation for Ia , and then multiply that result by the left-hand side,
we get the third term in (5.34), after negation and setting Ic = 1. Again, we have
consistency with the theory of calculating changes in impedance, but now Einc must
account for the presence of the host current, Ih , as well as the coil current, Ic . The
environment of each of the scattering currents is different, and that determines how
Einc is to be computed. Note that Einc in (5.33) contains two terms, the first being
the direct contribution of the coil acting in free-space and the second the effect of
the coil acting through the host current, Ih .
We can now derive some very useful results for analyzing and interpreting model
results based on the frequency response of the measured impedance. Return to (5.34)
and set Mca = Mha = 0, which corresponds to the unflawed host. Throughout this
discussion we will work with Z = Zin − Rc , which is done because Rc has nothing
to do with magnetic coupling of the coil to the host. Furthermore, we will work with
Z /ω Lc , which is the impedance normalized to the free-space reactance of the coil.
It is, of course, a dimensionless quantity.
We have, after rationalizing the third (coupling) term in (5.34),
2 L /L
Z ω Mch2 R
h ω 2 Mch h c
= + j 1 −
ω Lc Lc (R2h + ω 2 L2h ) R2h + ω 2 L2h
2
ω k2 Lh Rh Rh + ω 2 L2h − ω 2k2 L2h
= 2 +j , (5.35)
Rh + ω 2 L2h R2h + ω 2 L2h
where we have introduced the coupling-coefficient, k2 = Mch /Lc Lh [108, page 399].
A.2 A Coupled-Circuit Model of the Volume-Integral Equation 69
k2
1−
2
1−k2
R’
ω Lc
It will be advantageous to shift the origin of the coordinate system for the
normalized impedance to j(1 − k2 /2):
2
Z k2 ω k2 Lh Rh Rh + ω 2 L2h − ω 2 k2 L2h k2
− j 1− = 2 +j − 1−
ω Lc 2 Rh + ω 2 L2h R2h + ω 2L2h 2
k2 2ω Lh Rh R2h − ω 2 L2h
= + j , (5.36)
2 R2h + ω 2 L2h R2h + ω 2 L2h
This is an extremely important result and is the basis for interpreting both model
and experimental results. If the circuit parameters, Rh , Lh , and k are all frequency
independent, as in an “ideal” lumped circuit, then the locus of impedance vs.
frequency is a semicircle, the “ideal” Förster plot shown in Fig. 5.4. The diameter
of this semicircle directly measures the coupling coefficient, k2 . The terminal points
of the diameter correspond to ω = 0 and ω → ∞, where R /ω Lc vanishes.
Of course, field calculations of eddy-current phenomena do not lead to “ideal” √
lumped circuit elements. For example, we know that at high frequencies Rh ∼ ω
because of the skin effect.1 This means that R /ω Lc in (5.35) dies out more slowly
1 See [109, pp. 286–303] for a thorough discussion of skin effect on circuit elements.
70 5 Computing Network Immittance Functions from Field Calculations
0 0
−0.01
−0.05 −0.02
−0.03
Normalized X
Normalized X
−0.1
−0.04
−0.05
−0.15
−0.06
−0.2 −0.07
−0.08
−0.25 −0.09
0 0.01 0.02 0.03 0.04 0.05 0.06 0.07 0 0.01 0.02 0.03 0.04 0.05 0.06
Normalized R Normalized R
Fig. 5.5 Normalized impedance responses over the frequency range of 100 Hz to 1 MHz, for a
given coil and lift-off, but for two half-spaces with different conductivities. The half-space on the
left has σ = 3 × 107 S/m, and that on the right σ = 3 × 105 S/m
By a “nonstandard probe,” we mean one that is not the usual air-core circular coil
with a square cross-section, whose axis is normal to a plane surface, as in the
pancake coil of Appendix A.1 of Chap. 5, or with its axis coinciding with the axis
of a tube, as in the bobbin coils of Chap. 9. In this chapter, we develop a theory
that allows us to efficiently discretize the current density in such probes, thereby
allowing us to use transfer matrices (these will be discussed in Chap. 10) and similar
mathematical constructs for solving problems.
In particular, this chapter deals with probes that are characterized by their
magnetic-dipole moments, which means that all equivalent electric currents are
solenoidal (divergence-free). These models allow us to calculate inductive effects,
only, which, of course, is a main concern in eddy-current modeling. In the next
chapter, we analyze probes that are driven by electric dipoles, which allows us to
calculate capacitive effects as well as inductive effects. This allows the probe models
to be used over a greater frequency range.
We only need to compute the incident electric field at a flaw, because the incident
magnetic field can be expressed in terms of the electric field, as we know.
The Green function that we have worked with previously was that for which
the source and field points were in the same layer of the workpiece. Now, upon
referring to Fig. 2.1 of Chap. 2, we need the Green function, G(±q0) , in which the
source point is in region 0, as before, but the field point is in the qth layer above
region 0, or the qth layer below the region. The first situation carries the label +q0
and the second −q0.
We compute the Green functions using the same four-vector algebraic approach
that was used in Chap. 2, and then the resulting volume-integral differential equation
is derived exactly as is explained in Chap. 3. Because the Amperian currents,
which are of magnetic origin, are actually equivalent electric currents [because
they appear in the second of Maxwell’s equations (Ampere’s law)], we only need
to compute the electric–electric dyadic Green functions. The resulting differentio-
integral equations are
Ex (r) = G(q0)(1)(x − x , y − y ; z, z )Jx (r )dr
∂
+ G(q0)(2)(x − x , y − y; z, z )∇t · Jt (r )dr
∂x
1 ∂
+ 2 G(q0)(3) (x − x , y − y; z, z )∇ · J(r )dr
k0 ∂ x
Ey (r) = G(q0)(1)(x − x , y − y ; z, z )Jy (r )dr
∂
+ G(q0)(2)(x − x , y − y ; z, z )∇t · Jt (r )dr
∂y
1 ∂
+ G(q0)(3)(x − x , y − y ; z, z )∇ · J(r )dr
k02 ∂ y
Ez (r) = G(q0)(4)(x − x , y − y ; z, z )Jz (r )dr
1 ∂
+ G(q0)(5) (x − x , y − y; z, z )∇ · J(r )dr , (6.1)
k02 ∂ z
which holds for r in region q. If r is in region −q, then replace (q0) by (−q0).
The result of (6.1) is quite general and could be used, for example, in finding
the interaction between a flaw in region 0 and one in region ±q, as in the Spatial
Decomposition Algorithm of Chap. 10. The Amperian current produced by the core
of such a probe has zero-divergence, as does the current in the exciting coil of the
probe. Hence, the last term in each of the equations in (6.1) vanishes. The electric
currents that flow in the various coils of the probe lie in the transverse, (x, y), plane,
because of the orientation of the coils. This means that ∇t · J(r ) = 0 for these
currents.
Upon calling the coil current, J(c) (r), and the current of magnetic origin (the
Amperian current) J(m) (r), then (6.1) becomes
(c)
Ex (r) = G(q0)(1) (x − x , y − y ; z, z )Jx (r )dr
(m)
+ G(q0)(1) (x − x , y − y; z, z )Jx (r )dr
∂ (m)
+ G(q0)(2) (x − x , y − y ; z, z )∇t · Jt (r )dr
∂x
6.3 A Solenoidal Current Model 73
(c)
Ey (r) = G(q0)(1) (x − x , y − y ; z, z )Jy (r )dr
(m)
+ G(q0)(1) (x − x , y − y; z, z )Jy (r )dr
∂ (m)
+ G(q0)(2) (x − x , y − y; z, z )∇t · Jt (r )dr
∂y
(m)
Ez (r) = G(q0)(4) (x − x , y − y ; z, z )Jz (r )dr . (6.2)
The kernels in (6.1) and (6.2) are computed using the same four-vector approach of
Chap. 2, but because of their complexity will not be shown here.
Equation (6.2) is still quite rigorous, within the context of the nature of the
currents, J(c,m) (r). We shall see shortly that is possible for common eddy-current
problems to eliminate the Ez (r) term, also. This occurs when the coil region is air,
and the field point is in a highly-conducting region, which is the usual combination
in eddy-current problems. The contributions of the coil current in the x- and
y-equation of (6.2) are precisely the classical terms produced by circular coils of
rectangular cross-section that VIC-3D R
already computes. Of course, if the coils
are noncircular, but, say, are rectangular in the (x, y)-plane, then we must evaluate
the contributions in the same manner as in the case of the magnetic currents. See
[92] for a benchmark validation of this model.
(p)(m)
where TKLM,q denotes the qth spatial derivative of the pth-oriented edge element.
The parameters, δx , δy , δz , refer to the source grid, not the flaw grid.
74 6 Advanced Probe Models Based on Magnetic Dipoles and Ferrite Cores
(l+1)*dy (m+1)*dz
l*dy m*dz
y
z
(l-1)*dy (m-1)*dz
(k-1)*dx k*dx (k+1)*dx (k-1)*dx k*dx (k+1)*dx
x x
(z)(m) (z)(m)
Fig. 6.1 Circulations. Left: Tklm,x (r)ay − Tklm,y (r)ax , in the (x, y)-plane, about the edge joining
(y)(m) (y)(m)
the points (xk , yl , zm−1 ) and (xk , yl , zm ); right: Tklm,x (r)az − Tklm,z (r)ax , in the (x, z)-plane, about
the edge joining the points (xk , yl−1 , zm ) and (xk , yl , zm )
(z)(m) (z)(m)
The physical interpretation of the basis vectors is: Tklm,x (r)ay − Tklm,y (r)ax is
a circulation in the x–y plane, about the edge joining the points (xk , yl , zm−1 ) and
(y)(m) (y)(m)
(xk , yl , zm ). Tklm,x (r)az − Tklm,z (r)ax is a circulation in the x–z plane, about the
edge joining the points (xk , yl−1 , zm ) and (xk , yl , zm ). The circulation is confined to
four adjacent cells sharing the edge on the axis of circulation. These circulations are
plotted in Fig. 6.1.
The solution of (6.3) for the expansion coefficients is given by
yl xk
(y) 1 (z)
Mklm = Jc (x, y, zm )dxdy
δy yl−1 x0
zm xk
(z) 1 (y)
Mklm = − Jc (x, yl , z)dxdz , (6.4)
δz zm−1 x0
(y)
where x0 lies outside the source region. Now that we have MKLM = 0, MKLM ,
(z)
MKLM , we can use the transfer matrices of (10.4) to calculate incident fields on
flaws.
6.4 Response of a Rectangular Coil When Rotated Relative to a Crack 75
Coil
Coil
Y − length = 30mils
X − width = 70mils
thickness = 2mils
height = 0.2mils
turns = 1
30mils
0.3mil
X
Fatigue Crack
0.0004
0.0002
−0.0002
−0.0004
0 10 20 30 40 50 60 70 80 90
Probe Angle (degrees)
6.4 Response of a Rectangular Coil When Rotated Relative to a Crack 77
3
Y-index
0
0 2 4 6 8 10 12 14 16 18 20 22 24 26 28 30 32
X-index
Fig. 6.6 Projections onto the xy-plane of the currents in the top layer of the anomaly: real
component at 0◦ multiplied by 2.0 × 10−9
Figures 6.6–6.9 illustrate the projections onto the xy-plane of the anomalous
(scattering) currents in the top layer of the crack. The coil is oriented at either 0◦ or
90◦ relative to the crack, as explained above. In order to facilitate the interpretation
of these figures, we have converted Jx into units of cell width and Jy into units of
cell length. By doing this, a current that flows toward the corner of a cell will be
plotted as a vector that points toward the corner of the cell. It is clear, now, why the
response at 90◦ is much larger than at 0◦ ; the anomalous currents are an order of
magnitude larger with the 90◦ orientation than with the 0◦ . Furthermore, it should
be clear that the currents within the very narrow crack are non-solenoidal; i.e., they
have a nonzero divergence. In extreme cases of very thin, very long cracks this can
cause problems with the convergence of the conjugate-gradient solver unless special
preconditioners are used, and the development of such preconditioners for volume-
integral equations is an active research area. Finally, we note that the currents are
predominately oriented in the y-direction, as one might expect for a long, thin crack
with the shape shown in Fig. 6.3.
In order to eliminate the distortion due to the large cell aspect-ratio, we plot
the current in a single cell so that the plotted cell has its true aspect ratio.
Figures 6.10–6.13 are the result.
78 6 Advanced Probe Models Based on Magnetic Dipoles and Ferrite Cores
3
Y-index
0
0 2 4 6 8 10 12 14 16 18 20 22 24 26 28 30 32
X-index
Fig. 6.7 Projections onto the xy-plane of the currents in the top layer of the anomaly: imaginary
component at 0◦ multiplied by 2.0 × 10−8
3
Y-index
0
0 2 4 6 8 10 12 14 16 18 20 22 24 26 28 30 32
X-index
Fig. 6.8 Projections onto the xy-plane of the currents in the top layer of the anomaly: real
component at 90◦ multiplied by 2.0 × 10−10
6.4 Response of a Rectangular Coil When Rotated Relative to a Crack 79
3
Y-index
0
0 2 4 6 8 10 12 14 16 18 20 22 24 26 28 30 32
X-index
Fig. 6.9 Projections onto the xy-plane of the currents in the top layer of the anomaly: imaginary
component at 90◦ multiplied by 1.0 × 10−10
5
Y-index
0
4 5
X-index
Fig. 6.10 Projections onto the xy-plane of the currents in a single cell of the anomaly: real
component at 0◦ multiplied by 2.0 × 10−9
80 6 Advanced Probe Models Based on Magnetic Dipoles and Ferrite Cores
5
Y-index
0
1 2
X-index
Fig. 6.11 Projections onto the xy-plane of the currents in a single cell of the anomaly: imaginary
component at 0◦ multiplied by 2.0 × 10−8
5
Y-index
0
4 5
X-index
Fig. 6.12 Projections onto the xy-plane of the currents in a single cell of the anomaly: real
component at 90◦ multiplied by 2.0 × 10−10
We will use the theory of this chapter to model several benchmark experiments
that are described in the Ditchburn–Burke (D–B) paper referred to above. In
particular, we will show model results for several experiments that involve the 50-
turn rectangular spiral coil whose aspect ratio is 2:1. We use the same dimensions
as in the paper: width = 18.52 mm, length = 36.68 mm, thickness = 8.71 mm,
6.6 Modeling a Differential-Receive Ferrite-Core Probe 81
5
Y-index
0
4 5
X-index
Fig. 6.13 Projections onto the xy-plane of the currents in a single cell of the anomaly: imaginary
component at 90◦ multiplied by 1.0 × 10−10
height = 0.01 mm. This value of height is the same as the thickness of the individual
tracks of copper forming each spiral turn, ≈ 10 μm. With these values, we calculate
the freespace inductance of the coil to be L0 = 48.33 μH, whereas the measured
value is 50.25 μH, the error being less than 4 %.
Next, we model experimental results for the coil over defect-free aluminum and
steel plates. The conductivity of aluminum is 1.653 × 107 S/m, and for steel, 3.472 ×
106 S/m. The relative magnetic permeability of the steel plate is 85. In both cases,
we used the same “fitted lift-off” in the paper: for aluminum, 0.12 mm, and for steel,
0.05 mm.
The results for the aluminum plate are shown in Fig. 6.14 and for steel in
Fig. 6.15.
0.6
Ditchburn and Burke dX/X0
−0.2
−0.4
−0.6
−0.8
−1
100 1000 10000 100000 1e+006 1e+007
Frequency (Hz)
1.2
Ditchburn and Burke dX/X0
Normalized Impedance Change
We’ll apply a model of the split-D probe to the problem of characterizing sur-
face cracks within bolt holes in a benchmark test case. Figure 6.17 illustrates
the test cases, and Fig. 6.18 shows a photomicrograph of an actual through-wall
fatigue crack of the type that we are going to size. The bottom part of Fig. 6.18 shows
the EDM notch that is used for calibrating the data. The eddy-current instrument that
6.6 Modeling a Differential-Receive Ferrite-Core Probe 83
2mm
Transmit Coil:
IR = 9.34mm
OR = 18.4mm
HT = 9mm Receive Coils:
Turns = 408 IR = 7.34mm
OR = 8.34mm
HT = 9mm
Turns = 100
Cores:
R = 7.34mm
HT=9mm
μ=2000
Fig. 6.17 Illustrating experimental test cases for sizing surface fatigue cracks using a split-D probe
within a bolt hole in an aluminum substrate. The probe (small white circle) is embedded in a
mandrel, which is then rotated azimuthally about a vertical axis that is not shown. Simultaneously,
the mandrel is lifted vertically, thereby generating a two-dimensional raster scan. The top figure
(a) illustrates the cross-section of a through-wall crack in a 0.100 -thick host, the middle figure
(b) illustrates a mid-bore crack in a 0.250 -thick host, and the bottom figure (c) illustrates the
EDM notch in a 0.250 -thick host that is used to calibrate the test setup (see Fig. 6.18)
84 6 Advanced Probe Models Based on Magnetic Dipoles and Ferrite Cores
Fig. 6.18 Illustrating a photomicrograph of an actual fatigue crack extending from the surface
of a bolt hole within an aluminum substrate (top), together with an EDM notch that is used for
calibrating the entire test setup (bottom). The “nominal” width of the crack is 0.007 mm (0.28 mils)
and the width of the notch is 6.1 mils
6.6 Modeling a Differential-Receive Ferrite-Core Probe 85
is used in the tests measures “instrument volts,” which are then scaled to impedances
for further processing by VIC-3D R
using the β -transformation scheme described
in Sect. 15.2.
In order to reduce the burden of computing the response of very thin, long cracks, we
will resort to a “surrogate” interpolation algorithm. The theory behind the algorithm
is based on the following formula for the change in impedance of a probe due to
scattering from an anomaly, as determined by fundamental electromagnetic theory
(recall Chap. 5):
Za = E0 · Ja dV , (6.5)
Va
where we assume that the probe is excited with a current of one ampere. E0
is the incident field due to the probe in the absence of the anomaly, Ja is the
anomalous (scattering) current, and Va is the volume of the anomaly. E0 and Ja
are both bounded, with bounded derivatives, so that the integral in (6.5) vanishes
to polynomial order in Va as Va → 0. This certainly includes the case where the
cross-section of the anomaly remains fixed and the width, W → 0.
Thus, we can derive an interpolation scheme in W that takes advantage of
this fact. Assuming that W is “small enough,” we can postulate the second-order
approximation, Za = aW + bW 2 . If we compute values of Za at W = 0.125 and
0.25 mil, calling them Z0.125 and Z0.25 , respectively, then it is easy to show that
a = 16Z0.125 − 4Z0.25 , b = −64Z0.125 + 32Z0.25, which yields the final result
Blending Functions for Three Values of W Blending Functions for Three Values of W
0.004 0.03
0.25
0.125 0.25
0.003 0 0.02 0.125
0
0.002
Resistance (Ohms)
Reactance (Ohms)
0.01
0.001
0 0
−0.001
−0.01
−0.002
−0.02
−0.003
−0.004 −0.03
−60 −40 −20 0 20 40 60 −60 −40 −20 0 20 40 60
Probe Position (mils) Probe Position (mils)
Fig. 6.19 Blending functions for the interpolation table with nodes at W = 0, 0.125 and 0.25 mil.
The functions associated with the latter two nodes are the Z0.125 and Z0.25 of (6.6), respectively
0.015
R
Resistance/Reactance (Ohms)
X
0.01
0.005
−0.005
−0.01
−0.015
−0.02
−60 −40 −20 0 20 40 60
Probe Position (mils)
Fig. 6.20 Showing the scaled measured data taken by the split-D probe. These data will be
submitted to NLSE for inversion to determine W
The result of the inversion is W = 0.0793, and the model results that are obtained
when using the quadratic interpolation formula derived in (6.6) with W = 0.0793
are shown in Fig. 6.21, along with the scaled input data. Clearly, there is a good fit
in reactance. The model resistance shows a higher spatial frequency content (more
wiggles) than the data, but this is probably due to uncertainties in measuring the
resistance. In addition, there remains a slight “shoulder” in the scaled reactance data
that the model does not reproduce, and this may be due in part to the measurements,
6.7 Validation Test: Tangent Coil Over a Crack in a Thin Plate 87
0.015
R(Data)
Resistance/Reactance (Ohms)
X(Data)
0.01 R(Interp)
X(Interp)
0.005
−0.005
−0.01
−0.015
−0.02
−60 −40 −20 0 20 40 60
Probe Position (mils)
Fig. 6.21 Comparing the model result with W = 0.0793 mil and the scaled data input to NLSE
as well as to an imperfect model that does not include the effects of the finite edges
of the workpiece.
The ability to detect and size a fatigue crack whose width is of the order of
0.05–0.5 mil is enhanced through the use of a split-D probe, rather than a single
absolute coil. The split-D probe is essentially a mathematical differentiator, and
taking the “derivative” of the response past a crack will give a signal that is
sensitive to the width of the crack. More complex applications of surrogate models
in computational electromagnetics are given in [9].
One of the important new features in modeling probes is the ability to rotate coils
in a variety of directions, allowing us, for example, to model circular tangent coils.
Figure 6.22 illustrates a tangent coil over a flawed workpiece, comprising a through-
wall crack in a thin plate. Burke and Rose have used this arrangement to produce
benchmark data which we have used to validate the rotation feature (see [49]).
Figure 6.23 shows a comparison of the benchmark data and a model calculation,
when the tangent coil is scanned over the slot at 2 kHz, and Fig. 6.24 compares the
benchmark data and a calculation of a frequency scan of the tangent coil, when it
is located over the center of the slot. The frequency range extends from 251 Hz to
10 kHz.
88 6 Advanced Probe Models Based on Magnetic Dipoles and Ferrite Cores
COIL
0.90 FLAW
101.0
WIDTH OF SLOT = 0.30mm
Fig. 6.22 Illustrating a tangent coil over a flawed workpiece, corresponding to the benchmark test
of [49]
R(Vic3D)
10
−2
−4
−6
0 10 20 30 40 50 60 70 80 90
Scan position (mm)
Fig. 6.23 Comparison of calculations with the position-scan benchmark data of [49]
−2
−4
−6
0 1000 2000 3000 4000 5000 6000 7000 8000 9000 10000
f (Hz)
Fig. 6.24 Comparison of calculations with the frequency-scan benchmark data of [49]
Appendix
The split-D coil configuration shown in Fig. 6.16 comprises one transmit and two
receive coils. Another common split-D probe contains no separate transmit coil but
has two identical coils connected in a bridge circuit that serves as the driver of the
two coils.
Figure 6.25 shows a standard bridge circuit that drives the split-D coils, labeled
L1 and L2 , in the presence of a flaw, shown as an inductor, L3 , and resistor, R3 . The
inductance of the two coils is L; M is the mutual inductance between these two coils,
and M13 , M23 , are, respectively, the mutual inductance between L3 and L1 and L3
and L2 . The bridge is open-circuited, and its output voltage is V0 .
The equations for the circuit of Fig. 6.25 are
Vin I1 + − I2
V0
L 1 L 2
M13 L3 M23
R3
I3
jω M13 I1 + jω M23 I2
I3 = − .
R3 + jω L3
Now, if coil 1 and coil 2 are both considered to be transmit coils, then VIC-3D
R
ω 2 (M13
2 − M2 )
V0 = Zout = 23
. (6.9)
R2 + jω L3
One way to guarantee that the circuit of Fig. 6.25 will behave as a current source
such that I1 ≈ I2 is to make the balance resistance, R0 , much greater than any of
the other impedance elements over the frequency range of interest. In this case,
I1 ≈ I2 ≈ Vin /R0 .
M13 and M23 depend upon the scan position of the probe relative to the flaw. All
other parameters are independent of the probe position. It is clear from (6.9) that if
the flaw is symmetrically placed with respect to the probe, then M13 = M23 , which
means that V0 = 0. Furthermore, because the probe is symmetrical, it follows that V0
will be an antisymmetrical function of the probe position if the flaw is symmetrical
with respect to the probe position. This gives the usual antisymmetrical response
that we expect of a differential probe.
Chapter 7
Advanced Probe Models Based
on Electric Dipoles
7.1 Introduction
δx
Applied
current
facet
element
I = 1A
x
I = 1A
Fig. 7.1 A printed-circuit rectangular-spiral coil lying within an “anomalous region.” The electric
dipole produced by the driving-point current excites the metallic traces of the coil. The convention
in electric-circuit theory is that current enters the positive terminal of a passive load and leaves
through the negative terminal
be defined by a single magnetic dipole that is oriented normal to the planes. The
field produced by these sources, then, are incident upon the anomalous region of the
problem. This model works well for typical eddy-current problems but allows only
the computation of the inductance of the source. The solenoidal nature of the current
precludes any possibility of computing the driving-point capacitance of the coil. In
order to compute capacitance, and therefore the resonant frequency of the coil, we
must relax the assumption of solenoidal current-flow and assume that the excitation
of the coil is an electric dipole.
We are going to apply our volume-integral algorithm to the metallic traces
contained within the anomalous region shown in Fig. 7.1. These traces are the
“anomaly” of the problem (much like a ferrite core in a conventional eddy-current
probe), and the driving-point terminals form an electric dipole that will drive
the coil.
7.2 Development of the Electric-Dipole Mathematical Model 95
where the Q’s are tri-diagonal matrices, the G(0) ’s the infinite-space matrices, the
G(a) ’s the convolutional layered-space matrices, and the G(b) ’s the correlational
layered-space matrices. The infinite-space matrices are convolutional, also. The
superscript (ee) denotes electric–electric matrices and (em) denotes electric–
magnetic matrices. The J’s are the unknown electric currents and the M’s are
the unknown magnetic polarization vectors. The last block in (7.1) is simply a
shorthand representation of the three blocks above it, except that it represents
electric–magnetic interactions.
From here on we will ignore magnetic interactions and will combine the various
G-matrices into one. The incident electric field moments on the left-hand are due,
of course, to any independent sources, currents, or charges, as in the usual case of a
separate probe coil. We compute an effective incident electric field-moment due to
the dipole in the following way. Assume that the positive terminal in Fig. 7.1 carries
1A of current at cell k+ l + m+ and that the negative terminal carries 1A of current
(x)
at cell k− l − m− . Then in the configuration shown in Fig. 7.1, we have Jk+ l + m+ =
(y) (z) (x) (y) (z)
−1/δ yδ z, Jk+ l + m+ = Jk+ l + m+ = 0, and Jk− l − m− = 1/δ yδ z, Jk− l − m− = Jk− l − m− = 0,
where δ x, δ y, δ z are the dimensions of a cell in the anomalous region. Then, from
(7.1) we have for the incident field moments
96 7 Advanced Probe Models Based on Electric Dipoles
1
(ix) (xx) (xx)
Eklm = −Gklm,k+ l + m+ + Gklm,k− l − m− ×
δ yδ z
1
(iy) (yx) (yx)
Eklm = −Gklm,k+ l + m+ + Gklm,k− l − m− ×
δ yδ z
1
(iz) (zx) (zx)
Eklm = −Gklm,k+ l + m+ + Gklm,k− l − m− × . (7.2)
δ yδ z
This electric-dipole moment vector can be thought of as the “constraint electric
field” that arises because of the unit-constrained current at k+ l + m+ and k− l − m− .
Now that we have transformed the known current-density vectors at k+ l + m+ and
k− l − m− into the incident electric-field moments (this is akin to using the G-matrices
as “transfer matrices”), we are left with an anomalous current of zero at these
cells. After all, a current at a location is either known or unknown (anomalous),
but not both. Thus, if the current is known, then the anomalous current must be
zero. The way to force the anomalous current to vanish is to make the cells at
k+ l + m+ , k− l − m− to contain only host material. Hence, with this proviso, the
problem can be solved completely as a standard problem for any anomalous region,
using (7.2) as the incident field. In this case, the anomalous region of Fig. 7.1 will
contain the source dipole.
The position of the applied current facet elements (plus and minus) are included
within the grid for the anomalous region. The magnitude of the current density at
the boundary of the cell for the + terminal is −1/δ yδ z, and the magnitude of the
current density for the − terminal is 1/δ yδ z.
Once we have the J’s, we can compute the change in driving-point impedance
seen at the terminals of the unit-current source in the usual way as the dot-product
of the J’s with the incident field moments:
(x) (ix) (y) (iy) (z) (iz)
Δ Z = −Σklm Jklm Eklm + Jklm Eklm + Jklm Eklm
1 $
(x) (xx) (xx)
= Σklm Jklm −Gklm,k+ l + m+ + Gklm,k− l − m−
δ yδ z
(y) (yx) (yx)
+Jklm −Gklm,k+ l + m+ + Gklm,k− l − m−
%
(z) (zx) (zx)
+Jklm −Gklm,k+ l + m+ + Gklm,k− l − m− (7.3)
The driving-point impedance of the loaded dipole is the sum of (7.3) and (7.4).
7.3 An Example 97
7.3 An Example
Consider the 8 × 8 × 2 mm grid with 8 × 8 × 2 cells shown in Fig. 7.2. If the bottom
layer comprises only blue (host) cells, then the current is constrained to flow only
in the top layer, which means that there will be no z-directed currents at all within
the grid.
The current enters cell 3, which makes it the positive terminal, and leaves cell
1, making it the negative terminal. Thus, cells 1 and 3 make up an electric dipole
that will excite the remaining cells. In order to determine the excitation source, we
constrain the expansion coefficient for the x-directed current at the facet between
cells 3 and 4 to be −1 × 106 A/m2 and for the x-directed current at the facet between
cells 1 and 2 to be +1 × 106 A/m2 . Clearly, the current in cells 1 and 2 and 3 and 4
is constrained to flow only in the x-direction.
At frequencies that are low enough that this body is electrically large, we can
model this structure by the electrical equivalent circuit shown in Fig. 7.3. The
capacitor, C0 , is associated with the stored electric energy within the gap between
cells 1 and 3 of the grid, the resistor is associated with energy loss in the metal,
and the inductance is associated with the stored magnetic energy within the area
enclosed by the coil.
I I
X : σ = 0 S/m
(Host)
: σ = 1E7 S/m
1 3
(−) (+)
2 4
C0 L0
−(1)
Consider the same problem as before, but this time let the excitation current lie in
the y-direction, as shown in Fig. 7.4. In this case, the excitation currents lie in the
y-direction, and we show that the sources are limited to a single cell. Hence, for the
(−) terminal, the current is a “ramp” function, and in the (+) terminal, it is a “slide”
function. Thus, in computing the incident field vectors in (7.2), we would use only
the appropriate “ramp” or “slide” portions of the matrix elements.
The electric-dipole vector oriented in the y-direction is a little clearer now,
because of the continuity equation for electric charge:
∂ρ
+∇·J = 0 , (7.5)
∂t
from which we deduce that
1
ρ=− ∇·J
jω
109
=− Q/m3 (−) terminal
jω
109
=+ Q/m3 (+) terminal. (7.6)
jω
Figures 7.5–7.11 show the results of the dipole calculation for the example of
Fig. 7.2, but using ramp functions at the terminals (cells 1 and 3) instead of tent
functions. The tent function that spans cells 1 and 2, as well as the tent function that
7.5 Initial Results 99
Applied Currents
1 x 106 A/m2
Jy in cell(−) Jy in cell(+)
Fig. 7.4 Illustrating a square coil on an 8 × 8 × 2 grid. The excitation current lies in the y-direction
this time, and the nature of the excitation currents within their cells is shown
spans cells 3 and 4 are treated as unknowns to be computed. These are “freespace”
results—there is no workpiece, and the grid has 8 × 8 × 2 cells. The frequency of the
broad resonance increases as the gap increases, which suggests that the associated
capacitance decreases (see the discussion of Fig. 7.12). The sharp resonances, on the
other hand, do not change frequency with the gap, which suggests that they may be
transmission-line effects associated with the length of the loop.
Figure 7.12 is a plot of the terminal capacitance as a function of the distance
between the terminals, as deduced from the frequency of the zero-crossing of the
broad resonance in the reactance functions. This result is reasonable, because we
expect this capacitance to decrease with an increased separation of the terminals.
In addition, this figure includes the capacitance of the “naked terminals” as a
function of the gap between the terminals. These capacitances are computed from
the reactances (not shown) of the naked terminals at 108 Hz. The latter capacitances
are smaller than the former since they include only the ramp functions of the
terminal and not the tent functions that complete the terminal posts.
We can do a simple analysis of the circuit of Fig. 7.3 and estimate the parameters
from the frequency responses of the driving-point impedances that we have just
computed. The driving-point impedance is given by
1
Zin =
1
jω C0 +
R0 + jω L0
100 7 Advanced Probe Models Based on Electric Dipoles
Resistance vs Frequency
1
Gap: 6 cells
5 cells
4 cells
3 cells
Resistance (Ohm) 2 cells
0.1 1 cells
Closed loop
0.01
0.001
1e+007 1e+008 1e+009
Frequency (Hz)
Reactance vs Frequency
100
Gap: 1 cell
2 cell
3 cell
4 cell
Reactance (Ohm)
5 cell
6 cell
10 Closed loop
aircore
Fig. 7.5 Results of the dipole model calculation in freespace at frequencies below the first
resonance. The curve labeled “aircore” is the result of VIC-3D R
’s standard aircore calculation
for a single-loop coil enclosing the same area. Top: resistance. Bottom: reactance
R0 + jω L0
=
1 − ω 2L0C0 + jω R0C0
R0 + jω (L0 [1 − ω 2L0C0 ] − R20C0 )
= . (7.7)
(1 − ω 2L0C0 )2 + (ω R0C0 )2
7.5 Initial Results 101
Resistance vs Frequency
14000
Gap: 1 cell
12000
Resistance (Ohm)
10000
8000
6000
4000
2000
Reactance vs Frequency
8000
6000
Gap: 1 cell
4000
Reactance (Ohm)
2000
−2000
−4000
−6000
−8000
5e+009 6e+009 7e+009 8e+009
Frequency (Hz)
Fig. 7.6 Results for the 8 × 8 × 2 grid, one-cell-gap (1 mm) model. Top: resistance. Bottom:
reactance
Resonance occurs at that frequency for which the total reactance vanishes, or L0 (1 −
ωR2 L0C0 ) = R20C0 , from which
&
1 L0 − C0 R20
ωR =
L0C0 L0
1
≈ , (7.8)
L0C0
for R20C0 L0 .
102 7 Advanced Probe Models Based on Electric Dipoles
Resistance vs Frequency
10000
8000
Resistance (Ohm)
6000
Gap: 2 cells
4000
2000
Reactance vs Frequency
8000
6000
Gap: 2 cells
4000
Reactance (Ohm)
2000
−2000
−4000
−6000
−8000
5e+009 6e+009 7e+009 8e+009
Frequency (Hz)
Fig. 7.7 Results for the 8 × 8 × 2 grid, two-cell-gap (2 mm) model. Top: resistance. Bottom:
reactance
We can estimate R0 as follows: the coil conductivity is 1 × 107 S/m, and the cells
are cubes of 1 mm on a side. Thus, a path of N cells will have a resistance of
R0 = ρ × l/A
= 1 × 10−7 × N × 1 × 10−3/(1 × 10−6)
= N × 1 × 10−4 Ω , (7.9)
Resistance vs Frequency
10000
Gap: 3 cells
8000
Resistance (Ohm)
6000
4000
2000
Reactance vs Frequency
8000
6000
Gap: 3 cells
4000
Reactance (Ohm)
2000
−2000
−4000
−6000
−8000
5e+009 6e+009 7e+009 8e+009
Frequency (Hz)
Fig. 7.8 Results for the 8 × 8 × 2 grid, three-cell-gap (3 mm) model. Top: resistance. Bottom:
reactance
These results are in good agreement with the low-frequency resistances plotted
in Figs. 7.5 and 7.16.
We can estimate L0 from the reactance at 1 × 108 Hz in Fig. 7.5: L0 = XL /
(2π × 1 × 108) (Table 7.2).
Finally, we estimate the capacitance from the resonant frequencies as shown in
(7.8). These are the “Full Circuit” capacitance values that are plotted in Fig. 7.12
(Table 7.3).
Figure 7.13 shows the change in impedance when a 1 mm plate is placed below
the loop, and Figs. 7.14 and 7.15 show the results of a lift-off scan at 10 MHz. The
plotted results in the latter two figures are the change due to the plate.
104 7 Advanced Probe Models Based on Electric Dipoles
Resistance vs Frequency
10000
Gap: 4 cells
8000
Resistance (Ohm)
6000
4000
2000
Reactance vs Frequency
8000
6000
Gap: 4 cells
4000
Reactance (Ohm)
2000
−2000
−4000
−6000
−8000
5e+009 6e+009 7e+009 8e+009
Frequency (Hz)
Fig. 7.9 Results for the 8 × 8 × 2 grid, four-cell-gap (4 mm) model. Top: resistance. Bottom:
reactance
Figure 7.16 shows the low-frequency response of the refined grid, which should be
compared with the corresponding results of Fig. 7.5. Finally, we show in Fig. 7.17
the impedance response of the “naked” terminal-pair.
The result for the “naked” terminal pair is interesting. It represents a resistor,
whose resistance varies with frequency, in series with an object whose log-log
plot varies linearly (for the most part), with a negative slope. Thus, this object’s
7.6 Results With a Grid Of 16×16×4 Cells 105
Resistance vs Frequency
10000
Gap: 5 cells
8000
Resistance (Ohm)
6000
4000
2000
Reactance vs Frequency
8000
6000
Gap: 5 cells
Reactance (Ohm)
4000
2000
−2000
−4000
−6000
−8000
5e+009 6e+009 7e+009 8e+009
Frequency (Hz)
Fig. 7.10 Results for the 8 × 8 × 2 grid, five-cell-gap (5 mm) model. Top: resistance. Bottom:
reactance
Resistance vs Frequency
10000
Gap: 6 cells
8000
Resistance (Ohm)
6000
4000
2000
6000
Gap: 6 cells
4000
Reactance (Ohm)
2000
−2000
−4000
−6000
−8000
5e+009 6e+009 7e+009 8e+009
Frequency (Hz)
Fig. 7.11 Results for the 8 × 8 × 2 grid, six-cell-gap (6 mm) model. Top: resistance. Bottom:
reactance
Figure 7.18 shows the divergence of the current in the loop with a 1-cell (1 mm)
gap at 10 MHz, and Fig. 7.19 shows the same thing at 1 GHZ. In both of these
figures, Jx flows out of the terminal at 0 < x < 0.001, 0.002 < y < 0.003 and into
the terminal at 0 < x < 0.001, 0.004 < y < 0.005. The units are A/m3 . Figure 7.20
shows the divergence of the current in the closed-loop at 10 MHz. The two driving
tent functions span the cells in the region 0.002 < x < 0.005, 0 < y < 0.001 and
flow in the positive x-direction. The units are A/m3 .
7.7 Computation of the Divergence of the Currents 107
7e-014
6e-014
5e-014
4e-014
3e-014
2e-014
0 1 2 3 4 5 6 7
Terminal Gap Size (mm)
Fig. 7.12 Showing the “loaded” and “naked” terminal capacitance as a function of the distance
between the terminals
0.1
0.05
0
1e+007 1e+008 1e+009
Frequency (Hz)
aircore
100
50
−50
−100
1e+007 1e+008 1e+009
Frequency (Hz)
Note that the 1-mm gap coil and the closed-loop coil have comparable divergence
values at 10 MHz, but the response of the 1-mm gap coil at 1 GHz is orders-
of-magnitude larger, as we would expect. At this frequency resonance effects, as
manifested by the increased capacitive reactance, begin to assert themselves, and
we always associate capacitive effects with charge buildup, which is the same as
divergence of the current.
7.8 Three-Turn Spiral Coil 109
0.014
0.01
0.008
0.006
0.004
0.002
0
0.1 1 10 100 1000
Liftoff (mm)
−0.1
−0.15
−0.2
−0.25
−0.3
−0.35
−0.4
0.1 1 10 100 1000
Liftoff (mm)
Fig. 7.14 Change in impedance of the dipole model calculation over a 1 mm plate as a function of
lift-off at 10 MHz. Top: resistance. Bottom: reactance
Figure 7.21 shows a three-turn spiral coil, whose results are displayed in the next few
figures. Figure 7.22 shows the frequency scan of this coil in freespace, and Fig. 7.23
shows these same results below the resonance around 2 GHz. Finally, Figs. 7.24 and
7.25 illustrate the results of the dipole model of the three-turn spiral coil when a
1 mm plate is placed below the coil.
110 7 Advanced Probe Models Based on Electric Dipoles
dX vs dR
0
−0.15
−0.2
−0.25
−0.3
Fig. 7.15 Impedance-plane plot of the change in impedance of the dipole model calculation over
a 1 mm plate as a function of lift-off at 10 MHz
Figure 7.26 shows a two-layered spiral coil whose dimensions are 0.8 mm ×
0.8 mm × 0.15 mm. The low-frequency response of this coil is shown in Fig. 7.27,
and the response through resonance is shown in Fig. 7.28. Figure 7.29 shows the
high-frequency response of the two-layered coil, in which two weaker resonances
appear.
The conductivity of the metallic traces is σ = 1 × 107 S/m, as before. Given the
dimensions that were stated above, we compute the resistance of each voxel to be
Rvoxel = 0.002 Ω, and because there are 260 voxels connected in series in Fig. 7.26,
we compute the dc resistance of the coil to be R0 = 0.520 Ω, which is in excellent
agreement with the low-frequency resistance shown in Fig. 7.27.
Using the results of Figs. 7.27 and 7.28, we can estimate the other elements of the
equivalent circuit of Fig. 7.3. From Fig. 7.27 we estimate the inductance to be L0 =
1.59 × 10−8H, and from Fig. 7.28 we estimate the resonant frequency to be 6.4 ×
109 Hz, or ω0 = 40.2 × 109rad/sec. Using these results we compute C0 = 1/(L0 ×
ω02 ) = 3.89 × 10−14Fd. For a high-Q parallel resonant circuit of this type, the peak
resistance at resonance is given by Rres = L0 /(C0 R0 ) = 7.86 × 105 Ω, which is in
excellent agreement with the result shown in Fig. 7.28.
7.10 A Model of a THz Transmitter on GaAs 111
Resistance vs Frequency
0.1
Gap: 1 mm
Resistance (Ohm)
0.01
0.001
1e+007 1e+008 1e+009
Frequency (Hz)
Reactance vs Frequency
100
Gap: 1 mm
Reactance (Ohm)
10
Fig. 7.16 Results of the 16 × 16 × 4-cell 1-mm gap model at low frequencies. Top: resistance.
Bottom: reactance
10 Gap: 1 mm
Resistance (Ohm) 1
0.1
0.01
0.001
0.0001
1e-005
1e-006
2e+008
Frequency (Hz)
100000 Gap: 1 mm
Reactance (Ohm)
10000
1000
100
10
1
2e+008
Frequency (Hz)
Fig. 7.17 Results for the “naked” driving-point terminal response for the 16 × 16 × 4-cell 1-mm
gap model. Top: resistance. Bottom: reactance
The results of the impedance calculation are shown in Figs. 7.31 and 7.32. It
is clear from the high-frequency results shown in the bottom of Fig. 7.32 that the
impedance response can be modeled by the series-parallel circuit of Fig. 7.33 up to
a frequency of about 8 × 1011 Hz. We can estimate the parameters of Fig. 7.33 by
using the data shown in Figs. 7.31 and 7.32.
7.10 A Model of a THz Transmitter on GaAs 113
div(J) (A/m**3)
500
−500
0.008
0.007
0.006
0.005
0.004
0.003
0.002 x(m)
0.001
0.008 0.007 0.006 0.005 0
0.004 0.003 0.002 0.001 0
y(m)
−2
0.008
0.007
0.006
0.005
0.004
0.003 x(m)
0.002
0.001
0.008 0.007 0.006 0
0.005 0.004 0.003 0.002 0.001 0
y(m)
Fig. 7.18 Divergence of the current of the 1-mm gap model at 10 MHz. Top: real. Bottom:
imaginary
114 7 Advanced Probe Models Based on Electric Dipoles
div(J) (A/m**3)
5e+006
−5e+006
0.008
0.007
0.006
0.005
0.004
0.003
0.002 x(m)
0.008 0.001
0.007 0.006 0.005 0
0.004 0.003 0.002 0.001 0
y(m)
10000
5000
−5000
−10000
0.008
0.007
0.006
0.005
0.004
0.003 x(m)
0.002
0.001
0.008 0.007 0.006 0.005 0
0.004 0.003 0.002 0.001 0
y(m)
Fig. 7.19 Divergence of the current of the 1-mm gap model at 1 GHz. Top: real. Bottom: imaginary
7.10 A Model of a THz Transmitter on GaAs 115
div(J) (A/m**3)
500
−500
0.008
0.007
0.006
0.005
0.004 y(m)
0.003
0.002
0.001
0 0.001 0.002 0
0.003 0.004 0.005 0.006 0.007 0.008
x(m)
div(J) (A/m**3)
−2
0.008
0.007
0.006
0.005
0.004 y(m)
0.003
0.002
0.001
0 0.001 0.002 0
0.003 0.004 0.005 0.006 0.007 0.008
x(m)
Fig. 7.20 Divergence of the current of the closed-loop model at 10 MHz. Top: real. Bottom:
imaginary
116 7 Advanced Probe Models Based on Electric Dipoles
Fig. 7.21 A three-turn spiral coil. The bottom layer of the three-layer system contains the spiral
coil (red). The middle layer (upper right) contains a single via (red) that connects the bottom to
the top layer (upper left), which contains the lead (red) to the negative terminal on the left edge.
The positive terminal is at the upper left corner of the bottom-most layer. This makes the input gap
one-layer (the middle layer) high in the z-direction, and five cells wide along the left edge
1
C0 ≈
2π × 109 × 24,000
= 6.63 × 10−15 Fd
R ≈ 70 Ω
1
L≈
(2π × 3 × 1011)2 × 6.63 × 10−15
= 4.25 × 10−11 H
1
C1 ≈
(2π × 5 × 1011)2 × 4.25 × 10−11
= 2.38 × 10−15 Fd . (7.10)
7.10 A Model of a THz Transmitter on GaAs 117
Resistance vs Frequency
100
80 spiral coil
Resistance (Ohm)
60
40
20
−20
1e+007 1e+008 1e+009 1e+010
Frequency (Hz)
Reactance vs Frequency
0
Reactance (Ohm)
−4000
−6000
−8000
Fig. 7.22 Frequency scan of the three-turn spiral coil of Fig. 7.21 in freespace. Top: resistance.
Bottom: reactance
Resistance vs Frequency
0.3
spiral coil
0.25
0.15
0.1
0.05
0
1e+007 1e+008 1e+009
Frequency (Hz)
Reactance vs Frequency
600
500
spiral coil
Reactance (Ohm)
400
300
200
100
−100
1e+007 1e+008 1e+009
Frequency (Hz)
Fig. 7.23 Results of the dipole model calculation of the three-turn spiral coil in freespace below
the resonance around 2 GHz. Top: resistance. Bottom: reactance
being approximately 5 × 1011 Hz. The wiggles beyond 8 × 1011 are attributed to
parasitic elements that are not easily represented in a simple lumped equivalent
circuit. This response is contrary to that shown in the earlier models, in which the
parallel resonance occurs first. The reason for that is that the simple closed-loop
circuit of the earlier models has a significant inductance, which dominates the low-
frequency response, contrary to the THz transmitter, which is simply a pair of open
terminals.
7.10 A Model of a THz Transmitter on GaAs 119
80
Resistance (Ohm) Without Plate
With Plate
60
40
20
−20
1e+007 1e+008 1e+009 1e+010
Frequency (Hz)
Without Plate
Reactance (Ohm)
−4000
−6000
−8000
Fig. 7.24 Impedance of the dipole model calculation of the three-turn spiral coil in freespace and
over a 1 mm plate. Top: resistance. Bottom: reactance
120 7 Advanced Probe Models Based on Electric Dipoles
0.5
−0.5
1e+007 1e+008 1e+009
Frequency (Hz)
−100
spiral coil
−200
−300
−400
−500
1e+007 1e+008 1e+009
Frequency (Hz)
Fig. 7.25 Change in impedance of the dipole model calculation of the three-turn spiral coil over a
1 mm plate at low frequencies. Top: resistance. Bottom: reactance
7.10 A Model of a THz Transmitter on GaAs 121
Resistance vs Frequency
100
spiral coil
Resistance (Ohm)
10
0.1
1e+007 1e+008 1e+009
Frequency (Hz)
Reactance vs Frequency
10000
spiral coil
1000
Reactance (Ohm)
100
10
1
1e+007 1e+008 1e+009
Frequency (Hz)
Fig. 7.27 Results of the dipole model calculation of the two-layered spiral coil in freespace below
resonance. Top: resistance. Bottom: reactance
7.10 A Model of a THz Transmitter on GaAs 123
Resistance vs Frequency
900000
800000
spiral coil
700000
Resistance (Ohm)
600000
500000
400000
300000
200000
100000
0
5e+009 6e+009 7e+009 8e+009 9e+009 1e+010
Frequency (Hz)
Reactance vs Frequency
400000
spiral coil
Reactance (Ohm)
200000
−200000
−400000
Fig. 7.28 Results of the dipole model calculation of the two-layered spiral coil in freespace
through resonance. Top: resistance. Bottom: reactance
124 7 Advanced Probe Models Based on Electric Dipoles
Resistance vs Frequency
80000
70000
spiral coil
Resistance (Ohm) 60000
50000
40000
30000
20000
10000
0
2e+010 4e+010 6e+010 8e+010 1e+011
Frequency (Hz)
Reactance vs Frequency
40000
30000
spiral coil
20000
Reactance (Ohm)
10000
−10000
−20000
−30000
−40000
2e+010 4e+010 6e+010 8e+010 1e+011
Frequency (Hz)
Fig. 7.29 Results of the dipole model calculation of the two-layered spiral coil in freespace at
higher frequencies; two weaker resonances appear. Top: resistance. Bottom: reactance
7.10 A Model of a THz Transmitter on GaAs 125
GaAs
+ Metal
126 7 Advanced Probe Models Based on Electric Dipoles
Resistance vs Frequency
400
350
250
200
150
100
50
0
1e+009 1e+010 1e+011 1e+012
Frequency (Hz)
Reactance vs Frequency
5000
−5000
Reactance (Ohm)
−10000
−15000
−20000
−25000
−30000
1e+009 1e+010 1e+011 1e+012
Frequency (Hz)
Fig. 7.31 Results of the dipole-model calculation for the THz transmitter of Fig. 3 of [93]. Top:
resistance. Bottom: reactance
7.10 A Model of a THz Transmitter on GaAs 127
Reactance vs Frequency
10000
(-)Reactance (Ohm)
1000
100
1e+009 1e+010 1e+011 1e+012
Frequency (Hz)
Impedance vs Frequency
400
R
300 X
Impedance (Ohm)
200
100
−100
−200
Fig. 7.32 Results of the dipole-model calculation for the THz transmitter of Fig. 3 of [93]. Top:
low-frequency reactance. Bottom: high-frequency impedance
R L C0
Zin
Chapter 8
Planar and Conforming Arrays of Probes
The ZETEC X-Probe is used for the rapid inspection of steam generator tubes in
nuclear power plants.1 It comprises a periodic azimuthal array of sensors, whose
fundamental basis cell is shown in Fig. 8.1 (we have chosen the dimensions shown
1 We are indebted to Prof. Sung-Jin Song of Sungkyunkwan University for bringing this probe
design to our attention.
Axial Flaw
Y (Tube Axis)
R1 R2
Circumferential Flaw
R3
Scan Direction T2
R4
X (Circumferential Direction)
T1
Coordinates(mm): T1(0,0)
T2(−2.5,5)
R1(−2.5,10)
R2(2.5,10)
R3(7.5,5)
R4(10,0)
Fig. 8.1 Showing the basis cell of the X-Probe coil arrangement. “T” stands for a transmit coil
and “R” for a receive coil
in this figure). The array is designed for the inspection of steam generator tubes, and
operates by being pulled through the tube in the axial direction. The configuration
comprising T1R1 and T1R2 is sensitive to axial flaws, whereas T2R3 and T1R4
are sensitive to circumferential flaws. “T” stands for a transmit coil, and “R” for a
receive coil.
We have created a model for the interaction of the basis cell with a flaw of dimen-
sions 0.5 mm × 3 mm × 1.5 mm (width × length × height), located in an aluminum
half-space with conductivity σ = 2.801 × 107 S/m and excitation frequency of
100 kHz. The flaw is oriented in either the axial direction (the Y-axis in Fig. 8.1)
or the circumferential direction (the X-axis in Fig. 8.1). The long dimension of the
flaw determines its orientation.
We show the responses of T1R1 in the axial and circumferential directions
and T1R4 in the circumferential direction in Fig. 8.2. The response of T1R2 is
identical with that of T1R1 when the flaw is symmetrically placed with respect to
the Y -axis, so that we show only the T1R1 response. Clearly, T1R1 is sensitive to
axial flaws, while being relatively insensitive to circumferential flaws, while T1R4 is
sensitive to circumferential flaws, exactly as advertised. It is interesting to note that
the configuration T1R4 relative to the circumferential flaw is precisely the “parallel
scan” configuration discussed in [95].
8.1 Modeling a Circular Array 131
0.005
T1R1:axial
T1R1:circum
T1R4:circum
0.004
|Z|(Ohms)
0.003
0.002
0.001
0
−10 −5 0 5 10
Scan Position (mm)
Fig. 8.2 Comparing the responses of T1R1 and T1R4 to axial and circumferential cracks
We’ll do some modeling with the T1R4 configuration using various solid models
shown in Table 8.1. These solids all have equal volumes, are surface-breaking, and
have the same maximum depth, 1.5 mm, into the host. The A and B parameters
determine the extent of the solid in the transverse (X,Y ) plane. Our first exercise
is to determine the sensitivity of the T1R4 configuration to the X-coordinate of the
hole as the probe is scanned in the Y -direction. The X-coordinates of the hole are
1, 2, 3, 4, and 5 mm. Recall that the center of T1 is X = 0 mm and the center of
R4 is X = 10 mm, which means that a hole location of 5 mm is at the center of the
configuration. We do not need to go beyond 5 mm because of the obvious mirror
symmetry about this point.
We contrast two conditions for the experiment; the first is with a host of
aluminum, σal = 2.801 × 107 S/m, and the second with a host of stainless steel,
σss = 1 × 106 S/m. Both problems are run at a frequency of 100 kHz. The results
are shown in Fig. 8.3. Two things, in particular, stand out in these results. In the
132 8 Planar and Conforming Arrays of Probes
0.002 1
2
3
0.0015 4
5
|Z|(ohms)
0.001
0.0005
−0.0005
−10 −5 0 5 10
Scan Position (mm)
0.006
1
2
0.005 3
4
5
0.004
|Z|(ohms)
0.003
0.002
0.001
−0.001
−10 −5 0 5 10
Scan Position (mm)
Fig. 8.3 Response of the T1R4 configuration to the X-position of a hole as the probe is scanned in
the Y -direction. The coordinates of the hole are 1, 2, 3, 4, and 5 mm, and the frequency is 100 kHz.
Top: aluminum substrate. Bottom: stainless-steel substrate
first place, the response in the aluminum substrate is not a monotonic function of
the location of the hole, whereas it is in the stainless-steel substrate (locations 4
and 5 are virtually identical). This means that the parameters that define the mutual
inductive coupling between the hole and either coil depend upon the electromagnetic
properties of the structure as well as the frequency.
The second point is that the response is much larger in the stainless-steel substrate
than in the aluminum. The reason for this is that the skin depth, which defines the
8.1 Modeling a Circular Array 133
X
−0.004
−0.001
−0.005
−0.0012 Block
Hole
−0.0014 −0.006 SemiEllip
Cone
−0.0016 −0.007
−0.0018 −0.008
−10 −5 0 5 10 −10 −5 0 5 10
Scan Position (mm) Scan Position (mm)
Response of T1R4 to
Flaw Models in Stainless Steel
0.008
0.007
0.006 Block
Hole
0.005 SemiEllip
Cone
|Z|
0.004
0.003
0.002
0.001
0
−10 −5 0 5 10
Scan Position (mm)
Fig. 8.4 Response of each of the four solid models to the T1R4 configuration, when each is
centered at X = 5,Y = 0
penetration extent of the coil fields in any direction, is much smaller in aluminum
than in stainless steel. This means that more of the field of either the transmit or
receive coil in stainless steel interacts with the hole than in aluminum.
Models of the type that we have just computed are useful in making probability-
of-detection estimates. For example, by choosing frequencies correctly for a given
host material, so that the response with respect to flaw position is monotonic, as
in the bottom part of Fig. 8.3, we can choose the spacing between the transmit and
receive coils so that the response will have a given minimum value with respect to
flaw position. This allows one to estimate the probability of detection for the given
transmit–receive configuration.
The second exercise is to compute the response of the T1R4 configuration to each
of the solid models of Table 8.1, when each is located midway between T1 and R4,
i.e., at X = 5, Y = 0, in stainless steel, and the excitation frequency is 100 kHz. The
responses are plotted in Fig. 8.4.
134 8 Planar and Conforming Arrays of Probes
0 0
T2R3
−0.0002 −0.001 T1R4
T2R3
T1R4
−0.0004
−0.002
−0.0006
−0.003
R
X
−0.0008
−0.004
−0.001
−0.005
−0.0012
−0.0014 −0.006
−0.0016 −0.007
−10 −5 0 5 10 −10 −5 0 5 10
Scan Position (mm) Scan Position (mm)
Fig. 8.5 Response of the T1R4–T2R3 configuration as the probe is scanned past the hole in
stainless steel
The significance of these results is that, because the four solids have the same
volume, their reactance and magnitude responses are quantitatively and qualitatively
similar, the major differences being due to the different lateral extents of the four
flaws. It is the resistance function, however, that gives insight into the depth-features
of each of the flaws. This function is completely different in magnitude and shape,
depending upon the nature of the profile-in-depth of the flaw. This is a typical
result in eddy-current NDE; flaws of equal volume have similar reactance functions,
whereas the resistance functions almost uniquely determine the shape of the flaw
in depth. (This is not a theorem, but a reasonable rule of thumb!) This is why it
is absolutely crucial to record both R and X of a response if one is to accurately
reconstruct the flaw from impedance data. This holds true for any measurement
technique; two “channels” are required, not just magnitude, say, or reactance.
We can think of the cavity of the flaw as storing magnetic energy, so that, if
all conditions are equal, flaws of equal volume will store equal energy, and since
the time-rate-of-change of magnetic energy determines the inductive reactance of a
system, we can understand that the reactance function of equal-volume flaws will
be equal, or nearly so.
Figure 8.6 illustrates a planar T/R array that is designed to produce benchmark data
to validate voxel-based inversion algorithms. The algorithms are not of importance
here; we intend to develop an approach to characterizing such arrays through
techniques that have been developed for analyzing electrical (and microwave)
N-port networks. The 8 × 8 array of printed-circuit receive-sensors is shown
in Fig. 8.7.
The freespace response (reactance) of the sensor when drive 1 is excited at 1 MHz
is shown in Table 8.2 and the response when drive 2 is excited at 1 MHz is shown
in Table 8.3. The response at any other frequency is directly proportional to the
frequency. In particular note the symmetries of the responses.
The transmit coils are driven by the A-channel of an HP3577A network analyzer,
and the receive coils are connected to the B-channel of the analyzer, as shown in
Fig. 8.8. The electronic commutator at the transmit ports sequentially connects the
two transmit coils, represented as Ports A and B, to the A-channel of the network
analyzer. While connected to either of the transmit ports, the electronic commutator
will sequentially connect each of the receive coils, represented as Ports 1 through
16 (for the 4 × 4 array), to the B-channel of the analyzer. Thus, the operation
reduces to measuring a series of scattering coefficients for a simple 2-port network.
136 8 Planar and Conforming Arrays of Probes
For example, when transmit coil A is connected to the analyzer, we will measure the
sequence of scattering coefficients, SAA, SAi , SiA , Sii , for i = 1, . . . , 16. With these
results, we can calculate the set of open-circuit transfer impedances from port A to
port i via
2Z0 SAi
ZAi = , i = 1, . . . , 16, (8.1)
1 − SAA − Sii + (SAASii − SAiSiA )
Table 8.2 Freespace response (reactance) of the sensor when drive 1 is excited at 1MHz
Table 8.3 Freespace response (reactance) of the sensor when drive 2 is excited at 1 MHz
2
A
3
Channel A
HP3577A Transmit Ports Receive Ports Channel B
HP3577A
B
62
63
64
Fig. 8.8 Illustrating an N-port system representation of the T/R array and its connection to the
HP3577A network analyzer
Englewood Cliffs, 1964, Chap. 4. For a reciprocal network of the type that we
have, SAi = SiA . Open-circuit transfer impedances are the input data for voxel-based
inversion algorithms.
138 8 Planar and Conforming Arrays of Probes
In addition to the usual open-circuit impedance parameters, z11 , z12 , z21 , z22 ,
and short-circuit admittance parameters, y11 , y12 , y21 , y22 , there exists a set of
parameters, called the chain parameters, that specify the current and voltage at one
port in terms of those at the other. In this way, one can chain networks together very
conveniently. These parameters are labeled A, B, C, and D, where
V1 = AV2 − BI2
I1 = CV2 − DI2 . (8.3)
Table 8.4 gives the relationship between the three sets of two-port parameters.2
The vertical bars denote the determinant of the corresponding 2 × 2 matrix. It is a
2 Takenfrom N. Balabanian and T. Bickart, Electrical Network Theory, J. Wiley & Sons, Inc., New
York, 1969, Chap. 3.
8.3 N-Port Theory of the T/R Array 139
z12
simple matter to verify that the determinant of the chain matrix, AD − BC = =
z21
y12
= 1, for reciprocal two-ports. Hence, for such networks the inverse chain matrix
y21
always exists.
Two independent two-ports can be chained together to produce a composite third
two-port simply by multiplying their corresponding chain matrices:
A3 B3 A1 B1 A2 B2
= × . (8.4)
C3 D3 C1 D1 C2 D2
M1 = M3 × M−1
2 . (8.5)
If, then, we assume that the parasitic effects are unchanged whether the T/R array
is in freespace or over a host material, we simply left-multiply the new measured
composite chain matrix by M−1 1 to get the new model chain matrix that corresponds
to the current test configuration. Using the top two lines of Table 8.4 allows us
to transform from the A, B, C, D parameters of the model to the open-circuit
impedance parameters that VIC-3D R
uses.
If, as is likely, the parasitic effects change when the T/R array is over a host, we
redo the procedure described above by replacing freespace by an unflawed version
of the host, or a metal whose conductivity is close to that of the host.
There are 128 chain matrices for the two drive coils and 64 receive coils. We’ll
compute a few of them here. First, we note that the freespace driving-point
impedance of either drive coil is ZD = j562.64 Ω at 1 MHz, and the driving-point
impedance of each receive coil is ZR = j5.7559 Ω in freespace at 1 MHz. These
are model values computed by VIC-3D R
. The entries in Tables 8.2 and 8.3 are
the open-circuit transfer impedances between the appropriate drive coil and each of
the receive coils, also calculated by VIC-3D R
.
We’ll label each receive coil by its row-column indices, and each drive coil by
either A or B, with A corresponding to drive 1, and B to drive 2. Thus, our notation
140 8 Planar and Conforming Arrays of Probes
Table 8.6 Measured freespace scattering parameters at 1MHz taken with the
HP3577A
Connector 7 7 7 7
Pins 5 4 3 2
S11 Mag dBm −0.133 −0.133 −0.133 −0.133
Phase degrees −3.957 −3.992 −4.016 −4.044
S21 Mag dBm −62.4 −62.6 −60.3 −59.4
Phase degrees 162 156 159 158
S12 Mag dBm −62.4 −62.3 −60.2 −59.5
Phase degrees 162 157 159 158
S22 Mag dBm −1.101 −1.199 −1.280 −1.370
Phase degrees 160.70 0160.70 0160.50 0160.44
yields ZAA = ZBB = j562.64, Zi j,i j = j5.7559, for all i j, i j, and, for example, ZA,45 =
Z45,A = j0.11258. With this notation, we’ll compute the chain matrix for the coupled
pair (A, 45) using the relationships of Table 8.4:
⎡ ⎤
562.64 −562.64 × 5.7559 + (0.11258)2
⎢ 0.11258 ⎥
AB ⎢ j0.11258 ⎥
⎢
=⎢ ⎥
CD ⎥
(A,45) ⎣ 1 5.7559 ⎦
j0.11258 0.11258
4997.69 j28766.09
= . (8.6)
− j8.88257 51.1272
The dc resistance of the two drive coils are 14.0 Ω (yellow–orange) and 15.3 Ω (red–
brown). The dc resistance for several receive coils are listed in Table 8.5.
Table 8.6 shows measured freespace scattering parameters taken at 1 MHz with
the HP3577A, and Table 8.7 gives the scattering parameters expressed as complex
numbers.
Using (8.1) and (8.2) and Table 8.7, we compute the open-circuit driving-point
and transfer impedances at 1 MHz and list the results in Table 8.8. It is clear from
these results that the driver coil is operating beyond resonance because of the large
negative (capacitive) reactance in S11 , as well as a large resistance component.
8.3 N-Port Theory of the T/R Array 141
Table 8.7 Measured freespace scattering parameters at 1 MHz expressed as complex numbers
Connector 7 7 7 7
Pins 5 4 3 2
S11 0.9848 0.9848 0.9848 0.9848
0.9976 − j0.069 0.9976 − 0.0696 0.9975 − j0.070 0.9975 − j0.0705
S21 7.5858 × 10−4 7.4131 × 10−4 9.661 × 10−4 10.715 × 10−4
−0.951 + j0.3090 −0.9135 + j0.4067 −0.9336 + j0.3584 −0.9272 + j0.3746
S12 7.5858 × 10−4 7.4131 × 10−4 9.661 × 10−4 10.715 × 10−4
−0.951 + j0.3090 −0.9135 + j0.4067 −0.9336 + j0.3584 −0.9272 + j0.3746
S22 0.8809 0.8711 0.8630 0.8541
−0.9438 + j0.3305 −0.9438 + j0.3305 −0.9426 + j0.3338 −0.9423 + j0.3348
This has an effect on the other parameters. For example, from Tables 8.2 and 8.3,
we expect the reactance of Z12 to have much smaller values than those shown in
Table 8.8.
By using the data of Table 8.8 in the two-port relations of Table 8.4, we compute
the ABCD-matrices for the array at 1 MHz. The results are shown in Table 8.9. The
integer at the left is the pin number, and the complex number beneath each matrix is
the determinant of the matrix. Note that each determinant is equal to the theoretical
value of unity.
Now, we’ll apply the chain-matrix algorithm described in (8.4) to the measured
data. In this example, we’ll let M3 correspond to the pin 5 measured data and M2
correspond to the Drive 2(1,1) model data. When we apply (8.4) to these matrices,
we compute M1 to be as shown in Table 8.10. When we substitute the Z-parameter
data for M1 in Table 8.10 into the generic T-network 2-port of Fig. 8.9, we obtain the
equivalent 2-port T-network for the parasitic structure shown in Fig. 8.10. Though
we do not need to show the parasitic network, since we are not trying to synthesize
it (in fact, we’re trying to eliminate it!), it is interesting to note that it is physically
realizable with passive elements.
142 8 Planar and Conforming Arrays of Probes
Table 8.9 ABCD matrices for the array based upon the HP3577A measurements at 1 MHz
2 −0.166034D+04 −0.322446D+03 −0.396387D+04 −0.155253D+05
−0.272074D−01 −0.122143D+01 0.103496D+02 −0.518038D+01
0.10000D+01 0.11595D−11
3 −0.185639D+04 −0.326783D+03 −0.423618D+04 −0.170986D+05
−0.565164D−01 −0.135132D+01 0.113315D+02 −0.559892D+01
0.10000D+01 −0.15228D−11
4 −0.240460D+04 −0.556934D+03 −0.380976D+04 −0.223168D+05
0.199263D−01 −0.176077D+01 0.149665D+02 −0.607118D+01
0.10000D+01 −0.22027D−12
5 −0.240655D+04 −0.295666D+03 −0.533580D+04 −0.213383D+05
−0.165187D+00 −0.170783D+01 0.139209D+02 −0.696156D+01
0.10000D+01 0.18492D−11
The integer at the left denotes the pin, and the complex number beneath each array is the value of
the determinant of the matrix (ABCD matrix HP3577A.f90)
Figure 8.11 shows the networks associated with M1 and M2 coupled together
between the new ports 1 and 2. The coupled network has the same Z-parameters
as originally tabulated for Pin 5 in Table 8.8. (We’ll leave this as an exercise for
the reader.) This shows the consistency between the algebra of ABCD-matrices and
simple network theory.
8.3 N-Port Theory of the T/R Array 143
−0.00547
j0.00372
Fig. 8.10 Showing the equivalent T-network 2-port for the parasitic structure
M1 M2
−0.00547
Port1 −j0.04808 Port2
j0.00372
Fig. 8.11 Showing the coupled networks that correspond to M1 (the parasitic elements) and M2
(the model data for Drive 2(1,1)). The composite network extending from Port1 to Port2 has the
same Z-parameters as shown in Table 8.8 for Pin 5
Let M3 = M1 × M2 , where
and let
9.1 Introduction
There are a number of rather complicated geometries that appear in the inspection
of steam generator tubing by means of eddy currents. Figures 9.1–9.3 illustrate
several of them, and Fig. 9.4 illustrates a number of different flaws that must be
modeled. The model must not only contend with these geometries, but it must deal
with ferromagnetic bodies, as well.
Laser weld
therefore solenoidal in the electric variables, it will be easier to use an “all electric”
model for the coupled system, following the same ideas that we developed in the
context of Amperian currents in Chap. 3. The electric–electric Green function is
simple, and because all electric variables are solenoidal, the matrices should be well
conditioned for all frequencies.
9.3 Coupled Ferromagnetic Integral Equations 147
1
B(r, z) = − ∇×E
jω
1 (ee)
= B(i) (r, z) − ∇ × aφ 2π Gφ φ (r, z; r , z )Ja (r , z )r dr dz
jω flaw
1 (ee)
− ∇ × aφ 2π Gφ φ (r, z; r , z )(∇ × Ma )φ r dr dz . (9.3)
jω flaw
E(i) and B(i) are the incident fields produced by the exciting coil. The curl operation
divided by − jω in (9.3) defines a magnetic–electric operator.
Upon multiplying E by the anomalous conductivity, σa (r) = jω (ε̂ (r) − ε̂h ), we
μ (r)μh
get the anomalous current, Ja . Similarly, B(r) = Ma (r). Hence, when
μ (r) − μh
we make these substitutions in (9.3), we get the coupled integral equations for
axisymmetric, ferromagnetic problems:
Ja (r, z) (ee)
= E (i) (r, z) + 2π Gφ φ (r, z; r, z )Ja (r , z )r dr dz
σa (r, z) flaw
(ee)
+ 2π Gφ φ (r, z; r, z )(∇ × Ma )φ r dr dz
flaw
μ (r, z)μh 1 (ee)
Ma (r, z) = B(i) (r, z) − ∇ × aφ 2π Gφ φ (r, z; r, z )Ja (r , z )r dr dz
μ (r, z) − μh jω flaw
1 (ee)
− ∇ × aφ 2π Gφ φ (r, z; r, z )(∇ × Ma )φ r dr dz .
jω flaw
(9.4)
We have dropped the vector components from the electrical variables; the scalar
equation in (9.4) is the φ -component of the electric field integral equation.
lm
where π1m is the unit pulse function, which starts at the mth node, and π2m is the
triangular function, which starts at the mth node. Ja (r, z), on the other hand, can be
expanded in pulse functions
(9.7)
where
(i)
Elm = E (i) (r, z)π1l (r)π1m (z)rdrdz
(ee)
Glm,LM = 2π rdrdzπ1l (r)π1m (z) G(r, z; r , z )π1L (r )π1M (z )r dr dz
(em)(r)
Glm,LM = 2π rdrdzπ1l (r)π1m (z) G(r, z; r , z )π1L (r )π2M
(z )r dr dz
(ee) (ee)
Glm,LM − Glm,LM+1
=
δz
(em)(z)
Glm,LM = 2π rdrdzπ1l (r)π1m (z) G(r, z; r , z )π2L
(r )π1M (z )r dr dz
(ee) (ee)
Glm,LM − Glm,L+1M
= . (9.8)
δr
The discretized version of (9.4)(b) is
(+) (−)
rl + rl
∑ QmM MlM + ∑ Glm,LM JLM + ∑ Glm,LM
(i)(r) (r) (r) (me)(r) (mm)(rr) (r)
Blm = δr MLM
2 M LM LM
−∑
(mm)(rz) (z)
Glm,LM MLM
LM
(9.9)
150 9 Multilayered Media with Cylindrical Geometries
where the tri-diagonal matrices are symmetric and have the following nonzero
entries:
(r) (−)
QmM = ν (rl , z)π2M (z)π2m (z)dz
⎧
νlm if M = m − 1
δz ⎨
= 2(νlm + νlm+1 ) if M = m
6 ⎩
νlm+1 if M = m + 1
(z)
QlL = ν (r, mδ z)π2L (r)π2l (r)rdr
⎧
⎪ 1 l + r0 /δ r
⎪
⎪ νlm + if L = l − 1
⎪
⎪ 12 6
⎪
⎪
⎪
⎪
⎪
⎨
1 l + r0 /δ r 10 l + r0 /δ r
= δr 2
νlm + + νl+1,m + if L = l
⎪
⎪ 4 3 24 3
⎪
⎪
⎪
⎪
⎪
⎪
⎪
⎪ 1 l + r0 /δ r
⎩ νl+1,m + if L = l + 1
4 6
(9.10)
(−) (+)
where r0 is the radial starting point of the grid. rl = r0 + l δ r and rl = r0 +
(l + 1)δ r. The first entry in each of these matrices is the lower diagonal, the second
the main diagonal, and the third the upper diagonal.
The other matrices are given by
(i)(r)
Blm = B(i) (r, z) · aρ π1l (r)π2m (z)rdrdz
(i) (i)
1 Elm − Elm+1
=−
jω δz
(i)(z)
Blm = B(i) (r, z) · az π2l (r)π1m (z)rdrdz
(i) (i)
1 Elm − El+1m
=
jω δr
(me)(r) 2π
Glm,LM = rdrdzπ1l (r)π2m (z) G(r, z; r , z )π1L (r )π1M (z )r dr dz
jω
(ee) (ee)
1 Glm,LM − Glm+1,LM
=
jω δz
9.5 Calculation of Matrices 151
(mm)(rr) 2π
Glm,LM = rdrdzπ1l (r)π2m (z) G(r, z; r , z )π1L (r )π2M
(z )r dr dz
jω
(ee) (ee) (ee) (ee)
Glm,LM − Glm+1,LM − Glm,LM+1 + Glm+1,LM+1
=
jωδ z2
(mm)(rz) 2π
Glm,LM = rdrdzπ1l (r)π2m (z) G(r, z; r , z )π2L
(r )π1M (z )r dr dz
jω
(ee) (ee) (ee) (ee)
Glm,LM − Glm+1,LM − Glm,L+1M + Glm+1,L+1M
=
jωδ rδ z
(me)(z) 2π
Glm,LM = rdrdzπ2l (r)π1m (z) G(r, z; r , z )π1L (r )π1M (z )r dr dz
jω
(ee) (ee)
1 Glm,LM − Gl+1m,LM
=
jω δr
(mm)(zr) 2π
Glm,LM = rdrdzπ2l (r)π1m (z) G(r, z; r , z )π1L (r )π2M
(z )r dr dz
jω
(ee) (ee) (ee) (ee)
Glm,LM − Gl+1m,LM − Glm,LM+1 + Gl+1m,LM+1 (mm)(rz)
= = GLM,lm
jωδ rδ z
(mm)(zz) 2π
Glm,LM = rdrdzπ2l (r)π1m (z) G(r, z; r , z )π2L
(r )π1M (z )r dr dz
jω
(ee) (ee) (ee) (ee)
Glm,LM − Gl+1m,LM − Glm,L+1M + Gl+1m,L+1M
= . (9.11)
jωδ r2
Keep in mind that the Green function that appears in the integrands of (9.8) and
(ee)
(9.11) is really Gφ φ (r, z; r , z ). We will derive expressions for this function later,
but for the present we will use the results.
Refer to Fig. 9.5, which shows an axisymmetric layered tube with concentric
layers. A coaxial filamentary current source of unit strength is located at r = r in the
layer bounded by rs−1 < r < rs −1 . The wave parameter in the ith layer is given by
αi = (ki2 − h2)1/2 , where h is the one-dimensional spatial Fourier transform variable,
and ki2 = ω 2 μi ε̂i = ω 2 μi εi − jω μi σi . The eigenvectors, V1m , V2m , V3m , V4m , shown
in Fig. 9.5 are defined in Appendix A.1.
152 9 Multilayered Media with Cylindrical Geometries
Axis of Tube
r=0
V(1)=a1V11+c1V31 Region 1 α=α1
r1
V(2)=a2V12+b2V22+c2V32+d2V42 Region 2 α=α2
r2
..
..
rs-1
Region s
V(s)=asV1s+bsV2s+csV3s+dsV4s α=αs
Source Point
r=r | αs=αs|
|
V(s’)=as’V1s’+bs’V2s’+cs’V3s’+ds’V4s’ Region s
α=αs|
rs| -1
..
..
r2 |
|
Region 2
V(2’)=a2’V12’+b2’V22’+c2’V32’+d2’V42’ α=α2|
r1|
V(1’)=b1’V21’+d1’V41’ |
Region 1
α=α1|
(2)
παs V11 (s)e− jαs Δs e− jαs rs −1 J1 (αs r ) + V21(s)e jαs rs−1 H1 (αs r )
d1 = · , (9.12)
j2 e− j2αs Δs U22 (s )V11 (s) − U12 (s )V21 (s)
μs (2)
Ẽφ (r, r ) = − jω d1 U12 (s )e− jαs (rs−1 −r) j1 (αs r) + U22 (s )e jαs (rs−1 −r) h1 (αs r)
αs
r > r
μs (2)
= − jω c1 V11 (s)e− jαs (rs −1 −r) j1 (αs r) + V21(s)e jαs (rs −1 −r) h1 (αs r)
αs
r < r . (9.15)
ω μs (2)
Ẽφ (r, r ) = −π H (αs r )J1 (αs r) , r < r
2 1
ω μs (2)
Ẽφ (r, r ) = −π H (αs r)J1 (αs r ) , r < r . (9.16)
2 1
Alternatively, we can treat (9.16) as an “incident Green function,” which is then
scattered from the various layers, producing the total function in (9.15).
The spatial-domain Green function, in which the field point, r, is in region
m, m , s, and the source point in region s is given by the inverse Fourier transform of
the corresponding expression in (9.13)–(9.15)
∞
1
G(r, z; r , z ) = Ẽ(r, r )e jh(z−z ) dh . (9.17)
(2π )2 −∞
The G(r, z; r , z ) that appear in (9.8) and (9.11) have r, r in the same (source)
region, so it corresponds to (9.15) being substituted into (9.17).
The U’s and V ’s are transfer coefficients that will be defined recursively shortly.
Note that when s = 1, i.e., when there are no layers above the source point in Fig. 9.5,
then V11 (1) = 1, V21 (1) = 0; when s = 1 , i.e., when there are no layers below the
source point, then U22 (1 ) = 1, U12 (1 ) = 0.
It will be more convenient to compute the infinite-space Green function by using
∞ −αs |z−z |
j ω μs e
G(r, z; r , z ) = − J1 (r l)J1 (rl)ldl , (9.18)
2π 0 2αs
which is the same as the free-space Green function that is used in the ferrite-
core module. Equations (9.17) and (9.18) produce the same result for a filamentary
current loop in infinite space, even though they correspond to different coordinate
systems.
When computing matrix elements that involve G(ee) in the source region, it will
be advantageous to separate the incident, or infinite-space, term, (9.16), from the
layered-space terms, as in Chap. 2. Because (9.15) includes both terms, we subtract
(9.16) from (9.15) to get the layered-space terms.
154 9 Multilayered Media with Cylindrical Geometries
(2)
U12 (s )e− jαs (rs−1 −r) j1 (αs r) +U22 (s )e jαs (rs−1 −r) h1 (αs r)
r > r
π
= − jω μs ·
j2
⎡ j αs Δ s ⎤
U12 (s ) j2αs Δs e− jαs rs −1 J (α r ) + e jαs rs−1 H (2) (α r )
⎢ U22 (s )V11 (s) e 1 s e s ⎥
⎢ V11 (s) 1
⎥·
⎣ 1 − Rs ,s e j2α s Δ s ⎦
(2)
V11 (s)e− jαs (rs −1 −r) j1 (αs r) +V21 (s)e jαs (rs −1 −r) h1 (αs r)
r < r . (9.21)
⎡
U12 (s ) − jαs (2rs−1 −(r+r ))
e j1 (αs r ) j1 (αs r)
π ⎢
⎢ U22 (s )
Ẽφ (r, r ) = − jω μs
j2 ⎣ 1 − Rs,s e j2αs Δs
(2) (2)
H1 (αs r)J1 (αs r ) + Rs,s e j2αs Δs H1 (αs r )J1 (αs r)
+
1 − Rs,s e j2αs Δs
⎤
V21 (s) (2) (2)
e jαs (2rs −1 −(r+r )) h1 (αs r )h1 (αs r) ⎥
V11 (s) ⎥
+ ⎦
1 − Rs,s e j2αs Δs
r > r
⎡
U12 (s ) − jαs (2rs−1 −(r+r ))
e j1 (αs r ) j1 (αs r)
π ⎢
⎢ U22 (s )
= − j ω μs
j2 ⎣ 1 − Rs,s e j2αs Δs
(2) (2)
H1 (αs r )J1 (αs r) + Rs ,s e j2αs Δs H1 (αs r)J1 (αs r )
+
1 − Rs,s e j2αs Δs
⎤
V21 (s) (2) (2)
e jαs (2rs −1 −(r+r )) h1 (αs r )h1 (αs r) ⎥
V11 (s) ⎥
+ ⎦
1 − Rs,s e j2αs Δs
r > r (9.22)
Note the similarity between these results and the (a) and (b) terms (convolution
and correlation) in Chap. 2. Of course, because the scaled Bessel and Hankel
functions are not independent of (r, r ), the results of (9.23) are not rigorously
convolutions and correlations in (r, r ). The expression (9.23), however, does make
clear the presence of incident and reflected fields at the layers, r and r , just as in
Chap. 2.
In a similar manner, we can derive expressions for the transmitted fields above
and below the source region. In region m (rm−1 < r < rm ) above the source region,
we have, upon substituting (9.12) into (9.13):
αs π
Ẽφ (r, r ) = − jω μm
αm j2
(2)
ρs e− jαs Δs e− jαs (rs −1 −r ) j1 (αs r ) + e jαs (rs−1 −r ) h1 (αs r )
×
1 − ρs ,s e− j2αs Δs
V21 (m) jαm (rm −r) (2) V11 (m) − jαm (rm −r)
× e h ( α r) + e j ( α r) ,
V11 (s)e− jαs Δs
m
V11 (s)e− jαs Δs
1 1 m
(9.24)
and in region m (rm < r < rm −1 ) below the source region:
αs π
Ẽφ (r, r ) = − jω μm
αm j2
(2)
ρs e− jαs Δs e jαs (rs−1 −r ) h1 (αs r ) + e− jαs(rs −1 −r ) j1 (αs r )
×
1 − ρs,s e− j2αs Δs
U12 (m )
× e− jαm (rm −r) j1 (αm r)
U22 (s )e− jαs Δs
U22 (m ) j αm (rm −r) (2)
+ e h (α r) . (9.25)
U22 (s )e− jαs Δs 1 m
By writing
V11 (m)
= Rm
V11 (s)
U12 (m ) U22 (m ) U12 (m )
= ·
U22 (s ) U22 (s ) U22 (m )
U22 (m )
= R , (9.26)
U22 (s ) m
in region m (rm < r < rm −1 ) below the source region. We will return to a
discussion of the computation of the reflection coefficients (the R’s) shortly, but
right now we’ll get into details of the computation of the matrix elements.
Consider
(ee)
Glm,LM = 2π rdrdzπ1l (r)π1m (z) G(r, z; r , z )π1L (r )π1M (z )r dr dz
where
G(r, z; r , z ) = G(0) (r, z; r , z ) + G(s) (r, z; r , z ) . (9.29)
158 9 Multilayered Media with Cylindrical Geometries
G(0) (r, z; r , z ) is the infinite-space Green function and G(s) (r, z; r , z ) is the layered-
space Green function:
∞ −αs |z−z |
jω μs e
G(0) (r,z;r ,z ) = − J1 (r l)J1 (rl)ldl
2π 0 2αs
ω μs ∞ jh(z−z )
G(s) (r,z;r ,z ) = − e
8π −∞
(2) (2)
ρ e− jαs (2rs −1 −(r+r )) j1 (αs r ) j1 (αs r) + ρs e jαs (2rs−1 −(r+r )) h1 (αs r )h1 (αs r)
× s
1 − ρs ,s e− j2αs Δs
⎤
(2) (2)
ρs ,s e− j2αs Δs e− jαs (r −r) h1 (αs r ) j1 (αs r) + e jαs (r −r) h1 (αs r) j1 (αs r )
+ ⎦ dh .
1 − ρs ,s e− j2αs Δs
(9.30)
where
⎧ ' (2
⎪
⎪ 1 − e− α s δ z
⎪
⎪ e α s δ z(m−M+1) , if m − M ≤ −1
⎪
⎪ αs δ z
⎪
⎪ ' (
⎪
⎨ e−αs δ z − 1 + αsδ z
F(m − M) = δ z2 2 , if m − M = 0
⎪
⎪ (αs δ z)2
⎪
⎪ ' (2
⎪
⎪
⎪ −αs δ z(m−M−1) 1 − e−αsδ z
⎪
⎪
⎩e , if m − M ≥ 1
αs δ z
b
I (a, b) = zJ1 (z)dz . (9.32)
a
Note that F(m − M) = F(M − m); i.e., it is symmetrical in (m, M), and corresponds
to the F3 function in (5.29).
The layered-space matrix element is given by:
' (
(s)(ee) ω μs 2 ∞ 2 − e− jhδ z − e jhδ z
Glm,LM =− δz e jh(m−M)δ z
4 −∞ h2 δ z2
I (l(αs δ r), (l + 1)(αs δ r)) · I (L(αs δ r), (L + 1)(αs δ r))
× R(1) (αs )
αs4
H (l(αs δ r), (l + 1)(αs δ r)) · H (L(αs δ r), (L + 1)(αs δ r))
+ R(2) (αs )
αs4
9.5 Calculation of Matrices 159
I (l(αs δ r), (l + 1)(αs δ r)) · H (L(αs δ r), (L + 1)(αs δ r))
(3)
+ R (αs )
αs4
H (l(αs δ r), (l + 1)(αs δ r)) · I (L(αs δ r), (L + 1)(αs δ r))
+ dh ,
αs4
(9.33)
where
ρs e− jαs 2rs −1
R(1) (αs ) =
1 − ρs,s e− j2αs Δs
ρs e jαs 2rs−1
R(2) (αs ) =
1 − ρs,s e− j2αs Δs
ρs ,s e− j2αs Δs
R(3) (αs ) =
1 − ρs,s e− j2αs Δs
z2
(2)
H (z1 , z2 ) = H1 (z)zdz. (9.34)
z1
Now we calculate the field produced by a bobbin coil that extends from l0 ≤ z ≤
l1 , r0 ≤ r ≤ r1 , and is located in region c. First, we note that the current density
flowing within such a coil is Nc I, where Nc is the density of turns (i.e., turns per unit
area), and I is the current flowing within the coil. Hence, we have
l1 r1
E (i) (r, z) = 2π Nc I dz r dr G(r, z; r , z )
l0 r0
∞ l1 r1
Nc I − jhz
= e jhz
dh e dz r dr Ẽ(r, r ) . (9.35)
2π −∞ l0 r0
Ẽ(r, r ) is given by (9.27) if the field point is above region c, and by (9.28) otherwise.
Hence,
∞ jh(z−l0 ) r
(i) Nc I e − e jh(z−l1) 1
E = r dr Ẽ(r, r ) dh . (9.36)
2π −∞ jh r0
(2)
× Rm e jαm (rm −r) h1 (αm r) + e− jαm(rm −r) j1 (αm r) dh
where
The first result in (9.37) would be used for an encircling-coil, because in that case
the entire tube is “above” the coil. Similarly, the second result would be used for an
internal-coil, because the entire tube is “below” the coil.
Because the dependency on r and r , as well as z and z , is separated in (9.37), it
is easy to compute the moments of the incident field, as given in (9.8):
(i) Nc I ∞ αc V11 (m)
Elm = ω μm
4 −∞ αm V11 (c)e− jαc Δc
e− jh(l0 −(m+1)δ z) − e− jh(l0 −mδ z) − e− jh(l1 −(m+1)δ z) + e− jh(l1 −mδ z)
×
h2
I (αc r0 , αc r1 ) H (αc r0 , αc r1 )
× R(4) (αc ) + R (5)
( αc )
αc2 αc2
H (l αm δ r, (l + 1)αm δ r) − jαm rm I (l αm δ r, (l + 1)αm δ r)
× Rm e j αm rm + e dh
αm2 αm2
rm−1 < r < rm (above coil)
9.5 Calculation of Matrices 161
∞
Nc I αc U22 (m )
= ω μm
4 −∞ αm U22 (c )e− jαc Δc
e− jh(l0 −(m+1)δ z) − e− jh(l0 −mδ z) − e− jh(l1 −(m+1)δ z) + e− jh(l1 −mδ z)
×
h2
H (αc r0 , αc r1 ) I (αc r0 , αc r1 )
× R(6) (αc ) + R (7)
( αc )
αc2 αc2
− j α r m
I (l αm δ r, (l + 1)αm δ r) j α r m
H (l αm δ r, (l + 1)αm δ r)
× R m e m +e m dh
αm2 αm2
The next thing that we must do with the coil is compute its driving-point impedance.
We will use the reaction-formula
E · J
Z=− , (9.40)
I2
where J is the current density in the coil (= Nc Iaφ ), and I is the total current. Thus,
2π Nc I
Z=− E(r, z)rdrdz
I2
(2π )2 (Nc I)2
=− rdr dz r dr dz G(0) (r, z; r , z )
I2
)
(s)
+ G (r, z; r , z ) , (9.41)
where G(0) (r, z; r , z ) and G(s) (r, z; r , z ) are the infinite-space and layered-space
Green functions given in (9.30). Note that the integral over G(0) gives the free-
space impedance, Z (0) , of the coil, and the integral over G(s) gives the change in
impedance, ΔZ, due to the presence of the scattering body (the tube). Hence,
162 9 Multilayered Media with Cylindrical Geometries
Z = Z (0) + ΔZ
∞ (c)
F (l1 − l0 ) I 2 (r0 h, r1 h)
= jω μc 2π Nc2 · dh
0 2αc h3
∞
' (
π 2 − e− jh(l1−l0 ) − e jh(l1 −l0 )
+ ω μc Nc2
2 −∞ h2
I 2 (αc r0 , αc r1 ) H 2 (αc r0 , αc r1 )
× R(1) (αc ) + R (2)
(αc )
αc4 αc4
(3) I (αc r0 , αc r1 )H (αc r0 , αc r1 )
+2R (αc ) dh , (9.42)
αc4
−1 + e −(l1−l0 )αc + α (l − l )
where F (c) (l1 − l0 ) = 2
c 1 0
is the specialization to the
αc2
coil of the F-function of (9.32).
Appendix
∇ × E = − jω μ H − M(i)
∇ × H = jω ε̂ E + J(i) , (9.43)
where M(i) and J(i) are impressed current sources, M(i) being magnetic, and J(i)
electric.
In cylindrical coordinates, the component forms of (9.43) are
A.1 Cylindrical Eigenvectors and the Green Function 163
1 ∂ Ez ∂ Eφ (i)
− = − jω μ Hr − Mr
r ∂φ ∂z
∂ Er ∂ Ez (i)
− = − jω μ Hφ − Mφ
∂z ∂r
1 ∂ (rEφ ) 1 ∂ Er (i)
− = − jω μ Hz − Mz
r ∂r r ∂φ
1 ∂ Hz ∂ Hφ (i)
− = jω ε̂ Er + Jr
r ∂φ ∂z
∂ Hr ∂ Hz (i)
− = jω ε̂ Eφ + Jφ
∂z ∂r
1 ∂ (rHφ ) 1 ∂ Hr (i)
− = jω ε̂ Ez + Jz . (9.44)
r ∂r r ∂φ
2π ∞
1
F̃(r, m, h) = e− jmφ d φ e− jhz F(r, φ , z )dz , (9.45)
(2π )2 0 −∞
where the second equation is the direct transform from physical space to frequency
(wave-number) space.
Upon transforming (9.44), we get, after cancelling the common factor, e jnφ e jhz :
jn (i)
Ẽz − jhẼφ = − jω μ H̃r − M̃r
r
d Ẽz (i)
jhẼr − = − jω μ H̃φ − M̃φ
dr
1 d(rẼφ ) n (i)
− j Ẽr = − jω μ H̃z − M̃z
r dr r
jn (i)
H̃z − jhH̃φ = jω ε̂ Ẽr + J˜r
r
d H̃z (i)
jhH̃r − = jω ε̂ Ẽφ + J˜φ
dr
1 d(rH̃φ ) n (i)
− j H̃r = jω ε̂ Ẽz + J˜z . (9.46)
r dr r
164 9 Multilayered Media with Cylindrical Geometries
The first and fourth equations are free of derivatives. Hence, they allow us to
eliminate H̃r and Ẽr in favor of the transverse components, H̃z , H̃φ , Ẽz , Ẽφ :
(i)
n h M̃r
H̃r = − Ẽz + Ẽφ −
ωμr ωμ jω μ
(i)
n h J˜r
Ẽr = H̃z − H̃φ − . (9.47)
ω ε̂ r ω ε̂ jω ε̂
d Ẽz (i)
jhẼr − = − jω μ H̃φ − M̃φ
dr
1 d(rẼφ ) n (i)
− j Ẽr = − jω μ H̃z − M̃z
r dr r
d H̃z (i)
jhH̃r − = jω ε̂ Ẽφ + J˜φ
dr
1 d(rH̃φ ) n (i)
− j H̃r = jω ε̂ Ẽz − J˜z . (9.48)
r dr r
d Ẽz jhn jh2 jω μ h ˜(i) (i)
= H̃z + − + (rH̃φ ) − Jr + M̃φ
dr ω ε̂ r ω ε̂ r r ω ε̂
d(rẼφ ) jn2 jnh n ˜(i) (i)
= − jω μ r H̃z − (rH̃φ ) − Jr − rM̃z
dr ω ε̂ r ω ε̂ r ω ε̂
2
d H̃z jhn jh jω ε̂ (i) h (i)
=− Ẽz + − (rẼφ ) − J˜φ − M̃r
dr ωμr ωμr r ωμ
d(rH̃φ ) jn2 jnh (i) n (i)
= − + jω ε̂ r Ẽz + (rẼφ ) + rJ˜z − M̃r . (9.49)
dr ωμr ωμr ωμ
A.1 Cylindrical Eigenvectors and the Green Function 165
⎡ ⎤ ⎡ ⎤⎡ ⎤
Ẽz 0 0 a11 (r) a12 (r) Ẽz
d ⎢ ⎥
⎢ rẼφ ⎥ =
⎢0
⎢ 0 a21 (r) a22 (r) ⎥ ⎢
⎥ ⎢ rẼφ ⎥
⎥
dr ⎣ H̃z ⎦ ⎣ b11 (r) b12 (r) 0 0 ⎦ ⎣ H̃z ⎦
rH̃φ b21 (r) b22 (r) 0 0 rH̃φ
⎡ ⎤
J˜r
⎡ ⎤
u11 0 0 0 u15 0 ⎢ ⎢
⎥
J˜φ
⎥
⎢ u21 ⎢ ⎥
0 0 0 0 u26 ⎥ ⎢ J˜z
⎥
+⎢
⎣ 0
⎥⎢ ⎥, (9.50)
u32 0 u34 0 0 ⎦ ⎢ M̃r ⎥
⎢ ⎥
0 0 u43 u44 0 0 ⎣ M̃φ ⎦
M̃z
where
⎡ ⎤
0 0 0 0
dM M ⎢ 0 0 − j2ω μ 0⎥
= − +⎢
⎣
⎥.
dr r 0 0 0 0⎦
j2ω ε̂ 0 0 0
Then
166 9 Multilayered Media with Cylindrical Geometries
d2V dM dV
2
= V+M
dr dr dr
dM
= V + M2 V
dr
⎤ ⎡
0
1 dV n2 ⎢ − j2ω μ V3 ⎥
=− + 2 + h2 − ω 2 μ ε̂ V + ⎢
⎣
⎥.
⎦ (9.52)
r dr r 0
j2ω ε̂ V1
When we recall (9.47), we see that H̃r = 0, Ẽr = 0, for axisymmetric problems,
which means that we have TE and TM fields with respect to the radial direction, as
well.
In order to compute the Green function, recall Fig. 9.5. We use eigenvectors V1
and V3 in the region next to r = 0, because these two vectors contain J0 (α1 r), which
(2)
is continuous at r = 0. Conversely, V2 and V4 contain H0 (α1 r), which, because it
represents an out-going cylindrical wave, satisfies the radiation condition at infinity,
which includes Region 1 .
Our strategy is to eliminate (a2 , b2 , c2 , d2 ), . . . , (ak , bk , ck , dk ) in favor of (a1 , c1 ),
and (a2 , b2 , c2 , d2 ), . . . , (ak , bk , ck , dk ) in favor of (b1 , d1 ), and then use jump
conditions at r = r to determine (a1 , c1 , b1 , d1 ). The intermediate steps are
accomplished by equating the four-vectors at each interface.
For example, at r = r1 , we have a1 V11 + c1 V31 = a2 V12 + b2 V22 + c2 V32 + V42 ,
which produces the two systems
(2)
a1 J0 (α1 r1 ) = a2 J0 (α2 r1 ) + b2H0 (α2 r1 )
α2 ε̂1 (2)
a1 J0 (α1 r1 ) = a2 J0 (α2 r1 ) + b2H0 (α2 r1 )
α1 ε̂2
(2)
c1 J0 (α1 r1 ) = c2 J0 (α2 r1 ) + d2H0 (α2 r1 )
α2 μ1 (2)
c1 J (α1 r1 ) = c2 J0 (α2 r1 ) + d2H0 (α2 r1 ) . (9.57)
α1 μ2 0
(2) (2)
b1 H0 (α1 r1 ) = a2 J0 (α2 r1 ) + b2 H0 (α2 r1 )
α2 ε̂1 (2) (2)
b 1 H0 (α1 r1 ) = a2 J0 (α2 r1 ) + b2 H0 (α2 r1 )
α1 ε̂2
(2) (2)
d1 H0 (α1 r1 ) = c2 J0 (α2 r1 ) + d2 H0 (α2 r1 )
α2 μ1 (2) (2)
d 1 H0 (α1 r1 ) = c2 J0 (α2 r1 ) + d2 H0 (α2 r1 ) . (9.58)
α1 μ2
In general, at r = rM ,
produces
(2) (2)
aM J0 (αM rM ) + bM H0 (αM rM ) = aM+1 J0 (αM+1 rM ) + bM+1 H0 (αM+1 rM )
αM+1 ε̂M αM+1 ε̂M (2)
aM J0 (αM rM ) + bM H0 (αM rM )
αM ε̂M+1 αM ε̂M+1
(2)
= aM+1 J0 (αM+1 rM ) + bM+1 H0 (αM+1 rM )
(2) (2)
cM J0 (αM rM ) + dM H0 (αM rM ) = cM+1 J0 (αM+1 rM ) + dM+1H0 (αM+1 rM )
αM+1 μM αM+1 μM (2)
cM J0 (αM rM ) + dM H0 (αM rM )
αM μM+1 αM μM+1
(2)
= cM+1 J0 (αM+1 rM ) + dM+1 H0 (αM+1 rM ) . (9.59)
Upon referring, once again, to Fig. 9.5, (9.61) is seen to produce the system
170 9 Multilayered Media with Cylindrical Geometries
⎡ ⎤ ⎡ (2) ⎤
J0 (αs r ) H0 (αs r )
⎢ ⎥ ⎢ ⎥
⎢ 0 ⎥ ⎢ 0 ⎥
⎢
(as − as ) ⎢ ⎥ ⎢ ⎥
0 ⎥ + (bs − bs ) ⎢ 0 ⎥
⎣ ω ε̂s ⎦ ⎣ ω ε̂ ⎦
(2)
−j r J0 (αs r ) −j
s
r H0 (αs r )
αs αs
⎡ ⎤ ⎡ ⎤
0 0
⎢ ω μs ) ⎥ ⎢ ω μs (2) ⎥
⎢j r J (α r ⎥ ⎢j r H0 (αs r ) ⎥
+(cs − cs ) ⎢ ⎥ ⎢ α ⎥
s
⎢ αs
0
⎥ + (d s − d )
s ⎢ s ⎥
(2)
⎣ J0 (αs r ) ⎦ ⎣ H0 (αs r ) ⎦
0 0
⎡ ⎤
0
⎢ 0 ⎥ − jhz
⎢ ⎥e
=⎢
⎢−1
⎥
⎥ (2π )2 . (9.62)
⎣ r ⎦
0
a s − a s = 0
b s − b s = 0
(2)
H1 (αs r )παs e− jhz
c s − c s = −
j2 (2π )2
J1 (αs r )παs e− jhz
d s − d s = . (9.63)
j2 (2π )2
(2)
H0,1 (z) = e− jz h0,1 (z)
c = ce jz
d = de− jz . (9.65)
(2)
cM j0 (αM rM ) + dM
h0 (αM rM ) = cM+1 e− jαM+1 ΔM+1 j0 (αM+1 rM )
(2)
+ dM+1 e jαM+1 ΔM+1 h0 (αM+1 rM )
αM+1 μM αM+1 μM (2)
cM
j1 (αM rM ) +dM h1 (αM rM )
αM μM+1 αM μM+1
(2)
= cM+1 e− jαM+1 ΔM+1 j1 (αM+1 rM ) +dM+1
e jαM+1 ΔM+1 h1 (αM+1 rM ) , (9.66)
where we have used the derivative identities defined below (9.63). We have also
defined
ΔM+1 = rM+1 − rM
ΔM +1 = −(rM +1 − rM ) , (9.67)
where ΔM +1 is defined with a negative sign, because rM +1 < rM .
We express (9.66) in matrix notation
cM+1 T11 (M + 1, M) T12 (M + 1, M) cM
= , (9.68)
dM+1 T21 (M + 1, M) T22 (M + 1, M) dM
where
172 9 Multilayered Media with Cylindrical Geometries
παM+1 rM e− jαM+1 ΔM+1 αM+1 μM
T21 (M + 1, M) = j0 (αM+1 rM ) j1 (αM rM )
j2 αM μM+1
− j1 (αM+1 rM ) j0 (αM rM )]
παM+1 rM e− jαM+1 ΔM+1 αM+1 μM (2)
T22 (M + 1, M) = j0 (αM+1 rM )h1 (αM rM )
j2 αM μM+1
(2)
− j1 (αM+1 rM )h0 (αM rM ) . (9.69)
In arriving at the final version of (9.69), we used the fact that the Wronskian for
(2)
unscaled Bessel functions holds also for scaled Bessel functions; j0 (z)h1 (z) −
(2)
j1 (z)h0 (z) = j2/π z.
The T-matrix is called the transition matrix from region M to M + 1. The
boundary condition on (9.68) is d1 = 0, c1 = 0.
The corresponding transition matrix from M to M + 1 (i.e., below the source) is:
cM +1 T11 (M + 1, M ) T12 (M + 1, M ) cM
= , (9.70)
dM +1 T21 (M + 1, M ) T22 (M + 1, M ) dM
where
A.1 Cylindrical Eigenvectors and the Green Function 173
(2) αM +1 μM
− h0 (αM +1 rM ) j1 (αM rM )
αM μM +1
(2) (2) αM +1 μM
− h0 (αM +1 rM )h1 (αM rM )
αM μM +1
παM +1 rM e jαM +1 ΔM +1 αM +1 μM
T21 (M + 1, M ) = j0 (αM +1 rM ) j1 (αM rM )
j2 αM μM +1
παM +1 rM e jαM +1 ΔM +1 αM +1 μM (2)
T22 (M + 1, M ) = j0 (αM +1 rM )h1 (αM rM )
j2 αM μM +1
(2)
− j1 (αM +1 rM )h0 (αM rM ) . (9.71)
παs − jαs rs−1 e− jhz
ds − e jαs (rs −1 −rs−1 ) ds = e J1 (αs r ) . (9.73)
j2 (2π )2
Note that all exponentials have exponents that are proportional to the difference
of nearby radii. For example, rs −1 − rs−1 is the width of the source region,
(2)
and e jαs rs−1 H0,1 (αs r ) → e− jαs (r −rs−1 ) , where r > rs−1 , and e− jαs rs −1 J0,1 (αs r ) →
e− jαs (rs −1 −r ) , where rs −1 > r . Hence, the right-hand sides of (9.73) vanish to
174 9 Multilayered Media with Cylindrical Geometries
exponential order for ℜαs → ∞, thereby ensuring proper behavior of the scaled
coefficients. Note that the unknowns for the entire system of equations are scaled
coefficients.
Now
cs c1
= T (s, s − 1) · · · T (2, 1)
ds 0
cs 0
= T (s , s − 1) · · · T (2 , 1 ) , (9.74)
ds d1
where T is a 2 × 2 matrix, with entries T11 , T12 , T21 , T22 , as defined in (9.69) and
(9.71).
Call
then
In general
Couple (9.76) with (9.73), and call rs −1 − rs−1 = Δs ; the results are then
παs jαs r (2) e− jhz
e jαs Δs U12 (s )d1 − V11(s)c1 = − e s −1 H1 (αs r )
j2 (2π )2
παs − jαs rs−1 e− jhz
U22 (s
)d1 − e jαs Δs V21 (s)c1 = e J1 (αs r ) , (9.78)
j2 (2π )2
A.1 Cylindrical Eigenvectors and the Green Function 175
(2)
παs V11 (s)e− jαs rs−1 J1 (αs r )+V21 (s)e jαs rs −1 e jαs Δs H1 (αs r ) e− jhz
d1 = (9.79)
j2 U22 (s )V11 (s)−e j2αs Δs U12 (s )V21 (s) (2π )2
This is the same as (9.12), with which we started the calculation of matrices.
Chapter 10
Some Special Topics in Computational
Electromagnetics
Region A Flaw
Region B
(e)
Ex (r) = G(q0)(1) (x − x , y − y; z, z )Jx (r )dr
(m)
+ G(q0)(1) (x − x , y − y; z, z )Jx (r )dr
∂ (e)
+ G(q0)(2) (x − x , y − y ; z, z )∇t · Jt (r )dr
∂x
∂ (m)
+ G(q0)(2) (x − x , y − y ; z, z )∇t · Jt (r )dr
∂x
1 ∂
+ 2 G(q0)(3) (x − x , y − y ; z, z )∇ · J(e) (r )dr
k0 ∂ x
(e)
Ey (r) = G(q0)(1) (x − x , y − y; z, z )Jy (r )dr
(m)
+ G(q0)(1) (x − x , y − y; z, z )Jy (r )dr
∂ (e)
+ G(q0)(2) (x − x , y − y; z, z )∇t · Jt (r )dr
∂y
∂ (m)
+ G(q0)(2) (x − x , y − y; z, z )∇t · Jt (r )dr
∂y
1 ∂
+ G(q0)(3) (x − x , y − y; z, z )∇ · J(e) (r )dr
k02 ∂ y
(e)
Ez (r) = G(q0)(4) (x − x , y − y; z, z )Jz (r )dr
(m)
+ G(q0)(4) (x − x , y − y; z, z )Jz (r )dr
1 ∂
+ G(q0)(5)(x − x , y − y ; z, z )∇ · J(e) (r )dr , (10.1)
k02 ∂ z
10.1 Spatial Decomposition Algorithm 179
which hold for r in region q and r in region 0. The notation for the Z-layering,
including the source region, is shown in Fig. 2.1 of Chap. 2. If r is in region −q,
then replace (q0) by (−q0). In deriving the final version of (10.1), we used the fact
that ∇ · J(m) (r) = 0.
The kernels that appear in (10.1) are given by
∞
(q0)(1) j ω μq B̃q0 (E)
G =− G̃q0 (z, z )J0 (rl)ldl
2π 0 2λ0
⎡ ⎤
(E) μ0 λ q (M)
∞ ⎢ B̃q0 G̃q0 (z, z ) − D̃q0 G̃q0 (z, z )
j ω μq μq λ 0 ⎥
G(q0)(2) =− ⎢ ⎥ J0 (rl)ldl
2π 0 ⎣ 2λ0 l 2 ⎦
∞
j ω μq μ0 λq D̃q0 (M)
G(q0)(3) = − G̃q0 (z, z )J0 (rl)ldl
2π 0 μq λ 0 2 λ 0
∞
(q0)(4) j ω μq μ0 D̃q0 (M)
G =− G̃q0 (z, z )J0 (rl)ldl
2π 0 μq 2 λ 0
∞
j ω μq μ0 λ0 D̃q0 (M)
G(q0)(5) = − G̃q0 (z, z )J0 (rl)ldl
2π 0 μq λ q 2 λ 0
∞
jω μ−q Ã−q0 (E)
G(−q0)(1) = − G̃ (z, z )J0 (rl)ldl
2π 2λ0 −q00
⎡ ⎤
(E)
(z, ) − μ0 λ−q C̃ (M)
(z, )
à G̃
−q0 −q0 z G̃
−q0 −q0 z
jω μ−q ∞ ⎢
⎢ μ−q λ0 ⎥
⎥ J0 (rl)ldl
G(−q0)(2) =− ⎣ ⎦
2π 0 2 λ 0 l 2
∞
jω μ−q μ0 λ−q C̃−q0 (M)
G(−q0)(3) = − G̃−q0 (z, z )J0 (rl)ldl
2π 0 μ−q λ0 2λ0
∞
(−q0)(4) jω μ−q μ0 C̃−q0 (M)
G =− G̃−q0 (z, z )J0 (rl)ldl
2π 0 μ−q 2λ0
∞
jω μ−q μ0 λ0 C̃−q0 (M)
G(−q0)(5) = − G̃−q0 (z, z )J0 (rl)ldl , (10.2)
2π 0 μ−q λ−q 2λ0
where r = [(x − x )2 + (y − y)2 ]1/2 , and the source point, z , is in region 0, and the
field point, z, is in region ±q.
180 10 Some Special Topics in Computational Electromagnetics
The G̃’s that appear in (10.2) are given by the product of terms that involve the
source parameters only and those that involve the field parameters:
⎡ ⎤
e −λ0 (z0 −z ) + R(E) e−λ0 (z0 +z −2z−1 )
(E) ⎦ e−λq (z−zq ) + R(E) eλq (z−zq )
G̃q0 (z, z ) = ⎣
0(−)
(E) (E) q(+)
1 − R0(+)R0(−) e−2λ0 Δ0
⎡ ⎤
(M)
(M)
e−λ0 (z0 −z ) + R0(−)e−λ0 (z0 +z −2z−1 )
G̃q0 (z, z ) = ⎣ ⎦ e−λq (z−zq ) + R(M) eλq (z−zq )
(M) (M) q(+)
1 − R0(+)R0(−) e−2λ0 Δ0
⎡ ⎤
e −λ0 (z0 −z ) − R(M) e−λ0 (z0 +z −2z−1 )
(M)
G̃q0 (z, z ) = ⎣
0(−) ⎦ e−λq (z−zq ) − R(M) eλq (z−zq )
(M) (M) q(+)
1 − R0(+)R0(−) e−2λ0 Δ0
⎡ ⎤
−λ0 (z −z−1 ) (E) −λ0 (2z0 −z −z−1 )
(E)
e + R e
G̃−q0 (z, z ) = ⎣ ⎦
0(+)
(E) (E) −2λ Δ
1 − R0(+)R0(−) e 0 0
(E)
eλ−q (z−z−(q+1) ) + R−q(−)e−λ−q (z−z−(q+1) )
⎡ ⎤
e −λ0 (z −z−1 ) + R(M) e−λ0 (2z0 −z −z−1 )
(M)
G̃−q0 (z, z ) = ⎣ ⎦
0(+)
(M) (M) −2λ Δ
1 − R0(+)R0(−) e 0 0
(M)
eλ−q (z−z−(q+1) ) + R−q(−)e−λ−q (z−z−(q+1) )
⎡ ⎤
e −λ0 (z −z−1 ) − R(M) e−λ0 (2z0 −z −z−1 )
(M)
G̃−q0 (z, z ) = ⎣ ⎦
0(+)
(M) (M)
1 − R0(+)R0(−) e−2λ0 Δ0
(M)
eλ−q (z−z−(q+1) ) − R−q(−)e−λ−q (z−z−(q+1) ) . (10.3)
Chap. 4, into (10.1), and then testing with the electric facet elements, (4.2) of
Chap. 4. The result is:
∑ ∑ ∑ Tklm,KLM
(mm)(xx)(q0) (x) (mm)(xy)(q0) (y) (mm)(xz)(q0) (z)
+ Tklm,KLM MKLM + Tklm,KLM MKLM + MKLM
KLM KLM KLM
∑ ∑ ∑ Tklm,KLM
(y) (me)(yx)(q0) (x)(e) (me)(yy)(q0) (y)(e) (me)(yz)(q0) (z)(e)
Bklm = Tklm,KLM JKLM + Tklm,KLM JKLM + JKLM
KLM KLM KLM
∑ ∑ ∑ Tklm,KLM
(mm)(yx)(q0) (x) (mm)(yy)(q0) (y) (mm)(yz)(q0) (z)
+ Tklm,KLM MKLM + Tklm,KLM MKLM + MKLM
KLM KLM KLM
∑ ∑ ∑ Tklm,KLM
(z) (me)(zx)(q0) (x)(e) (me)(zy)(q0) (y)(e) (me)(zz)(q0) (z)(e)
Bklm = Tklm,KLM JKLM + Tklm,KLM JKLM + JKLM
KLM KLM KLM
∑ ∑ ∑ Tklm,KLM
(mm)(zx)(q0) (x) (mm)(zy)(q0) (y) (mm)(zz)(q0) (z)
+ Tklm,KLM MKLM + Tklm,KLM MKLM + MKLM,
KLM KLM KLM
(10.4)
(q0) (q0)
where the transfer matrices satisfy Tklm,KLM = Tk−K,l−L;m,M ; i.e., they are Töplitz
(2D-convolution) in (X,Y ). Thus, one can still use two-dimensional FFTs to
efficiently execute the computations. The superscript (ee) on the transfer ma-
trices denotes “electric–electric,” which means that these matrices “transfer” an
electric current in region 0 into an electric field in region q. Similar interpreta-
tions attach to “electric–magnetic,” “magnetic–electric,” and “magnetic–magnetic”
transfer matrices; all transfer a source in region 0 into a field in region q. The other
transfer matrices can be deduced from the electric–electric ones by virtue of
Maxwell’s equations.
182 10 Some Special Topics in Computational Electromagnetics
For example, the magnetic-field moments shown on the left-hand side of the
bottom three equations in (10.4) can be easily computed in terms of the electric-
field moments:
(i)(x) (x)(m)
Bklm = B(i) (r) · Tklm (r)ax dr
1 (x)(m)
=− ∇ × E(i) (r) · Tklm (r)ax dr
jω
1 (x)(m)
=− E(i) (r) · ∇ × Tklm (r)ax dr
jω
!
1
=− E(i) (r) · π1k (x)π2l (y)π2m
(z)ay
jω
"
−π1k (x)π2l (y)π2m (z)az dr
(y) (y) (z) (z)
1 Eklm − Eklm+1 Eklm − Ekl+1m
=− −
jω δz δy
(z) (z) (x) (x)
(y) 1 Eklm − Ek+1lm Eklm − Eklm+1
Bklm =− −
jω δx δz
(x) (x) (y) (y)
(z) 1 Eklm − Ekl+1m Eklm − Ek+1lm
Bklm =− − . (10.5)
jω δy δx
The master discrete equations, (4.9) and (4.10), are augmented by the transfer
matrices when there is a second grid containing anomalous currents, yielding the
following equations for a coupled problem between grids q and 0:
⎡ ⎤ ⎡
E(i)(q) Q(ee)(qq) − G(ee)(qq) −G(em)(qq)
⎢ B(i)(q) ⎥ ⎢ G(me)(qq) Q(mm)(qq) + G(mm)(qq)
⎢ ⎥ ⎢
⎣ E(i)(0) ⎦ = ⎣ −T(ee)(0q) −T(em)(0q)
B(i)(0) −T (me)(0q)
−T(mm)(0q)
⎤ ⎡ (e)(q) ⎤
−T(ee)(q0) −T(em)(q0) J
−T(me)(q0) −T(mm)(q0) ⎥ ⎢ M(q) ⎥
⎥⎢ ⎥
Q(ee)(q0) − G(ee)(00) −G(em)(00) ⎦ ⎣ J(e)(0) ⎦. (10.6)
G(me)(00) Q(mm)(00) + G(mm)(00) M(0)
10.1 Spatial Decomposition Algorithm 183
Field Grid
Fig. 10.2 Showing a flaw that spans two regions and a grid in the probe region
Consider Fig. 10.2, which shows a flaw spanning two regions, together with its
“source grid,” and a “field grid” in the upper-region where the probe is located.
The objective is to calculate the scattered magnetic field in the probe region due to
the anomalous current, Ja , of the flaw, because this is the variable that a magnetic
sensor, such as a Hall-effect probe or a giant magnetoresistive (GMR) sensor, would
measure.
We assume that we have used (10.4) to compute the magnetic-field moments on
the “Field Grid” of Fig. 10.2, so now we want to compute the B-field, itself. To do
this, we expand the B-field in terms of the magnetic edge elements:
∑ bKLM TKLM
(x) (x)(m)
B(x) (r) = (r)
KLM
∑ bKLM TKLM
(y) (y)(m)
B(y) (r) = (r)
KLM
∑ bKLM TKLM
(z) (z)(m)
B(z) (r) = (r), (10.7)
KLM
= δ x ∑ Qlm,LM bkLM
(yz) (x)
LM
184 10 Some Special Topics in Computational Electromagnetics
(r)dr = δ y ∑ Qkm,KM bKlM
(y)(m) (xz) (y)
B(y) (r)Tklm
KM
= δ z ∑ Qkl,KL bKLm.
(z)(m) (xy) (z)
B(z) (r)Tklm (r)dr (10.8)
KL
LM KM
+ ∑ Qkl,KL bKLm+1 − bKLm ,
(xy) (z) (z)
(10.9)
KL
where the first equality comes from (10.5). Hence, the divergence condition is met
in the discrete sense involving moments.
2e-007
1e-007
-1e-007
-2e-007
30
25
20
15 y index
10
5
0 5 10 0
15 20 25 30
x index
2e-007
1e-007
-1e-007
-2e-007
0
5
10
15
x index 20
25
30 20 25 30
5 10 15
0
y index
Fig. 10.3 Plot of imaginary part of B(x) (top) and imaginary part of B(y) (bottom)
186 10 Some Special Topics in Computational Electromagnetics
4e-007
2e-007
0
30
25
20
15 y index
10
5
0 5 10 15 0
20 25 30
x index
Fig. 10.4 Plot of imaginary part of B(z) . The four dimples occur at the corners of the flaw
Then we have
(z)
Zfaraday = 100 × j2π × 100 × Bavg × π ((1 × 10−3)2 + (1.1 × 10−3)2 )/2
= (−1.329 × 10−8, 8.272 × 10−12),
5e-011
-5e-011
30
25
20
15 y index
10
5
0 5 10 0
15 20 25 30
x index
’sd18By.txt’ index 1 using 1:2:5
5e-011
-5e-011
0
5
10
15
x index 20
25
30 20 25 30
5 10 15
0
y index
2e-010
1.5e-010
1e-010
30
25
20
15 y index
10
5
0 5 10 15 0
20 25 30
x index
Fig. 10.5 Real part of Bx (top: preceding page), B(y) (bottom: preceding page), and B(z) (this page)
2e-008
-2e-008
25
20
15 y index
10
5
0 5 10 0
15 20 25 30
x index
Fig. 10.6 Plot of the “normalized divergence” expression on the left of (10.9). Top: real part;
bottom: imaginary part
10.1 Spatial Decomposition Algorithm 189
0.1
0.05
25
20
15 y index
10
5
0 5 10 0
15 20 25 30
x index
0.5
-0.5
-1
25
20
15 y index
10
5
0 5 10 0
15 20 25 30
x index
Fig. 10.7 A “pseudo expression” for the divergence in terms of the expansion coefficients, only
190 10 Some Special Topics in Computational Electromagnetics
where Δh A = A(h)− A(4h), Δ4h A = A(4h)− A(16h), and Δ16h A = A(16h)− A(64h).
10.2 A Theory of Error Extrapolation 191
Upon substituting this result into (10.11), we get an expression for the error, e(h),
which, when substituted into (10.10) gives the final result
1261 83 1
f (0) = A(h) + ΔhA − Δ4h A + Δ16hA
2835 2835 2835
Resistance (Ohms)
based on equation (10.14), 7 Data
are also shown. Top: 6
resistance; bottom: reactance 5
4
3
2
1
0
0 0.2 0.4 0.6 0.8 1
Probe position (inch)
Extrapolation Results
25
2x128x32
2x64x16
2x32x8
20 2x16x4
Extrapolated
Reactance (Ohms)
Data
15
10
-5
0 0.2 0.4 0.6 0.8 1
Probe position (inch)
the need to apply the magnetic-dipole model of nonstandard probes to the racetrack
transmit and receive coils, as well as model the nickel mesa anomalous region.
This can result in a problem with a large number of unknowns. To help alleviate
this problem, we use extrapolation techniques to generate increasingly accurate
solutions, while simultaneously conserving computer resources and time. We
present the results of some numerical experiments that demonstrate the application
of extrapolation techniques to this particular sensor problem.
Table 10.1 lists the convergence results for the racetrack T/R coils in freespace.
Note that the results converge nicely with Nz , for a fixed Nx × Ny . In each of the first
three cases, the change in going from Nz = 8 to Nz = 16 is 0.51 %, so this suggests
that Nz = 16 is sufficient to model this sensor.
As for the extrapolation analysis, if we use the three-point formula,
Resistance (Ohms)
results, based on the two 6
coarsest grids,
5
2 × 16 × 4 and 2 × 32 × 8, are
also shown. Top: resistance; 4
bottom: reactance 3
2
1
0
0 0.2 0.4 0.6 0.8 1
Probe position (inch)
Linear Extrapolation Results
25
Extrapolated
Data
20
Reactance (Ohms)
15
10
-5
0 0.2 0.4 0.6 0.8 1
Probe position (inch)
on the first three bottom entries (with Nz = 16), we get an “ideal” solution of
0.107308, which is higher by less than 0.2 % than the computed solution for
32×256×16. If we apply the four-point formula, (10.14), to the bottom four entries,
we get 0.108 for the new “ideal” solution. This result differs from 0.10711 by 0.8 %.
In order to determine the convergence profile for modeling the anomalous region
within the host material, we ran a number of different models, whose results are
shown in Table 10.2. The first four entries demonstrate convergence under the
fixed conditions of frequency and host conductivity, while varying the cell-density
count, Ny × Nz , keeping Nx , the density across the width of the mesa, fixed. If
we use the three-point extrapolation over the first three entries, we get an ideal
result of f (0) = (−8.0617 × 10−7 , 3.3687 × 10−6), which should be compared
to (−8.197 × 10−7, 3.3913 × 10−6 ), the value corresponding to the finest grid
with 8 × 64 × 32 cells. The error in the real part is 1.68 %, and in the imaginary
0.67 %. If we use the four-point formula over the first four entries, we get f (0) =
(−8.3576 × 10−7, 3.4295 × 10−6).
194 10 Some Special Topics in Computational Electromagnetics
2in
Receive Coil 0.062 in
Circuit Board
Transmit Coil
Air 0.25 in
1.5 in
Air
Nickel Mesa
1 in
Host σ=1.15E7 S/m
2 in
Air
1.625 in
0.1875 in
Fig. 10.10 A proximity sensor that utilizes racetrack transmit and receive coils
Table 10.2 Convergence profile (iterations/% error in residuals) for iterative solver
Cells No. unknowns Frequency σhost Profile Solution
8×8×4 768 1 MHz 10 1389/1.00 % (−6.0793E−7,2.8248E−6)
8 × 16 × 8 3072 1 MHz 10 1643/1.0 0 % (−7.0964E−7,3.0987E−6)
8 × 32 × 16 12288 1 MHz 10 2016/1.00 % (−7.791E−7,3.2928E−6)
8 × 64 × 32 49152 1 MHz 10 3756/1.00 % (−8.197E−7,3.3913E−6)
8 × 64 × 32 49152 100 kHz 10 500/8.62 % (−1.6599E−6,−1.9207E−6)
8 × 64 × 32 49152 100 kHz 100 500/5.96 % (−1.7655E−6,−1.9889E−6)
8 × 64 × 32 49152 1 MHz 10 500/6.41 % (−7.1739E−7,3.0734E−6)
8 × 64 × 32 49152 1 MHz 100 500/4.36 % (−3.7532E−7,3.1366E−6)
10.2 A Theory of Error Extrapolation 195
TOP VIEW X
μ, σ
μ, σ
FRONT VIEW
Table 10.3 Results, using a cubic error term, for grids of increasing
refinement in a nonconducting ferrite core
Grid(Nx × Ny × Nz ) Z Grid(Nx × Ny × Nz ) Z
4×4×8 −j1.5078 4 × 4 × 16 −j1.7702
8×8×8 −j1.6049 8 × 8 × 16 −j1.9335
16 × 16 × 8 −j0.92719 16 × 16 × 16 −j0.98303
32 × 32 × 8 −j1.1119 32 × 32 × 16 −j1.0733
Extrapolated Value −j1.2139 Extrapolated Value −j1.1413
The last four entries in the table indicate convergence profiles as a function of
either frequency or host conductivity. Clearly, for a given frequency, convergence
is better at the higher host conductivity, whereas for a given host conductivity,
convergence is better at the higher frequency.
A ferrite core probe is shown in Fig. 10.11. Consider the ferrite core problem with
σ = 0, μ = 4300, ε = 1, and grids of increasing refinement, with a cubic error term
as shown in Table 10.3. Since the core is nonconducting, the impedances are purely
imaginary. The extrapolated values differ by 6.4 %, which is reasonable for this
computation. The value for the 32×32×8 grid differs by 9.2 % from its extrapolated
value, whereas that for 32 × 32 × 16 differs by 6.3 %, which makes sense since the
finer mesh should produce a better estimate.
196 10 Some Special Topics in Computational Electromagnetics
X
-1.4
-1.6
-1.8
-2
0 10 20 30 40 50 60 70
Relative Fineness of Grid
It is clear that the complex shape of the probe produces an oscillating result for
the impedance as the grid is refined (see Fig. 10.12), contrary to what we saw in
Fig. 10.8 for a regular slot, for which the convergence was monotonic. In the case of
the slot, the finer mesh produces a more accurate estimate of the anomalous currents,
while doing nothing to improve the accuracy of the representation of the geometry of
the anomaly. In the case of the probe, however, we need a refined grid to accurately
represent both the anomaly and the anomalous currents.
It may be argued that the data for the four coarsest grids in Table 10.3 unduly
skews the extrapolation, or, to put it another way, these data may be outside the range
of a cubic error analysis. This argument may justify a simple linear extrapolation
over the results for grids 32 × 32 × 16 and 16 × 16 × 8:
A(h) − A(8h)
f (0) = A(h) +
7
8A(h) − A(8h)
=
7
8(1.0733) − 0.92719
= −j
7
= − j1.0942. (10.16)
The result for the finer grid differs by less than 2 % from the linearly extrapolated
value. Clearly, there is a numerical advantage in using finer grids, but the disadvan-
tage is in computational time. In any case, this analysis gives us an estimate of the
discretization error when using any of these grids.
These results hold for computations in freespace and suggest that using a
relatively coarse grid in X,Y may not be serious from a numerical viewpoint. When
the probe is placed over a host that includes a flaw, it is important to have enough
10.2 A Theory of Error Extrapolation 197
Offset Coil
Y
14mm
12.6mm X
Al: σ=1.92 E7
cells in the transverse plane to produce an accurate incident field on the flaw, since
it is this field that will not only couple the probe to the host, but will produce the
flaw “signature” that will be used in an inversion process.
Figure 10.13 illustrates the configuration for the tilted notch benchmark problem
with a 30◦ EDM notch. Data are taken at 7 kHz, which, with a conductivity of
1.92 × 107 S/m for the host, yields a skin depth of 1.38 mm. This implies that the
problem is quite large electrically, given the dimensions and extent of the EDM
notch anomaly in Fig. 10.13. Thus, we will apply extrapolation theory with cell
sizes of 32 × 32 × 32, 32 × 32 × 64, 64 × 32 × 64, and 64 × 32 × 128. The model
results, using the previous notation, are, therefore, A(h), A(2h), A(4h), and A(8h),
and the extrapolation formula, using a cubic error term, becomes
64A(h) − 56A(2h) + 14A(4h) − 1A(8h)
f (0) = . (10.17)
21
We tabulate in Table 10.4 the results for |Z| at 7 kHz for scans over the 30◦ notch
with these four grids.
These results are plotted in Figs. 10.14 and 10.15. The convergence is monotonic
from below, in contrast to the previous results shown in Fig. 10.12, confirming,
once again, that the geometry is well defined with a rather coarse grid and that the
refinement of the grid goes into creating a better approximation of the anomalous
currents. The asymmetry in the extrapolated model response is clearly evident, with
198 10 Some Special Topics in Computational Electromagnetics
Table 10.4 Extrapolated results for |Z| at 7 kHz for the 30◦ notch using four grids
Grid Size
Scan Pt 32 × 32 × 32 32 × 32 × 64 64 × 32 × 64 64 × 32 × 128 Extrapolated Data
−28 0.000315 0.000408 0.00055 0.00069 0.000893 0.013287
−21 0.002506 0.003369 0.004699 0.006004 0.007894 0.016114
−14 0.021642 0.030513 0.044342 0.057872 0.077437 0.069345
−7 0.070115 0.105 0.16076 0.21561 0.29506 0.312066
0 0.092742 0.14237 0.22237 0.30166 0.41685 0.426732
7 0.069253 0.11037 0.17197 0.23366 0.32380 0.316074
14 0.023636 0.035744 0.053971 0.0719 0.09790 0.080692
21 0.002819 0.004053 0.005868 0.007689 0.01035 0.008785
28 0.000344 0.000472 0.000658 0.000845 0.001119 0.008045
0.2
0.15
0.1
0.05
0
-30 -20 -10 0 10 20 30
Scan Point (mm)
the larger signal occurring for positive values of x (the slot is located at x = 0),
where the slot is tilted towards the surface. The symmetry is less noticeable in the
actual data, probably because of a slightly non-flat surface. At lower frequencies,
this effect probably will be less of a problem.
Figure 10.16 illustrates the measured and extrapolated-model responses in the
impedance plane of the measured data for the 30◦ notch. The enclosed areas are
roughly equal, which is a measure of the asymmetry associated with the 30◦ tilt.
We continue with the problem defined in Fig. 10.13, except that the 5-mm long slot
is tilted 10◦ from the vertical. We use two grids, 16 × 32 × 32 and 32 × 32 × 64, and
extrapolate with a linear error term over these two grids. The formula is
10.2 A Theory of Error Extrapolation 199
Z
0.2
0.15
0.1
0.05
0
-30 -20 -10 0 10 20 30
Scan Point (mm)
Fig. 10.16 The measured Measured and model data for 30 degree notch
and extrapolated-model 0.45
responses in the impedance
0.4
plane for the 30◦ notch
0.35
0.3
Measured
Model
0.25
X
0.2
0.15
0.1
0.05
-0.05
-0.02 0 0.02 0.04 0.06 0.08 0.1 0.12 0.14 0.16 0.18 0.2
R
4A(h) − A(4h)
f (0) = , (10.18)
3
and the corresponding data are shown in Table 10.5. It is clear, when comparing
the extrapolated data of Table 10.4 with the corresponding data of Table 10.5, that
the asymmetry associated with the 30◦ slot is considerably greater than that with the
10◦ slot, as we would expect intuitively.
200 10 Some Special Topics in Computational Electromagnetics
Table 10.7 Extrapolated results for |Z| at 7 kHz for the 20◦ notch using four grids and algorithm
(7.5)
Grid Size
Scan Pt 8 × 32 × 16 16 × 32 × 32 32 × 32 × 64 64 × 32 × 128 Extrapolated Value
−28 0.000157 0.000263 0.000614 0.000845 0.000938
−21 0.001046 0.002053 0.005365 0.007518 0.008379
−14 0.006663 0.017137 0.051309 0.073263 0.082031
−7 0.013207 0.052769 0.18542 0.27134 0.305687
0 0.013525 0.068195 0.25435 0.3763 0.425
7 0.013263 0.054705 0.19471 0.28571 0.322102
14 0.007017 0.018769 0.058905 0.084589 0.094842
21 0.001132 0.002282 0.006216 0.008869 0.009934
28 0.000167 0.000286 0.000692 0.000968 0.001079
We’ll use the same linear extrapolation formula, (10.18), as for the 10◦ notch, but
with grids of 64 × 32 × 32 and 64 × 32 × 128 cells. Table 10.6 gives the results.
Table 10.4 showed the results for a cubic interpolation using four rather fine grids.
If we use the cubic interpolation algorithm of (7.5), we can get by with coarser grids,
as shown in Table 10.7 for the 20◦ notch.
10.2 A Theory of Error Extrapolation 201
0.25
Z
0.2
0.15
0.1
0.05
0
-30 -20 -10 0 10 20 30
Scan Point (mm)
25
(Asymmetry Ratio -1) (%)
20
15
10
0
0 5 10 15 20 25 30
Tilt Angle (degrees)
Figure 10.17 compares the |Z| response to EDM notches at 0◦ , 10◦ , 20◦ , and 30◦
tilts. It is difficult to eyeball these results and draw conclusions as to the slant of each
notch, because each notch has the same volume. If, however, we refer to the raw data
in Tables 10.4–10.6, and consider the “asymmetry ratio” |Z(14)|/|Z(−14)|, then the
results tabulated in Table 10.8 give a clear estimate of the tilt. Indeed, it is a virtually
linear relation, with a slightly increasing slope with angle (see Fig. 10.18).
202 10 Some Special Topics in Computational Electromagnetics
0.35
0.3
0
10
0.25 20
30
X
0.2
0.15
0.1
0.05
0
0 0.02 0.04 0.06 0.08 0.1 0.12 0.14 0.16 0.18 0.2
R
Fig. 10.19 Illustrating the increase in area with tilt angle enclosed by the impedance-plane plots
of each of the notches
The impedance-plane plots of the responses for the 0◦ , 10◦ , 20◦ , and 30◦ notches
are shown in Fig. 10.19. As we would expect, the area enclosed by each curve
increases with the tilt angle, indicating a greater asymmetry. Note, further, that the
phase angle of the peak response decreases with tilt angle.
In this section, we will address the question of how to estimate the required grid
resolution to accurately solve a model problem, or, to put it another way, how
do we determine a priori how many cells will be required to solve the problem
accurately. We can always do the recommended thing, and that is to test several grid
resolutions, chosen in a systematic manner, and then check the rate of convergence
of the solution. This process can be followed by an extrapolation, as we have shown
above. Nevertheless, we want to gain insight into the requirements without resorting
at the outset to this empirical approach.
Ja (r)
E0 (r) = − G(r, r ) · Ja (r ), (10.19)
σa (r)
E0 (r) is the incident electric field, which is known, as is the anomalous conductivity,
σa (r) (for a forward problem). G(r, r ) is the dyadic Green’s function, and Ja (r) is
the unknown anomalous current.
The gridding question can be answered by analyzing the spatial-frequency
content of Ja (r). This requires taking the spatial Fourier transform of (10.20), which
involves the product of two spatially-varying functions, σa (r), and the total electric
field within the brackets. Once we get an estimate of the highest spatial frequencies
that must be present in Ja (r), we can estimate the size of the cells that are required
to produce these frequencies.
We are going to apply Fourier analysis to study the question, using the notion of
spatial frequency in the same manner that temporal frequency is used to answer
resolution questions in system theory. Equation (10.20) shows that we need to
determine the spatial Fourier transform of the product of two functions. This is given
by the convolution in the spatial-frequency domain of the individual transforms, as
we’ll now prove.
)
1 jk ·r
e− jk·r f (r)g(r)dr = e− jk·r f (r) e G̃(k
)dk
dr
(2π )3
)
1 − j(k−k )·r
= G̃(k ) e f (r)dr dk
(2π )3
1
= G̃(k )F̃(k − k )dk. (10.21)
(2π )3
Consider, for example, that g(r) is uniform in space, with a value of g. Then G̃(k) =
(2π )3 gδ (k), so that the right-hand side of (10.21) is equal to gF̃(k), as expected.
To simplify our lives a bit, we’ll ignore the integral term in (10.20), and just work
with the approximation, ja (r) ≈ E0 (r)σa (r), in determining the frequency content
of ja (r):
1 ,0 (k )Σ̃a (k − k )dk.
J̃a (k) ≈ E (10.22)
(2π )3
204 10 Some Special Topics in Computational Electromagnetics
σ=1.92E7S/m
0.28mm
4.7463mm
σ=0
X
1.9906mm
We will call Σ̃a (k) the spectrum of the conductivity scene, Ẽ0 (k) the spectrum
of the electromagnetic scene, and the convolution in (10.22) the composite spectral
scene. As a simple example of what to expect, let the conductivity and electromag-
netic spectral scenes in (10.22) each be a “baseband” unit pulse extending from 0
to Kmax , then the composite spectrum will be a unit tent function extending from
0 to 2Kmax . Clearly, the spatial-frequency content has been doubled for J̃a , when
compared to the original spectra.
The VIC-3D R
-generated bounding box for the 20◦ notch problem of Sect. 10.2.2.2
is shown in Fig. 10.20. We’ll start by computing the spectrum of the conductivity
scene, Σ̃a (kx , kz ). Since our “computational universe” is bounded by the rectangle
X0 = 1.9906, Z0 = 4.7463, we will use a Fourier series representation with discrete
spatial frequencies for all spectra. For the conductivity spectrum we write
X0 Z0
1
Σ̃a (kx , kz ) = σa (x, z)e− j2π (kx x+kz z) dxdz (10.23)
X0 Z0 0 0
10.3 Determining Grid Resolution Requirements 205
The number 0.153 is the ratio of the area of the notch in the (X, Z) plane to the total
area of the bounding box.
We complete the calculation of the iterated integral by substituting (10.25) into
(10.23):
X0
sin(0.153π nz) 1
Σ̃a (kx , kz ) = −2.94 × 106e− j2π nz e j2π nz (0.502x) e− j2π kx x dx
0.153π nz X0 0
sin(0.153π nz) 1 X0 − j2π x(kx −0.502nz )
= −2.94 × 106e− j2π nz e dx
0.153π nz X0 0
− j2π (kx −0.502nz )X0 − 1
6 − j2π nz sin(0.153π nz) e
= −2.94 × 10 e
0.153π nz − j2π (nx − nz )
sin(0.153 π n ) e − j2π (nx −nz ) − 1
= −2.94 × 106e− j2π nz
z
0.153π nz − j2π (nx − nz )
206 10 Some Special Topics in Computational Electromagnetics
= 0, if nx = nz
sin(0.153π nx)
= −2.94 × 106e− j2π nx , if nx = nz. (10.26)
0.153π nx
The fact that nx = nz means, from (10.24), that the sinusoidal expansion functions
for σa (x, z) have precisely the same spatial frequencies in the x and z directions.
This means that in order to reconstruct the conductivity, we must sample equally
finely in the x and z directions. (This holds only for characterizing the conductivity
scene.) To get an idea of what the sampling requirements are, consider Table 10.9.
At a spatial frequency of 100, the magnitude of the spectrum in (10.26) is reduced
to slightly less than 2 %. Hence, if we are to reconstruct σa (x, z) to this precision we
will need 200 samples in each direction to avoid aliasing. For a flaw with four times
the area, however, we can get by with one-tenth the number of samples for the same
precision.
Now, we’ll turn our attention to the electromagnetic scene for the 20◦ notch problem.
Let the coil lie in region 1, above the workpiece, and let region 2 be the workpiece,
which we assume is a homogeneous half-space. The appropriate axisymmetric
Green’s functions are [27]:
∞
1 Re−α0 (z+z ) +e−α0 |z−z |
G11 (r, z; r , z ) = J1 (r l)J1 (rl)ldl, (r , z ) ∈ 1, (r, z) ∈ 1
2π 0 2α0
∞
1 T
G21 (r, z; r , z ) = e(α1 z−α0 z ) J1 (r l)J1 (rl)ldl, (r , z ) ∈ 1, (r, z) ∈ 2, (10.27)
2π 0 2α0
where
α0 − α1 2α0
R= , T= (10.28)
α0 + α1 α0 + α1
10.3 Determining Grid Resolution Requirements 207
Consider
Nz
Nz max
Then
Ẽ2 (l, kz ) ≈ .
Z0 2α0 α0 l2 2π nz
(l 2 + jω μ0 σ )1/2 + j
Z0
(10.32)
2π
a frequency of 7 kHz, we have (ω f μ0 σ )1/2 = 1030, and = 1319, which yields
Z0
Nzmax = 39. At this frequency, the bounding box is about 3.5 skin-depths deep, so
we can argue that we need about 10 cells per skin depth just to account for the
electromagnetic scene. We’ll come back to this.
We can estimate the extent of the bounding box in (kx , ky , kz )-space by simply
2π Nxmax 2π Nymax
postulating that = lmax , = lmax , where, as before, X0 and Y0 are
X0 Y0
the lengths of the physical bounding box in the x and y directions, respectively. This
X0 lmax Y0 lmax
gives the plausible results that Nxmax = , Nymax = , from which we also
2π 2π
Nxmax Nymax
conclude that the cells should satisfy a “similarity relationship”: = .
X0 Y0
Now we can put everything together, and say that, as far as gridding requirements
for Ja (r) are concerned, this can be determined by convolving the spectrum of the
conductivity scene with the spectrum of the EM scene, or J̃a (k) = Σ̃a (k) ⊗ Ẽ2 (l, kz ).
From our previous comments about the extension of the frequency response that
comes from convolutions, we can estimate the composite frequency response to be
(Nxmax
c + Nxmax
e , Nymax
c + Nymax
e , Nzmax
c + Nzmax
e ), where the superscript “c” denotes the
conductivity scene, and “e” denotes the electromagnetic scene. In other words, each
scene contributes to the number of cells required for Ja (r).
The gridding requirement for an accurate impedance calculations requires an
extra convolution, as the following argument shows. We have
δZ = Ja (r) · E0 (r)dr
≈ σa (r)E20 (r)dr, (10.34)
where we are using the same approximation as in (10.22). Thus, in order to get
an accurate estimation of the integrand in (10.34), we should sample at the rate of
(Nxmax
c + 2Nxmax
e , Nymax
c + 2Nymax
e , Nzmax
c + 2Nzmax
e ), which will determine the number
of cells in the grid.
Let σ (r) be the conductivity when the flaw is present. Outside the background
region, σ (r) is equal to the host conductivity. Inside the background region, but
outside the flaw, σ (r) = σb (r). Inside the flaw, σ (r) is not equal to σb (r), but varies
with position.
First, consider the unflawed background region. The anomalous current for this
problem satisfies
Next, consider the flawed background region and define anomalous currents
From (10.35), (10.36), and (10.39) we obtain the uncoupled integral equations
Jb (r)
Ein (r) = − E(r) [Jb ] (10.40)
σb (r) − σh
J f (r) ! "
Ein (r) = − E(r) J f (10.41)
σ (r) − σb (r)
Jint (r)
Eef (r) = − E(r) Jint , (10.42)
σ (r) − σh
where the effective incident field, Eef (r), is given by
σb (r) − σh ! " σ (r) − σb (r)
Eef (r) = E Jf + E [Jb ]. (10.43)
σ (r) − σh σ (r) − σh
We construct Jint (r) from two currents, Jintf (r) and Jb (r), defined on the flaw
int
f (r)
Jint b (r)
Jint
Eef (r) = + − E(r) Jint − E(r) Jint
b . (10.44)
σ (r) − σh σ (r) − σh f
b (r) and J f (r) in terms of tent functions that are defined on the
We will expand Jint int
Q f b . G f b (Gb f ) is the transfer matrix that multiplies the expansion coefficients for
b (r) (J f (r)) to give the moments of the fields that they produce, evaluated on the
Jint int
where
⎡ ⎤
V x,b f 0 0
Vbf = ⎣ 0 V y,b f 0 ⎦. (10.51)
0 0 V z,b f
V b f is exactly analogous to V f f , except that the ratios are averaged over the
background cells instead of the flaw cells.
In evaluating the contribution of Jb (r) to Eef,b , we approximate the tent functions
that are defined on the background grid in terms of the tent functions that are defined
on the flaw grid. Then we have
The sparse matrix, Ab f , holds the expansion coefficients and has the structure
10.5 Notch at a Bolt Hole Benchmark Problems 213
⎡ ⎤
Ax,b f 0 0
Ab f = ⎣ 0 Ay,b f 0 ⎦. (10.53)
0 0 Az,b f
After computing Eef,b and Eef, f , we solve the two coupled equations given in
(10.45) and (10.46) using an iterative procedure. First, we rewrite these equations in
the following form:
We first solve (10.55), assuming that Jint( f ) = 0. Then we use the solution, Jint(b) , to
compute the right-hand side of (10.56), and solve this equation for Jint( f ) . Then we
use this solution to update the right-hand side of (10.55) and re-solve. We continue
this process until the norm of the residual vector
is sufficiently small.
The algorithm for eliminating the background is:
1. Solve (10.40) for Jb using a coarse background grid, Gb
2. Solve (10.41) for J f using a fine grid, G f , covering only the flaw
3. Solve (10.42) for Jint using (10.55) and (10.56)
and then compute the change in the probe impedance, due to the flaw, as
where the dot products are the usual expressions for impedances. An attractive
feature of this system is that the flaw and background may be gridded separately
in a manner appropriate to their size and characteristic.
The advantage of using dual integral equations to compute the change in impedance
due to a small flaw in the presence of a larger background region is to avoid the more
straightforward, but less accurate, method of simply computing Z with and without
the flaw and subtracting the two nearly equal values to obtain the small difference.
214 10 Some Special Topics in Computational Electromagnetics
Coil H1
25 mm 8 mm
Aluminum Alloy 4 mm 3 mm
Fig. 10.24 Benchmark Problem 2: notch in a bolt hole through a 4 mm thick aluminum plate.
The coil for the test, H1, is shown roughly to scale
1400
1200
1000
Ohms
800
R
X
600
400
200
0
100 1000 10000 100000 1e+06
Freq
The response of C1 in air is shown in Fig. 10.25 and of H1 in air in Fig. 10.26.
Though it can’t be inferred from Fig. 10.25 because of the scale of the ordinate,
the slope of the reactance curve at low frequencies yields an equivalent inductance
of 0.463 μH for C1, and the same analysis of Fig. 10.26 yields a low-frequency
inductance of 0.10052 H for H1, as shown in Table 10.10. The low-frequency
resistance values for each coil are also given by the data plotted in Figs. 10.25
and 10.26 and are also shown in Table 10.10. Clearly, the small inductance of C1
has pushed its resonant frequency beyond the limits shown in Fig. 10.25, whereas
the large inductance of H1 produces a relatively low-frequency resonance point.
Consider the typical equivalent circuit for a coil in freespace shown in Fig. 10.27.
Because there is only a single input port for the model, we refer to this as a
“one-port circuit model.” By applying the usual algorithm, we can determine the
216 10 Some Special Topics in Computational Electromagnetics
200000
150000
100000
R
X
Ohms 50000
-50000
-100000
-150000
100 1000 10000 100000 1e+06
Freq
Yp = G + j B
1
Zair =
1
Yp +
R0 + jω L0
1
=
1
G p + jB p +
R0 + jω L0
R0 + G p (R20 + ω 2 L20 )
=
(1 + R0G p − ω L0 B p )2 + (ω L0 G p + R0B p )2
10.5 Notch at a Bolt Hole Benchmark Problems 217
b Rair(ω) + jXclutter(ω)
Zwkpc δZ
ω L0 (1 − ω L0B p ) − R20 B p
+j
(1 + R0G p − ω L0 B p )2 + (ω L0 G p + R0B p )2
= Rair (ω ) + jXair (ω )
= Rair (ω ) + j[Xair (ω ) − ω L0 ] + jω L0
= Rair (ω ) + jXclutter (ω ) + jω L0, (10.59)
where we have singled out jω L0 for purposes of introducing the coupled circuit
representation in the next step.
When the coil is coupled to the workpiece, we have the situation depicted in
Fig. 10.28. The top part of the figure is the standard coupled-circuit representation,
whereas the bottom part is the VIC-3D R
representation in which δ Z is “air
balanced.” Thus, in order to get the circuit representation to agree with VIC-
3D R
we must subtract Rair (ω ) + jXclutter (ω ) from the measured impedance, Zwkpc .
We could attach another shunt element, Yq , to the driving-point terminals of
Fig. 10.28 to account for loading effects on the probe that are not accounted for
by inductive coupling. For example, it is possible that the presence of the metallic
host may cause a capacitive coupling that is not accounted for by Yp in the freespace
model of the probe. If that is the case, we can determine Yq by repeating the same
process that was used to determine Yp , but this time with a known δ Z in Fig. 10.28.
Figure 10.29 shows Yp for coils C1 and H1. The slope of a line joining the origin
to the extreme point for B( f ) in the top of Fig. 10.29 yields an equivalent capacitance
of 13.85 pF for C1, which is a reasonable approximation for frequencies greater
than 150 kHz, or so. The slope of B( f ) in the bottom of Fig. 10.29 for 0 ≤ f ≤
320 kHz yields an equivalent capacitance of 72.8 pF. These are the values listed in
Table 10.10.
218 10 Some Special Topics in Computational Electromagnetics
4e-05
3e-05
2e-05
G
Siemens
B
1e-05
-1e-05
-2e-05
-3e-05
0 50 100 150 200 250 300 350 400 450 500
Freq(kHz)
Yp for Coil H1
0.00025
0.0002
0.00015
Siemens
0.0001
G
B
5e-05
-5e-05
0 50 100 150 200 250 300 350 400 450 500
Freq(kHz)
10.5 Notch at a Bolt Hole Benchmark Problems 219
2 NLSE, a nonlinear least-squares estimator, will be defined and described in Chap. 12. The material
in this section can be deferred until after reading that chapter, or can simply be accepted on faith,
“in anticipation.”
220 10 Some Special Topics in Computational Electromagnetics
dR/f (Ohm/Hz)
over the entire frequency
0.0004
range
0.0003
0.0002
0.0001
0
-0.0006
-0.0008
-0.001
-0.0012
-0.0014
-0.0016
-0.0018
1000 10000 100000
Frequency (Hz)
When we use the coil data of Table 10.10, and the results of data set M1 of
Benchmark 1 in Table 10.11 to compute the forward model, and then plot the
results over the entire frequency range of 500 Hz to 500 kHz, we obtain Fig. 10.30.
(The data are normalized with respect to frequency.) That there is a breakdown in
the measured resistance data at low frequencies (dR goes negative!) was anticipated
because of the considerable uncertainty in Yp for coil C1 at low frequencies in
Fig. 10.29. The corresponding results for M1 of Benchmark 2 in Table 10.12
are shown in Fig. 10.31. The results for the other data sets in Tables 10.11 and
10.12 lie on top of the results shown in Figs. 10.30 and 10.31. The high-frequency
breakdown in the resistance and reactance in Fig. 10.31 was anticipated because of
the uncertainty in Yp for H1 at high frequencies in Fig. 10.29. Note, however, that
the normalized model reactance, dX/ f , due to NLSE in Fig. 10.31 “saturates” at
high frequencies, as expected by classical coupled-circuit theory and Förster plots.
The asymptote is a measure of the coupling of the coil to the host. The asymptotic
value of dR/ f = 0 is reached at a higher frequency in Fig. 10.31, as is also predicted
by coupled-circuit theory.
10.5 Notch at a Bolt Hole Benchmark Problems 221
dR/f (Ohm/Hz)
over the entire frequency
0.04
range
0.02
0
-0.02
-0.04
-0.06
1000 10000 100000
Frequency (Hz)
-0.1
-0.2
-0.3
-0.4
-0.5
-0.6
1000 10000 100000
Frequency (Hz)
The final results of our validation studies for the multiscale algorithm are plotted
in Fig. 10.32 for Benchmark Problem 2 and Fig. 10.33 for Benchmark Problem 1.
Numerical experiments show that accurate results for Benchmark Problem 2 require
a background grid that is much finer near the notch. We can obtain results that
effectively use a graduated grid which becomes progressively finer in regions closer
to the notch. It is clear, however, from varying the number of notch cells that the
impedances we are trying to compute have not converged with respect to number of
cells along the x, y, or z directions for a grid of 32 × 2 × 16 cells.
It is also clear from varying the number of background (bolt-hole) cells that the
impedances have not converged with respect to number of cells along the x, y, or z
directions for a grid of 16× 16 × 4 cells. We can, however, improve these impedance
values by combining them with values obtained from an 8 × 8 × 2 cell graduated
grid for the bolt hole and a 16 × 2 × 8 cell notch grid. This allows us to perform a
linear extrapolation to zero cell dimensions. That is the result labeled “Extrapol’d”
in Fig. 10.32.
222 10 Some Special Topics in Computational Electromagnetics
dR (Ohm)
hole in Benchmark -2
Extrapol’d
Problem 2. The scan is along -3
x1.21
Y =0 -4
-5
-6
-7
-30 -20 -10 0 10 20 30
Probe x position (mm)
6 x1.12
dR (Ohm)
hole in Benchmark -0.1
Problem 1. The scan is along Expt
Y =0 -0.2
Extrapol’d
-0.3
x1.39
-0.4
-0.5
-30 -20 -10 0 10 20 30
Probe x position (mm)
0.8 Expt
Extrapol’d
dX (Ohm)
0.6
x1.17
0.4
0.2
0
-30 -20 -10 0 10 20 30
Probe x position (mm)
224 10 Some Special Topics in Computational Electromagnetics
Now that we have the electromagnetics down pat, let’s look ahead to see where that
leads us. In this part of the book, a general inverse method process is introduced that
provides a framework for incorporating the numerical forward models discussed in
Part I in an algorithm that estimates an unknown set of values of the test. A diagram
of a general inverse method process is presented in Fig. 11.1. The heart of the
inversion process is the inverse method step. Typical parameters to be evaluated
include crack dimensions (e.g., surface length or depth), material properties (e.g.,
conductivity) or state(s) of the nondestructive evaluation system (e.g., probe lift-
off or tilt.) Here, parameter values are estimated using an iterative scheme such as
nonlinear least-squares estimation (NLSE) which equates the model results with
experimental data through adjusting the model parameter values. Such general
inverse methods will be presented in detail in Chap. 12. To ensure accurate
estimation results, statistical estimators can be used to evaluate convergence to
the solution and quantify the error bounds on the results. Noise and uncertainty
analysis is first introduced in Sect. 11.2.2, with a more complete study developed in
Sects. 11.6 and 11.7. The Cramer–Rao Lower Bound, another stochastic metric,
is introduced in Sect. A.1 of Chap. 12. The theory and use of robust statistical
estimation metrics will be presented in Chap. 13.
Although inverse methods have been studied for many years, to some degree
there has been limited progress to transition them to practical applications. There
are three primary reasons for the difficulty in transitioning inverse methods to
NDE applications in particular: long solution times for accurate NDE measure-
ment models, the inherent ill-posedness of certain inverse problems in NDE, and
the lack of robustness of the inverse method schemes to the presence of noise
and uncontrolled variation with in-field NDE measurements. To address these
challenges, a comprehensive approach is proposed in Fig. 11.1 for model-based
inversion design to ensure the reliability of the inverse method step through (a)
key parameter assessment and model parameterization, (b) model reduction through
surrogate models, (c) model calibration, (d) noise removal, (e) data registration, and
(f) feature extraction techniques.
First, the inverse method development process must always begin with a
full assessment of the key factors controlling the experimental data and driving
inversion performance. All critical factors for the NDT technique, part material,
part geometry, and discontinuity characteristics that control signal and noise must
be identified and well understood. Clearly, the flaw characteristics such as the crack
dimensions are the key factors of interest in the forward and inverse models. This
step provides a critical opportunity to reduce the complexity of the inverse problem
by parameterizing the flaw morphology, potentially addressing some levels of ill-
posedness by simplifying the inverse problems. It is important to also consider all
varying conditions in the NDE measurement that affect the measurement response,
for example lift-off or scan resolution. If such key factors cannot be controlled
during measurements, these parameters should also be incorporated in the inverse
problem with the characteristics of the flaw. An excellent case study supporting this
point is the inversion of inner diameter (ID) pits presented in Sect. 18.3.
Although the volume integral method introduced in Part I is quite efficient
numerically for the simulation of eddy current NDE measurements, for inverse
problem applications it can still be time consuming to perform such repeated
numerical calculations within an inversion scheme. Surrogate models or structures
can be created from the results of the numerical model to greatly improve the
solution time and performance of inverse methods. Frequently, data or look-up
tables are constructed and interpolation schemes are applied in order to provide
a fast means of sampling the model over a wide range of conditions. The use of
surrogate structures, first mentioned in Sects. 1.4 and 6.6, is formally introduced
in Chap. 12. Emerging stochastic simulation methods, such as the probabilistic
11.1 Overview of the General Inversion Process for Parameter Estimation 229
a
INPUT SYSTEM OUTPUT
[KNOWN] [KNOWN] [TO BE DETERMINED]
b
SYSTEM OUTPUT
INPUT
[KNOWN] [KNOWN]
[TO BE DETERMINED]
c
INPUT SYSTEM OUTPUT
[KNOWN] [TO BE DETERMINED] [KNOWN]
Fig. 11.2 System representation of direct and inverse problems: (a) the direct problem; (b) the
signal-detection (communication) problem; (c) the inverse problem
inverse problem will be ill-posed, while reducing calculation time and improving the
inverse solution performance. Several examples demonstrating the important of data
registration and feature extraction steps are included in the remainder of the book.
Nondestructive evaluation (NDE) is to materials and structures what computer-
aided tomography (CAT) scanning is to the human body—an attempt to look inside
without opening the body. It is in nature an inverse problem. Figure 11.2 illustrates
a system representation for three important problems: (a) a direct problem, in
which the input and system are known and the output is to be determined; (b) a
signal-detection (communication) problem, in which the system (a communication
channel) and output are known, and the problem is to determine the input signal;
and (c) the inverse problem, in which the input and outputs are known, and we must
determine the system.
The results that we showed in Part I of this text were direct problems; we
assumed knowledge of the probe and flaw and determined the response of the
probe, namely the driving-point or transfer impedances. The second problem is
dealt with in communication and information theory texts, such as [50], and has
a close relation to inverse problems. In the typical inverse problem of eddy-current
NDE, we assume that impedances (or instrument voltages) are measured, given the
probe and measurement conditions, and the intent is to determine the anomaly that
produced the output. By this we mean to quantify the size, shape, and location of a
flaw, or other departure from an assumed-known condition.
The results of Part I have given us an engine, VIC-3D R
, that is designed
to solve the forward problem, and, by doing so, will provide the engine for
11.2 Thickness Measurements with Eddy-Current Probes 231
solving the inverse problem. In this chapter, we will demonstrate the notion of
eddy-current inverse problems by solving three basic problems: (a) a thickness
measurement, (b) determining a conductivity profile, and (c) a combination of the
two in the semiconductor chip industry. These problems are simple examples of the
application of eddy-current NDE to process control and in-service inspection, and
they also give us our first examples of model-based inversion, in which the physical
problem can be modeled with just a few unknown parameters. We will defer to the
next chapter a detailed discussion of the mathematical algorithms that will be used
in this chapter and throughout the remainder of the book.
The problem is illustrated in Fig. 11.3 [52]. The thickness, t, of the plate is to be
determined from measurements with an air-core probe. The probe lift-off, l, is often
poorly known. This can be due to a nonconducting layer of unknown thickness on
the surface, as illustrated in Fig. 11.3b, or simply because the probe geometry is not
well known. Since the probe response can be very sensitive to this parameter, it is
desirable to have an inversion scheme that does not require the lift-off to be known.
We use a scheme that determines this parameter as well as the plate thickness.
The impedance of the probe is a complicated nonlinear function of several
variables:
Z = g(coil, workpiece, t, l, f ) , (11.1)
where “coil” denotes coil parameters and “workpiece” denotes workpiece
parameters. The plate thickness, probe lift-off and frequency are denoted by
t, l, and f , respectively. The problem is to invert (11.1), obtaining t and l when
Z, f , and the coil and plate parameters are known.
The left-hand side of (11.2) is taken to be the measured values of resistance and
reactance. An initial value, (t0 , l0 ), is chosen, and (11.2) is solved for (t, l). Iteration
then proceeds by replacing (t0 , l0 ) with (t, l), and re-solving (11.2) until t and l
converge, and
Rmeas ( f ) R̃( f ,t, l)
= . (11.4)
Xmeas ( f ) X̃( f ,t, l)
⎡ ⎤
∂R ∂R
( f1 ) ( f1 )
⎢ ∂t ∂l ⎥
⎢ ⎥
⎡ ⎤ ⎢ ⎥
ΔR( f1 ) ⎢ ∂X ∂ ⎥
⎢ ( f1 )
X
( f1 ) ⎥
⎢ ΔX( f1 ) ⎥ ⎢ ⎥
⎢ ⎥ ⎢ ∂t ∂l ⎥
⎢ .. ⎥ ⎢ . . ⎥ t − t0 Δt
⎢ . ⎥ =⎢ .. .. ⎥ ≡A
⎢ ⎥ ⎢ ⎥ l − l0 Δl
⎣ ΔR( fN ) ⎦ ⎢ ∂R ∂R ⎥
⎢ ( f ) ( f ) ⎥
ΔX( fN ) ⎢ ∂t N
∂l
N ⎥
⎢ ⎥
⎢ ⎥
⎣ ⎦
∂X ∂X
( fN ) ( fN )
∂t ∂l (t0 ,l0 )
Iteration proceeds as before, except that now we solve the overdetermined system by
finding the least-squares solution that minimizes the norm of the residuals (Gauss–
Newton):
Residual2 ≡ ∑ ΔR( fi )2 + ΔX( fi )2 . (11.6)
i
we choose frequencies in the range 20–40 kHz, where the variation with thickness
is largest. We note that if the primary interest had been in determining the lift-off,
then higher frequencies would be preferable. O. Baltzersen [55] gives an empirical
formula, attributed to Förster and Libby [56], for the optimum frequency:
700
fopt = , (11.10)
σ tAeff
where f is in kHz, σ is the conductivity of the plate in mega/ohm-m, t is the plate
thickness in mm, and Aeff is essentially the coil diameter in mm. In this case the
formula gives 29 kHz, in good agreement with our choice of frequencies.
Six frequencies are chosen to cover the range of 20–40 kHz. Table 11.1 shows
the thickness and lift-off that are obtained from our inversion procedure. The initial
guess for both the thickness and lift-off is 0.2 mm. Convergence is achieved in three
iterations.
The uncertainties in the thickness that is obtained from the inversion procedure
that we have described can be estimated from the uncertainties in the measured
probe impedances. For small changes in resistance and reactance, the change in the
deduced thickness is given by
N
∂t ∂t
Δt = ∑ ΔRi + ΔXi. (11.11)
i=1 ∂ R i ∂ Xi
If ΔRi and ΔXi have Gaussian distributions with half-width σRi and σXi , then Δt is
Gaussian with width
1/2
∂t 2 2 ∂t 2 2
σt = ∑ σRi + σXi . (11.12)
i ∂ Ri ∂ Xi
236 11 Examples of Basic Inverse Problems
In the simple case that σRi /Ri = σXi /Xi = σZ /Z, the half-width for the thickness is
given by
2 2 1/2
σt ∂ t Ri ∂ t Xi σZ
t
= ∑ ∂ Ri t
+
∂ Xi t Z
. (11.13)
i
If the errors in Test Case 1 are systematic, with ΔRi /Ri = ΔXi /Xi = ΔZ/Z < 0.05,
then the errors in thickness and lift-off, using six frequencies, are given by
11.2 Thickness Measurements with Eddy-Current Probes 237
Δt ΔZ
= −7.36
t Z
Δl ΔZ
= −1.85 .
l Z
If the analysis includes only the single frequency, 30271.2 Hz, then Δt/t is about the
same, while Δl/l is twice as large. Thus, if only one frequency is used in the analysis,
then the systematic errors and statistical errors make comparable contributions to the
thickness error, while the inclusion of six frequencies makes the statistical errors
much less significant. In the six-frequency analysis, changes in reactance at high
and low frequencies produce opposite effects on the deduced lift-off. In this case,
systematic errors can be cancelled by an appropriate choice of frequencies.
We now consider a second test case, for which we have experimental measurements.
The problem is illustrated in Fig. 11.8. The measurements are those of Burke
(1992, private communication) and cover the range 100–10,000 Hz. The computed
derivatives of impedance with respect to thickness and lift-off are shown in
Figs. 11.9 and 11.10. From these plots, the range 1–2 kHz is chosen for determining
the thickness. The thickness and lift-off values obtained from a number of analyses,
each using a different number of frequencies, are shown in Table 11.2.
In Fig. 11.11 the ratio of the deduced thickness and lift-off to the measured
values are plotted versus the number of frequencies used in the analysis. The
improvement observed in thickness determination with the addition of frequencies
to the analysis suggests that there exists a statistical component of the errors in the
measured impedance values. The remaining 1 % difference between the measured
and deduced thickness is likely due to systematic errors in the measurements or
calculations. The probe lift-off is not determined quite as well as the thickness. This
could be anticipated from Figs. 11.9 and 11.10, which show the impedance to be
less sensitive to the lift-off.
238 11 Examples of Basic Inverse Problems
11.2.4 Conclusions
Fig. 11.11 Test Case 2: thickness and lift-off versus the number of frequencies used
3e-09 8e+07
region 1 7e+07 True
Conductivity (S/m)
2.5e-09
region 2 6e+07 Initial
2e-09
region 3 5e+07 Final
1.5e-09 region 4 4e+07
1e-09 3e+07
2e+07
5e-10
1e+07
0 0
10000 100000 1e+06 0 0.1 0.2 0.3 0.4 0.5
Frequency (Hz) Depth (mm)
Fig. 11.12 A model problem with three layers over a half-space. Derivative of normalized
impedance with respect to conductivity, versus frequency (left), and conductivity profile. Region 1
lies between 0 and 0.1 mm; region 4 lies below 0.3 mm
Suppose a flat, layered structure has layers with the conductivities shown (solid)
in Fig. 11.12 [54]. We can reconstruct this profile using impedance measurements
from a simple air-core probe. To model this problem, we use VIC-3D R
to compute
the impedance seen by an air-core probe resting on our layered structure. Our probe
has 400 turns, is 1.73 mm in height, and has inner and outer radii of 2.54 and
4.27 mm. To determine the best frequencies for our probe, we look at the derivative
of the normalized impedance with respect to the conductivities of the four regions
of our structure. For all regions, this derivative peaks between 3 kHz and 3 MHz, as
shown in Fig. 11.12. We use this frequency range.
We use the same nonlinear least-squares algorithm (Gauss–Newton) that was
used in the thickness measurements to recover the conductivities from the
impedances. The starting point for our iterative algorithm is a uniform conductivity
11.3 Conductivity Profile Measurements 241
Impedance Noise
0.01
Resistance
0.008 Reactance
0.006
Fractional Noise 0.004
0.002
0
- 0.002
- 0.004
- 0.006
- 0.008
- 0.01
1000 10000 100000 1e+06 1e+07
Frequency (Hz)
Fig. 11.13 Perturbations in the computed impedances at 21 frequencies. This is the input data that
represents the effects of “noise”
of 4 × 107 S/m. Three to six frequencies, in a fairly narrow band, are sufficient
to exactly reconstruct the conductivities, if our computed “noiseless” impedances
are used. To investigate the effect of “noise” in the data, we perturb the com-
puted impedances, which are our input data, by the fractional amounts shown in
Fig. 11.13. This is not representative of any particular noise model, but simply
illustrates the effects that perturbations have on the reconstructions. Additional
frequencies are now required to produce a satisfactory inversion. The conductivities
obtained using 21 frequencies are shown (dashed) in Fig. 11.12. The frequencies
chosen are simply those that form a geometric progression between 3 kHz, and
3 MHz. Ten iterations were required.
We can better understand the quality of the reconstruction we can expect from
our measurements by referring to the derivatives shown in Fig. 11.12. Clearly, at
low frequencies, the impedance is much more sensitive to the conductivity of the
half-space, while at high frequencies it is much more sensitive to that of the top
layer. Hence, our recovered conductivity profile is most accurate in these regions.
Our modeling allows us to test our inversion algorithm, determine the proper
frequencies to measure, establish noise requirements for our measurements, and
understand the quality of reconstruction we can expect.
We believe that this is an important point to stress. Eddy-currents have long been
applied to the measurement of thickness and conductivity, but the techniques have
generally been quite empirical, and used discrete standards. The results of blindly
using the nearest available instrument have often been disappointing. Modeling must
be performed before equipment is purchased, or expensive experiments undertaken.
242 11 Examples of Basic Inverse Problems
σ = 800000 μ = 16 8 mils
σ = 730000 μ = 1
Fig. 11.14 A model configuration to study the problem of simultaneously reconstructing magnetic
permeabilities and electrical conductivities. For this study, the relative permeability of the top layer
varies from 1 to 16
1 Recall
that a Förster diagram is a plot of normalized reactance to normalized resistance, with the
normalizing factor being the freespace reactance of the probe.
11.4 Conductivity and Permeability Profile Measurements 243
Fig. 11.16 Förster plots of the unperturbed configuration of Fig. 11.14 (circles) and the config-
uration in which the conductivity of the top layer is slightly perturbed (triangles). The frequency
range is the same as in Fig. 11.15
Fig. 11.17 Förster plots of the unperturbed configuration of Fig. 11.14 (circles) and the configu-
ration in which the permeability of the top layer is slightly perturbed from its nominal value of two
(triangles). The frequency range is the same as in Fig. 11.15
σ = 800000 μ = 16 8 mils
σ = 730000 μ=1
Fig. 11.18 A two-layer ferromagnetic system, whose conductivities and permeabilities are to be
reconstructed using multifrequencies. The frequency range extends from 50 kHz to 1 MHz in 19
equal steps
σ = 744700 μ = 0.9968
Fig. 11.19 Reconstruction of the system shown in Fig. 11.18, when 20 frequencies are used from
50 kHz to 1 MHz. The impedance data are noiseless
This problem, which is shown in Fig. 11.18, addresses the issue of reconstructing
magnetic permeabilities, as well as electrical conductivities.
The impedance data are taken over the frequency range of 50 kHz to 1 MHz,
in 19 equal intervals. These data, without the addition of noise, are used to
reconstruct both permeability and conductivity, with the results shown in Fig. 11.19.
The starting points for the iterations were σL = σS = 750, 000, μL = μS = 2, where
the subscript “L” denotes the layer and “S” denotes the substrate.
The geometry of the problem is illustrated in Fig. 11.20 (TENCOR, Inc., 1996,
private communication):
246 11 Examples of Basic Inverse Problems
t0 Silicon t3
Statement No. 1
Given:
(a) t0 , the total thickness of the package
(b) σAl = 3.57 × 107 S/m, the conductivity of aluminum
(c) σEpi = 10 S/m, t2 = 10 × 10−6 m
Determine:
(a) t1 , the thickness of the aluminum layer
(b) σSi , the conductivity of the silicon substrate
Statement No. 2
Given:
(a) t3 = 550 × 10−6 m, the thickness of the silicon substrate
(b) σAl = 3.57 × 107, the conductivity of aluminum
(c) σEpi = 10 S/m, t2 = 10 × 10−6 m
Determine:
(a) t1 , the thickness of the aluminum layer
(b) σSi , the conductivity of the silicon substrate
The only difference between the two statements lies in the information that is
given about the thickness. Mathematically, the two problems are identical, because
one piece of thickness information is given in each of the statements.
11.5 A Problem in the Semiconductor Chip Industry 247
We will follow the approach described in Sect. 11.2, but now there are two
approaches that we can take to gather the data. As before, we can use multifrequency
excitation of a coil that is placed at a known lift-off above the aluminum layer
and invert the impedance data at these frequencies to determine t1 and σSi , jointly.
The second possibility is to vary the lift-off of the coil that is excited at a fixed
frequency, measure the impedance at each lift-off value, and invert these data
to determine t1 and σSi , jointly. The latter is a realistic possibility, because the
manufacturing and measurement processes for semiconductor chips are performed
in clean rooms, with well-defined conditions and instruments. The lift-off can be
controlled and measured accurately by ultrasonic means.
We’ll briefly examine the multifrequency case and discuss more thoroughly
numerical experiments with the multi lift-off case. It is convenient for analysis
purposes to normalize impedances with respect to the free-space reactance of
the coil, when data are taken over a broad frequency range. Figure 11.21 shows
derivatives of the normalized resistance and reactance with respect to t1 and σSi ,
as a function of frequency over the range 106 –109 Hz. The nominal values of the
unknowns are t1 = 560 Å = 5.6 × 10−5 mm and σSi = 2 × 104 S/m, with a (known)
lift-off of 0.5 mm.
The results indicate that care must be taken in choosing an operating frequency.
The thin aluminum layer shields the probe coil from the silicon substrate at
frequencies greater than 100 MHz, which is why we see a significant decrease in
the conductivity sensitivity coefficients (the derivatives with respect to σSi ) over
this frequency range in the bottom part of Fig. 11.21.
We’ll turn now to the second possibility, in which lift-off is varied over 16
values, starting at 0.0762 mm and ending at 1.01614 mm in equal logarithmic
increments. We use three different frequencies, 10, 25 and 100 MHz. The sensitivity
coefficients (derivatives) for each of these frequencies are shown in Figs. 11.22–
11.24. It is interesting to note that dR/d σSi , as a function of lift-off, is positive for
10 MHz and “pivots” around the zero-value at a lift-off of 1 mm and is negative
at 100 MHz. It appears that at a frequency somewhat higher than 25 MHz, this
derivative is actually zero for all lift-off values, which suggests that at this frequency
the resistance is insensitive to changes in conductivity.
The input data for the multi lift-off inversion problem are obtained by scanning
a coil of inner radius 1 mm, outer radius 1.2 mm, height 0.2 mm, with 18 turns
above the system shown in Fig. 11.20 at 16 lift-off values, and then recording the
change in impedance from freespace conditions. The freespace resistance is zero, of
course, and the freespace reactance is 73.0313 Ω at 10 MHz, 182.578 Ω at 25 MHz,
and 730.314 Ω at 100 MHz, as computed by VIC-3D R
. The model data are shown
in Table 11.4.
248 11 Examples of Basic Inverse Problems
50
d(Z/Xfree)/dt (1/mm)
0
-50
-100
-150
-200
-250
1e+06 1e+07 1e+08 1e+09
Frequency (Hz)
5e-07
d(Z/Xfree)/ds (m/S)
-5e-07
-1e-06
-1.5e-06
1e+06 1e+07 1e+08 1e+09
Frequency (Hz)
The matrix, A, of (11.5) is called the Jacobian of the system, and the derivatives
shown in Figs. 11.22–11.24 form the columns of A:
A = [t, σ ], (11.15)
11.6 Noise Analysis: The Covariance Matrix 249
d(Z/Xfree)/dt (1/mm)
200
100
-100
-200
0.01 0.1 1 10
Lift-Off (mm)
1.5e-06
1e-06
d(Z/Xfree)/ds (m/S)
5e-07
-5e-07
-1e-06
-1.5e-06
-2e-06
-2.5e-06
0.01 0.1 1 10
Lift-Off (mm)
200
d(Z/Xfree)/dt (1/mm)
100
-100
-200
-300
-400
-500
-600
0.01 0.1 1 10
Lift-Off (mm)
0
d(Z/Xfree)/ds (m/S)
-1e-06
-2e-06
-3e-06
-4e-06
-5e-06
0.01 0.1 1 10
Lift-Off (mm)
t·t t·σ
F= . (11.16)
σ ·t σ ·σ
d(Z/Xfree)/dt (1/mm)
-200
-400
-600
-800
-1000
0.01 0.1 1 10
Lift-Off (mm)
-1e-06
-1.5e-06
-2e-06
-2.5e-06
-3e-06
0.01 0.1 1 10
Lift-Off (mm)
C = F −1 σmeas
2
. (11.17)
where σmeas is the variance in the measurement noise. There should be no confusion
between σ implying a variance and σ representing electrical conductivity. From
(11.16) and (11.17) we get
252 11 Examples of Basic Inverse Problems
C11 C12 1 σ · σ −t · σ
= × σmeas
2
. (11.18)
C21 C22 (t · t)(σ · σ ) − (t · σ )(σ · t) −σ · t t · t
and the correlation between the two parameters, which lies between 0 and 1, is the
negative cosine of the angle between the two column-vectors of A:
C12 −t · σ
cor(t1 , σSi ) = √ = = − cos(θ ) (11.20)
C11 × C22 (t · t) × (σ · σ )
Notice that the variances are smallest when the correlation is small, in which case
the variance of each parameter is roughly proportional to the inverse of the derivative
of the impedance with respect to that parameter.
Consider, for example, the conditions at the nominal values of the parameters:
t1 = 5.6 × 10−5 mm and σSi = 2 × 104 S/m. From Figs. 11.22–11.24 we compute
the covariance matrix at f = 10, 25, and 100 MHz to be:
0.1392 × 10−4(mm)2 −0.1716 × 104(S/m)(mm)
C[10] = × σmeas
2
−0.1716 × 104(mm)(S/m) 0.2280 × 1012(S/m)2
11.6 Noise Analysis: The Covariance Matrix 253
0.1629 × 10−5(mm)2 −0.2185 × 103(S/m)(mm)
C[25] = × σmeas
2
−0.2185 × 103(mm)(S/m) 0.3764 × 1011(S/m)2
0.3028 × 10−6(mm)2 −0.6073 × 102(S/m)(mm)
C[100] = × σmeas
2
−0.6073 × 102(mm)(S/m) 0.3019 × 1011(S/m)2
(11.21)
Table 11.5 Results of inversions with four different input noise sets at 10 MHz
Noise set t1 (mm) σSi (S/m) C
0.1403×10−4 −0.1727×104
1 5.72505 × 10−5 19794.9
−0.1727×104 0.2292×1012
0.1423×10−4 −0.1751×104
2 5.99395 × 10−5 19543.5
−0.1751×104 0.2320×1012
0.1412×10−4 −0.1739×104
3 5.80033 × 10−5 19671.8
−0.1739×104 0.2305×1012
0.1360×10−4 −0.1678×104
4 5.22920 × 10−5 20441.1
−0.1678×104 0.2236×1012
and
for 100 MHz. It is clear that there is significant improvement in noise reduction at
the higher frequencies. Furthermore, t1 and σSi become less correlated at higher
frequencies, which leads to a more reliable inversion.
(11.20)–(11.22), we obtain the results shown in Table 11.6. The closeness of the
values in the last two columns of Table 11.6 indicates that the Monte Carlo process
was reasonably well represented with these four data sets.
Appendix
AT Ax = AT b. (11.26)
x = Db, (11.27)
Thus
m m
Δxi Δx j = ∑ ∑ Dik D jl Δbk Δbl . (11.30)
k=1 l=1
Assuming that the errors Δbk are distributed so as to give cancellation of the terms
with k = l (e.g., they have a normal distribution), the expected values of Δxi Δx j are
then given by
m
E(Δxi Δx j ) = ∑ Dik D jk E(Δbk Δbk )
k=1
m
= ∑ Dik D jk σb2k
k=1
m
= ∑ [DS]ik [DS]Tkj
k=1
= Ci j , (11.31)
where S is a diagonal matrix with Skk = σbk , the standard deviation for bk , and C is
the covariance matrix
When S = σmeas I, i.e., the standard deviation for all measured values is the same,
then
and the correlation between xi and x j , which lies between 0 and 1, is given by
Ci j
cor(xi , x j ) = . (11.35)
CiiC j j
Chapter 12
NLSE: Parameter-Based Inversion Algorithm
12.1 Introduction
Let
Z = g(p1 , . . . , pN , f ) , (12.1)
where p1 , . . . , pN are the N parameters of interest, and f is a control parameter at
which the impedance, Z, is measured. f can be frequency, scan-position, lift-off,
etc. It is, of course, known; it is not one of the parameters to be determined. To be
explicit during our initial discussion of the theory, we will call f “frequency.”
where the superscript (q) denotes the qth iteration and the partial derivatives are
computed numerically by the software. The left side of (12.3) is taken to be the
measured values of resistance and reactance. We rewrite (12.3) as
0 ≈ r+Jp , (12.4)
The left-hand side of (12.6) can be approximated to the first order in σ by the first-
order Taylor expansion:
⎡ ⎤T
∂ f1 ∂ f2M
f1+ · · · + f2M
⎢ ∂ x1 ∂ x1 ⎥
1 ⎢ ⎢ .. ⎥
⎥
=
r(x) ⎢
⎣ ∂f
. ⎥
⎦
1 ∂ f2M
f1 + · · · + f2M
∂ xN ∂ xN
⎡
∂ f1 ∂ f1 ⎤
···
⎢ ∂ x1 ∂ xN ⎥
T
r(x) ⎢ ⎢ ⎥
= .. ⎥
r(x) ⎢
⎣∂f
. ⎥
2M ∂f ⎦ 2M
···
∂ x1 ∂ xN
= eT (x) · J , (12.8)
where the superscript T denotes the transpose of a matrix (or vector) and e(x) =
r(x)/r(x) is a unit vector. Thus, (12.7) becomes
The factor multiplying σ in (12.9) is the dot product of the two vectors, e(x) and
J · v. Hence, its value is less than the product of the magnitude of each vector, which
means that (12.9) becomes
r(x∗ + σ v) − r(x∗) ≤ σ J · v , (12.10)
because e(x) has unit magnitude. Upon equating the right-hand side of (12.10) to
the right-hand side of (12.6), we obtain the first-order estimate of σ :
r(x∗ )
σv = ε . (12.11)
J(x∗ ) · v
260 12 NLSE: Parameter-Based Inversion Algorithm
r(xi) − r(x*i )
r(x*i )
xi
−σI −σS x*i σS σI
Fig. 12.1 Showing sensitivity parameters for two system responses to xi . Response S is sensitive
to xi at x∗i , whereas response I is not
Note that if J(x∗ ) · v is small compared to r(x∗ ), then σ is large and the residual
norm is insensitive to changes in the linear combination of the parameters specified
by v. If v = ei , the ith column of the N × N identity matrix, then (12.11) produces
σi , the sensitivity bound for the ith parameter. Since σi will vary in size with the
magnitude of x∗i , it is better to compare the ratios σi /x∗i for i = 1, · · · , N before
drawing conclusions about the fitness of a solution.
The importance of these results is that we now have metrics for the inversion
process: Φ = r(x∗ ), the norm of the residual vector at the solution, tells us how
good the fit is between the model data and measured data. The smaller this number
the better, of course, but the “smallness” depends upon the experimental setup and
the accuracy of the model to fit the experiment. Heuristic judgement based on
experience will help in determining the quality of the solution for a given Φ .
The sensitivity coefficient, σ , is more subtle, but just as important. It, too, should
be small, but, again, the quality of the “smallness” will be determined by heuristics
based upon the problem. If σ is large in some sense, it suggests that the solution
is relatively independent of that parameter, so that we cannot reasonably accept the
value assigned to that parameter as being meaningful, as suggested in Fig. 12.1,
which shows a system, S, for which the system is sensitive to variable, xi , at the
solution point, x∗i , and another system, I, for which the system is insensitive to xi .
An example occurs when one uses a high-frequency excitation, with its attendant
small skin depth, to interrogate a deep-seated flaw. The flaw will be relatively
invisible to the probe at this frequency, and whatever value is given for its parameters
will be highly suspect. When this occurs we will either choose a new parameter to
characterize the flaw, or acquire data at a lower frequency.
12.2 NLSE: Nonlinear Least-Squares Parameter Estimation 261
These metrics are not available to us in the current inspection method, in which
analog instruments acquire data that are then interpreted by humans using hardware
standards. The opportunity to use these metrics is a significant advantage to the
model-based inversion paradigm that we propose in this book.
If the residual norm is relatively insensitive to changes in some linear com-
bination of the parameters, then the Jacobian matrix at the solution is nearly
rank-deficient, and it may be useful to determine a set of linearly-independent
parameters. The covariance matrix (J T J)−1 can be used for this purpose.
The problem defined above leads to the global optimization problem: finding the
lowest minimizer of a nonlinear function of several variables that has multiple local
minimizers. In the stochastic approach to global optimization, one applies a strictly
descent local search procedure to a subset of a sample of starting points drawn from
a uniform distribution over R, so as to find all the local minima of r that are
potentially global [58–60]. One such stochastic approach, the multilevel, single-
linkage, will, with probability one, find all relevant local minima of the objective
function with the smallest possible number of local searches [59, 61]. We do not
implement the multilevel, single-linkage approach in this book, but use a uniform
distribution of starting points (in each coordinate direction), as in the multilevel
approach. We do not reduce the sample size, however, but use the entire sample,
comprising, perhaps, 500 points in each variable. We have found that, even with four
variables, the procedure is so fast with modern machines, that it is quite efficient.
Hence, our global optimization algorithm starts by generating a uniform dis-
tribution of 500 points in each coordinate, and then immediately applying the
Gauss–Newton iteration to each of these points. The result is 500 local minima,
which are then ordered to give the smallest to largest values of the norm of the
residual. The location of the smallest of these local minima is presumed to be the
global minimum.
PZ,p = PZ (Z)P(p|Z)
= Pp (p)P(Z|p)
= Pp,Z (p, Z) , (12.12)
where PZ (Z) and Pp (p) are called a priori probability density functions, and
P(p|Z), P(Z|p) are called conditional, or a posteriori, probability density func-
tions. The variable to the right of the vertical bar in these latter functions is
called the conditioning variable; for example, P(p|Z) is called the a posteriori
probability density for p, conditioned on the fact that Z was received (or measured).
This a posteriori probability density function is the object of our interest, as we
shall now see.
From (12.12) we want to maximize
PZ (Z)P(p|Z) = Pp (p)P(Z|p) (12.13)
That is, we want to choose that “message,” p, that is best associated with the
“received signal,” Z.
If the measured Z of (12.2) is corrupted by “noise,” of density Pn , then we
can replace the conditional density P(Z|p) by Pn (Z1 − g1 , Z2 − g2 , . . . , ZM −
gM |p), where g1 = g(p, f1 ),. . . , gM = g(p, fM ) in (12.2). Because the noise in the
measurements is independent of the transmitted message, p, we can ignore the
conditioning variable, and replace (12.14) by
P(p|Z) ∝ Pp (p)Pn (Z − g(p)) . (12.15)
If we have no prior knowledge of p, or if all transmitted messages are a priori
equally likely, then we can ignore Pp (p) in (12.15) and work with the “likelihood
function,” PN (Z − g(p)). Maximizing the likelihood function over p is called
“maximum likelihood estimation.” We usually have some prior knowledge of p,
however, so we incorporate that knowledge in the a priori function, Pp (p).
A.1 Cramer–Rao Lower Bound 263
If p and n are jointly Gaussian processes, then we have the classical problem of
communicating a Gaussian signal in Gaussian noise, which reduces to a classical
least-squares problem. Typically, we work with the negative logarithm of the a
posteriori density in (12.15), so we need to minimize
− ln P(p|Z) = − ln Pp (p) − ln Pn (Z − g(p)) + f (Z) . (12.16)
Now, we let the M components of the noise vector, n, be statistically independent,
zero-mean, Gaussian random variables, each with variance, σn2 , and the N com-
ponents of p be statistically independent Gaussian random variables, with mean
values p, and each with variance, σ p2 . Then, upon taking the negative logarithm of
the appropriate density functions, and discarding unimportant factors, we replace
(12.16) with the objective function
1 M 1 N
Φ (p|Z) = ∑
2σn2 i=1
(Zi − g i (p)) 2
+ ∑ (pi − pi )2
2σ p2 i=1
|Z − g(p)|2 |p − p|2
= + , (12.17)
2σn2 2σ p2
which is to be minimized over p. Multiply (12.17) by 2σn2 to get the final expression
for the objective function
σn2
Φ (p|Z) = |Z − g(p)|2 + |p − p|2 , (12.18)
σ p2
where we use the same notation for the objective function.
The ratio, σ p2 /σn2 , of the variances is called the signal-to-noise ratio and is
usually known. In the context of least-squares problems, this ratio is known as the
Levenberg–Marquardt parameter, and in mathematical inverse theory it is called the
Tichonov–Miller parameter.
Setting the value of the LM-parameter determines the certainty with which we
choose to assert the a priori constraint on p. If σ p is very small, then the LM-
parameter is large, and we are more certain to impose the constraint. Unless we
know the variances, we cannot determine the LM-parameter at the outset. There are
numerical techniques, such as ridge regression [53] and cross-validation [62], that
can be useful in selecting the LM-parameter.
Appendix
The Cramer–Rao lower bound (CRLB), being the lower bound on the variance of
any unbiased estimator, plays a role in statistical estimation theory [64, 65] that is
similar to the sensitivity (upper) bound, σ , of (12.11). As with σ , the CRLB finds
application in electromagnetic scattering and inverse problems [66].
264 12 NLSE: Parameter-Based Inversion Algorithm
Another tool used in numerical linear algebra for sensitivity analysis is singular
value decomposition (SVD). SVD essentially provides a measure of sensitivity
of measurements to perturbations in the unknown parameters [106]. To evaluate
the sensitivity of an inverse problem for a set of measurements to changes in fit
parameters, SVD can be applied to the Jacobian matrix where
⎡ ⎤
∂ Z1 ∂ Z1
···
⎢ ∂ θ1 ∂ θn ⎥
⎢ . ⎥
J=⎢
⎢ ..
.. .
. .. ⎥ = UΣ V .
⎥ (12.22)
⎣ ∂Z ∂ Zm ⎦
m
···
∂ θ1 ∂ θn
The condition number (CN) of the matrix is defined as the ratio of the largest and
smallest singular values resulting from SVD. For inversion, CN has been used to
quantify the well-posedness of the inverse problem for select parameters. The ability
to estimate parameters independently increases as the condition number approaches
unity. It should be noted that SVD does not incorporate noise; it depends only on the
noiseless relationship between the measurement output and the parameter changes.
Fig. 12.2 (a) Distribution of source data in impedance plane, and (b) corresponding estimated
values in lift-off-thickness parameter space
The calculations required for the CRLB involve taking numerical derivatives of
the impedance changes with respect to the parameter changes from the forward
model. These calculations thus require far less computational expense with respect
to Monte-Carlo simulation. Following (12.20), the Fisher information for this
particular case is given by:
2 + J2
J11 J12 J11 + J22J21
I= 21 . (12.24)
J11 J12 + J21J22 2 + J2
J12 22
The covariance matrix is then calculated from the Fisher information (by (12.19)):
C = σ 2 I−1 . (12.25)
The Jacobian is also decomposed into its singular values and singular vectors in the
form of the right-hand side of (12.22). The ratio of the smallest to largest singular
values provides the condition number.
Figure 12.3 shows the CRLB of the estimation of the thickness and lift-off of a
1 mm thick plate and 1 mm lift-off for multiple frequencies. The agreement between
the CRLB and the Monte-Carlo approach is quite good. This analysis demonstrates
that there is an optimal frequency to achieve highest accuracy in the estimation of
thickness. Estimating conductivity and thickness simultaneously is typically more
ill-conditioned than estimating thickness and lift-off simultaneously. The CRLB
for conductivity and thickness estimation along with the condition number and
correlation number as a function of frequency are all displayed in Fig. 12.4. The
behavior of the CRLB as a function of frequency for estimating conductivity and
thickness simultaneously follows a similar trend and this is expected since the
impedance changes due to conductivity and thickness are similar. The condition
number reaches a maximum around 95 kHz which implies that selectivity is good
and the correlation is zero at this frequency which further confirms that point.
A.1 Cramer–Rao Lower Bound 267
Fig. 12.3 Comparison of variance with varying frequency using CRLB and Monte Carlo methods
for estimating (a) lift-off and (b) thickness, respectively
Fig. 12.4 Comparison of inversion metrics with varying frequency: (a) CRLB variance for
thickness and conductivity estimation and (b) correlation and condition number
Figure 11.5 of Chap. 11 shows the variation with frequency of the derivative of
impedance with respect to thickness for Test Case 1 (shown in Fig. 11.4). There is a
relatively broad peak in this derivative extending from roughly 20 to 50 kHz. Thus,
we would expect that the optimum range of frequencies to be used for determining
the thickness should lie in this range and that was essentially confirmed when
frequencies in the range of 20–40 kHz were found to give good results.
Figure 12.5(top) illustrates the CRLB thickness response for this same test case,
and we see that there is a relatively broad minimum centered near 50 kHz, but
generally including the same frequency range chosen in executing the test case. This
confirms the consistency between the CRLB result and the information contained in
the first derivative.
268 12 NLSE: Parameter-Based Inversion Algorithm
Fig. 12.5 Comparison of variance with varying frequency for Test Case 1 in Chap. 11, using
CRLB and Monte Carlo methods for estimating (top) thickness and (bottom) lift-off, respectively
As for the lift-off problem in Test Case 1, we see from Fig. 11.6 that the derivative
with respect to lift-off continues to increase with frequency even beyond 100 kHz,
which corresponds to what we might expect for the CRLB associated with lift-off,
A.1 Cramer–Rao Lower Bound 269
Fig. 12.6 Comparison of variance with varying frequency for Test Case 2 in Chap. 11, using
CRLB and Monte Carlo methods for estimating (top) thickness and (bottom) lift-off, respectively
as shown in the bottom part of Fig. 12.5. In the latter figure, we see that the CRLB
never does achieve a true minimum, even out to 150 kHz, but the decrease is leveling
off at this frequency.
270 12 NLSE: Parameter-Based Inversion Algorithm
Similar results are obtained for Test Case 2 of Chap. 11, except that the
frequencies are much lower (see Figs. 11.9 and 11.10). The derivative with respect to
thickness for a 1.0 mm-thick brass plate has a maximum in the vicinity of 2 kHz, and
the derivative with respect to lift-off peaks at approximately 10 kHz, for a nominal
lift-off of 2.0 mm. These results are consistent with the CRLB results shown in
Fig. 12.6. In particular, note that the CRLB for lift-off is virtually flat beyond 8 kHz,
which corresponds with the levelling off of the peak derivative shown in Fig. 11.10.
13.1 Introduction
“Robust” estimators are resistant to outliers in data, contrary to the usual classical
least-squares estimator such as NLSE. We will describe two robust estimators in
this chapter and give an example of the application of them to pre-processing some
experimental data. Both of the robust estimators are taken from [63], and we closely
follow those authors in our presentation.
The mathematical regression problem that we are addressing is based on the linear
model given by yi = θ xi + ei for i = 1, . . . , N. yi are the data points corresponding to
the dependent variable, xi the data points corresponding to the independent variable,
ei the error term, and θ the parameter to be determined.
Central to the notion of robust estimation is the concept of the “breakdown point,”
which we define following [63]: Take any sample X of N data points (xi , yi ), and any
estimator T of the parameter θ . Let β (M, T, X) be the supremum of T (X )− T (X)
for all corrupted samples X , where any M of the original points of X are replaced
by arbitrary values. Then the breakdown point of T at X is defined as
M
εN∗ (T, X) = min ; β (M, T, X) is infinite . (13.1)
N
In words, εN∗ is the smallest fraction of contaminated data that can cause the
estimator to take on values arbitrarily far from T (X). For least-squares, only one
bad observation is needed to cause breakdown, so that εN∗ (T, X) = 1/N, which tends
to 0 % when the sample size N becomes large.
Definition 13.1. Let (x1 , y1 ), . . . , (xn , yn ) be a sample of regression data. For each
slope, θ , we obtain residuals ri (θ ) = yi − θ xi . Then the LMS-estimator θ̂ that is
defined by
1 n
∑ ρ (ri /s) = K ,
n i=1
(13.3)
Definition 13.3. Let (x1 , y1 ), . . . , (xn , yn ) be a sample of regression data. For each
slope, θ , we obtain residuals ri (θ ) = yi − θ xi , from which we can calculate the
dispersion s(ri (θ ), . . . , rn (θ )) by (13.3), where ρ satisfies (R1) and (R2). Then the
S-estimator θ̂ is defined by
x2 x4 x6
ρ (x) = − 2 + 4 for |x| ≤ c
2 2c 6c
(13.6)
c2
= for |x| ≥ c.
6
Estimator function
0.05
-0.05
-0.1
-0.15
-0.2
0 2 4 6 8 10 12
S (scale)
We sketch, in Fig. 13.1, the estimator function, 1n ∑ni=1 ρ (ri /s) − K, for a typical
data set.
Finally, let us say something about the asymptotic behavior of S-estimators at
the central Gaussian model, where (xi , yi ) are independent, identically distributed
random variables satisfying
yi = xi θ0 + ei . (13.9)
θ̂n → θ0 a.s.
σ̂n → σ0 a.s. (13.10)
13.5 An Application of Robust Estimation 277
n1/2 (σ̂n − σ0 ) → N 0, (ρ (y) − K)2d Φ (y)/( yψ (y)d Φ (y))2 . (13.11)
These theorems guarantee that, while the S-estimators do not break down easily
when the data are contaminated, they continue to behave well when the data are
not contaminated, that is, when they satisfy the classical assumptions [63]. The
Asymptotic Normality Theorem defines the asymptotic convergence rate that was
alluded to earlier.
Though our main interest in robust estimators is in their application to the voxel-
based inversion algorithm that will be treated in the second volume of this series,
we will here apply the LMS-estimator to the problem of determining a “baseline”
for experimental data that will be the input to the actual inversion algorithm.
The problem is to determine a constant (the “baseline”) from a set of data that
are virtually constant, except for isolated signals due to flaws, tube supports, and
removal of the probe from the end of the tube. Our interest is in the application to
eddy-current NDE of steam generator tubes in nuclear power plants. The isolated
signals constitute “impulsive noise” and may be considered to be the outliers of the
data set. Clearly, if we averaged the data, the resulting baseline would be strongly
affected by these signals, which would hardly result in a baseline. The current
approach in the industry is simply to “eyeball” the data to produce the baseline,
but this is hardly efficient, and is still subject to errors. What is needed is a reliable
mathematical approach that will automatically produce the answer with no human
intervention or preprocessing.
The baseline estimator was exercised on a set of experimental data, supplied
by the Electric Power Research Institute (EPRI). The data consisted of a set of
approximately 12,000 measurements, taken at regular intervals along the axis of
a tube, using eddy-current instrumentation. The measurements were made at four
different frequencies, and at each frequency, the “baseline” had a different value due
to the way the instrumentation behaved. One of the post-processing procedures for
278 13 Robust Statistical Estimators
these data required that the data at the different frequencies have the same baseline
(the data at two or more different frequencies were scaled and mixed to produce a
composite signal). We used the baseline estimator to “normalize” all frequencies to
the same baseline. The figures below show the experimental data, before and after
using the baseline estimator. The results are shown for the full scan, as well as for
smaller portions of the full scan (Figs. 13.2–13.4).
Fig. 13.2 Full scan of experimental data, at four different frequencies (approximately 12,000
points), before (top) and after (bottom) any “balancing” using the baseline estimator
13.5 An Application of Robust Estimation 279
Fig. 13.3 Portion of the full scan, data points 0–1,000, of experimental data, before (top) and after
(bottom) any “balancing” using the baseline estimator
280 13 Robust Statistical Estimators
Fig. 13.4 Portion of the full scan, data points 1,000–2,000, of experimental data, before (top) and
after (bottom) any “balancing” using the baseline estimator
Chapter 14
Some Special Signal-Processing Algorithms
In this chapter we sketch some rather elegant mathematical theorems that have had a
significant impact on computational aspects of electrical engineering. We have used
them over the years in performing eddy-current inversions and believe that they will
have an expanded role to play in the future development of eddy-current NDE.
The mathematical foundation of the theorems is the notion of projection onto convex
sets (POCS) in a Hilbert space, which is proving to be quite valuable in engineering
analysis and design [71–78]. We recall that a Hilbert space is simply a vector space
in which an inner-product is defined.1 Our entire development of volume-integral
equations is done in a Hilbert space, though we don’t emphasize that fact. A set, S ,
is called convex if for every pair of vectors, V and U that lie in S , the straight line
joining these two vectors, V + λ (U − V), 0 ≤ λ ≤ 1, also lies in S . Hyperplanes,
spheres, and cubes are simple examples of convex sets. The projection of a vector, V,
onto a convex set, S , is that point, U, on S that is closest to V. We write U = PS V,
where PS is the projection operator.
Projection operators offer a computationally convenient algorithm for deter-
mining the intersections of sets, which can be translated into solving systems of
linear and nonlinear equations. Consider Fig. 14.1, which shows two convex sets
intersecting at the single point, v∗ . If we define the composition operator2
T = P2 P1 , (14.1)
then the alternating projection scheme shown will converge to v∗ ,
for an arbitrary
starting point, v. Because of the tangency of the two boundaries, however, the
convergence will not be geometric in the vicinity of v∗ , so we use over-relaxation
1 The discussion in this section follows [73], which should be consulted for the details.
2 The order of the Pi ’s is not important.
v
C1
P1v
TvT2vT3v v*
C2
Fig. 14.1 Illustrating projections onto two convex sets, and the convergence of the iterates to v∗ ,
the intersection of the sets. The starting point, v, is arbitrary
by extending the projections beyond the boundaries of the sets. This is done by
selecting appropriate “relaxation” parameters, r1 , r2 within the interval 0 < r < 2,
replacing P1 and P2 by
T1 = 1 + r1(P1 − 1)
T2 = 1 + r2(P2 − 1) , (14.2)
and then defining
T = T2 T1 , (14.3)
before proceeding with the scheme shown in Fig. 14.1. We will give in the next
section an example of this over-relaxation procedure in the context of Kaczmarz’
algorithm for solving large systems of linear equations. Following that, we will give
examples of the use of POCS for signal processing and inversions with incomplete
or noisy data.
H3 H2 H1
(1)
(2) x
x
(0)
x
(3)
x
(5)
x (4)
(6)
x
x
Fig. 14.2 Geometric interpretation of ART for three equations and two cycles of iterations. Each
equation defines a hyperplane. At each stage, the current solution is projected onto the next
hyperplane. Results after two cycles of iterations with unity relaxations are shown
λ x(0) = AT z(0) ,
284 14 Some Special Signal-Processing Algorithms
step:
where
(k)
bi − λ zi − (ai , x(k) )
d (k) = rk ,
λ 2 + ai2
and ai is the ith row of A, ei is the ith row of the identity matrix I, i = (k mod m) + 1,
and 0 ≤ lim inf rk ≤ lim sup rk ≤ 2. d (k) is the term which represents a generalization
of projections onto hyperplanes, z(k) is a dual variable, and rk is a relaxation
parameter. For each iteration of the algorithm, 1 ≤ i ≤ m. The method converges
for a given range of LM parameters and a given range of relaxation parameters. It is
clear from the algorithm that the role of the LM parameter is to stabilize d (k) when
ai 2 is very small.
In voxel-based inversion algorithms (which will be the subject of volume 2 of
this series) we acquire data via an array of sensors, each of which can be excited to
serve as a transmitter as well as a receiver. This yields the system (14.5)
..
.
Theorem 14.2. If any analytic function in the (kx , ky )-plane is known exactly in
an arbitrarily small (nonzero) region of that plane, then the entire function can be
found (uniquely) by means of analytic continuation.
In order to use these theorems we recall that the support of g(x, y) is bounded.
Thus, by Theorem 14.1 G(kx , ky ) is analytic in the (kx , ky )-plane. If we have
only limited information about G, say only its values at low spatial-frequencies,
Theorem 14.2 tells us that we can uniquely extend G to the entire (kx , ky )-plane.
Once we have continued G to the entire (kx , ky )-plane, we can take the inverse
Fourier transform to recover g. The following algorithm is an application of POCS
that performs analytic continuation under the conditions just stated:
ALGORITHM
Let the function G(kx , ky ) be given over a prescribed region L , and let F be the
Fourier transform. Then starting with:
f0 (x, y) = F −1 [G(kx , ky )];
r = 0;
REPEAT
(1)
fr = P1 fr ;
(2) (1)
fr = P3 fr ;
(1) (2)
Fr+1 = F [ fr ];
(1)
Fr+1 = P2 Fr+1 ;
−1
fr+1 = F [Fr+1 ];
r = r + 1;
UNTIL CONVERGENCE OCCURS.
The important operations in the Algorithm, in addition to the Fourier and inverse
Fourier transforms, are the various projection operators, Pi :
f , (x, y) ∈ S , S = support of f ,
P1 f =
0, otherwise.
G(kx , ky ), (kx , ky ) ∈ L ,
P2 F =
F(kx , ky ), (kx , ky ) ∈ L , where F(kx , ky ) = F [ f (x, y)].
⎧
⎨ −1, if f (x, y) < −1
P3 f = f (x, y), if − 1 ≤ f (x, y) ≤ 0 (14.7)
⎩
0, if f (x, y) > 0.
From its definition, −1 ≤ g(x, y) ≤ 0, which is the reason for the definition of
P3 . Hence, the algorithm successively applies the known properties of the function,
g(x, y). The analyses of [72,73,78] prove that the algorithm converges. In Volume 2
we will apply this algorithm to a problem of flaw reconstruction in graphite-epoxy
advanced composites. Other applications of the algorithm can be found in [85–91].
14.4 Reconstructing Network Functions 287
R3
(PH2,PD2)3
R2
(PH2,PD2)2
R1
(PH2,PD2)1
Rmeas
X
Xmeas X3 X2 X1
Fig. 14.3 Illustrating an iterative method, based on the idea of projections onto convex sets
(POCS), for improving resistance data during the inversion process. It is likely that the points
(R1 , X1 ), (R2 , X2 ), (R3 , X3 ) lie on another convex set (or simply a locally-convex set) that
approaches X = Xmeas , but this has not been proved
In electric circuit theory a network function is the ratio of the Fourier or Laplace
transform of a response to that of an excitation, when the network is initially relaxed.
Typical network functions are driving-point and transfer impedances or admittances.
The inverse problems described in this text use impedances as input data, so it is
important to know that the quality of these impedance data is sufficient to allow
a good reconstruction. It is abundantly clear throughout this text that measured
impedance data usually have a reasonably good reactance, but the resistance is much
noisier, generally because it is much smaller. Even though the resistance may be
small compared to the reactance, it still plays a crucial role in determining the shape
and location of a flaw. Thus, it is important to develop a means of “reconstructing”
the resistance from the known reactance data in order to complete the inversion
process. The POCS concept, together with NLSE, allows us to do that, as illustrated
in Fig. 14.3.
The nomenclature in Fig. 14.3 refers to a corrosion topology problem that will
be described in Chaps. 16 and 17. PH2 and PD2 refer, respectively, to the height
and diameter of a “pillbox” model of a pit in layer 2 of a double-layer system.
The input data for this problem are the impedances obtained from a 245-point
scan of a probe past the pit. Each point on the X-axis, therefore, corresponds to
a 245-vector of reactance data, with Xmeas corresponding to the original measured
288 14 Some Special Signal-Processing Algorithms
data. A similar interpretation holds for the R-axis and Rmeas . The composite (X, R)
space can be called a state-vector space, or simply a function space. The hyperplane
X = Xmeas is a convex set in this space (any two points in the hyperplane can be
joined by a straight line lying within the plane), and the arrowed trajectories intersect
it orthogonally, so that part of the algorithm is a “projection onto a convex set,” but
we have not demonstrated that the other trajectories are projections onto a convex
set. Nevertheless, because of the similarity between this figure and Fig. 14.1, we
will refer to the iterative algorithm of Fig. 14.3 as the “POCS” algorithm later in the
book. If we must distinguish between Figs. 14.1 and 14.3 to avoid confusion, we
will do so at the appropriate time.
The algorithm proceeds iteratively through a series of steps that sequentially
calls NLSE to perform an inversion, VIC-3D R
to perform a direct calculation, and
an orthogonal projection to return the impedance computed by VIC-3D R
back to
the straight line, X = Xmeas , thus assuring that all further input data to NLSE will
always have X = Xmeas as its reactance data. This ensures that the algorithm will be
grounded in the more reliable component of the original measured data.
Starting with the original measured data, (Rmeas , Xmeas ), NLSE0 computes the
first approximation to the pit dimensions, (PH2, PD2)1 , which is then fed to
VIC-3D R
to produce the first update to the impedance data, (R1 , X1 ). X1 , however,
is projected (orthogonally) back onto the hyperplane, X = Xmeas , so that the new
impedance data for NLSE1 are (R1 , Xmeas ). The output of NLSE1 is (PH2, PD2)2 ,
which is then fed into VIC-3D R
to produce (R2 , X2 ), and then (R2 , Xmeas ) after
projection. The process is continued until some stopping point. We have chosen
the stopping point to be when the β -scale factor in β × Zk+1 − Zk = 0 is unity,
or very close. This criterion states that there is not much difference (in the least-
squares sense) between Zk+1 = (Rk+1 , Xk+1 ) and Zk = (Rk , Xk ), so there is no sense
in running another NLSE step. A by-product of this algorithm for reconstructing the
measured resistance is the reconstruction of the pit, which is what we wanted in the
first place, and for that reason the algorithm will play an important role in Part III of
the text.
Chapter 15
Preprocessing Data and Transformation
of Signal Vectors
By “clutter” we mean a large background signal of known origin, such as probe lift-
off, or a large background signal due to systematic errors of unknown origin, such
as a variation in the material properties of the sample, or variations in the subsurface
structure. In any case, clutter is usually more pernicious to the reliable detection and
reconstruction of a flaw than random (electronic) noise that can be often eliminated
by averaging. Hence, the ability to model clutter and reject it is paramount to the
successful application of inverse methods and is the subject of the first part of this
chapter.
Data were taken on an Inconel 600 plate, which is nonmagnetic (relative permeabil-
ity, μ = 1), and has a conductivity of σ = 9.86 × 105 S/m. A notch is introduced
into the “outer diameter” of the plate, which means the surface opposite to the
probe. A coil is excited at 100 kHz and is scanned transversely to the slot, such that
impedances are measured at 250 points, each separated by 4 mils (see Fig. 15.1).
The OD of the coil is 0.112 inch, the ID is 0.034 inch, the height is 0.048 inch, and
there are 131 turns in the coil. The experiment is run with a coil lift-off of 0.015
inch above the plate.
The impedance data were measured by Warren Junker of the Westinghouse
Research Labs, using a Hewlett-Packard 4194A impedance analyzer, and recorded
to six significant digits, which implies a dynamic range of 120 dB. The “host-
only” impedance was measured to be Zhost = 4.05115 + j10.8770 Ω. When Zhost
is subtracted from all 250 data points, the result is the “anomaly signal” shown
in Fig. 15.2. Note the significant “clutter” in the resistance data, produced by a
systematic error.
Fig. 15.1 Illustrating the setup for the experiment and model calculations
0.004 0.006
0.005
0.003
Resistance (Ohms)
Reactance (Ohms)
0.004
0.002
0.003
0.001
0.002
0
0.001
-0.001 0
-0.002 −0.001
−0.5 −0.4 −0.3 −0.2 −0.1 0 0.1 0.2 0.3 0.4 0.5 −0.5 −0.4 −0.3 −0.2 −0.1 0 0.1 0.2 0.3 0.4 0.5
Probe Position (in) Probe Position (in)
Fig. 15.2 Original data supplied by Warren Junker of the Westinghouse Research Labs. Note the
significant “clutter” in the resistance data, produced by a systematic error
In order to determine the “flaw signal,” we subtract the clutter from the anomaly
signal. This is done by first modeling the clutter with a simple mathematical
expression, in this case by the piecewise linear functions:
0.0045 − 0.00225 × (pos+ 0.5) if pos ≤ −0.1
Rclutter =
0.0036 + 0.0066 × (pos+ 0.1) otherwise
Xclutter = −0.0005 + 0.0005 × (pos+ 0.5), (15.1)
where “pos” designates the probe position. When the clutter signal of (15.1) is
subtracted from the data of Fig. 15.2, we get the “processed” data of Fig. 15.3.
It is now clear that the original data resulted from a scan over a notch-type flaw.
We assume that the flaw is a rectangular paralleliped notch that breaks the back
surface of the workpiece, i.e., the surface away from the probe. The problem then
is to determine the size of the notch, namely its length, width, and height. In order
to do that, we apply the processed data of Fig. 15.3 as the input to NLSE (recall
Chap. 12). The three parameters are, of course, the length, width, and height of the
notch. The results are: length = 0.2277 inch, width = 0.0049 inch, and height =
0.025 inch. When we repeat the inversion experiment, this time holding the height
of the flaw to 0.024 inch, we get length = 0.2424, and width = 0.0050. In the first
case (height = 0.025) the final norm of the residuals is 0.8112711 × 10−2, whereas
15.1 Clutter Modeling and Rejection 291
Reactance (Ohms)
−0.001
0.004
−0.002 0.003
0.002
−0.003
0.001
−0.004
0
−0.005 −0.001
−0.5 −0.4 −0.3 −0.2 −0.1 0 0.1 0.2 0.3 0.4 0.5 −0.5 −0.4 −0.3 −0.2 −0.1 0 0.1 0.2 0.3 0.4 0.5
Probe Position (in) Probe Position (in)
Fig. 15.3 Showing the original data with the clutter removed
0.005
Resistance (Ohms)
Reactance (Ohms)
−0.001
0.004
−0.002 0.003
0.002
−0.003
0.001
−0.004 0
−0.001
−0.005 −0.5 −0.4 −0.3 −0.2 −0.1 0 0.1 0.2 0.3 0.4 0.5
−0.5 −0.4 −0.3 −0.2 −0.1 0 0.1 0.2 0.3 0.4 0.5
Fig. 15.4 Showing the effects of inverting the processed original data to reconstruct the flaw
with the height = 0.024 the norm is equal to 0.8522527 × 10−2, which is 5 % larger.
The effect of this is discussed next.
When we use VIC-3D R
in the “forward mode” to compute the scanned
impedances, given notches of the size just determined, and compare these results
with the input data of Fig. 15.3, we get the results shown in Fig. 15.4. Clearly,
the clutter-removal algorithm, together with VIC-3D R
, has been effective. The
resistance curve for the case of the notch with height = 0.025 falls closer to the
negative peak of the input data than does the curve for the other notch. The reactance
curves of the two notches, however, are virtually indistinguishable. This suggests
that it is the better fit of the resistance data that gives the 0.025-inch notch the smaller
residual-norm compared to the 0.024-inch notch.
Table 15.1 compares the nominal, measured, and computed data for these
notches.
292 15 Preprocessing Data and Transformation of Signal Vectors
Table 15.1 Comparison of dimensions of the reconstructed A4 notches with two different depths.
The measured data were determined from salastic molds
Length Length Length Width Width Width Depth Depth Depth
Nominal Measured Computed Nominal Measured Computed Nominal Measured Computed
(inch) (inch) (inch) (inch) (inch) (inch) (inch) (inch) (inch)
0.25 0.252 0.2277 0.005 0.007 0.0049 0.020 0.0238 0.025
0.25 0.252 0.2424 0.005 0.007 0.0050 0.020 0.0238 0.024
0.008
0.006
0.004
0.002
-0.002
-0.004
-0.5 -0.4 -0.3 -0.2 -0.1 0 0.1 0.2 0.3 0.4 0.5
Probe position (inches)
We continue with another example from the data provided by the Westinghouse
Research Labs. In this example we consider an “outer-diameter” flaw comprising 5
notches and 4 gaps, which could model a crack with periodic contact points between
the surfaces, i.e., a partially closed crack.
The notches are each 0.041 inches long, 0.0065 inches wide, and 0.0138 inches
deep. The plate is 0.048 inches thick, so this would be an outer-diameter flaw that
extends 28.75 % into the host. The gaps (ligatures) are 0.0066 inches long. Hence,
the crack is 0.2314 inches long and is 11.4 % closed.
The original measured impedance data and VIC-3D R
model results at 200 kHz
are shown in Fig. 15.5. We model the clutter signal by the piecewise linear functions:
15.1 Clutter Modeling and Rejection 293
Reactance (Ohms)
0.001 0.003
0.0005 0.002
0 0.001
−0.0005 0
−0.001 −0.001
−0.5 −0.4 −0.3 −0.2 −0.1 0 0.1 0.2 0.3 0.4 0.5 −0.5 −0.4 −0.3 −0.2 −0.1 0 0.1 0.2 0.3 0.4 0.5
Probe Position (in) Probe Position (in)
Results of Processing
0.002
Processed Data
VIC Results
0.0015
Resistance (Ohms)
0.001
0.0005
−0.0005
−0.001
−0.5 −0.4 −0.3 −0.2 −0.1 0 0.1 0.2 0.3 0.4 0.5
Probe Position (in)
Fig. 15.6 Measured data with the clutter removed, together with the same VIC-3D R
model
results. Top left: resistance, top right: reactance. Bottom: Measured resistance data with the clutter
removed and then passed through a Bezier filter
100.5 dB for reactance. Hence, we have significantly gained dynamic range by the
process of clutter removal, and what was clearly an undetectable flaw-resistance
value in Fig. 15.5 becomes quite detectable in Fig. 15.6. The results of Fig. 15.6 can
then be used successfully in an inversion process.
If we believe in the “principle of conservation of dynamic range,” that dynamic
range in data can be neither created nor destroyed, then we must conclude that the
Hewlett-Packard 4194A impedance analyzer had the requisite dynamic range of
100dB, but that the original data of Fig. 15.5 obscured this fact because of the clutter.
Once the clutter was removed the “true-anomaly data” were then manifest.
The preceding clutter removal process was heuristic and required user interaction.
Now, we want to develop an automated, model-based clutter removal algorithm
that uses a more general higher-order polynomial fit. The proposed clutter rejection
algorithm begins with an assessment of the model data. A series of N simulated
impedance data vectors, consisting of M probe positions, are defined as Zk,i (x j ),
where Ri (x j ) is the resistance component, and Xi (x j ) is the reactance component,
representing the ith model data set and jth probe position. To evaluate background
regions of interest in the experimental data, the model data are evaluated for regions
of scans where the response is invariant to changes in flaw parameters. Figure 15.7a
shows a series of simulated data vectors for the reactance component of eddy-current
impedance measurements for varying pit dimension. It is desired to have model data
with pertinent flaw parameters ranging over all expected levels in application.
A model-based variation measure, which is used to estimate regions of invari-
ance, is given by:
σk (x j )
Gk (x j ) = 1 − , (15.3)
max(σk )
where
&
1 N ! "2
σk (x j ) = ∑
N − 1 i=1
Zk,i (x j ) − Z k (x j ) ; (15.4)
Fig. 15.7 Plots of (a) simulated data for the reactance component of impedance measurements
for varying pit dimensions, (b) a model-based variation measure of the reactance components as a
function of position, and (c) a threshold function defining a select region for background fit
Using the result of the model-based background region assessment, a fit of the
background clutter can now be performed. To properly represent the background
clutter found in experimental data, a 3rd-order polynomial representation,
fcr (x) = a3 x3 + a2 x2 + a1 x + a0 , (15.5)
is used. The coefficients of the polynomial are evaluated in a least-squares sense
with the experimental data fit over the background data point regions. This function
can then be applied to the experimental data to subtract the estimated error due to
background clutter.
Figure 15.8 presents a plot of the resistance component at f = 1.3 kHz for
the transformed experimental data (exp0 ) for the case of a subsurface artificial
pit (0.062 diameter, 0.062 height) on the back side of the first layer of a two-
layer stack-up of aluminum panels of 0.125 in thickness. Figure 15.8 also displays
296 15 Preprocessing Data and Transformation of Signal Vectors
Fig. 15.9 Illustrating the results of applying the automated clutter removal algorithm to various
measured data sets
the curve, fcr (x), representing a polynomial fit to background clutter, and the
resulting experimental data with clutter removal (expcr ). Clearly, the application
of the automated clutter removal algorithm is beneficial in eliminating the severe
background variance across both the background and flaw regions. Figure 15.9
shows some results of applying the algorithm to various measured data.
15.2 Transformations of Signal Vectors 297
Suppose that we run two N-point scans with a probe, one at frequency f1 and the
other at frequency f2 . We arrange the output of each scan as the two complex
N-vectors, Z1 and Z2 , respectively. That is, these vectors represent the complex
impedance associated with the two scans. We wish to determine a coupling
coefficient, or scale factor, β , such that
β Z1 + Z2 = 0 . (15.6)
If we form the dot-product of (15.6) with Z1 , we easily solve for β :
Z1 · Z2
β =− , (15.7)
|Z1 |2
where
N
Z1 · Z2 = ∑ Z1∗ (n)Z2 (n) , (15.8)
n=1
and the asterisk denotes the complex-conjugate. This transformation can be used
in a number of ways. The use of multifrequencies was introduced as a method
of eliminating background signals due to tube-support plates in the inspection of
heat-exchanger tubes in nuclear power plants. We use it extensively at a single
frequency to transform instrument voltage readings into an equivalent impedance,
thereby effectively transforming the instrument into an impedance analyzer.
If we wish to mix four frequencies, we first collect the impedance vectors, Z1 , Z2 ,
Z3 , Z4 , at the four frequencies, and then determine the three parameters, β1 , β2 , β3 ,
such that
β1 Z1 + β2 Z2 + β3 Z3 + Z4 = 0 . (15.9)
where
d1 (c22 c33 − |c23 |2 ) + d2 (c13 c∗23 − c12 c33 ) + d3(c12 c23 − c13c22 )
β1 =
Δ
d1 (c∗13 c23 − c∗12 c33 ) + d2(c11 c33 − |c13|2 ) + d3(c∗12 c13 − c11c23 )
β2 =
Δ
d1 (c∗12 c∗23 −c∗13 c22 )+d2 (c12 c∗13 −c11 c∗23 )+d3 (c11 c22 −|c12 |2 )
β3 = , (15.12)
Δ
where Δ = c11 (c22 c33 − |c23 |2 ) + c∗12(c13 c∗23 − c12 c33 ) + c∗13 (c12 c23 − c13 c22 ).
that would allow the reconstruction of the small flaw. This would obviate the need
to use a different piece of hardware, while still retaining the ability to satisfactorily
invert the original measured data. That this concept is mathematically feasible will
be demonstrated in this section.
Assume that we know the “bounding-box” of the flaw, which will allow us to
establish a flaw grid, but we don’t know what is in the box; i.e., we don’t know
what the flaw is, but only its position and maximum extent. With this assumption,
(L)
we can compute the incident-field moments, E0 , due to the large coil. Similarly,
(s)
we can compute E0 , the incident-field moments produced by the small coil, once
we decide what the parameters of the small coil should be. With these results, we
transform the large-coil incident-field moments into the small-coil moments via our
usual least-squares β -relationship:
(L) (s)
β E0 = E0 , (15.13)
which has for its solution
(L)H (s)
E0 · E0
β= (L)
, (15.14)
E0 2
with the symbol, H, denoting the Hermitian (complex-conjugate) transpose of a
vector or matrix.
If the anomalous-current response of the volume-integral equation to a flaw with
(L) (L) (s)
E0 as the driver is Ja , then the response to the same flaw with E0 as the driver
(s) (L)
is Ja = β Ja . This holds without knowing the nature of the flaw and is a property
of the linearity of the volume-integral equation. Thus, the impedance measured by
the virtual (small) probe is
(s) (s) (s)
Za = Ja · E0
(L) (L)
= β 2 Ja · E0
(L)
= β 2 Za ; (15.15)
(L)
Za is the known (measured) impedance using the large probe. This relationship
must be computed at each scan position, so we assume that the small probe follows
the same scan as the large probe.
We’ll apply the ABCD-matrix theory of Chap. 8 to the situation of a single probe,
rather than a T/R system. A single probe is usually represented as a one-port, but
in order to use ABCD-matrix theory, we must transform it into a two-port. If the
driving-point impedance of the probe is Z2 , then an appropriate two-port is shown
300 15 Preprocessing Data and Transformation of Signal Vectors
1 2
1
1 1 Z2
−
Z3 Z2
Y3 − Y2 Y2
M1 = M−12 × M3
1 0 1 0
= ×
−1/Z2 1 1/Z3 1
1 0
= , (15.18)
−1/Z2 + 1/Z3 1
from which we conclude that
15.3 Further Developments and Applications of Scaling and Transformations 301
1 2
1
1 1 Z2 + δZ2
−
Z3 Z2
Y2
Y3 − Y2
1 + Y2 δZ2
Fig. 15.11 Showing a network representation of one-ports when the virtual probe is loaded,
producing a change in the driving-point impedance of δ Z2 . Top: representation in terms of
impedances. Bottom: representation in terms of admittances
⎡ 1 1 ⎤
⎢ 1/Z3 − 1/Z2 1/Z3 − 1/Z2 ⎥
Z1 = ⎣ 1 1 ⎦. (15.19)
1/Z3 − 1/Z2 1/Z3 − 1/Z2
This result yields the network representation labeled “1” in the top part of Fig. 15.10.
Under load, the virtual probe has a driving-point impedance of Z2L = Z2 + δ Z2 ,
where δ Z2 is the change in impedance induced by the load. This situation is
shown in Fig. 15.11. Under this condition, the “real” probe has a loaded admittance
given by
(L) Y2
Y3 = Y3 − Y2 +
1 + Y2δ Z2
Y3 + (Y2Y3 − Y22 )δ Z2
= , (15.20)
1 + Y2δ Z2
from which we get for the change in driving-point impedance of the real probe
1 + Y2δ Z2 1
δ Z3 = −
Y3 + (Y2Y3 − Y22 )δ Z2 Y3
Y22 δ Z2
= . (15.21)
Y3 [Y3 + (Y2Y3 − Y22 )δ Z2 ]
We can use this in several ways: given a measured δ Z3 due to a complex probe,
perhaps with a conducting core, and a model result, Y2 δ Z2 , for the same situation
that produced δ Z3 , we have a free parameter, Y3 , that allows us to use (15.21) to
replace the actual probe with the simpler virtual probe. Note that if we ignore the
parenthetical term in the denominator of (15.21), then we have a linear relation
between δ Z3 and δ Z2 , with Y22 /Y32 playing the role of β in (15.6) or (15.13).
302 15 Preprocessing Data and Transformation of Signal Vectors
Reactance
−0.3 0.6
−0.4 deltaZ3 0.4
−0.5 deltaZ2
0.2
−0.6
−0.7 0
−10 −5 0 5 10 −10 −5 0 5 10
Probe Position (mm) Probe Position (mm)
Fig. 15.12 Response of two probes over a flaw. Probe 2 is an air-core coil, and Probe 3 is the same
coil enclosing a ferrite core
Results for DeltaZ3 Results for DeltaZ3
0.1 1.2
0
1
−0.1
0.8
−0.2
Resistance
VIC-3D
Reactance
Transform
−0.3 0.6
−0.4
VIC-3D 0.4
Transform
−0.5
0.2
−0.6
−0.7 0
−10 −5 0 5 10 −10 −5 0 5 10
Probe Position (mm) Probe Position (mm)
15.3.4 An Example
Given the values of Y3 and Y2 in this example, together with the fact that |δ Z2 | is
small, we can ignore the parenthetical expression in the denominator of (15.21), so
that the ratio δ Z3 /δ Z2 ≈ [(4.4197 + j57.982)/(2.8603 + j43.9184)]2 ≈ 1.746. For
this same problem, we get β = 1.7053 − j0.0620, which differs from 1.746 by less
than 3 %.
In developing the idea of the virtual probe, we came across a transformation matrix
that is more general than the simple product of a vector with the scalar, β , yet
should be very easy to implement numerically. The mathematical tool that is basic
to our idea is the Householder transformation matrix that finds significant use in
computational linear algebra.1 Our presentation follows the footnoted reference
closely.
The Householder matrix, P, has the form
P = I − 2w · wH , (15.22)
where w is a complex unit vector. As written, the second term in (15.22) is the outer
product of the two vectors.2 The inner or scalar product of two vectors, a and b is
written aH · b. The matrix P is clearly Hermitian and orthogonal, because
P2 = (I − 2w · wH ) · (I − 2w · wH )
= I − 4w · wH + 4w · (wH · w) · wH
=I. (15.23)
R0
L0
where a is a unit vector that points in the direction of the image vector, P · x, and we
have used the fact that PH = P = P−1 , which implies that |P · x| = |x|. φ is the angle
between x and its image and is a measure of the amount that x has been rotated by
the orthogonal transformation. Equation (15.24) is the law of cosines in hyperspace.
Now, let u be any vector, and rewrite P as
u · uH
P = I− , (15.25)
N
1
where N = |u|2 . We’ll use this operator to rotate a vector, x, into another vector
2
oriented in the direction of the unit vector, a. To that end, let u = x + |x|a. Then
u
P·x = x− · (x + |x|a)H · x
N
u
= x − (xH · x + |x|2 cos α )
N
u
= x − xH · x(1 + cos α )
N
2uxH · x(1 + cos α )
= x−
(x + (xH · x)1/2a)H · (x + (xH · x)1/2 a)
2uxH · x(1 + cos α )
= x−
(xH · x + 2(xH · x) cos α + xH · x)
= x−u
= −|x|a . (15.26)
15.4 A General Transformation Matrix 305
The problem that we are addressing is shown in Fig. 16.1. A probe, comprising a
coil with a rather complex ferrite core, is scanned past a truncated cylindrical pit of
diameter, PD, embedded in a quarter-inch aluminum slab. The pit extends only into
the upper-half of the slab and has a height of PH1. For this analysis we will assume
that the frequency of excitation is 2,200 Hz.
It will facilitate running the many forward models that will be required for inverting
data if we can replace the ferrite core in the probe model with a simpler probe
that does not have a ferrite core. We have run a number of NLSE calculations
to determine that coil whose impedance most closely approximates the measured
results of a pit that is 75 % through-wall (PH1 = 2.38125 mm in Fig. 16.1), with a
diameter of 0.125 in (PD = 0.125 in. in Fig. 16.1). The results, which are shown
in Fig. 16.2, were obtained with a coil whose inner radius is 0 mm, outer radius is
2.75 mm, height of 2.54 mm, containing 2,500 turns, and tilted at an angle of 2◦
Z
Core
Scan
Coil
X
PD
Pit PH1
Aluminum : σ = 30% IACS 6.35mm
3.175mm
Comparison of Measured and Model Data Comparison of Measured and Model Data
0.02 0.16
measured measured
model 0.14 model
0
0.12
-0.02
Resistance (Ohms)
Reactance (Ohms)
0.1
-0.04 0.08
-0.06 0.06
0.04
-0.08
0.02
-0.1
0
-0.12 -0.02
-15 -10 -5 0 5 10 15 -15 -10 -5 0 5 10 15
Probe Position (mm) Probe Position (mm)
Fig. 16.2 Comparison of measured and model data for a probe that is tilted at an angle of 2◦
about the Y -axis of Fig. 16.1. This probe model captures the peaks and valleys of
the measured data reasonably well and will be used when creating the interpolation
table for inverting the data for the class of problems defined in Fig. 16.1.
The experimental data for this setup are recorded as A/D counts (or “voltage”) in
two channels that are 90◦ out of phase with each other. We have done some initial
processing on this data by first defining the “baseline value” of each channel and
then subtracting this baseline from the scanned data of each channel. The baseline
value is taken to be the left-hand value of the data, because there are no other pits or
anomalies to interact with this value (we are well away from any edges of the host
plate). We call these data the “host-balanced” data, and they are shown in Fig. 16.3.
The measured data are clearly asymmetrical, even though the pit and probe
are symmetrical. As we discussed above, we have created an accurate, though
simplified, model of the probe, and have shown that if the probe is tilted about
2◦ about the Y -axis of Fig. 16.1, then the model also produces asymmetric lobes
similar to those of Fig. 16.3. It remains, now, to transform data, measured in A/D
counts or “voltage” into impedances, so that they can be further processed using
VIC-3D R
. We use the transformation of signal vectors described in Chap. 15
to accomplish this. The actual data used to determine β were for the same flaw
(PD = 0.125 in, PH1 = 2.38125 mm) as above, and the value obtained was
β = 6.5276 × 10−6 + j8.2605 × 10−6. When the measured data of Z1 (see Fig. 16.3)
are multiplied by β , we get the “β -normalized” measured data shown in Fig. 16.4.
312 16 Modeling Corrosion and Pitting Problems
-50
Channel 1 or Channel 2
-100
-150
-200
-250
-300
-350
-400
-450
-15 -10 -5 0 5 10 15
Probe Position (mm)
0.002
0.001
-0.001
-0.002
-0.003
-15 -10 -5 0 5 10 15
Probe Position (mm)
The inverse problem can be addressed in terms of Fig. 16.1, in which two 1/8-inch
panels are placed back-to-back, with corrosion pits emanating from the bottom of
the first panel upward, and from the top of the second panel downward. The first pit
is called pit 1, and the second, pit 2. The problem is to determine the height of each
pit (PH1 and PH2), as well as the diameter (PD1 and PD2). We confine ourselves
here to pits that emanate upward only; hence, PH2 = PD2 = 0.
16.3 The Inverse Problem 313
1.40
1.05
0.7
From here on we work with only the radius and height of the pit. The interpolation
table is defined with respect to the grid shown in Fig. 16.5. Because this is a 5 × 5
grid, we can use polynomial splines of order 1–4 for interpolation in each variable.
The interpolating functions (sometimes called “blending” functions) associated
with each of the 25 nodes of Fig. 16.5 are shown in Figs. 16.6–16.10. These
functions were computed using the same probe-flaw model that produced the model
data of Fig. 16.2.
It is clear that the challenge in the inversion process comes not only from the
nonlinear relationship between the impedances and the two parameters, PD and
PH1, but also from the significant change in appearance of the resistance scan-data
with these two parameters. It appears that the inversion might be less confounding
if we restricted ourselves to only reactance data; this will be discussed shortly.
We have run a number of inversions, and we show six that are typical of the results.
In each of the following tabulated results, we show the norm of the residuals, Φ ,
the computed values of pit radius, R, and pit height, H, and the sensitivity of the
solution, ΣR , ΣH , to each of these variables. These data are tabulated as a function
of the order of the interpolating polynomial. Because there are five nodes in each
variable, the interpolating polynomial in each variable can range in order from 1
to 4.
Table 16.1 shows the results of the inversion using the higher-order spline
interpolator when applied to the model input data of Fig. 16.2.
314 16 Modeling Corrosion and Pitting Problems
Reactance (Ohms)
0.003
0.006
0.0025
0.002
0.004
0.0015
0.001 0.002
0.0005
0
0
-0.0005 -0.002
-15 -10 -5 0 5 10 15 -15 -10 -5 0 5 10 15
Probe Position (mm) Probe Position (mm)
Fig. 16.6 Interpolating (“blending”) functions for the case PH1 = 0.7. The numbers indicate the
radius of the corresponding node in Fig. 16.5. Left: resistance; Right: reactance
0.006
Reactance (Ohms)
0.02
0.004
0.015
0.002
0.01
0
0.005
-0.002 0
-0.004 -0.005
-15 -10 -5 0 5 10 15 -15 -10 -5 0 5 10 15
Probe Position (mm) Probe Position (mm)
Fig. 16.7 Interpolating (“blending”) functions for the case PH1 = 1.2. The numbers indicate the
radius of the corresponding node in Fig. 16.5. Left: resistance; Right: reactance
The original data of Fig. 16.11, pd-031 ph1-25 ph2-0 f1, show a clear system-
atic error that we call generically “clutter” (see Chap. 15). The inverted results when
using these cluttered data as input to NLSE are shown in Table 16.2. In order to
determine the “flaw signal,” we must remove the clutter from the original measured
data. This is done, as described in Chap. 15, by modeling the clutter as a simple
mathematical expression:
Rclutter = −0.001 + (0.0085/17.907) × (pos+ 10.033)
Xclutter = 0.00026 − (0.00246/17.907) × (pos+ 10.033) , (16.1)
where “pos” denotes the probe position.
16.3 The Inverse Problem 315
Reactance (Ohms)
0.005
0.04
0
0.03
-0.005
0.02
-0.01
0.01
-0.015 0
-0.02 -0.01
-15 -10 -5 0 5 10 15 -15 -10 -5 0 5 10 15
Probe Position (mm) Probe Position (mm)
Fig. 16.8 Interpolating (“blending”) functions for the case PH1 = 1.7. The numbers indicate the
radius of the corresponding node in Fig. 16.5. Left: resistance; Right: reactance
Reactance (Ohms)
-0.01
0.08
-0.02
-0.03 0.06
-0.04
0.04
-0.05
0.02
-0.06
0
-0.07
-0.08 -0.02
-15 -10 -5 0 5 10 15 -15 -10 -5 0 5 10 15
Probe Position (mm) Probe Position (mm)
Fig. 16.9 Interpolating (“blending”) functions for the case PH1 = 2.2. The numbers indicate the
radius of the corresponding node in Fig. 16.5. Left: resistance; Right: reactance
After removing the clutter by subtracting (16.1) from the measured data, we get
the results shown in Fig. 16.12, which are a clear improvement. The inverted results
for the new data are shown in Tables 16.3 and 16.4.
In Fig. 16.13, we compare the original pd-031 ph1-50 data with the recon-
structed data computed from the fourth-order result of Table 16.5. Clearly, the
reactance data, which are much cleaner, drive the inversion.
Now, we will test pd-125 ph1-75 ph2-75 f1 as the first part of a two-step
strategy to size the corrosion in layer 2. We will treat this input, since it is at the
high frequency of 2,200 Hz, as containing only a pit in layer 1, so we will use
the machinery already set up for a two-parameter problem, PH1 and PD, to invert
316 16 Modeling Corrosion and Pitting Problems
Reactance (Ohms)
-0.05
0.15
-0.1
0.1
-0.15
0.05
-0.2
-0.25 0
-15 -10 -5 0 5 10 15 -15 -10 -5 0 5 10 15
Probe Position (mm) Probe Position (mm)
Fig. 16.10 Interpolating (“blending”) functions for the case PH1 = 2.7. The numbers indicate the
radius of the corresponding node in Fig. 16.5. Left: resistance; Right: reactance
these data (Table 16.6). Then we will use the parameters obtained for PH1 and PD
as constraints for determining PH2. For this purpose we will need pd-125 ph1-
75 ph2-75 f2, which are data at the lower frequency of 1,300 Hz. Further, we will
develop an entirely new interpolation table for operation at this frequency, as will
be described in the next section.
The results for pd-125 ph1-75 ph2-75 f1 are shown in Table 16.7 and confirm
the reasonableness of our strategy of using high frequencies to reconstruct flaws in
layer 1.
Our intention is to use the lower frequency to reconstruct flaws in layer 2. To this
end we introduce a three-parameter interpolation grid and table for PH1, PH2, and
PD. The grid is a three-dimensional version of Fig. 16.5, in which the third variable,
16.3 The Inverse Problem 317
Fig. 16.11 Original data for Data for PD=.031, PH1=25, PH2=0
PD = 0.03125 in., PH1 = 0.008
25% through-wall = resistance
reactance
0.79375 mm
0.006
Resistance/Reactance (Ohms)
0.004
0.002
-0.002
-0.004
-15 -10 -5 0 5 10 15
Probe Position (mm)
PH2, is normal to the page. This variable will have the same five nodes as PH1,
which means that the grid will have five copies of Fig. 16.5, giving a total of 125
nodes.
Our first result is for pd-125 ph1-75 ph2-75 f2 and is shown in Table 16.8. The
sensitivity parameter gives us insight into the quality of the results. The smallest
318 16 Modeling Corrosion and Pitting Problems
Processed Data for PD=.031, PH1=25, PH2=0 Processed Data for PD=.031, PH1=25, PH2=0
0.0025 0.002
R X
0.002
0.0015
0.0015
Reactance (Ohms)
0.001
Resistance (Ohms)
0.001
0.0005
0 0.0005
-0.0005
0
-0.001
-0.0015
-0.0005
-0.002
-0.0025 -0.001
-15 -10 -5 0 5 10 15 -15 -10 -5 0 5 10 15
Probe Position (mm) Probe Position (mm)
Fig. 16.12 Data of Fig. 16.11 after the systematic error “clutter” has been removed. Left:
resistance; Right: reactance
value of Φ occurs with the first-order polynomial, and the results for PD and PH1
are reasonable. The error in PD is 6.7%, and in PH1 6.9%. Note that the sensitivity
of these two values at the solution is considerably smaller than that for PH2. This
indicates that PD and PH1 are reasonably well determined, but PH2 is not. The
solution is insensitive to PH2 because the frequency is too high for PH2 to be
resolved. This is what we suspected at 2,200 Hz and holds true also for 1,300 Hz.
The result for pd-031 ph1-25 ph2-25 f2, however, is a little bit more encourag-
ing, as shown in Table 16.9. The values given with the fourth-order polynomial for
the three unknowns are in reasonable agreement with their nominal values, but the
sensitivities for these parameters, though a little large due to the quality of the input
data, are all comparable. This suggests that the value of PH2 can be resolved, even
at 1,300 Hz, because PH1 is small.
16.3 The Inverse Problem 319
Comparison of Original and Reconstructed Data Comparison of Original and Reconstructed Data
0.0035 0.007
Original Original
0.003 Reconstructed 0.006 Reconstructed
0.0025 0.005
Resistance (Ohms)
Reactance (Ohms)
0.002
0.004
0.0015
0.003
0.001
0.002
0.0005
0.001
0
0
-0.0005
-0.001 -0.001
-0.0015 -0.002
-15 -10 -5 0 5 10 15 -15 -10 -5 0 5 10 15
Probe Position (mm) Probe Position (mm)
Fig. 16.13 Comparison of original pd-031 ph1-50 data with the reconstructed data computed
from the fourth-order result of Table 16.5. Left: resistance; Right: reactance
If we take the (first-order) solution for R in Table 16.8 and average it with the
(fourth-order) solution in Table 16.4, we get 1.5719, which is in error by less than
1% from the nominal value of 1.5875 mm. Similarly, when we average the solutions
for PH1 from the same tables, we get 2.3405, which is in error by 1.7% from the
320 16 Modeling Corrosion and Pitting Problems
Table 16.8 Results of higher-order spline interpolator when applied to the input data of pd-
125 ph1-75 ph2-75 f2. ΣV denotes the sensitivity of the solution to the variable, V . The first-order
result has an error in radius of 6.7% and in PH1 of 6.9%. PH2 is essentially undetermined
Polynomial order Φ R (mm) ΣR PH1 (mm) Σ PH1 PH2 (mm) Σ PH2
1 0.12293 1.7015 0.05962 2.2169 0.12605 0.7 3.5745
2 0.12344 1.7091 0.04978 2.2255 0.16515 0.7 3.5464
3 0.12346 1.7136 0.04671 2.2147 0.16735 0.7 3.5473
4 0.12358 1.7232 0.03667 2.2120 0.18108 0.7 3.5622
Table 16.9 Results of higher-order spline interpolator when applied to the input data of pd-
031 ph1-25 ph2-25 f2. ΣV denotes the sensitivity of the solution to the variable, V . The fourth-
order result has an error in radius of 11.8%, in PH1 of 11.6% and 4.8% in PH2
Polynomial order Φ R (mm) ΣR PH1 (mm) Σ PH1 PH2 (mm) Σ PH2
1 0.011785 0.35 0.8655 0.7019 0.8907 0.8817 4.9636
2 0.01180 0.35 0.9287 0.7095 1.0495 0.7434 5.1100
3 0.01182 0.3580 0.8385 0.7 1.1195 0.8822 5.01
4 0.011779 0.35 0.6277 0.7016 1.7667 0.8337 5.3144
nominal value of 2.38125 mm. This indicates that we can often use “frequency
mixing” to get improved results when compared to the results at the individual
frequencies.
The resistance data for pd-031 ph1-25 ph2-00, shown in Fig. 16.11, are clearly quite
noisy and play a significant role in producing the bad results shown in Table 16.2
for the estimate of the radius. This suggests that we remove the resistance data and
invert only the reactance data. The result of this experiment is shown in Table 16.10.
The estimates of the height and height-sensitivity remain about the same as before,
but the estimates of the radius and radius-sensitivity are improved considerably.
We should not think, however, that the resistance should be discarded under
all conditions. In several of our experiments, we found that discarding “good”
16.4 A New Flaw Model 321
resistance data degraded otherwise “good” solutions. It appears that resistance data
stabilize the inversion process, especially in determining the height of a pit. Later,
we will give an example of the use of the projection algorithm, POCS, to clean up
the resistance and yield an improved inversion.
The pits that have been used in obtaining data that are described in this chapter
are the result of drilling into the aluminum hosts. The drill bits have pointed tips,
so the result of the drilling is to produce a pit that will have a right-conical tip at
its top. The resulting departure from the assumed truncated cylinder geometries is
not significant in most cases, but will be quite significant if the desired “pit” is quite
shallow. In this case we must modify our model of the pit to accommodate this
geometry. We will call this a new flaw model.
Figure 16.14 illustrates the new flaw “tip” model together with an equivalent
truncated cylinder “pit,” in the case that the pit is not deeper than the conical tip.
The truncated cylinder is assumed to have the same volume and centroid as the
conical tip.
322 16 Modeling Corrosion and Pitting Problems
In order to get good resolution for very deep-lying pits in the 0.125-inch panels, it
is necessary to go to lower frequencies than we have used earlier. In this section we
will describe the results of experiments that use larger-diameter probes of 0.44 and
0.57 inch, excited at 360 and 720 Hz for this purpose. We will carry out a few model
calculations to determine such things as the dynamic range required of instruments
in this frequency range.
Figure 16.15 shows two cases that we will consider. The dimensions of the pits
(mm) are calculated using the “equivalent-pit” model of conical-tip flaw shapes,
as was described in the previous section. The coil diameter, D, is either 0.44 inch
or 0.57 inch. The flaw model on the left is called HAL 5 7 44 57.vic, and the
one on the right is called HAL 5 13 44 57.vic, which follows the notation of the
experiments.
Table 16.11 lists the experimental conditions and the names of the corresponding
VIC-3D R
models. The pit dimensions are nominal lengths; the actual dimensions
of the pits (hole diameter and lengths) are calculated as explained in the preceding
section.
Figure 16.16 illustrates the frequency response of the isolated pit in layer 1 as
a function of the coil diameter. Note that we have extended the scan range from
−19 mm to +19 mm in order to start and finish over unflawed host material.
We use the results of Fig. 16.16 to first determine the dynamic range required to
detect the isolated pit. To do this we compare the peak-to-peak response of the probe
as it scans the pit to the impedance of the probe over the unflawed host material. The
base-10 logarithm of this number is roughly a measure of the number of significant
digits that must be recorded by the test instrument, and is called the dynamic range.
D
0mm
1.458
P1 P1 0.5512
−3.175
P2 2.126
1.5444
−6.35
Fig. 16.15 Showing two low-frequency flaw models. The coil, whose diameter, D, is either
0.44 inch or 0.57 inch, is scanned past two independent flaw systems. The dimensions of the
cylindrical pits (mm) are calculated using the equivalent-pit model. The model on the left is
called HAL 5 7 44 57.vic, and the one on the right is HAL 5 13 44 57.vic, in accordance with
the notation of the experiments
16.5 Low-Frequency Models and Experiments 323
Response_5_13_44_57 Response_5_13_44_57
0.001 0.006
5.588/360 5.588/360
7.239/360 7.239/360
5.588/720 0.005 5.588/720
0.0005 7.239/720 7.239/720
Resistance (Ohms)
0.004
Reactance (Ohms)
0
0.003
0.002
-0.0005
0.001
-0.001
0
-0.0015 -0.001
-20 -15 -10 -5 0 5 10 15 20 -20 -15 -10 5 0 5 10 15 20
Fig. 16.16 Response of HAL 5 13 44 57. The title denotes coil diameter, 0.44 or 0.57 inch,
whereas the legend denotes coil radius in mm, and frequency, 360 or 720 Hz
Table 16.12 Dynamic range requirements in dB for measuring response of the isolated pit in layer
1, as a function of coil diameter and frequency
Radius (mm) Freq (Hz) Zhost Zpeak−peak Dynamic range (dB)
5.588 360 1.9617 + j37.691 0.0007 + j0.0007 69 + j94.7
7.239 360 4.1607 + j51.4588 0.00125 + j0.0013 70.4 + j92.0
5.588 720 5.7508 + j73.3232 0.0015 + j0.0028 71.7 + j88.4
7.239 720 11.445 + j98.2759 0.0021 + j0.005 74.7 + j85.9
This number can also be expressed as A/D bits by dividing by 0.3. The dynamic
range is sometimes expressed in decibels by dB = 20 log(ratio).
Table 16.12 tabulates the results of the calculations. The dynamic range require-
ments are not particularly severe, with the clear advantage going to the 0.57 inch coil
excited at 720 Hz. The dynamic range in reactance is a rather modest 85.9 dB, which
translates into 4.3 significant digits or an A/D converter with 14.3 bits of precision.
Typical instruments start at 16 bits, which give about 4.8 significant digits, or 96 dB
dynamic range.
Figures 16.17–16.20 illustrate the responses of the other model experiments.
Figure 16.20 compares the responses at the single frequency of 720 Hz and coil
324 16 Modeling Corrosion and Pitting Problems
Response_5_6_44_57 Response_5_6_44_57
0.0015 0.01
5.588/360 5.588/360
7.239/360 0.009 7.239/360
0.001 5.588/720 5.588/720
7.239/720 0.008 7.239/720
0.0005 0.007
Resistance (Ohms)
Reactance (Ohms)
0 0.006
0.005
-0.0005
0.004
-0.001 0.003
-0.0015 0.002
0.001
-0.002
0
-0.0025 -0.001
-20 -15 -10 -5 0 5 10 15 20 -20 -15 -10 -5 0 5 10 15 20
Probe Position (mm) Probe Position (mm)
Fig. 16.17 Response of HAL 5 6 44 57. The title denotes coil diameter, 0.44 inch or 0.57 inch,
whereas the legend denotes coil radius in mm, and frequency, 360 Hz or 720 Hz
Response_5_8_44_57 Response_5_8_44_57
0.004 0.016
5.588/360 5.588/360
7.239/360 0.014 7.239/360
0.003 5.588/720 5.588/720
7.239/720 7.239/720
0.012
0.002
Reactance (Ohms)
Resistance (Ohms)
0.01
0.001
0.008
0
0.006
-0.001
0.004
-0.002 0.002
-0.003 0
-0.004 -0.002
-20 -15 -10 -5 0 5 10 15 20 -20 -15 -10 -5 0 5 10 15 20
Probe Position (mm) Probe Position (mm)
Fig. 16.18 Response of HAL 5 8 44 57. The title denotes coil diameter, 0.44 inch or 0.57 inch,
whereas the legend denotes coil radius in mm, and frequency, 360 Hz or 720 Hz
diameter of 0.57 inch. This figure shows that the responses are well distinguished,
which means that at this frequency, and using this coil configuration, we should be
able to invert the measured impedance data and resolve the layer-two pits. For this
reason we recommend the use of the 0.57-inch coil operating at 720 Hz.
Finally, note in all of these figures, especially in the reactive components, the
significant dip that occurs when the coil is centered over the flaw. This is another
manifestation of the response that we have seen earlier, in that a very large coil will
have a very significant valley when the coil is so positioned. This valley will serve
as a “pointer” to the pit and should help resolve nearby pits in the lateral direction.
16.5 Low-Frequency Models and Experiments 325
Response_5_7_44_57 Response_5_7_44_57
0.005 0.02
5.588/360 5.588/360
0.004 7.239/360 7.239/360
5.588/720 5.588/720
0.003 7.239/720 0.015 7.239/720
Reactance (Ohms)
Resistance (Ohms)
0.002
0.001 0.01
-0.001 0.005
-0.002
-0.003 0
-0.004
-0.005 -0.005
-20 -15 -10 -5 0 5 10 15 20 -20 -15 -10 -5 0 5 10 15 20
Fig. 16.19 Response of HAL 5 7 44 57. The title denotes coil diameter, 0.44 inch or 0.57 inch,
whereas the legend denotes coil radius in mm, and frequency, 360 Hz or 720 Hz
Reactance (Ohms)
0.002
0.01
0.001
0 0.005
-0.001
0
-0.002
-0.003 -0.005
-20 -15 -10 -5 0 5 10 15 20 -20 -15 -10 -5 0 5 10 15 20
Probe Position (mm) Probe Position (mm)
Fig. 16.20 Response of second-layer pits. The first-layer pit is fixed at a nominal value of 25%
through-wall. The legend denotes percent through-wall of the second-layer pits. The coil diameter
is fixed at 0.57 inch, and the frequency is fixed at 720 Hz
Now, we return to the situation depicted in Fig. 16.15 and invert measured data taken
on this system. We label the various configurations by their nominal values, in which
the diameter of each pit is “fixed” at 0.0625 inch. In truth, the diameters vary with
depth, as indicated in Sect. 16.4, but since the “β -scaling” is done using the usual
nominal values, that’s what we’ll use. The depth of P1 is fixed at 25% through-
326 16 Modeling Corrosion and Pitting Problems
cd-57_f720_ph2-75 cd-57_f720_ph2-75
0.0004 0.002
measured measured
0.0003 model model
0.0002 0.0015
Resistance (Ohms)
Reactance (Ohms)
1e-04
0 0.001
-0.0001
-0.0002 0.0005
-0.0003
-0.0004 0
-0.0005
-0.0006 -0.0005
-20 -15 -10 -5 0 5 10 15 20 -20 -15 -10 -5 0 5 10 15 20
Probe Position (mm) Probe Position (mm)
Fig. 16.21 Comparison of model and measured data for the ph2-75 pit. The data are taken with
the 0.57-inch probe at 720 Hz. Left: resistance; Right: reactance
wall, which is equal to 0.79375 inch, and the depths of P2 are 0% through-wall
(0 mm), 25% through-wall (0.79375 mm), 50% through-wall (1.5875 mm), and 75%
through-wall (2.38125 mm). Data were taken at 360 and 720 Hz using ferrite-core
probes of nominal diameters 0.44 and 0.57 inch. We’ll describe the results for the
0.57-inch diameter probe at 720 Hz, which we label cd-57 f720, here.
In order to generate the most accurate interpolation table, which will be used
with NLSE to invert the data, we tuned the model by using the more accurate data
of Sect. 16.4 for the largest pit, and introducing a simple “fast-probe” model to
facilitate generating the interpolation table. The inner radius of the probe is 0 mm,
the outer radius is 5.588 mm, the height is 2.54 mm, and there are 1,075 turns. These
parameters were chosen to give the best fit to the ph2-75 model and measured data.
The actual parameters for the various pit-models are shown in Table 16.13.
Figure 16.21 shows the result of tuning the probe and ph2-75 model to the
measured data. We emphasize that only the ph2-75 measured data were used in
determining the fast-probe model. Figures 16.22–16.24 demonstrate that the other
measured data were scaled consistently, so that they agree with the fast-probe
model and the actual pit parameters tabulated in Table 16.13. No attempt has been
made to model the asymmetries in the measured impedance data (especially in the
resistance), as was done earlier when we “tilted” the probe model.
We assume that P1 is known (perhaps from UT data), so the problem is to
reconstruct P2. We do this by first generating an interpolation table based on the grid
16.5 Low-Frequency Models and Experiments 327
cd-57_f720_ph2-00 cd-57_f720_ph2-00
0.0002 0.0008
measured measured
model 0.0007 model
1e-04
0.0006
0
Resistance (Ohms)
Reactance (Ohms)
0.0005
-0.0001 0.0004
-0.0002 0.0003
0.0002
-0.0003
0.0001
-0.0004
0
-0.0005 -0.0001
-20 -15 -10 -5 0 5 10 15 20 -20 -15 -10 -5 0 5 10 15 20
Probe Position (mm) Probe Position (mm)
Fig. 16.22 Comparison of model and measured data for the ph2-00 pit. The data are taken with
the 0.57-inch probe at 720 Hz. Left: resistance; Right: reactance
cd-57_f720_ph2-25 cd-57_f720_ph2-25
0.0002 0.0012
measured measured
model model
1e-04 0.001
Resistance (Ohms)
0 0.0008
Reactance (Ohms)
-0.0001 0.0006
-0.0002 0.0004
-0.0003 0.0002
-0.0004 0
-0.0005 -0.0002
-20 -15 -10 -5 0 5 10 15 20 -20 -15 -10 -5 0 5 10 15 20
Probe Position (mm) Probe Position (mm)
Fig. 16.23 Comparison of model and measured data for the ph2-25 pit. The data are taken with
the 0.57-inch probe at 720 Hz. Left: resistance; Right: reactance
of Fig. 16.25. In creating this table, we fixed P1 using the parameters of Table 16.13,
while allowing the P2 parameters to vary according to Fig. 16.25. The results are
shown in Tables 16.14–16.17.
In order to validate our interpolation table, we inverted the model data of
Figs. 16.23 and 16.24, and obtained the results shown in Tables 16.18 and 16.19,
respectively. The results, especially for the model ph2-50 data, indicate that the table
is correct. The discrepancies in the results for the ph2-25 data may be due to the fact
that we did not put enough cells in the grid of the smaller pit while computing the
response with VIC-3D R
.
328 16 Modeling Corrosion and Pitting Problems
cd-57_f720_ph2-50 cd-57_f720_ph2-50
0.0004 0.0016
measured measured
0.0003 model 0.0014 model
0.0002 0.0012
Resistance (Ohms)
Reactance (Ohms)
1e-04 0.001
0 0.0008
-0.0001 0.0006
-0.0002 0.0004
-0.0003 0.0002
-0.0004 0
-0.0005 -0.0002
-20 -15 -10 -5 0 5 10 15 20 -20 -15 -10 -5 0 5 10 15 20
Probe Position (mm) Probe Position (mm)
Fig. 16.24 Comparison of model and measured data for the ph2-50 pit. The data are taken with
the 0.57-inch probe at 720 Hz. Left: resistance; Right: reactance
P Q R S T
2.1
K L M N O
1.4
F G H I J
0.7
0.0 A B C D E PD2(mm)
0.4 0.8 1.2 1.6 2.0
Table 16.14 Results for ph2-00 57 720. σ denotes the sensitivity of the
parameter at the solution point
Order Φ PD2/σ PH2/σ
1 0.2183248(−2) 1.200/0.9015 0.1514/0.4280
2 0.2179565(−2) 1.414/1.1797 0.1107/0.3084
3 0.2178009(−2) 1.473/1.2648 0.1024/0.2840
4 0.2172408(−2) 1.974/2.0086 0.0640/0.1757
Tables 16.20–16.23 display the results of inverting the measured data, but with
the resistance data suppressed.
16.5 Low-Frequency Models and Experiments 329
The fourth-order result of Table 16.14 indicates that PH2 is quite small, which
agrees with the actual parameter for ph2-00 in Table 16.13. The values of PD2 in
Table 16.14 are not well determined, of course, because of the small value of PH2.
The results for ph2-75, shown in Table 16.17, also agree quite well with the actual
parameters listed in Table 16.13, for both PD2 and PH2.
The results for PD2 and PH2 for ph2-25 and ph2-50, given in Tables 16.15
and 16.16, respectively, though not accurate, display an interesting feature: the
volume of each of the pits is reproduced accurately. Consider the fourth-order result
for ph2-25 in Table 16.15; we have for the volume, PH2∗ π ∗PD22 /4 = 0.9547 mm3 .
The “true” volume for this pit is computed from Table 16.13 to be 0.9203, which
differs by 3.6% from the “measured” volume. Similarly, the “measured” volume of
ph2-50, as given by the fourth-order result of Table 16.16, is equal to 2.5940 mm3 ,
whereas the true volume is equal to 2.45 mm3 , and differs from the measured volume
by 5.9%.
Tables 16.20–16.23 show that suppressing the resistance data (which are quite
ugly when compared to the reactance data in Figs. 16.21–16.24) has a moderately
stabilizing effect on PD2, in that all of the results cluster closely to each other
(except for ph2-00), as is the case with the actual data of Table 16.13, but, of course,
the measured PD2’s are not close to the actual values. The PH2’s follow an orderly
ascent for ph2-00 to ph2-75, but for the latter four the values are off by a factor of
about 2. So, it is difficult to conclude precisely the role played by the resistance data,
except to say that it is useful, and it needs to be measured more accurately than it
is now.
An alternative inverse problem is to choose a value of PD2 for each flaw, and then
fix that parameter during the inversion and determine the optimum estimate of the
height, PH2, of each flaw. This corresponds to our notion, expressed earlier in this
section, of fixing a priori the lateral resolution of a simple or complex pit, and then
determining the height of the pit. The question is how to determine the resolution.
One way of doing this, if we simply don’t want to postulate a lateral resolution, is to
16.5 Low-Frequency Models and Experiments 331
take, say, 500 random samples of PD2 and PH2, and determine Φ for each of these
samples. Then we order Φ from smallest to largest in the ensemble, choose PD2 for,
say, the smallest 10 values of Φ , and average the results to give us a prior estimate
of PD2. The inversion is then completed as a one-dimensional problem for PH2. We
will call this the “exhaustive-search” method.
We did this for ph2-25 and ph2-50, and got for ph2-25 values of PD2 = 1.324,
PH2 = 0.5997, and for ph2-50, PD2 = 1.24, PH2 = 2.052. The results for ph2-25
are not in bad agreement with Table 16.13, being off by about 10% in diameter and
about 9% in height. The errors in PD2 and PH2 for ph2-50 are, respectively, 23%
and 53%. Clearly, the measured data for ph2-50 are not in good agreement with the
data that should correspond to the pit whose dimensions are given in Table 16.13.
In general, however, we can conclude that the use of a 0.57-inch probe, operating
at 720 Hz is well suited for characterizing second-layer pits in a system comprising
two 0.125-inch panels.
As we have stated many times before, the resistance data are the least reliable and
most prone to producing poor reconstructions. This is certainly true of the data set
ph2-50 57 720. The POCS algorithm of Fig. 14.3 was developed to ameliorate this
problem and was applied to this data set.
The results of running the POCS algorithm on ph2-50 57 720 are listed in
Table 16.24.
The input impedance data generated by the POCS algorithm are shown in
Fig. 16.26. Note that the resistance data approach the “model” with each iteration.
The reactances behave in a similar manner, but do not change as strongly from
iteration to iteration. Only the measured reactance data are used in each iteration. It
332 16 Modeling Corrosion and Pitting Problems
Reactance (Ohms)
1e-04 0.001
0 0.0008
-0.0001 0.0006
-0.0002 0.0004
-0.0003 0.0002
-0.0004 0
-0.0005 -0.0002
-20 -15 -10 -5 0 5 10 15 20 -20 -15 -10 -5 0 5 10 15 20
Probe Position (mm) Probe Position (mm)
Fig. 16.26 Input impedance data generated by the POCS algorithm applied to ph2-50. The
“measured” and “model” data are repeated from Fig. 16.24. The numbered data correspond
to (PH2, PD2)1 , (PH2, PD2)2 and (PH2, PD2)3 given in Table 16.24. Left: resistance; Right:
reactance
is clear that the improvement in the reconstruction is due to the improved resistance
data during the iterations. Finally, we stress that we never use the “model” data
because the model is not known to us a priori. Indeed, determining the model is the
reason for doing the inversion in the first place.
Response of Feature Vectors for ph2-25 Response of Feature Vectors for ph2-25
0.0002 0.0012
1D 1D
2D 2D
1e-04 model 0.001 model
Resistance (Ohms)
0 0.0008
Reactance (Ohms)
-0.0001 0.0006
-0.0002 0.0004
-0.0003 0.0002
-0.0004 0
-0.0005 -0.0002
-20 -15 -10 -5 0 5 10 15 20 -20 -15 -10 -5 0 5 10 15 20
Probe Position (mm) Probe Position (mm)
Fig. 16.27 Comparison of processed data with model data when 1D and 2D feature vectors are
used. Left: resistance; Right: reactance
random-PD2 algorithm twice, getting PD2 = 1.4866 the first time, and PD2 =
1.4758 the second. The results for PH2 and Φ in each case were PH2 = 0.5444,
Φ = 0.1200699 × 10−2 for the first, and PH2 = 0.5508, Φ = 0.1200684 × 10−2 for
the second. These results are in excellent agreement with the nominal values of PD2
= 1.458 and PH2 = 0.5512 that are listed in Table 16.13. The corresponding results
for the 1D-data of Sect. 16.5.1 were PD2 = 1.324, PH2 = 0.5997.
Furthermore, when we force PD2 = 1.458 in NLSE, we get a value of PH2 =
0.5193 when using the 1D-data, and PH2 = 0.5618 when using the 2D-data. The
former error is 5.79%, whereas the latter error is 1.92%, so we see that the 2D-data
are much better.
Now we’ll go back and apply the projection (POCS) algorithm of Fig. 14.3 to the
2DFV data. The result of the first iteration is shown in Table 16.25. We will use
the second-order result for generating the input for the second step, because the
334 16 Modeling Corrosion and Pitting Problems
first-order result is “trapped” at the node PH2 = 0.7 of the interpolation grid of
Fig. 16.25. This often happens with first-order results because the interpolation
table is linear in this order, which means that there will be discontinuities in the
derivative at each node. This discontinuity is suspected to be the culprit in trapping
the solution at a node. The third- and fourth-order results are artificially constrained
at the boundary PD2 = 2.0 of the grid, so we reject those solutions, as well.
When we use PD2 = 1.8294, PH2 = 0.3834 to generate our next input
impedance, which is then projected back to the original reactance data and fed into
NLSE, we get the results shown in Table 16.26.
The third iteration, which uses the results for PD2 = 1.7256, PH2 = 0.4241 as its
input data (after projection), is shown in Table 16.27. Clearly, the 2nd-order results
are converging slowly to the nominal values of PD2 = 1.458, PH2 = 0.5512.
In addition to the examples using artificial pits as described in the earlier sections of
this chapter, validation of this methodology has also been made using a sample with
real corrosion pitting. The pits were generated by an accelerated corrosion process:
ASTM B117 (salt fog test chamber at 95 ◦ F, modified Exco solution, pH 3.2) in
7075-T6 aluminum. Figure 16.28 shows the methods of imaging these pits.
Figure 16.29 shows various images of pits as obtained by the methods shown
in Fig. 16.28. One of the more significant pitting regions was selected from a set
16.6 Results of a 2D Circumferential Sweep Feature Extraction Algorithm 335
Fig. 16.28 Sketch illustrating the methods of imaging real corrosion samples. (a) Laser pro-
filometry, (b) ultrasonic/eddy-current (far 1st layer), (c) eddy-current/ultrasonic (near 2nd layer),
(d) eddy-current/ultrasonic (near 2nd layer with 1st layer simulated pit)
Fig. 16.29 Real corrosion pits. Top left: photograph, Top right: laser profilometry image, Bottom
left: ultrasonic image (pits in far side of single layer), Bottom right: eddy-current image (pits in
near side of second layer, no corrosion in first)
of samples evaluated using laser profilometry (top right of Fig. 16.29). Ultrasonic
time-of-flight data were also acquired to evaluate the pit height and diameter (bottom
left of Fig. 16.29). The pit height was determined to be 0.020 by laser profilometry
and 0.026 using an ultrasonic method. However, there may be some error in the
laser profilometry results due to the presence of corrosion by-product. From the
image data for each method, an estimate of equivalent pit diameter was also made,
varying from 0.040 by laser profilometry and 0.100 using ultrasonic method. It is
336 16 Modeling Corrosion and Pitting Problems
Fig. 16.30 Real corrosion images. Left: eddy-current (far 1st layer at 6.2 kHz), Right: eddy-current
(near 2nd layer at 1.7 kHz)
likely that the pit diameter is closer to the value determined by laser profilometry
given beam spread in the ultrasonic measure. Using the proposed methodology for
second-layer pit characterization with eddy-current data at a single frequency of
6.2 kHz (bottom right of Fig. 16.29), the pit depth was estimated to be 0.029 and
the pit diameter was estimated to be 0.058. These results are in good agreement
with the known values for the pit and provide validation of this general approach. In
particular, the pairing of the iterative inversion (POCS) scheme for noise rejection
with an exhaustive search varying initial conditions was found to be beneficial for
achieving these results (Fig. 16.30).
Chapter 17
Applications to Aerospace Structures
Probe
0mm
6.35mm
1mm Head_Ti: σ = 1.098E6 S/m
7075 :
σ = 2.325 E7 S/m
4.064mm 2.032
Gap:
2.282
Shank_Ti:σ =1.098E6 S/m
Clip_2024:
σ = 1.728 E7 S/m 7.628mm
4.822
Gap:
5.072
Skin_2024:
σ = 1.728 E7 S/m
8.628
and a shank connecting all layers [102]. In the model, the layers and gaps make
up a uniform background or host (in x and y directions) with the fastener being an
“anomaly” requiring a local volume element mesh. Because the anomaly extends
through several layers of material with different electromagnetic properties, we must
use VIC-3D R
that has been augmented with the spatial-decomposition algorithm
(SDA) of Chap. 10.
Two model calculations were run based upon Fig. 17.1, the first with both
gaps filled with air, and the second with the top gap filled with 7075-aluminum
(σ = 2.325 × 107), and the other gap filled with 2024-aluminum (σ = 1.728 × 107).
The probe was a simple air-core coil with inner and outer radii of 3.02 and 5.14 mm,
respectively, and a height of 2.48 mm. Both model calculations were run with a
coil lift-off of 1.0 mm, and a frequency of 2,500 Hz. Five spatial-decomposition
grids were used to represent the “anomaly” regions, each with 8 × 8 × 2 cells,
yielding a total of 1,920 unknowns. The results are shown in Figs. 17.2 and 17.3.
The plot labeled “Gap” in Fig. 17.2 corresponds to the air-filled gaps of Fig. 17.1,
and the plot labeled “NoGap” corresponds to the case in which the gaps are filled
17.2 The Cessna Sandwich 339
Fig. 17.2 Model results for Model Results for the Cessna Double-Sandwich
the Cessna Double Sandwich
1
of Fig. 17.1
R:Gap
0.8 R:NoGap
X:Gap
Resistance/Reactance (Ohms)
X:NoGap
0.6
0.4
0.2
-0.2
-0.4
-0.6
-25 -20 -15 -10 -5 0 5
Probe Position (mm)
0.6 X:4x4x2
SDA grid. There are five grids X:8x8x2
in all. A linear extrapolation 0.4 X:16x16x2
of the 4 × 4 × 2 and 8 × 8 × 2
results is within a percent or 0.2
so of the 16 × 16 × 2 results
0
-0.2
-0.4
-0.6
-0.8
-25 -20 -15 -10 -5 0 5
Probe Position (mm)
with aluminum. The solution time for the “Gap” run with 26 probe scan points is
about 6 min on an AMD/Athlon machine, whereas the “NoGap” run took about
2 min. Figure 17.3 shows convergence of the Cessna-Sandwich Problem with grid
refinement.
340 17 Applications to Aerospace Structures
Fig. 17.4 Diagrams of fastener site FEM model with two gap layers between three aluminum
panels
Fig. 17.5 FEM and Results for Cessna Double-Sandwich with Ti Insert
VIC-3D R
model results for 0.6
the Cessna Sandwich with
two airgaps 0.5 VIC3D:R
Resistance/Reactance (Ohm)
X
0.4 Opera3D:R
X
0.3
0.2
0.1
-0.1
-0.2
-0.3
-25 -20 -15 -10 -5 0 5 10 15
Probe Position (mm)
Balance Coil
Shield
σ = 4.06E6
μ = 40
Cap Cap
Test Coil
Fig. 17.6 The ring probe
17.3 Simulated Studies on Crack Characteristics: The Cessna Setups 343
σ = 4.06 E6 S/m
μ = 40 0.112"
0.1635"
0.05" Al
0.00196"
0.00196"
SETUP 0
0.054"
0.05" Al Crack
Width=0.0107"
0.00196"
0.00196"
SETUP 1
The test (bottom) and balance (top) coils of the ring probe shown in Fig. 17.6 are
connected in two arms of a bridge within the eddy-current test instrument, as shown
in Fig. 17.13. Zt is the driving-point impedance of the test coil in the presence of the
workpiece (the “loaded impedance”) and Zb is the driving-point impedance of the
balance coil. By the construction of the ring probe, as shown in Fig. 17.6, the cap
344 17 Applications to Aerospace Structures
0.05" Al
0.00196"
0.00196"
SETUP 2 0.1397"
0.0620"
0.05" Al Crack
Width=0.0087"
0.00196"
0.00196"
SETUP 3
0.05" Al
0.00196"
0.00196"
SETUP 4 0.1472"
separating the test and balance coils isolates them electrically, so we assume that the
mutual inductance between the two coils is small enough to ignore. That is, the cap
is a “flux-shunt” that shunts the flux produced by either coil from linking the other
coil. For the same reason, the balance coil is isolated from the workpiece, because
the workpiece is beneath the cap. This means that the driving-point impedance of
the balance coil will always be the freespace impedance, no matter where the probe
is positioned.
17.3 Simulated Studies on Crack Characteristics: The Cessna Setups 345
R0 R0
+
Ein Vout
−
+ −
Zt Zb
The bridge circuit of Fig. 17.13 is designed so that R0 is much larger than either
Zt or Zb . Under this condition, and the condition of zero mutual inductance between
the two coils, it is a simple matter to compute the response of the bridge circuit to be
Ein
Vout ≈ (Zt − Zb ) , (17.1)
R0 + 2Rin
which demonstrates that the bridge circuit acts as an impedance differentiator, and
that the impedance measured is balanced against the freespace impedance of the test
coil. This follows because the freespace impedance of the test and balance coils are
equal, because the coils are identically situated when in freespace. The output of
the test instrument is the “total signal” minus the “freespace signal,” which, in the
terminology of VIC-3D R
, corresponds to the “host + flaw impedance.”
The output of VIC-3D R
is already balanced against both the freespace and host
medium, so it appears that in order to compare the test-output with the VIC-3D R
output, we would have to add the host-only impedance, which will be computed in a
separate model. (All of the VIC-3D R
calculations will be done without the need to
model the balance coil.) Then we will scale the experimental data so that they agree
with the VIC-3D R
-generated model data and produce a scale-factor, β , which will
then be used on all other problems at the same frequency.
The practical upshot of all of this is that because the subtraction (balancing) of
the impedances is done analogically (in a bridge circuit), the probe requires two
identical coils, with identical shielding. If the balancing were done numerically, one
would need only the test coil, which would result in a much less expensive probe.
Data for determining the presence of a flaw are obtained by varying the lift-off of the
ring probe over the range 0, 0.25, 0.7493, and 0.9398 mm. By “lift-off” we mean
the position of the bottom of the probe above the workpiece. The solution of the
346 17 Applications to Aerospace Structures
Fig. 17.14 Final scaled Final Scaled Model Curves for SETUPS 0-5
model results for Setups 0–5,
0.35
using Setup 0 as the 0
reference. The horizontal and 1
vertical axes are both 0.3 2
3
measured in instrument volts, 4
V. The parametric points on 0.25 5
the curves are lift-off values
Channel 1 (V)
of 0, 0,25, 0.7493, and
0.9398 mm (right to left) 0.2
0.15
0.1
0.05
0
-3 -2.5 -2 -1.5 -1 -0.5 0 0.5
Channel 2 (V)
problem with 27,648 unknowns takes about 35 min for each lift-off, giving a total
solution time of about 2.3 h for each setup. About 1Gb of storage for auxiliary files
was required.
VIC-3D R
computes impedances, but the industry typically uses analog
instruments to measure data, so the final results of the computation are transformed
into “instrument volts,” as shown in Fig. 17.14. Channel 1 is called the vertical
channel (on the oscilloscope) and indicates the difference signal between flawed
and unflawed rivet. Channel 2 is the horizontal channel and corresponds to lift-
off. The markers correspond to the four lift-off values of 0(mm), 0.25, 0.7493, and
0.9398, from right to left. Both channels (axes) are measured in instrument volts.
We showed in Chap. 15 how impedance calculations can be transformed into
voltage measurements using calibration data with known samples (see also [102]).
Setup 0 is the reference, so it is zero for each lift-off along the negative real axis.
The remaining curves indicate the presence of a flaw and yield information that can
be used in an inversion procedure.
It is clear that Setup 2 produces the largest signal, even though the flaw is in the
second layer. This is due to the fact that the flaw in this setup has a larger volume
than any of the other flaws. Furthermore, it is clear that the flaws in Setups 1 and
3, even though they are in the upper layer, are more difficult to distinguish than the
other two flaws. This is due to the fact that these two flaws have a smaller volume
than those in Setups 2 and 4. Setup 5 is the same as Setup 4, except that the length of
the crack is 0.0986 in., instead of 0.1472 in. The response of Setup 5 is about 0.67
times the response of Setup 4, which is consistent with the fact that the response is
roughly proportional to the volume of the defect.
17.4 Further Tests with the SDA and “Weak-Host” Layers 347
The setup for this exercise is shown in Fig. 17.15. The insert can be ferromagnetic,
and we have run tests with σ = 1.79 × 106 S/m and μ = 200. In those runs, we
used three SDA grids, each with 32 × 32 × 4 cells, yielding over 62,000 unknowns.
The convergence to 1 % took 251 iterations. With a nonpermeable insert, the number
of unknowns is reduced to 33,024, and convergence to 1 % required 431 iterations.
Apparently the permeability yields a better conditioned system. In those runs, the
middle layer conductivity was σ3 = 0, which made it a “weak-host” layer.
Our interest, now, is in testing the effect of varying σ3 over several decades
in order to determine the effect of a nonzero conductivity in speeding up the
convergence of a problem in which σ = 0. For this exercise we made the insert
nonpermeable, keeping the conductivity equal to 1.79 × 106 S/m, and used a
convergence criterion of 1 %. The convergence results with respect to σ3 are shown
in Table 17.1.
348 17 Applications to Aerospace Structures
8x8
σ=1.876Ε7
μ=1 0.8128
Insert
σ, μ
σ3 0.25
μ=1
σ=1.876Ε7
μ=1
0.8128
The immediate observations are that the model is robust and consistent, in the
sense that there is a monotonic increase in the number of iterations required to
achieve 1 % convergence. This is an indication that the problem becomes slightly
less well conditioned as σ3 decreases, which is to be expected. Further indication of
the consistency in the results is that the magnitude of δ Z increases monotonically
with a decrease in σ3 , as we would expect because the electrical contrast of the
anomaly increases with a decrease in σ3 in the middle slab.
The most important observations, however, are that the phase of δ Z is virtually
independent of the value of σ3 , which means that the magnitude of δ Z, alone, is
sufficient to determine the effects of changing the conductivity of the slab, and
furthermore, the magnitudes change by less than 1 %, which means that we could
safely replace a slab of air with one of σ3 = 105 S/m and get the same answer, while
achieving a significant reduction of computation time. This could be extremely
important when doing a large scan. We shall see how this can be very effective
in the next example.
This sandwich is shown in Fig. 17.16. Each of the three SDA grids has 16 × 16 × 2
cells, which produces a total of 768 electric unknowns and 498 magnetic unknowns.
17.4 Further Tests with the SDA and “Weak-Host” Layers 349
0.25 x 1.25
σ=1.876E7
μ=1 0.156
σ3 Insert 0.004
σ=1.875E6
σ=1.876E7
μ=200
μ=1 0.58
Fig. 17.16 Setup for Half-Sandwich No. 2. All dimensions are in inches
y2 = 4a + 4b + c
y3 = 9a + 6b + c
y4 = 16a + 8b + c. (17.2)
where c is the sought-for solution for σ3 = 1, which is close enough to zero for our
purposes.
Referring to Table 17.2, we take y4 to be the entry for either δ R or δ X for σ3 =
104 , and so on for y3 and y2 . Using these values, we compute the entries for δ R, δ X,
and δ Z in the fifth row “(Extrap.)”. Our strategy, then, is to do the VIC-3DR
model
scans using σ3 = 10 S/m, and then multiplying these results by the ratio
4
8.4427e− j143.63
= 1.3419e− j0.82
6.2916e− j142.81
= 1.3418 − j0.0192 . (17.4)
This is the same as Half-Sandwich No. 1, except that the gap thickness is 0.025 mm,
and the top and bottom halves of the sandwich are 0.9253 mm thick, yielding the
same overall height of the sandwich of 1.8756 mm. Our interest is in determining
the effect of gap thickness on convergence. For Half-Sandwich No. 3, after 1,001
iterations, the convergence criterion was reduced to only 6.559 %, which clearly
shows the deleterious effect that a very thin gap has on convergence. This supports
our contention that the slow convergence of Half-Sandwich No. 2 is largely due to
the very thin airgap.
We have shown that the volume-integral approach has significant advantages over
other numerical methods such as the finite-element method, in that the formulation
of the numerical model is much simpler in the former than in the latter. Furthermore,
the solution time for VIC-3D R
is extremely short for many problems in NDE,
17.5 Comments and Conclusions 351
because the formulation in terms of the Galerkin variant of the method of moments
on a regular grid results in operators that have very special structures; they are
either three-dimensional convolutions, or two-dimensional convolutions and one-
dimensional correlations, which means that we can use three-dimensional FFT’s
to evaluate them in a conjugate-gradient search algorithm. The use of a highly
irregular mesh in the finite-element technique does not allow a similar advantage
in the solution process.
This advantage accrues from the very different nature of the physics that goes
into the formulation of the mathematical models. In volume-integral equations, as
well as boundary-integral equations, the unknowns are anomalous currents that are
supported in a compact domain, namely the domain of the anomaly; in the example
of the Cessna Series, the anomaly is the rivet or rivet and crack. In finite-element or
finite-difference methods the unknowns are the electric and magnetic fields, which
extend to infinity. This has two disadvantages; it increases the number of unknowns
and requires some method of approximating the “boundary-at-infinity” in order to
truncate the problem domain. This increases the complexity in simply defining the
model and presents an extreme challenge to prospective users who are not skilled in
computational electromagnetics. Furthermore, the FEM method is not particularly
well suited to solve typical problems in NDE, because the anomalies, such as rivets
or cracks, require a very complicated mesh, with a large number of very irregular
cells.
Finally, we conclude from the model calculations of the Cessna Series that the
spatial decomposition algorithm is a very efficient method of solving problems
in which an anomaly extends through layers with different electric and magnetic
properties. The spatial decomposition algorithm formulation was also demonstrated
to be valid for cases where a conducting anamoly is present in a nonconducting layer
such as air. This approach greatly expands the capability of volume-integral equation
methods for complex problems in computational electromagnetics, and in particular
nondestructive evaluation. Lastly, rapid solution of the forward problem for com-
putational electromagnetics will also be beneficial for the practical application of
advanced inverse method techniques for quantitative nondestructive evaluation of
material discontinuities.
Chapter 18
Applications to Nuclear Power
Fig. 18.1 Photomicrographs showing the effects of long-term, under deposit pitting corrosion of
a heat exchanger tube whose nominal wall thickness was 49–51 mils. Upper-left: little corrosion;
Upper-right: remaining wall of 29 mils; Lower-left: remaining wall of 20 mils; Lower-right: actual
pits, with an indication of 88 % through-wall within the middle third of the sample shown, based
on nominal 0.049 inch
cooling heat exchanger tube whose nominal wall thickness is 49–51 mils.1 They
further suggest that we can construct a reasonable model of corrosion pits as
three-dimensional semiellipsoids that originate in the inner-radius of tube walls.
Figure 18.2 illustrates the model, together with its three defining parameters, the
semiaxes, A, B, and C. By modeling the pit in this manner, we cast the problem into
the model-based inversion paradigm. In this case, we hope to determine the three
semiaxes, thereby solving the inverse problem. Clearly, C determines the depth of
the pit.
Furthermore, the photographs give us insight into the sizes of typical pits, which
can then be translated into values of A, B, and C for typical pits. Indeed, from
these photographs we see that the pits appear to be circular, when looking into the
inner surface of the tube, with a nominal radius of 0.0625 inch. Hence, we model
these pits by setting A and B equal to 0.0625. We then choose C to be equal to
the difference between 0.049 inch (the nominal wall thickness) and the minimum
measured wall thickness, as given in the photomicrographs. We assume that the pits
are empty, but, it would be quite easy to model pits filled with corrosion deposits,
whether conducting or magnetic (or both).
1 These photographs and the data concerning them are courtesy of Exelon PowerLabs.
18.1 Modeling Pitting and Corrosion Phenomena in Heat-Exchanger Tubes 355
TUBE AXIS
B
TUBE WALL
B
INNER SURFACE
OF TUBE WALL
Fig. 18.2 Illustrating a model of pits as three-dimensional semiellipsoids that originate in the
inner-radius of a tube wall. The ellipsoid is defined by its three semiaxes, A, B, and C
Table 18.1 Measured pit data for a component cooling heat exchanger tube
Minimum Calculated % thru-
measured wall based on Eddy current
Indication remaining wall (in) Pit depth (in) nominal 0.049 in % thru wall
No. 3 0.023 0.026 53 59
No. 4 0.015 0.034 69.4 88
No. 5 0.020 0.029 59.2 89
No. 7 0.016 0.033 67.3 65
No. 8 0.015 0.034 69.4 97
No. 9 0.017 0.032 65.3 100
The pit data of Table 18.1 are measurements taken on the component cooling heat
exchanger tubes of Fig. 18.1. The first calculations that we make using VIC-3D R
are to determine the impedance signature versus depth for the various pits labeled
“Indication Nos. 3–9,” omitting No. 8, because it has the same depth as No. 4. In all
cases, we fixed A and B to be equal to 0.0625 inch, as explained above. The results
are shown in Fig. 18.3.
356 18 Applications to Nuclear Power
Reactance (Ohms)
five, and C is equal to the pit 0.2
depth 0.1
-0.1
-0.2
-0.3
-0.4
-0.5
-0.05 -0.04 -0.03 -0.02 -0.01 0 0.01 0.02 0.03 0.04 0.05
Resistance (Ohms)
Table 18.2 Estimate of A, B, and C for the semiellipsoidal representation of the pits
labeled Indication Nos. 3–9
Indication Original model data (in) Reconstructed (inverted) data (in)
No. 3 A = 0.0625, B = 0.0625, C = 0.026 A = 0.0643, B = 0.0634, C = 0.0263
No. 4 A = 0.0625, B = 0.0625, C = 0.034 A = 0.0644, B = 0.0639, C = 0.0325
No. 5 A = 0.0625, B = 0.0625, C = 0.029 A = 0.0645, B = 0.0637, C = 0.0285
No. 7 A = 0.0625, B = 0.0625, C = 0.033 A = 0.0645, B = 0.0639, C = 0.0316
No. 8 A = 0.0500, B = 0.0700, C = 0.034 A = 0.0550, B = 0.0686, C = 0.0328
No. 9 A = 0.0625, B = 0.0625, C = 0.032 A = 0.0645, B = 0.0639, C = 0.0309
Now we consider the inversion problem in detail. We take the model data of
Fig. 18.3 and use them to simulate measured impedance data, which are then
submitted to NLSE. The output of the estimator is an estimate of the semiaxes,
A, B, and C of the ellipsoidal pit, as displayed in Table 18.2. It only takes a few
minutes to compute a table for the inversion algorithm, which is then used for all
pits. Subsequently, it takes about 2–4 s to compute the results for each pit.
Indication No. 8 was handled differently from the others. Because the X-Y
cross-section of this pit was not circular, we found it more reliable to acquire data
for it using a surface pancake probe scanned in a two-dimensional raster pattern.
These data were then submitted to the estimator exactly as were the other five.
The results displayed in Table 18.2 indicate that the algorithm for nonlinear
estimation is reliable and robust. The maximum error occurs in A for Indication
No. 8, and is only 10 %, which is quite tolerable for this dimension (the width
of the pit). The important point to be made from this study is the reliability in
reconstructing C, the semiaxes that determines the depth of the pit. Indeed, the
results of the algorithm indicate that the computation is more sensitive to depth
then to the lateral dimensions, which is exactly what we’re looking for!
Tube Axis
Bobbin
Coil
0.0625 in 0.0625 in 0.0625 in 0.0625 in
0.026 in
0.029 in 0.034 in
Tube Wall
Fig. 18.4 Illustrating a complex pit that comprises three intersecting semiellipsoids, whose
dimensions are shown
5
0.5
0.4 4
0.3 3
0.2
2
0.1
1
0
-0.1 0
-0.2 -1
-0.5 -0.4 -0.3 -0.2 -0.1 0 0.1 0.2 0.3 0.4 0.5 -0.5 -0.4 -0.3 -0.2 -0.1 0 0.1 0.2 0.3 0.4 0.5
Probe Position (in) Probe Position (in)
Fig. 18.5 The model input impedance data to NLSE at 10, 20, 40, 70, 200, and 700 kHz: resistance
(left), reactance (right)
Stainless−Steel Block
Semi−Ellipsoidal Pit
Coil
Acrylic Box
Fig. 18.6 A multifrequency validation test. The coil is excited at 10 frequencies between 0.1 and
1.0MHz
0.2
0.1
0.05
-0.05
-0.1
-0.15
-0.2
-0.25
0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1
Frequency (MHz)
of the pit. The dimensions of the pit semiaxes are: A = 0.0737 in., B = 0.0737 in.,
C = 0.0279 in. The solution is least sensitive to the lateral coordinates of the probe
relative to the pit, and most sensitive to the depth of the pit, as we noted before.
This validates the NLSE inversion process, as it is applied to semiellipsoidal pit
models. Further, this process illustrates another method for obtaining data when
a physical scan is difficult or impossible. It would be impossible to infer the
solution for five parameters by using analog instruments and checking traces on
an oscilloscope. This, again, demonstrates the value of model-based inversion in
eddy-current NDE.
0.049"
" "
0.2 0.2 0.2"
0.326"
Tube Axis
3.5
Resistance/Reactance (Ohms)
2.5
1.5
0.5
-0.5
-1
-0.8 -0.6 -0.4 -0.2 0 0.2 0.4 0.6 0.8
Probe Position (mm)
is to determine the radii of the three rings. The model input data are the scanned
impedances associated with the configuration shown in the figure, and are shown
in Fig. 18.9.
In order to invert these data using up to a fourth-order polynomial, we first
establish a 5 × 5 × 5 interpolation table, in which each wall-loss variable has nodes
at 0.0005, 0.003, 0.0055, 0.008, and 0.0105 inch. The blending functions associated
with the 1st and 125th nodes are shown in Fig. 18.10, along with the input data of
Fig. 18.9.
The results of the NLSE inversion are tabulated in Table 18.4.
362 18 Applications to Nuclear Power
Blending Functions and Input Data Blending Functions and Input Data
0.2 6
Data
BF1
0 5
BF125
Resistance (Ohms)
Reactance (Ohms)
-0.2 Data
BF1 4
BF125
-0.4
3
-0.6
2
-0.8
-1 1
-1.2 0
-0.8 -0.6 -0.4 -0.2 0 0.2 0.4 0.6 0.8 -0.8 -0.6 -0.4 -0.2 0 0.2 0.4 0.6 0.8
Probe Position (mm) Probe Position (mm)
Fig. 18.10 Illustrating the 1st and 125th blending functions, together with the input data
Table 18.4 Results of the NLSE inversion for the wall-thinning problem
Order Φ Wall-loss1(in)/sensitivity Wall-loss2(in)/sensitivity Wall-loss3(in)/sensitivity
1 0.05056 0.004028/0.6127(−4) 0.008/0.8104(−4) 0.004028/0.6127(−4)
2 0.03609 0.003939/0.3948(−4) 0.00797/0.5103(−4) 0.003939/0.3948(−4)
3 0.04160 0.0039175/0.4593(−4) 0.00795/0.5578(−4) 0.0039175/0.4593(−4)
4 0.04734 0.0039085/0.5220(−4) 0.007945/0.5905(−4) 0.0039085/0.5220(−4)
3 Themeasured data and parameters of this experiment were provided by J. Liu and N. Trepal of
the Naval Surface Warfare Center, Carderock Division.
18.2 Model-Based Inversion of Measured Impedance Data: The Carderock Problem 363
7.2 Model
Exp 22.6 Model
7.15 Exp
22.4
Resistance(Ohms)
7.1
Reactance(Ohms)
7.05
22.2
7
22
6.95
6.9 21.8
6.85
21.6
6.8
6.75 21.4
-0.6 -0.4 -0.2 0 0.2 0.4 0.6 -0.6 -0.4 -0.2 0 0.2 0.4 0.6
Probe Position (in) Probe Position (in)
Fig. 18.11 Model and experimental data for a 0.25 inch, through-wall round hole model of a pit
at 88 kHz
Balanced Results for the 0.25-in Hole Balanced Results for the 0.25-in Hole
6.9 22.8
Reactance(Ohms)
22.5
6.84
22.4
6.82
22.3
6.8
22.2
6.78
22.1
6.76 22
6.74 21.9
-0.6 -0.4 -0.2 0 0.2 0.4 0.6 -0.6 -0.4 -0.2 0 0.2 0.4 0.6
Probe Position (in) Probe Position (in)
Fig. 18.12 Model and experimental data for a 0.25 inch, through-wall round hole model of a pit
at 88 kHz after the host-only offset has been removed
without the cable.4 Figure 18.12 compares the responses when the model results are
“balanced” against the measured data by forcing the host-only offsets to be removed.
Now the l2 -norm error is less than 0.1 %.
A comparison of the differential-bobbin responses is shown in Fig. 18.13. It is
clear from this figure that noise in the measured data will produce a Lissajous figure
Fig. 18.13 Comparing the Differential Bobbin Results for the 0.25-in Hole
differential-bobbin 0.8
experimental and VIC-3DR
VIC-3D
responses Experiment
0.6
0.4
0.2
X(Ohms)
0
-0.2
-0.4
-0.6
-0.8
-0.15 -0.1 -0.05 0 0.05 0.1 0.15
R(Ohms)
11 12
0.3 9 10
5 6 7 8
0.2
1 2 3 4
0.1 Wall Loss (%)
25 50 75 100
that will not allow a very accurate assessment of the size and depth of the pit. This
is the reason for doing model-based inversion, as we alluded to in the preceding
section.
The first step in applying model-based inversion to this problem is to develop
an interpolation table for the variables of interest. Figure 18.14 illustrates one such
table for the variables of diameter and wall-loss (depth) of a round-hole. We use
VIC-3D R
to generate the standards (blending-functions) associated with each node
of the interpolation table. These functions are shown in Figs. 18.15–18.18.
18.2 Model-Based Inversion of Measured Impedance Data: The Carderock Problem 365
Blending Functions for Nodes 1-4 Blending Functions for Nodes 1-4
6.83 22.1
6.825
22.08 1
6.82 1 2
2 3
22.06 4
Resistance(Ohms)
6.815 3
Reactance(Ohms)
4
6.81
22.04
6.805
22.02
6.8
6.795 22
6.79
21.98
6.785
6.78 21.96
-0.6 -0.4 -0.2 0 0.2 0.4 0.6 -0.6 -0.4 -0.2 0 0.2 0.4 0.6
Probe Position (in) Probe Position (in)
Blending Functions for Nodes 5-8 Blending Functions for Nodes 5-8
6.88 22.5
22.45
6.86 5
6 22.4 5
7 6
6.84 8 22.35 7
Resistance(Ohms)
Reactance(Ohms)
8
22.3
6.82
22.25
22.2
6.8
22.15
6.78 22.1
22.05
6.76
22
6.74 21.95
-0.6 -0.4 -0.2 0 0.2 0.4 0.6 -0.6 -0.4 -0.2 0 0.2 0.4 0.6
Probe Position (in) Probe Position (in)
The results when the experimental data of Fig. 18.12 are used as the input to
NLSE are shown in Table 18.5. The heading labeled “Order” denotes the order of
the interpolating polynomial and Φ denotes the norm of the residuals. Recall that
“sensitivity” is the parameter that determines the sensitivity of Φ to that unknown—
the smaller the value of “sensitivity” the more sensitive to that parameter. In addition
to Φ , the sensitivity parameter is important in determining the quality of the
solution, especially in making a judgement as to the importance of an unknown
parameter, or in determining the confidence one has in the value of that parameter
at the solution.
The corresponding results when the model data of Fig. 18.12 are used as the
input to NLSE are given in Table 18.6. The purpose of showing these results is to
366 18 Applications to Nuclear Power
Blending Functions for Nodes 9-12 Blending Functions for Nodes 9-12
6.92 23.2
6.9 9
10 23
6.88 11 9
12 10
22.8
Resistance(Ohms)
Reactance(Ohms)
6.86 11
12
6.84
22.6
6.82
22.4
6.8
6.78 22.2
6.76
22
6.74
6.72 21.8
-0.6 -0.4 -0.2 0 0.2 0.4 0.6 -0.6 -0.4 -0.2 0 0.2 0.4 0.6
Probe Position (in) Probe Position (in)
Blending Functions for Nodes 13-16 Blending Functions for Nodes 13-16
7 23.8
23.6
6.95 13 13
14 23.4 14
15 15
16
Resistance(Ohms)
23.2
Reactance(Ohms)
6.9 16
23
6.85 22.8
22.6
6.8
22.4
22.2
6.75
22
6.7 21.8
-0.6 -0.4 -0.2 0 0.2 0.4 0.6 -0.6 -0.4 -0.2 0 0.2 0.4 0.6
Probe Position (in) Probe Position (in)
establish a standard for comparison. The fact that Φ in Table 18.5 is only about an
order of magnitude larger than that in Table 18.6 indicates that the results shown in
Table 18.5 are acceptable.
18.3 Model-Based Inversion of Measured Instrument Data... 367
Even though these results are acceptable, we can improve them by performing
the iterative-refinement algorithm that is based upon projections onto convex sets
(POCS) (recall Chap. 14, Fig. 14.3). We start the next iteration of POCS by
computing the impedance response to a hole that has a diameter of 0.2479 inch and a
depth of 0.0601 inch. These are the estimated parameters shown in Table 18.5. Then
we take the computed resistance as the next estimate of the “correct” resistance,
but replace the computed reactance by the original measured reactance. This keeps
us in touch with the original measurements. Using the “corrected resistance” and
the original reactance, we rerun NLSE and get the new inversion results shown in
Table 18.7. Clearly, this is a significant improvement over the previous results of
Table 18.5, both with respect to a smaller Φ , and for a more accurate estimate of
the diameter and depth of the hole, with smaller sensitivity values. Note that in all
of our inversion experiments, we find that the sensitivity in the depth parameter is
better (smaller) than in the diameter. This is typical for problems, such as corrosion
modeling, in which these are the two parameters of interest.
Figure 18.19 illustrates the model of the EPRI ID Pits test problem.5 The parameters
of the coil are estimates, only. Each coil is a bobbin coil, and their individual
responses are called “absolute.” When the two coils are connected in a bridge circuit,
they form a differential bobbin, whose response is essentially the derivative of the
5 The test setup and measured data for this section were supplied by K. Krzywosz of the Electric
Power Research Institute (EPRI) as part of its benchmark validation test for sizing inner-diameter
(ID) pit models.
368 18 Applications to Nuclear Power
0.028 ID PIT
0.472 OR
IR
SS: σ = 1.4E6 , μ = 1
Fig. 18.19 Schematic of the EPRI ID Pits test problem. The parameters of the coil are
estimates, only
Table 18.8 Labeling the Pit no. Diameter (in) Height (in) Wall loss (%)
999-series of round-bottom
pits 1 0.0625 0.007 25
2 0.0625 0.014 50
3 0.0625 0.021 75
4 0.0625 0.028 100
5 0.09375 0.007 25
6 0.09375 0.014 50
7 0.09375 0.021 75
8 0.09375 0.028 100
9 0.125 0.007 25
10 0.125 0.014 50
11 0.125 0.021 75
12 0.125 0.028 100
50
25 Diameter
1/32 1/16 3/32 1/8 (inches)
The measured data are instrument voltages, which means that they must be
transformed into impedances in order to be used as a source for inversion with VIC-
3D R
. Following the discussion in Chap. 15, we accomplish this by using a general
linear filter,
Z = β1 (C1 + jC2 ) − β2, (18.1)
The preprocessing and feature extraction of the experimental data are a key part of
the inversion process. The quality of the inversion results from NDE measurement
scans using in-field equipment can be improved by using the appropriate prepro-
cessing steps. Understanding the characteristics of the measurement data and how
they were acquired is important in designing the preprocessing steps. Here, a series
of steps were implemented to automate the extraction and filtering of the pit features
found in the EPRI scan data to locate the exact center of the pit responses. A cross-
correlation of the measured data with a selected calibration pit response curve was
used to locate the exact center of the pit responses. Subsequent steps addressing
background noise filtering and scan-step correction are presented in detail below.
Lastly, symmetry was used to average the experimental response about the pit.
A novel background clutter removal algorithm was developed to compensate for
systematic measurement noise in scan data. Typical sources for clutter and noise
found in eddy-current scans are probe lift-off variation, local changes in material
properties, sensitivity to sub-surface structures, thermal variation of measurement
components, and electrical noise. Figure 18.21a displays an example where the
background noise of the original experimental response is evaluated, using the
procedure described in Chap. 15, providing efficient removal of background clutter
from the experimental data. Compensation for the systematic background noise
will reduce the role that the tails of the pit response play in the least-squares
error minimization and emphasize the center pit region achieving a more accurate
model fit.
A method was also developed to correct for the significant variation in the scan
step size found between different pits in a scan. Since the data were acquired using
a constant rate of acquisition (1,000 Hz), it was observed that the pull velocity of the
probe through the tube actually varied by as much as 30 % with respect to the mean
rate. This resulted in uncertainty in the actual scan step size in the neighborhood
of each pit in the test sample. Thus, it was concluded that the local scan step size
must be estimated for each pit. To determine the local scan step size, an inversion
process was designed that iteratively fit both the equivalent depth and diameter of
the pit and the scan step size. Figure 18.21b displays the reactance response for pit 7
for the measured data with and without scan-step correction and the corresponding
simulated data. Since the scan step size more greatly affects the lateral fit of the data,
while the depth and diameter of the pit more greatly impact the fit in both magnitude
and at the pit center, the inversion of these three parameters simultaneously was
found to be well-posed over the parameter range tested in the EPRI study.
18.3 Model-Based Inversion of Measured Instrument Data... 371
Fig. 18.21 (a) Reactance response for pit 6 demonstrating background clutter removal.
(b) Reactance response for pit 7 demonstrating scan-step correction (for 008–999 at 1.0 MHz)
372 18 Applications to Nuclear Power
Table 18.9 Summary of Pillbox Model Results. “Multi” denotes a multifrequency combination of
all four frequencies, using unnormalized data. The sensitivity to the solution of the final (“Multi”)
estimates of D and H are also given
1 MHz 800 kHz 600 kHz 400 kHz Multi
Pit D H D H D H D H D H
Sensit Sensit
1 0.0433 0.0084 0.0469 0.007 0.0470 0.007 0.0438 0.007 0.0463 0.007
0.0082 0.0028
2 0.0550 0.0108 0.0554 0.0105 0.0558 0.0101 0.0534 0.0103 0.0557 0.0103
0.0044 0.0015
3 0.0578 0.0179 0.0574 0.0175 0.0570 0.0171 0.0538 0.0180 0.0576 0.0173
0.0053 0.0029
4 0.0699 0.028 0.0695 0.028 0.0692 0.028 0.0689 0.028 0.0695 0.028
0.0033 0.0022
5 0.0513 0.007 0.0510 0.007 0.0502 0.007 0.0484 0.007 0.0509 0.007
0.0083 0.0024
6 0.0897 0.0106 0.0904 0.0103 0.0952 0.0092 0.0964 0.0087 0.0911 0.0101
0.0051 0.0010
7 0.0999 0.0173 0.1002 0.0169 0.0985 0.0164 0.0975 0.0162 0.0990 0.0166
0.0033 0.0010
8 0.0992 0.028 0.0990 0.028 0.0987 0.028 0.0984 0.028 0.0990 0.028
0.0044 0.0017
9 0.1021 0.007 0.1015 0.007 0.1004 0.007 0.0973 0.007 0.1012 0.007
0.010 0.0014
10 0.1244 0.0132 0.125 0.0128 0.125 0.0124 0.125 0.0118 0.125 0.0127
0.0036 0.0006
11 0.125 0.0184 0.125 0.0180 0.125 0.0177 0.125 0.0178 0.125 0.0180
0.0046 0.0009
12 0.125 0.0268 0.125 0.0268 0.125 0.0267 0.125 0.0265 0.125 0.0267
0.0071 0.0015
We consider the inversion of the transformed (scaled) data at all four frequencies,
1MHz, 800kHz, 600kHz, and 400kHz, first individually, and then collectively in
one NLSE run. In the multifrequency run, the data are not normalized to account for
frequency differences. The results are summarized in Table 18.9.
The distribution of pit diameters and depths as listed in Table 18.9 are compared
to the nominal values in Fig. 18.22. It appears that the mechanism that produces
the round-bottom of the shallowest pits, namely 1, 5, and 9, does not allow a fully
developed pit to form, and this may account for the smaller than expected diameters
of these pits. This is reminiscent of the need for a “new flaw model” for pits, as
described in Sect. 16.4.
18.3 Model-Based Inversion of Measured Instrument Data... 373
0.03
0.14
nominal nominal
VIC-3D 0.025 VIC-3D
0.12
Pit Diameter (in)
0.08
0.015
0.06
0.01
0.04
0.005
0.02
0 0
-4 -2 0 2 4 6 8 10 12 14 -4 -2 0 2 4 6 8 10 12 14
Pit No. Pit No.
Fig. 18.22 Distribution of computed and nominal pit diameters and depths
The final results for each pit response at 1 MHz, when the diameter and height
shown in Table 18.9 under “Multi” are used, are plotted in the next figures.
0.15
0.02 Measured
Measured 0.1 Computed
Computed
Resistance(Ohms)
Reactance(Ohms)
0.01
0.05
0 0
-0.05
-0.01
-0.1
-0.02
-0.15
-0.03 -0.2
-0.5 -0.4 -0.3 -0.2 -0.1 0 0.1 0.2 0.3 0.4 0.5 -0.5 -0.4 -0.3 -0.2 -0.1 0 0.1 0.2 0.3 0.4 0.5
Probe Position (in) Probe Position (in)
Reactance(Ohms)
0.01 0.1
0 0
-0.01 -0.1
-0.02 -0.2
-0.03 -0.3
-0.04 -0.4
-0.5 -0.4 -0.3 -0.2 -0.1 0 0.1 0.2 0.3 0.4 0.5 -0.5 -0.4 -0.3 -0.2 -0.1 0 0.1 0.2 0.3 0.4 0.5
Probe Position (in) Probe Position (in)
374 18 Applications to Nuclear Power
0.4
0.06
Measured
Measured 0.3 Computed
0.04 Computed
Resistance(Ohms)
Reactance(Ohms)
0.2
0.02
0.1
0 0
-0.1
-0.02
-0.2
-0.04
-0.3
-0.06
-0.4
-0.08 -0.5
-0.5 -0.4 -0.3 -0.2 -0.1 0 0.1 0.2 0.3 0.4 0.5 -0.5 -0.4 -0.3 -0.2 -0.1 0 0.1 0.2 0.3 0.4 0.5
Probe Position (in) Probe Position (in)
Results for Pit 4 at 1MHz Results for Pit 4 at 1MHz
0.4 1.5
0.3
Measured 1
0.2 Computed Measured
Resistance(Ohms)
Computed
Reactance(Ohms)
0.5
0.1
0 0
-0.1
-0.5
-0.2
-1
-0.3
-0.4 -1.5
-0.5 -0.4 -0.3 -0.2 -0.1 0 0.1 0.2 0.3 0.4 0.5 -0.5 -0.4 -0.3 -0.2 -0.1 0 0.1 0.2 0.3 0.4 0.5
Probe Position (in) Probe Position (in)
0.03 0.15
Measured Measured
0.02 0.1
Computed Computed
Resistance(Ohms)
Reactance(Ohms)
0.01 0.05
0 0
-0.01 -0.05
-0.02 -0.1
-0.03 -0.15
-0.04 -0.2
-0.5 -0.4 -0.3 -0.2 -0.1 0 0.1 0.2 0.3 0.4 0.5 -0.5 -0.4 -0.3 -0.2 -0.1 0 0.1 0.2 0.3 0.4 0.5
Probe Position (in) Probe Position (in)
18.3 Model-Based Inversion of Measured Instrument Data... 375
0.08
0.6
Measured
0.06 Computed Measured
0.4 Computed
Resistance(Ohms)
Reactance(Ohms)
0.04
0.2
0.02
0 0
-0.02
-0.2
-0.04
-0.4
-0.06
-0.6
-0.08
-0.1 -0.8
-0.5 -0.4 -0.3 -0.2 -0.1 0 0.1 0.2 0.3 0.4 0.5 -0.5 -0.4 -0.3 -0.2 -0.1 0 0.1 0.2 0.3 0.4 0.5
Probe Position (in) Probe Position (in)
Results for Pit 7 at 1MHz Results for Pit 7 at 1MHz
0.15 1.5
Measured
0.1 Computed 1
Measured
Computed
Resistance(Ohms)
Reactance(Ohms)
0.05 0.5
0 0
-0.05 -0.5
-0.1 -1
-0.15 -1.5
-0.5 -0.4 -0.3 -0.2 -0.1 0 0.1 0.2 0.3 0.4 0.5 -0.5 -0.4 -0.3 -0.2 -0.1 0 0.1 0.2 0.3 0.4 0.5
Probe Position (in) Probe Position (in)
Results for Pit 8 at 1MHz Results for Pit 8 at 1MHz
0.8 2.5
2
0.6 Measured
Measured 1.5 Computed
0.4 Computed
Resistance(Ohms)
Reactance(Ohms)
1
0.2
0.5
0 0
-0.5
-0.2
-1
-0.4
-1.5
-0.6
-2
-0.8 -2.5
-0.5 -0.4 -0.3 -0.2 -0.1 0 0.1 0.2 0.3 0.4 0.5 -0.5 -0.4 -0.3 -0.2 -0.1 0 0.1 0.2 0.3 0.4 0.5
Probe Position (in) Probe Position (in)
376 18 Applications to Nuclear Power
0.1 0.4
Measured Measured
Computed
Resistance(Ohms)
Computed
Reactance(Ohms)
0.05 0.2
0 0
-0.05 -0.2
-0.1 -0.4
-0.15 -0.6
-0.5 -0.4 -0.3 -0.2 -0.1 0 0.1 0.2 0.3 0.4 0.5 -0.5 -0.4 -0.3 -0.2 -0.1 0 0.1 0.2 0.3 0.4 0.5
Probe Position (in) Probe Position (in)
Results for Pit 10 at 1MHz Results for Pit 10 at 1MHz
0.15 1.5
Measured
0.1 1 Computed
Measured
Computed
Resistance(Ohms)
Reactance(Ohms)
0.05 0.5
0 0
-0.05 -0.5
-0.1 -1
-0.15 -1.5
-0.5 -0.4 -0.3 -0.2 -0.1 0 0.1 0.2 0.3 0.4 0.5 -0.5 -0.4 -0.3 -0.2 -0.1 0 0.1 0.2 0.3 0.4 0.5
Probe Position (in) Probe Position (in)
0.2
1.5
0.15 Measured Measured
Computed 1 Computed
Resistance(Ohms)
Reactance(Ohms)
0.1
0.5
0.05
0 0
-0.05
-0.5
-0.1
-1
-0.15
-1.5
-0.2
-0.25 -2
-0.5 -0.4 -0.3 -0.2 -0.1 0 0.1 0.2 0.3 0.4 0.5 -0.5 -0.4 -0.3 -0.2 -0.1 0 0.1 0.2 0.3 0.4 0.5
Probe Position (in) Probe Position (in)
18.3 Model-Based Inversion of Measured Instrument Data... 377
0.8 Measured
3
Computed Measured
0.6 Computed
2
Resistance(Ohms)
Reactance(Ohms)
0.4
1
0.2
0 0
-0.2
-1
-0.4
-2
-0.6
-3
-0.8
-1 -4
-0.5 -0.4 -0.3 -0.2 -0.1 0 0.1 0.2 0.3 0.4 0.5 -0.5 -0.4 -0.3 -0.2 -0.1 0 0.1 0.2 0.3 0.4 0.5
Probe Position (in) Probe Position (in)
0.05756
0.0
Cavity
Host: σ = 1.4 E6
0.015
#3 0.017
#2 0.019
#1 0.021
0.028
We will develop a layered model6 of pit 3 in order to gain some insight into the
morphology of round-bottom pits. Figure 18.23 illustrates a three-layer model of
pit 3, with a resolution of 2 mils per layer. We used the data of Table 18.9 to
establish the size of the cavity and the position of the three layers. The solution
of the multifrequency inverse problem for the conductivities of the layers is shown
in Table 18.10.
0.05756
0.0
Cavity
0.01731
Host : σ = 1.4E6
0.017
σ = 0.5939E6 0.019
0.028
Fig. 18.24 Results of the inversion of the layered model of pit 3. The “round-bottom” is shown
as the amorphous shaded layer whose effective conductivity is 0.5939 × 106 S/m. The equivalent
pillbox model is shown as the dotted outline
Clearly, the bottom layer is host material, and the top layer is part of the pit cavity.
The middle layer with an effective conductivity of 5.939 × 105 S/m represents the
rounded bottom of the pit. We indicate in Fig. 18.24 this layer as an amorphous
structure, together with the equivalent pillbox model of the pit, using the Table 18.9
results. Upon applying the volume-fraction concept to the effective “round-bottom”
layer whose conductivity is 0.5939 × 106 S/m, we would find a total round-bottom
+ cavity volume that is virtually equal to the cavity volume of the equivalent pillbox
model. The important thing to note here is that the “rounded-bottom” that we have
just computed gives us a net depth of 0.019 inch for pit 3, which is in better
agreement with the nominal value listed in Table 18.8. (See, also, Fig. 18.22). This
shows the power and flexibility of model-based inversion in characterizing pits by
choosing appropriate parameters for modeling the pit.
Using the “Multi” results for D and H of pit 3 from Table 18.9, and the
results for the round-bottom model of pit 3 from Fig. 18.24, we plot in Fig. 18.25
the corresponding impedances at 1MHz. The round-bottom model gives a more
accurate estimation of the peak and slope through the origin of the resistance
scan. This is what we would expect, because the resistance data are important in
determining the depth of the pit. We have not applied this algorithm to the shallow
pits, 1, 5, and 9, because our interest was in determining the origin of the depth-
discrepancies of the deeper pits when the pillbox model was used.
A.1 Modeling Probes + Cables 379
0.4
0.06 Measured
Measured 0.3 Computed
0.04 Computed
Resistance(Ohms)
0.2
Reactance(Ohms)
0.02
0.1
0 0
-0.1
-0.02
-0.2
-0.04
-0.3
-0.06
-0.4
-0.08 -0.5
-0.5 -0.4 -0.3 -0.2 -0.1 0 0.1 0.2 0.3 0.4 0.5 -0.5 -0.4 -0.3 -0.2 -0.1 0 0.1 0.2 0.3 0.4 0.5
Probe Position (in) Probe Position (in)
Results for Pit 3 Layered Model at 1MHz Results for Pit 3 Layered Model at 1MHz
0.08 0.5
Measured 0.4
0.06 Computed
Measured
0.3 Computed
0.04
Resistance(Ohms)
0.2
Reactance(Ohms)
0.02
0.1
0 0
-0.1
-0.02
-0.2
-0.04
-0.3
-0.06
-0.4
-0.08 -0.5
-0.5 -0.4 -0.3 -0.2 -0.1 0 0.1 0.2 0.3 0.4 0.5 -0.5 -0.4 -0.3 -0.2 -0.1 0 0.1 0.2 0.3 0.4 0.5
Probe Position (in) Probe Position (in)
Fig. 18.25 Final results for pit 3 at 1MHz. Top: pillbox model. Bottom: layered (round-bottom)
model
Appendix
Data taken for a single probe of the differential-pair are given in Table 18.11.
Referring to the air measurements, we note that the resistance of the probe varies
with frequency, and the reactance is not proportional to frequency. This suggests
that there is a frequency-sensitive two-port network that connects the probe to the
impedance analyzer. It could be something as simple as a shunt self-capacitance of
the coil, but in this case it is a 25-foot (7.62 m), 5/16-inch coaxial cable. Cables
of this length (or more) are used to connect the probe to the measuring instrument
when inspecting heat-exchanger tubes.
The problem, now, is to characterize a probe coil whose parameters are not
known, and to characterize the cable that connects the coil to the impedance
analyzer. This can be done quite simply, though a bit tediously, with VIC-3D R
.
Trial and error on the turns and inner- and outer-radii of the coil suggest that a good
fit to the measured data is given by: Nturns = 55, IR = 0.19 in. and OR = 0.21 in.
The height of the coil is known to be 0.055 in. When we run VIC-3D R
for this
probe in freespace, we get an inductance of 49.55 μH, which is close to, but smaller
than, the inductance measured at 10 kHz. The freespace resistance of the model coil
is, of course, 0 Ω, so we manually enter a resistance of R0 = 4.5 Ω into the VIC-
3D R
file to see what that gives. We choose this value because it agrees with the
measured resistance at the lowest frequency.
Cable effects can be accounted for by defining four parameters: characteristic
impedance, capacitance per unit length, attenuation in dB/m and length. This
follows from transmission-line theory [98]. Let ZL be the load (terminating)
impedance of a coaxial cable, and Zin be the driving-point (input) impedance of
the loaded cable. Then
ZL + Z0 tanh( jβ l + α l)
Zin = Z0
Z0 + ZL tanh( jβ l + α l)
Z0 = (L/C) = characteristic impedance of the cable
ω
β =
vp
1
vp = ≈ 2 × 108 m/s. (18.2)
(LC)
A.1 Modeling Probes + Cables 381
L and C are, respectively, inductance and capacitance per unit length of the
cable. The value for phase velocity, vP , is a reasonable approximation for typical
transmission lines.
Typical values for coaxial cables that are used in eddy-current NDE are 50 Ω
for the characteristic impedance, and 100 pF/m for the distributed capacitance. Our
problem calls for a cable length of 7.62 m, which leaves only the attenuation to be
determined. Attenuation, α , is frequency dependent, but VIC-3D R
assumes it to
be constant, so it becomes necessary to run several tests to determine the frequency
response of the cable.
The results for tests of the probe+cable in freespace are shown in Table 18.12.
The diagonal entries in this table agree with the corresponding freespace results of
Table 18.11 within 2 %. This indicates that the attenuation varies with frequency as
shown in the left-hand column. This variation agrees with theory, in that attenuation
always increases with frequency.
When modeling the probe within the tube, we use the same cable parameters
as above, except that we use α = 0.0 only for f = 104 Hz, α = 0.005 only for
f = 5 × 104 , etc. The results for the model calculations for the coil within the tube
are given in Table 18.13. The differences between these values and those shown in
Table 18.11 are 3 %, or less, except for X104 and R105 .
It is difficult to determine an outer radius of a many-turn bobbin coil when it is
within a tube, because the windings on the outer layer of the coil nearest the tube
will not lie smoothly in that layer, as shown in Fig. 18.26. This gives the appearance
of an uneven spacing between the outer radius of the coil and the inner tube wall.
This is not as serious a problem when the coil is in the “pancake” aspect.
382 18 Applications to Nuclear Power
0.472 OR
IR
SS: σ = 1.4E6 , μ = 1
One possible scheme for determining the shape of a corrosion pit that is assumed to
have a fixed morphology would be to use the layered-pit model shown in Fig. 18.27.
The objective is to determine the conductivity of each layer, and from that result
infer the size, and perhaps shape, of the pit. We assume that the pit has a circular
cross-section, as before, so that the only parameter that defines the model is the
radius of the anomalous region. If the layer has a conductivity of 0S/m, then clearly
that layer is entirely filled by the pit, whereas if the layer has a conductivity of
1.4 × 106 S/m, then the layer is entirely free of the pit, being host material, only.
Anything in between requires further consideration, as we show next.
The unknowns in the inversion process are σ1 , . . . , σ4 , the conductivities of layers
1–4, with 1 corresponding to the inner wall and 4 to the outer wall, as well as
the diameter of the anomalous region. Once the conductivity of a layer has been
estimated by NLSE, we apply a simple volume-fraction computation to determine
the relative volume of cavity to host material within that layer. For example, suppose
layer L has a conductivity of 1.135 × 106 S/m. Then to see how much of the Lth
layer is occupied by the cavity, we compute the volume-fraction of the cavity:
1.135
VFL = 1 − = 0.1893. Thus, the cavity occupies less than 19 % of the Lth
1.40
layer.
Chapter 19
Coupled Problems in Heat-Exchanger Tubes
19.1 Introduction
The preceding chapters of this book dealt largely with problems in which the
host material and any anomalies were pure electrical conductors. The presence of
magnetic permeabilities was merely an interesting side effect to the creation of the
VIC-3D R
model and its solution. There are, of course, many problems in which
magnetic permeability is present with electrical conductivity and must be accounted
for, not only in creating the model but also in understanding the physics of the
solution and its effect on the NDE process. In this chapter we consider magnetic
effects in heat-exchanger tubes, focusing later on ferritic tubes, which are becoming
of increasing importance in the nuclear power industry.
In Figs. 19.1 and 19.2, we model “fretting” damage in a heat-exchanger tube [57].
The damage manifests itself as a “wear scar” surrounded by a “crust.” The crusts of
both figures are slightly magnetic, with that of Fig. 19.1 highly conducting, relative
to the host tube, and that of Fig. 19.2 having the same conductivity as the host tube.
The scars and crusts are modeled as semiellipsoids.
We model a seventeen-point scan over the anomalous region shown in Fig. 19.2
at three frequencies, 100 kHz, 400 kHz and 1.6 MHz, to reconstruct the conductivity,
σ , and permeability, μ , of the crust. The results are shown in Table 19.1. In all cases,
the result is more sensitive to μ than to σ . The computed impedance trajectories that
are the input to the inversion algorithm are shown in Fig. 19.3. The rotation of these
trajectories as a function of frequency is clearly shown.
TUBE AXIS
COIL
HIGH−COND
CRUST σ=0, μ=1
WEAR SCAR
Fig. 19.1 A semiellipsoidal “wear scar” and highly-conducting permeable “crust” in the outer
surface of a tube
TUBE AXIS
COIL
PERMEABLE
CRUST σ=0, μ=1
σ = 1Ε6, μ=1.1 WEAR SCAR
Fig. 19.2 A semiellipsoidal “wear scar” and permeable “crust” in the outer surface of a tube
19.3 Modeling Tube-Support Rings (TSR) 385
0
-0.01
-0.02
-0.03
-0.04
-0.05
-0.06
-0.07
-0.02 -0.01 0 0.01 0.02 0.03 0.04 0.05 0.06
Resistance (Ohms)
In this section and the next we model the tube support ring problem of Fig. 19.4 that
is quite common in the inspection of nuclear power plants. The major difference
between the problem here and in the next section is the nature of the tube, itself.
Here, it is inconel, which is nonmagnetic, while in the next section it is ferritic,
which introduces additional interesting challenges. In both cases, however the
TSP+tube structure is axisymmetric, which allows us to use the axisymmetric model
in cylindrical geometries that was developed in Chap. 9.
386 19 Coupled Problems in Heat-Exchanger Tubes
19.00
7.725 8.25
Differential Bobbin Coil
6.225
1.5 1.5
1.5
Inconel Tube
TSR
Fig. 19.4 Illustrating a typical tube support ring problem. All dimensions are in millimeters
The measured data for the problem are taken at 100, 200, 300 and 400 kHz, and
are presented here for convenience in Fig. 19.5.1 Note that the data are plotted
in the impedance plane (real vs. reactive), even though they are instrument A/D
counts (or “voltage”) in two channels that are orthogonal in time; i.e., they are
ninety degrees out of phase with each other. The “Lissajous figure” closed-loop
curve is typical of data that are taken with a differential-bobbin probe. The user
of the instrument is usually instructed to “rotate the impedance-plane plot” so that
it satisfies certain criteria, and this makes the interpretation of the measured data
more challenging. We will use a linear filter to transform the measured voltage data
into impedances in a manner similar to that which was done in the EPRI ID pits
problem of Sect. 18.3.1. There are several important features to glean from this
figure: note the clockwise rotation of the loops with frequency and note that the
loops get smaller with frequency. Indeed, the loop for 400 kHz, actually fits within
the loop for 100 kHz.
1 The statement of the problem described in this section, and the original measured data were
supplied by Prof. S-J. Song of Sungkyunkwan University, South Korea.
19.3 Modeling Tube-Support Rings (TSR) 387
Reactance (Ohms)
200
−200
−400
−600
−600 −400 −200 0 200 400 600
Resistance (Ohms)
0.2
-0.2
-0.4
-0.6
-0.8
-0.8 -0.6 -0.4 -0.2 0 0.2 0.4 0.6 0.8
Resistance (Ohms)
The model results (impedances) for Fig. 19.4 are shown in Fig. 19.6. We have not
yet applied the β -scaling that will bring the impedances to voltage or A/D-counts,
but the important features described above are already apparent. The loops decrease
in size with frequency and are rotated clockwise.
388 19 Coupled Problems in Heat-Exchanger Tubes
200
100
-100
-200
-300
-400
-500
-500 -400 -300 -200 -100 0 100 200 300 400 500
Resistance (Ohms)
Fig. 19.7 The differential-bobbin model data shown in Fig. 19.6 after scaling by the appropriate
β at each of the four frequencies
Following the usual procedure, we scale the model vector by a factor, β , that
will allow us to transform impedance into voltage (or A/D-counts), or vice-versa.
We have determined β for each of the data and model vectors at the four frequencies,
as listed in Table 19.2. We then apply β at each frequency to the corresponding
model data vector of Fig. 19.6 and derive the scaled-model data vectors shown in
the complex plane of Fig. 19.7. We now note the similarities to the measured data
of Fig. 19.5: the loops rotate clockwise and shrink with frequency, and the loop for
400 kHz fits within that for 100 kHz. Hence, we can conclude that there is consistent
agreement between model and measured data.
Up to this point, we scaled the model data of Fig. 19.6 to agree as closely as possible
(in the least-squares sense) with the measured data of Fig. 19.5. The result is shown
in Fig. 19.7.
19.3 Modeling Tube-Support Rings (TSR) 389
Now, we want to do the opposite and scale the measured data to agree as
closely as possible (in the least-squares sense) with the model data. Furthermore,
we are going to use a more general scaling approach in which four parameters,
β1 , β2 , β3 , and β4 , are used instead of only β1 and β2 . The four parameters are
defined by
R = β1C1 + β2C2
X = β3C2 + β4C1, (19.1)
where R and X are the vectors of the model results and C1 and C2 are vectors of
the measured data in each channel of the instrument (which are supposed to be
orthogonal in time).
Upon taking the dot-product of the first equation by, respectively, C1 and C2 , and
similarly the second equation, we get the two linear systems
(V M) (M) (M)
ΓR1 = β1Γ11 + β2Γ21
(V M) (M) (M)
ΓR2 = β1Γ12 + β2Γ22
(V M) (M) (M)
ΓX1 = β4Γ11 + β3Γ21
(V M) (M) (M)
ΓX2 = β4Γ12 + β3Γ22 , (19.2)
where, if these were random processes, the Γ s would be auto- and cross-correlation
functions. For our purposes, they are simply real numbers that are obtained by taking
the dot products of the appropriate vectors.
The solutions of (19.2) are tabulated for each frequency in Table 19.3.
An impedance-plane plot is shown in Fig. 19.8. Once again we see the strong
similarity between the scaled data and the model data of Fig. 19.6; the curves rotate
clockwise with frequency, and they become smaller and narrower.
When we compare the Lissajous figures of Figs. 19.5 and 19.7, we are liable to draw
the conclusion that the model data cannot be used to characterize the system, if the
only data that are known for the system are given by the measuring instrument,
390 19 Coupled Problems in Heat-Exchanger Tubes
0.2
-0.2
-0.4
-0.6
-0.8
-1 -0.8 -0.6 -0.4 -0.2 0 0.2 0.4 0.6 0.8 1
Resistance (Ohms)
Fig. 19.8 The differential-bobbin measured data shown in Fig. 19.5 after scaling by the appropri-
ate β1 –β4 at each of the four frequencies
because they are not “perfectly” matched. We will show that the model data and
measured data, after either has been scaled to agree with the other, can be used
together to characterize the system. We will do this by using both in an inverse
problem whose goal is to determine the conductivity and magnetic permeability of
the tube-support ring. The inverse problem will be solved using NLSE in impedance
space, which means that we will use the β -scaled measurements of Fig. 19.8.
When we have two unknowns, the conductivity and magnetic permeability, and
apply NLSE to the scaled measured data at each of the four frequencies, we get the
results shown in Table 19.4. The term “sens” denotes the sensitivity of the parameter
at the solution point: the smaller, the better. Φ is the norm of the residual vector at
the solution point: the smaller the better, again.
Note that the conductivity estimate increases somewhat with frequency. Because
conductivity is more accurately estimated at higher frequencies (see [57] as well
as Chap. 11 of this book) we will take σ = 6.55 × 106 S/m as our estimate of the
conductivity.
The estimate of the magnetic permeability, however, is somewhat suspect,
because the solution is rather insensitive to the parameter. This follows from the
large values of sens associated with μ that are shown in the table. Hence, we will
19.4 Modeling Direct and Inverse Problems in Ferritic Tubes 391
redo the inversion, this time fixing σ = 6.55 × 106, and solving for only the single
variable, μ . When we use NLSE for this one-variable problem, we get the results
shown in Table 19.5.
Now, we see that Φ is reduced, meaning that the solution is in better agreement
with the measured data, and the sensitivity of the solution for μ is much better.
We also see that the estimated value of μ decreases slightly with frequency. Because
μ is more accurately estimated at lower frequencies (see Sect. 11.4), we choose
μ = 42.517 for our best estimate. Thus, both σ and μ are well estimated, given
their nominal values of 6.7 × 106 and 50, respectively.
In performing this inversion, it is important to understand that we did not
rely upon any particular “feature” of the Lissajous figure at any one frequency.
The approach that utilizes features of measured data is likely to be subject to
significant errors, given the subjective nature of “observing” such features. This
is especially true if the Lissajous figure is corrupted by noise or is embedded in
significant background clutter. The mathematical approach that we have indicated
is much more reliable and objective and can be further enhanced by the application
of data-smoothing and clutter-rejection algorithms.
Ferritic stainless steels, such as Type 439 or SEACURE, are being increasingly
used in heat-exchanger tubes because of the increased resistance to chloride stress
corrosion and intergranular attack when compared to older alloys, such as Type 304.
This presents interesting modeling opportunities for the eddy-current nondestructive
evaluation of these tubes because their magnetic permeability is ∼60–100.2 In this
section,3 we present examples of direct and inverse problems involving such tubes,
2 The magnetic permeability is a function of the stress-state of the tube, as well as the frequency of
operation.
3 Details of the experiments that are described in this section are given in [112]. Related work on
0.75"
0.030"
0.0625"
,
1 2 3
0.878" Coil
SEACURE
μ=104
σ=1.552E6
0.75"
0.030"
0.0625"
,
1 2 3
0.878"
Coil
SEACURE
μ=68.18
σ=1.372E6
Fig. 19.9 Illustrating a ferritic tube, such as SEACURE, with a tube-support plate (TSP), and three
0.0625-inch through-wall holes. Hole 1 is well away from the TSP, hole 2 is centered under the
TSP, and hole 3 is centered at edge of the TSP. Top: High-frequency model. Bottom: Low-frequency
model. See text for details on parameters
then demonstrate conditions that are peculiar to ferritic tubes, and give insight into
the optimum methods for characterizing the tubes and flaws within them.
The problem configuration is shown in Fig. 19.9. It is similar to the model shown
in Fig. 19.4, except that the tube is ferritic, and there are three holes, one located
away from the tube support plate (TSP),4 one located under the center of the TSP,
4 Tube support plate and tube support ring are synonymous terms.
19.4 Modeling Direct and Inverse Problems in Ferritic Tubes 393
R0
L0
M0
Ζin
L1 R1
Lμ Mμ
Fig. 19.10 A coupled-circuit model of the coil in the presence of the ferritic tube. R0 and L0 are,
respectively, the resistance and self-inductance of the coil in freespace, and Lμ is the increased
inductance of the coil due to the permeability of the tube. L1 is the “virtual” secondary inductance
that accounts for induction effects within the tube, M0 is the mutual inductance between L0 and
L1 , and Mμ is the mutual inductance between Lμ and L1 . R1 is the effective “secondary resistance”
that is due to the electrical conductivity of the tube
and one located at the edge of the TSP. We will characterize the tube and then the
composite structure, including the TSP and holes in the following subsections.
ω 2 Mμ2 ω 2 M02
Zin = R0 + jω L0 + jω Lμ + + . (19.3)
R1 + jω L1 R1 + jω L1
From this we get the change in impedance due to the presence of the tube:
δ Zin = Zin − R0 − jω L0
ω 2 Mμ2 ω 2 M02
= jω Lμ + +
R1 + jω L1 R1 + jω L1
394 19 Coupled Problems in Heat-Exchanger Tubes
ω 2 Mc2 (R1 − jω L1 )
= jω Lμ +
R21 + ω 2L21
ω 2 Mc2 R1 ω 2 L1 Mc2
= 2 + jω Lμ − 2 (19.4)
R1 + ω 2 L21 R1 + ω 2 L21
where Mc2 = M02 + Mμ2 . Again, we see that at low frequencies the effects of
conductivity and permeability of the tube are uncoupled. The first term in (19.4)
is the effective coupled resistance as a function of frequency, whereas the second
term displays the coupled inductance as a function of frequency. Note, in particular,
the quadratic dependence on ω of the resistance at low frequencies, ω << R1 /L1 .
Furthermore, when ω is very small the coupled inductance is simply Lμ , the
inductance of the coil due solely to permeability. The coupling to R1 , which
corresponds to conductivity, is virtually nil. Depending upon the values of the
various parameters, the coupled inductance, while starting out as a positive number,
may go negative. If the permeability of the tube were that of freespace, then Lμ = 0,
and the change in coupled inductance would be negative at all frequencies, in
accordance with Lenz’ law.
Inversion Results Using 100Hz to 1kHz Data: A Inversion Results Using 100Hz to 1kHz Data: A
0.07 0.8
0.06 0.7
0.6
0.05 Measured
Model
0.5
R(Ohms)
X(Ohms)
0.04
Measured
Model 0.4
0.03
0.3
0.02
0.2
0.01 0.1
0 0
100 200 300 400 500 600 700 800 900 1000 1100 100 200 300 400 500 600 700 800 900 1000 1100
Frequency (Hz) Frequency (Hz)
Inversion Results Using 100Hz to 1kHz Data: B Inversion Results Using 100Hz to 1kHz Data: B
0.09 0.8
0.08 0.7
0.07 0.6
Measured
0.06 Model
X(Ohms)
0.5
R(Ohms)
0.05 Measured
Model
0.4
0.04
0.3
0.03
0.02 0.2
0.01 0.1
0 0
100 200 300 400 500 600 700 800 900 10001100 100 200 300 400 500 600 700 800 900 1000 1100
Frequency (Hz) Frequency (Hz)
Fig. 19.11 Comparing the model results, based on the results of Table 19.6, with the original
measured data over the frequency range of 100 Hz to 1 kHz. Top: location A. Bottom: location B
we are seeing the effects of Lμ , only, with no coupling to the conductivity in the
tube. This further confirms our model results and demonstrates the advantage of
working at low frequencies to distinguish conductivity and permeability effects.
When we use the inverted results of Table 19.6 for location A, and generate a
model response over the frequency range 10–100 kHz, we get the results shown
in Fig. 19.12. Because the resistance values are in good agreement over this
frequency range, but the reactances differ considerably, it is clear that Lμ in (19.4) is
underestimated, which in turn suggests that μ is too small. We do another inversion
to determine σ and μ , this time using the measured data for location A shown in
Fig. 19.12. The results of this high-frequency inversion are shown in Table 19.7 and
are plotted in Fig. 19.13.
396 19 Coupled Problems in Heat-Exchanger Tubes
Inversion Results Over 10kHz to 100kHz Inversion Results Over 10kHz to 100kHz
16 7.5
14 7
6.5 Measured
12
Measured Model
6
R(Ohms)
X(Ohms)
Model
10
5.5
8
5
6
4.5
4 4
2 3.5
10 20 30 40 50 60 70 80 90 100 10 20 30 40 50 60 70 80 90 100
Frequency (kHz) Frequency (kHz)
Fig. 19.12 Showing the high-frequency response at location A when the low-frequency parame-
ters of Table 19.6 are used
Inversion Results Using 10kHz to 100kHz Data Inversion Results Using 10kHz to 100kHz Data
18 7.5
16 7
14 6.5
Measured
Model
12 6
R(Ohms)
X(Ohms)
Measured
Model
10 5.5
8 5
6 4.5
4 4
2 3.5
10 20 30 40 50 60 70 80 90 100 10 20 30 40 50 60 70 80 90 100
Fig. 19.13 Comparing the model results, based on the results of Table 19.7, with the original
measured data for location A over the frequency range of 10–100 kHz
The fact that the norm of the residuals, Φ , is much larger here than in Table 19.6
is not surprising, given the much larger data values here, nor is it surprising that
the sensitivity coefficient for μ is much larger, given our observation that the two
variables have the smallest correlation at low frequencies. The important point to
note is that σ is changed by only 4 %, which is probably in reasonable agreement
with the data, but that μ increases by 15 %. It seems likely that the origin of the
ferromagnetic response is dispersive, which yields a permeability that is frequency
dependent. This frequency dependence is reasonable, given that dipolar effects,
whether electrical or magnetic, have their origins in electrons that are bound to the
19.4 Modeling Direct and Inverse Problems in Ferritic Tubes 397
Results Based on Trial and Error Results Based on Trial and Error
18 7.5
16 7
14 6.5
12 6
R(Ohms)
X(Ohms)
Measured
Model
10 5.5
Measured
Model
8 5
6 4.5
4 4
2 3.5
10 20 30 40 50 60 70 80 90 100 10 20 30 40 50 60 70 80 90 100
Frequency (kHz) Frequency (kHz)
Fig. 19.14 Comparing measured and model results over 10–100 kHz for location A, when the
model comprises σ = 1.372 × 106 S/m, as in the low-frequency regime, and μ = 85, chosen by
trial and error
Results Based on Trial and Error for Location B Results Based on Trial and Error for Location B
16 6.5
14 6
12
5.5
R(Ohms)
X(Ohms)
10
Measured
Model 5
8
4.5
6
4 4
2 3.5
10 20 30 40 50 60 70 80 90 100 10 20 30 40 50 60 70 80 90 100
Frequency (kHz) Frequency (kHz)
Fig. 19.15 Comparing measured and model results over 10–100 kHz for location B, when the
model comprises σ = 1.552 × 106 S/m, as in the low-frequency regime, and μ = 86, chosen by
trial and error
host lattice, and are not free to move as are those electrons in the conduction band
[101]. This suggests that dipoles should share the frequency response of the host
lattice, including resonance and similar dispersive phenomena.
An alternative strategy would be to keep the host conductivity constant at
σ = 1.372 × 106 S/m for all frequencies for location A, and then determine a
value of μ that is consistent with the measured data simply by trial and error.
We have done this, and found that μ = 85 gives excellent results, as shown in
Fig. 19.14. Figure 19.15 shows the corresponding results for location B, using
σ = 1.552 × 106 S/m and μ = 86. In either case, it is clear from Figs. 19.14
and 19.15 that the high-frequency reactance demonstrates the nonlinear behavior
shown in (19.4) that is due to Lenz’ law for coupled circuits.
398 19 Coupled Problems in Heat-Exchanger Tubes
0.7
0.3
0.6
0.25 A
0.5
R(Ohms)
X(Ohms)
B
BTSP
0.4
A 0.2 BF
B
0.3 BTSP
BF 0.15
0.2
0.1
0.1
0 0.05
10 20 30 40 50 60 70 80 90 100 10 20 30 40 50 60 70 80 90 100
Frequency (kHz) Frequency (kHz)
Fig. 19.16 Inner-coil responses at four probe locations. A: At end of tube. B: Near center of tube.
BF: At center with flaw (drilled hole). BTSP: Near center with tube support plate over it
We have taken multifrequency data using the inner coil at four locations along the
tube. Location A is at an end, B is near the center, in the absence of any anomalies,
BF is at the flaw, which is a drilled hole, and BTSP is near the center under the tube-
support plate. We show only the high-frequency data from 10 to 100 kHz, because
those are the most reliable.5 This is not a problem, because we are not interested in
distinguishing conductivity effects from permeability, as this has already been done
using the outer coil.
Figure 19.16 shows the impedance responses at these four locations. It is clear
that the response at A is distinguished because of its location near the end of the
tube. It is also clear from the resistance response that the tube-support plate is not as
clearly distinguished from its background (B) as is the flaw. This is reasonable, given
that the skin-effect is pronounced at these frequencies, which effectively shields the
TSP from the coil. The flaw, being a surface-breaking inner-diameter hole, is not
similarly shielded.
We have noted that over the frequency range of 10–100 kHz the permeability
does not remain fixed at its low-frequency value. By setting μ = 104 over this
frequency range, and letting the conductivity retain its low-frequency value of
5 Theinner coil is much smaller than the outer, with fewer turns, which means that it has a smaller
inductance, and is, therefore, prone to produce noisier low-frequency data.
19.4 Modeling Direct and Inverse Problems in Ferritic Tubes 399
0.8 0.35
0.7
0.3
0.6
0.25
R(Ohms)
0.5 Model
X(Ohms)
Model B-data
B-data
0.4 0.2
0.3
0.15
0.2
0.1
0.1
0 0.05
10 20 30 40 50 60 70 80 90 100 10 20 30 40 50 60 70 80 90 100
Frequency (kHz) Frequency (kHz)
Fig. 19.17 Showing the model inner-coil response over the frequency range of 10–100 kHz,
together with the measured data for location B. The model parameters used to obtain this result are
σ = 1.552 × 106 S/m, and μ = 104
σ = 1.552 × 106 for location B, we generate the model response for the inner coil
shown in Fig. 19.17, along with the measured data for location B. The responses are
in good agreement within a few percent over this frequency range.6
Now that the host ferritic tube has been characterized, we proceed with the main
thrust of this section: characterizing the composite structure consisting of the inner
coil, tube, through-wall hole, and tube support plate. We’ll start with some model
calculations, because these results will allow us to interpret the measured data that
will be described in the next section. The configuration that produces the model
results of this section is the same as that shown in the top of Fig. 19.9. It uses the
same values of σ = 1.552 × 106 S/m and μ = 104 for the host parameters as in
Fig. 19.17.
In Fig. 19.18, we show the results of a scan past the isolated hole at five
frequencies, 1, 10, 50, 75, and 100 kHz. The important feature here is the monotonic
increase (in the negative direction for R) of the responses with frequency. These
results should be contrasted with the corresponding ones for the isolated TSP, shown
in Fig. 19.19. The frequency responses of the TSP are much more convoluted, in
6 One possible explanation for the difference in the values of the permeabilities of Figs. 19.15
and 19.17 is due to material stress inhomogeneities through the wall thickness. See [101] for a
discussion of magnetoelastic effects on permeability.
400 19 Coupled Problems in Heat-Exchanger Tubes
0 0.003
0.0025
−0.0005 1kHz
1kHz 10kHz
10kHz 0.002 50kHz
R(Ohms)
X(Ohms)
−0.001 50kHz 75kHz
75kHz 100kHz
100kHz
0.0015
−0.0015
0.001
−0.002
0.0005
−0.0025 0
−0.003 −0.0005
−2 −1.5 −1 −0.5 0 0.5 1 1.5 2 −2 −1.5 −1 −0.5 0 0.5 1 1.5 2
Scan Position (in) Scan Position (in)
Fig. 19.18 Inner-coil responses at five frequencies for the isolated through-wall hole
0.0001 0.001
0.0008 1kHz
0
10kHz
0.0006 50kHz
R(Ohms)
X(Ohms)
−0.0001 75kHz
1kHz 100kHz
10kHz 0.0004
−0.0002 50kHz
75kHz
100kHz 0.0002
−0.0003
0
−0.0004 −0.0002
−0.0005 −0.0004
−2 −1.5 −1 −0.5 0 0.5 1 1.5 2 −2 −1.5 −1 −0.5 0 0.5 1 1.5 2
Scan Position (in) Scan Position (in)
Fig. 19.19 Inner-coil responses at five frequencies for the isolated tube-support plate
the sense that they “flip” at low frequencies. This is typical of an outer-diameter
anomaly and is due to the “competition” between μ and σ in the ferritic tube wall, as
we have seen before. Furthermore, the TSP becomes invisible at higher frequencies
(at least as far as the inner coil is concerned), due to the reduced skin depth. This
reduction with frequency is pronounced due to the large magnetic permeability of
the tube.
The important contrast between Figs. 19.18 and 19.19 is in the low- and high-
frequency responses. At low frequencies, 1–10 kHz, the TSP response overwhelms
the hole response, but the situation is completely reversed in the high-frequency
regime at 50 kHz and above, for which the hole dominates the TSP. This suggests
that any protocol for detecting pits in the presence of tube-support plates will
necessarily require choosing the right frequency range in which to operate.
19.4 Modeling Direct and Inverse Problems in Ferritic Tubes 401
0 0.003
1kHz
10kHz 0.0025
−0.0005 50kHz
75kHz
0.002
R(Ohms)
X(Ohms)
100kHz
−0.001
0.0015 1kHz
−0.0015 10kHz
0.001 50kHz
75kHz
−0.002 100kHz
0.0005
−0.0025 0
−0.003 −0.0005
−2 −1.5 −1 −0.5 0 0.5 1 1.5 2 −2 −1.5 −1 −0.5 0 0.5 1 1.5 2
Scan Position (in) Scan Position (in)
Fig. 19.20 Inner-coil responses at five frequencies for the tube-support plate with the hole
centered beneath it
−0.0005
0.001
−0.001
R(Ohms)
X(Ohms)
0.0005
−0.0015
−0.002
0
−0.0025
−0.0005
−0.003
−0.0035 −0.001
−2 −1.5 −1 −0.5 0 0.5 1 1.5 2 −2 −1.5 −1 −0.5 0 0.5 1 1.5 2
Fig. 19.21 Inner-coil response at 100 kHz for the hole of Fig. 19.18 surrounded by a through-wall
collar whose conductivity is 1.552 × 106 S/m and permeability is 100. The diameter of the collar
is one inch
We plot the combined response of a hole centered under the TSP in Fig. 19.20.
It is clear that the TSP completely dominates the hole response at frequencies
of 10 kHz and lower, but that the hole dominates the high-frequency response, in
accordance with the isolated results of Figs. 19.18 and 19.19.
Later, when interpreting measured scan data, we will make use of a model,
whose result is shown in Fig. 19.21, comprising the through-wall hole of Fig. 19.18
surrounded by a through-wall collar whose permeability is 100 and conductivity is
1.522 × 106 S/m. The purpose of this model is to illustrate the significant effect
that even a small material inhomogeneity, such as magnetic permeability, can
have on the impedance response. Contrast, for example, the reactance response
at 100 kHz in Fig. 19.18 with the reactance response in Fig. 19.21. The former
402 19 Coupled Problems in Heat-Exchanger Tubes
Fig. 19.22 Change in resistance (left) and reactance (right) as a function of frequency and scan
position when the probe is scanned past the hole, located at 0.00”, and TSP, located between 1.580”
and 2.30”. The frequencies are 1, 10, 50, 75, and 100 kHz
does not change sign with scan, whereas the latter does. As we will see in the next
section, the background clutter associated with permeability inhomogeneities must
be accounted for in developing an inspection protocol for ferritic tubes.
The model results presented in the last section are useful in interpreting measured
data in which the inner-coil is scanned past the hole and TSP. Figure 19.22 shows
the response of the inner coil when it is scanned past the hole, located at 0.00”, and
the TSP, which starts at 1.580” and ends at 2.300”. The frequencies are 1, 10, 50,
75, and 100 kHz.
These results, while exhibiting considerable coherent background clutter, still
show qualitative agreement with our model computations. For example, consider the
reactance function shown in the right-hand side of Fig. 19.22. The hole portion of
this response shows two negative-going “descenders” bracketing the upward-going
response over the hole. This agrees with the model reactance response shown in
Fig. 19.21, which suggests, but certainly does not prove, that the source of the
coherent clutter could be an inhomogeneous distribution of magnetic permeability.
Furthermore, note that the reactance response associated with the hole increases
monotonically with frequency, exactly as we showed in Fig. 19.18.
19.4 Modeling Direct and Inverse Problems in Ferritic Tubes 403
The reactance function response of the TSP in the right-hand side of Fig. 19.22
is also in agreement with the model calculations; the 1 and 10 kHz responses have
opposite signs, which agrees with the results shown in Fig. 19.19. The fact that
each of these two measured responses has the opposite sign of the corresponding
model responses is due to choosing the model TSP parameter, σTSP , arbitrarily to
be 6.76 × 106 S/m; a more reasonable value of 1.0 × 106 S/m will be demonstrated
shortly. Furthermore, note that the two lowest measured TSP signals (1 and 10 kHz)
are much larger than the measured hole signals at 1 and 10 kHz and is the reason
that we must not use such low frequencies to try to locate a hole in the presence
of the TSP. This, too, was predicted in the model calculations, as shown in
Figs. 19.18–19.20. It is clear that the higher frequency signals, namely 50, 75, and
100 kHz, are associated with an inner-diameter anomaly, since they monotonically
increase with frequency. In fact, judging from the shape of these signals—they are
miniversions of the corresponding through-wall hole signals—the anomaly may
well be a small pit. It is likely, however, that these signals are simply part of the
background clutter, whose source needs to be studied further.
Further validation is given in the series of responses shown in Figs. 19.23–19.26,
which display the results for the isolated hole, the hole centered under the TSP, and
the hole placed at an edge of the TSP. These data were taken with an inner coil that
had more turns, and was somewhat larger than that used in the earlier experiments.
Thus, with the increased sensitivity of the coil, the responses are larger than those
shown previously, but retain the same qualitative features. In all of these figures, the
hole is located at 0.00 inch.
The results shown in these figures confirm our previous conclusions. Note that
the high-frequency response of the hole + TSP system approaches the isolated hole
response, no matter where the hole is located relative to the TSP. On the other hand,
the low-frequency response is dominated by the TSP, no matter where the hole is
located, and furthermore, the low-frequency response exhibits the usual “flipping”
phenomenon that we observed earlier. This is made quite clear in Figs. 19.25
and 19.26.
In our discussion of the results of Figs. 19.19 and 19.22, we determined that
the permeability of the TSP was in error, and this was due to the fact that we
had not characterized the TSP, but had simply used a previous model, as shown
at the top of Fig. 19.9. In order to rectify this situation, we reran some of the
models with the TSP and hole at low frequencies, namely 1, 10, and 19 kHz.
The new runs were performed with low-frequency permeabilities of 68.18 and
73.43, instead of the high-frequency permeability of 104. Furthermore, we ran
different host conductivities, namely 1.372 × 106 S/m and 1.552 × 106 S/m, which
we computed earlier to be correct for two measurement points with the outer coil.
404 19 Coupled Problems in Heat-Exchanger Tubes
Fig. 19.23 Change in resistance as a function of frequency and scan for (a) no TSP, (b) TSP over
hole center, (c) TSP over hole edge
Because we had not characterized the TSP independently of the tube, we varied
the TSP parameters during these trials, using combinations of μTSP = 50, and 100,
together with σTSP = 6.76 × 106, 1 × 106, 5 × 105 S/m. The combination of all of
these parameters that gave a good qualitative fit to the measured data is shown in
Table 19.8. These parameters are used in the lower configuration shown in Fig. 19.9.
The results of the runs are shown in the top row of Fig. 19.27, and when they
are compared with the corresponding measured data shown in the bottom row, we
see that there is a good qualitative agreement and that the relative magnitudes of the
reactance responses are in reasonable agreement with the measured data.
19.4 Modeling Direct and Inverse Problems in Ferritic Tubes 405
Fig. 19.24 Change in reactance as a function of frequency and scan for (a) no TSP, (b) TSP over
hole center, (c) TSP over hole edge
Fig. 19.25 Change in resistance as a function of frequency and scan for (a) no TSP, (b) TSP over
hole center, (c) TSP over hole edge (low frequencies, only)
Fig. 19.26 Change in reactance as a function of frequency and scan for (a) no TSP, (b) TSP over
hole center, (c) TSP over hole edge (low frequencies, only)
Appendix
Following the discussion of inverse method quality metrics and the Cramer–Rao
Lower Bound in Chap. 12, we optimize the estimation of conductivity and per-
meability of the ferritic tube. Figure 19.28 presents changes in the normalized
impedance plane for varying permeability (red) and conductivity (black) levels at
408 19 Coupled Problems in Heat-Exchanger Tubes
0.001
−0.001
X(Ohms)
1kHz
10kHz 0.0005
−0.0015 19kHz
0
−0.002
−0.0025 −0.0005
−0.003 −0.001
−2 −1.5 −1 −0.5 0 0.5 1 1.5 2 −2 −1.5 −1 −0.5 0 0.5 1 1.5 2
Fig. 19.27 Low-frequency model response of the TSP and hole when the data in Table 19.8 are
used. Top: Model. Bottom: Measured
four frequencies: (a) 100 Hz, (b) 1.0 kHz, (c) 10 kHz, and (d) 100 kHz. Several
observations can be made from these plots. First, it is helpful to increase frequency
in order to get sensitivity to conductivity changes. At very low frequencies, sensitiv-
ity to changes in permeability is dominant in the reactance term. As the frequency
approaches 100 kHz, it is impossible to distinguish changes in permeability and
conductivity in the eddy-current measurements.
Figure 19.29 presents a comparison of four different estimation metrics for this
inversion problem: (a) CRLB for variation in conductivity estimation, (b) CRLB
for variation in permeability estimation, (c) condition number, and (d) correlation.
Based on the CRLB for optimal sensitivity to permeability, lower frequencies are
better. Note, there is really not much benefit from using 100 Hz versus 1.0 kHz.
A.1 Inverse Method Quality Metrics for the Ferritic Tube 409
Fig. 19.28 Changes in normalized impedance plane for varying permeability and conductivity
levels at four frequencies: (a) 100 Hz, (b) 1.0 kHz, (c) 10 kHz, and (d) 100 kHz. Center values
considered in the study were σ = 1.4 × 106 S/m and μ = 60. The plus sign denotes a change in
response with respect to a change in permeability, and the box denotes a change in response with
respect to a change in conductivity
Fig. 19.29 Comparison of estimation metrics: (a) CRLB for variation in conductivity estimation,
(b) CRLB for variation in permeability estimation, (c) condition number, and (d) correlation
Chapter 20
Applications to NDE of Coatings
β−Phase Aluminide
(MCrAlY)
(Ni-based superalloy)
Fig. 20.1 An as-coated PWA286 coating on a GTD111 substrate (left) and the same coating after
2,400 h (right)
Conductivity (x105)
8.4
7.8
7.0
3.32
Depth
0
Zone2 Beta Zone1 IZ GTD111
Fig. 20.2 Standard model of the PWA286 thermal barrier coating. The conductivities are deter-
mined from measurements on known samples
Inverting the impedance data via the eight-layer algorithm means assigning a
value to L0 , the thickness of the top layer of Fig. 20.4, and assigning conductivities to
the remaining seven layers, each of which has the same thickness, L, that defines the
resolution of the reconstruction. This is done by a process of nonlinear least-squares
(NLSE), in which the eight variables are chosen to give the best fit to the impedance
data. The inversion is done in several steps, starting with the computation of a table
of data, from which the final solution is determined by interpolation. The process is
quite fast and is quite conservative in computational resources.
The results of the inversion process are fitted to the “standard model” of the
PWA286 TBC, that is shown in Fig. 20.2. The values of the conductivity of the
various layers are determined from inversions on known samples. In this manner,
we are able to determine the thickness of each zone. The results are tabulated in
Tables 20.1 and 20.2.
In several cases, the final computed conductivity profiles were obtained by a
process of “focussing” the eight-layer inversion algorithm. This is accomplished
by starting with a rather coarse grid, which allows the Beta Aluminide Zone and
Interdiffusion and Inner Beta Depleted Zone 1 to be generally located relative to the
GTD111 Substrate. Then we refine the grid between the Beta Aluminide Zone and
GTD111 Substrate to get a more precise value for the various zone boundaries, as
well as a more precise value for the conductivity of Zone 1 and the Interdiffusion
Zone. In this manner, we were able to locate the Interdiffusion Zone and determine
its conductivity for the “as-coated” samples, B1B and B1T. This is a fairly difficult
computation to carry out without such a “multigrid rezoning” technique.
The tabulated results confirm that the inversion algorithm produces excellent
reconstructions, with very good resolutions. In particular, we note the good agree-
ment between the computed and measured thickness of the beta-zone of each
sample. This is the critical datum, when it comes to determining the remaining life
of the thermal barrier coating.
414 20 Applications to NDE of Coatings
0.12
Reactance
−0.25
0.1 −0.3
−0.35
0.08
−0.4
0.06
−0.45
0.04 −0.5
0 5 10 15 20 25 30 35 40 45 50 0 5 10 15 20 25 30 35 40 45 50
Frequency (MHz) Frequency (MHz)
0.12 −0.25
Reactance
0.1 −0.3
−0.35
0.08
−0.4
0.06
−0.45
0.04 −0.5
0 5 10 15 20 25 30 35 40 45 50 0 5 10 15 20 25 30 35 40 45 50
Frequency (MHz) Frequency (MHz)
Fig. 20.3 Real (left) and imaginary (right) parts of the normalized change in impedance for the
“top” PWA286 samples (upper two figures) and “bottom” PWA286 samples (lower two figures)
20.1 Assessing Thermal Barrier Coatings 415
σ=0 Al2O3 L0
σ1 L
σ2 L
σ3 L
σ4 L
σ5 L
σ6 L
σ7 L
Fig. 20.4 The eight-layer inversion algorithm, in which L is the given resolution, and the objective
is to determine L0 and σ1 through σ7
We’ll demonstrate the eight-layer algorithm on the White-5 Top (“w5t”) sample.
The numbers that we get may be a little bit different than what are shown in
Table 20.2 because we’ll follow slightly different paths to the final solution.
The data for the problem are: coil IR = 12.5 mils, coil OR = 45 mils, coil HT = 8
mils, and coil turns = 23. The position of the coil within the probe coordinate system
is [0, 0, 5.7874] mils. VIC-3D R
requires a minimum value for the z-coordinate
of 4 mils, and the coil was recessed 1 mil beneath the surface of its container.
The remaining 0.7874 mils is numerical fitting to give the best response to known
lift-off conditions. The probe origin position is [0, 0, Zone2]; i.e., we model the Zone
2 (“Outer Beta Depleted”) height by the z-coordinate in the probe origin position.
We use the 21 even-valued frequencies between 10 and 50 MHz for the model and
measured input data.
Our first calculation uses Fig. 20.4 with L = 20 microns, so that the grid extends
140 microns into the coating. The result of the inversion produced a value of Zone
2=48.8 microns, but the conductivities of the seven layers were not well determined.
The second calculation uses an eight-layer grid, but now with all eight layers
lying within the structure. Since each layer is 20 microns tall, the grid extends
160 microns into the structure. The coil coordinate system is as before, but the
probe coordinates are fixed at [0, 0, 48.8] microns. The results of this inversion are:
σ1 = 897969.603411648, σ2 = 748852.078029096, σ3 = 576648.447023520, σ4 =
821222.023798830, σ5 = 888060.286078721, σ6 = 300016.292476106,
σ7 = 300026.618414368, σ8 = 796133.184675437.
416 20 Applications to NDE of Coatings
Table 20.1 Comparison of measured and computed coating thicknesses for the
PWA286 test
Average Computed
Sample Coating Coating Thickness Thickness
Identification Location Identification (microns) (microns)
1950F-“as coated” Top Side Inner Diffusion 10.6 10.0
(Blue-1) Inner Beta Depleted – –
Beta Phase Zone 132.0 132.0
Outer Beta Depleted – –
Bottom Side Inner Diffusion 10.1 10.0
Inner Beta Depleted – –
Beta Phase Zone 140.9 142.0
Outer Beta Depleted – –
1950F-250 cycles Top Side Inner Diffusion 22.4 20.0
(Blue-2) Inner Beta Depleted 8.9 9.4
Beta Phase Zone 122.4 120.0
Outer Beta Depleted 14.5 17.7
Bottom Side Inner Diffusion 22.2 22.6
Inner Beta Depleted 8.1 9.1
Beta Phase Zone 114.2 112.5
Outer Beta Depleted 11.6 20.0
1950F-500 cycles Top Side Inner Diffusion 24.4 26.3
(Blue-3) Inner Beta Depleted 14.9 14.4
Beta Phase Zone 94.6 93.0
Outer Beta Depleted 17.9 19.2
Bottom Side Inner Diffusion 24.8 28.2
Inner Beta Depleted 16.8 15.3
Beta Phase Zone 98.0 94.5
Outer Beta Depleted 19.7 19.7
Even though there is considerable oscillation in the results (as is typical for these
problems), we use an averaging procedure, together with mixture theory, to make
sense of them. One reason for using a large number of layers is to allow averaging.
Note that the average of σ3 and σ4 is 6.99 × 105 S/m, which means that Zone 1
covers layers 3 and 4 (at least) (see Fig. 20.2). Furthermore, when we get such
agreement, it means that Zone 1 probably ends at a depth of 80 microns.
We can get more information about the start of Zone 1, and all of Beta by using
mixture theory on σ1 and σ2 . Let L be the volume-fraction associated with σ = 8.4×
105 S/m. Then the law for combining Beta and Zone 1 is: 8.9797 × 105 + 7.4885 ×
105 = 16.4682 × 105 = 16.8 × 105L+ 14 × 105(1 − L), or L = 0.8815. Since the total
distance covered is two layers, the length of Beta is 0.8815 × 40 = 35.26 microns.
The remaining 4.74 microns of the second layer belong to Zone 1, which means that
Zone 1 = 4.74 + 40 = 47.4 microns. Because of the oscillations in σ5 through σ7 ,
we must use a third calculation to compute IZ. It is pretty clear that σ8 is associated
with the GTD111 substrate. Another thing that makes us suspicious of the results
20.1 Assessing Thermal Barrier Coatings 417
Table 20.2 Comparison of measured and computed coating thicknesses for the
PWA286 test
Average Computed
Sample Coating Coating Thickness Thickness
Identification Location Identification (microns) (microns)
1950F-1000 cycles Top Side Inner Diffusion 25.8 26.4
(Blue-4) Inner Beta Depleted 34.0 33.0
Beta Phase Zone 62.6 63.8
Outer Beta Depleted 37.5 41.5
Bottom Side Inner Diffusion 23.7 25.0
Inner Beta Depleted 38.1 40.1
Beta Phase Zone 52.6 53.0
Outer Beta Depleted 33.2 34.8
1950F-1938 cycles Top Side Inner Diffusion 23.6 24.0
(Blue-5) Inner Beta Depleted 118.7 119.0
Beta Phase Zone – –
Outer Beta Depleted – 34.3
Bottom Side Inner Diffusion 28.4 25.0
Inner Beta Depleted 61.5 64.0
Beta Phase Zone 15.2 16.0
Outer Beta Depleted 48.5 46.3
1850F-3500 cycles Top Side Inner Diffusion 13.5 13.7
(White-5) Inner Beta Depleted 46.0 45.2
Beta Phase Zone 35.2 34.8
Outer Beta Depleted 50.2 44.6
Bottom Side Inner Diffusion 14.2 15.1
Inner Beta Depleted 45.9 45.5
Beta Phase Zone 27.6 29.6
Outer Beta Depleted 46.9 53.1
for σ5 through σ7 is that σ6 and σ7 are very close to the lower nodal value in the
two-point interpolation table. Thus, the NLSE algorithm may have “hit the stops”
for these two values.
At this point, we have concluded that Zone 2 is 48.8 micrometers thick, the Beta
Zone is 35.26 micrometers thick, and Zone 1 is 44.74 micrometers thick. We will
apply the linear-filter algorithm to determine the thickness of the remaining zone,
IZ, and to do this we start with the four-layer model shown in Fig. 20.5.
We first define the 16 nodes of an interpolation table in Table 20.3, which shows
the conductivities (divided by 105 ) of layers 1–4, respectively.
418 20 Applications to NDE of Coatings
0mm
Beta: σ = 8.4E5
0.03526
Zone 1: σ = 7E5
0.080
Layer 1: σ = 3.32E5, 7.8E5
0.095
Layer 2: σ = 3.32E5, 7.8E5
0.110
Layer 3: σ = 3.32E5, 7.8E5
0.125
Layer 4: σ = 3.32E5, 7.8E5
0.140
GTD111: σ=7.8E5
Fig. 20.5 Illustrating the four-layer algorithm that will be applied to the determination of IZ of
the thermal barrier coating
In calculating β according to (15.6), we use the “White 5 Top” data of Fig. 20.3
as the measured (input) data for Z1 , and the model data corresponding to each of the
nodes in Table 20.3 for Z2 . The results are shown in Table 20.4.
The data in the third column of Table 20.4 are entered into an interpolation table
for NLSE. The nonlinear least-squares problem now becomes: min pi |β (pi ) − 1|,
where pi , i = 1, .., 4 are the four conductivities of Fig. 20.5. The result of the NLSE
inversion is shown in Table 20.5. The results agree well with our standard model
of Fig. 20.2, except that p4 = 3.754 × 105 is too small. Because of the relatively
large value of σ4 , however, we conclude that the solution is fairly insensitive to this
parameter, so that we can set its value to 7.8 × 105 to agree with the standard model.
20.1 Assessing Thermal Barrier Coatings 419
Table 20.4 Results for β corresponding to the nodes of Table 20.3. The measured data are the
“White 5 Top” data of Fig. 20.3
Node β |β − 1| Node β |β − 1|
1 1.0032 + j0.0071912 0.00787 9 1.0014 + j0.005348 0.005528
2 1.0025 + j0.0009032 0.002658 10 1.0008 − j0.0005944 0.000997
3 1.0018 + j0.00266 0.003212 11 1.0001 + j0.0011611 0.001165
4 1.0009 − j0.0030114 0.00314 12 0.99929 − j0.0042202 0.00428
5 1.0015 + j0.0041487 0.004412 13 0.99975 + j0.0026443 0.002656
6 1.0008 − j0.0016615 0.00184 14 0.99912 − j0.0028707 0.0030
7 1.000 + j0.00012423 0.000124 15 0.99832 − j0.0011249 0.00202
8 0.99912 − j0.0051704 0.005245 16 0.99756 − j0.0061376 0.0066
Table 20.5 Results of min pi |β (pi ) − 1|. σ denotes the sensitivity of the solution
to the conductivity. The conductivities are normalized to 105
Φ p1 /σ1 p2 /σ2 p3 /σ3 p4 /σ4
0.4899173(−3) 3.334/0.4589 7.8/0.5568 7.535/0.8178 3.754/1.3211
This implies that IZ exactly occupies Layer 1 of Fig. 20.5, and the remaining
three layers belong to the GTD111 substrate. Hence, IZ is 15 micrometers thick.
We tabulate the computed results for “White 5 Top” along with the average of
the experimentally measured results in Table 20.6. The results computed by the
β -scale method agree well with the measurements and the results computed earlier
by another method, as can be seen in Table 20.2.
Chapter 21
Model-Assisted Probability of Detection
21.1 Introduction
An overview of the MAPOD process is presented in this section based on prior work
[121,124–128]. A block diagram of the model building process for a model-assisted
POD evaluation is presented in Fig. 21.1. This process was developed primarily
through the efforts of Bruce Thompson and Chuck Annis and can be found in parts
in the Appendix H of the MIL-HNBK 1823A [121]. For any POD study, the scope
of the POD study must be assessed and all critical factors for the NDT technique,
part material, part geometry, and discontinuity characteristics that control signal
and noise must be identified. As well, the amount of variability (as distribution
functions) and the impact must be assessed through expert knowledge, experimental
results, and/or model-based studies.
Once key factors have been determined, an assessment can be made on whether
to pursue a model-based POD evaluation based on the cost of a fully experimental-
based study and the performance of available NDE simulation tools. Knowledge
of the accuracy and speed of NDT simulations is a critical part of this assessment
process. Thus, an evaluation of the model quality is often necessary at this stage.
For many applications, there will be a mix of simulated and empirical studies that
will provide the greatest coverage of the key factors for an inspection technique.
For example, certain parameters such as the dimensions and location of a crack in
the test part can be more easily controlled and varied through simulated studies.
As well, rare events or factor conditions can be readily assessed and incorporated
in the MAPOD evaluation. Alternatively, noise data from material specimens and
Fig. 21.1 Model-Assisted POD model building process with complete approach to uncertainty
propagation in MAPOD from MIL-HNBK 1823A, Appendix H (2009) [121]
21.2 MAPOD Process 423
measurement system noise can be evaluated quickly through low cost experimental
studies. From this perspective, a model-assisted POD evaluation has the potential
to not only reduce the cost of POD study through reduced test sample requirements
but also to improve the quality of the assessment by including physics-based models
that fully address the factors driving the evaluation process.
A critical component of a MAPOD evaluation is the assessment and propagation
of variability in parameter conditions and the uncertainty in the POD assessment.
Experimental-based POD evaluations must consider the impact of limited samples
on the confidence bounds for a POD model fits. Labels were added to Fig. 21.1 to
present a complete perspective on uncertainty propagation during the POD study.
As part of quantifying the effects of changing the key factors, the amount of
variation of each factor must be well understood [124, 125]. Variability in NDE
measurements is prevalent due to varying part geometries, material properties,
surface conditions, flaw morphology, NDE hardware, and human factors. A typical
representation of variability for an input factor or condition is as a probability
density function (pdf). In addition, confidence bounds for the pdf may be considered
as variation measures of the pdf distribution parameters. Multivariate distributions
would thus include model parameters and a variation (covariance) matrix.
Before models can be directly applied in POD evaluations, two steps are needed
to ensure their performance. First, model calibration involves model adjustments
to variables in a way that mimics the NDE technique procedure. In many cases,
gains and thresholds are set based on the desired response to a known calibration
standard. Parameters of the model are essentially fit to obtain the best match with a
limited set of empirical data acquired according to a calibration procedure. Bayesian
calibration approaches have been implemented to facilitate this analysis with limited
data [137]. Second, model validation ensures that models are in agreement with
well-controlled studies for the appropriate range of conditions expected in practice.
Note, uncertainty bounds on the measurements and numerical model error [129]
must be tracked and extended to any model-assisted evaluation.
Stochastic numerical models provide a means to efficiently represent random
variables in physical systems without excessive computational overhead. New
efficient methods such as the probabilistic collocation methods (PCM) will en-
able the greater use of stochastic studies with existing NDE models [130].
Fundamentally, this POD evaluation including uncertainty bounds becomes a two-
level analysis. Parameter variability is associated with the inherent variability or
randomness of a factor. Alternatively, uncertainty is associated with imperfect
knowledge, often requiring the need for more or better quality data. In practice,
parameter variability and uncertainty can be represented by the random variables
of a statistical distribution and their associated confidence intervals. Methods such
as second order probability analysis can be applied [131]. MAPOD demonstrations
using this approach have recently been performed [134, 136].
424 21 Model-Assisted Probability of Detection
Lastly, at the final stage of the POD evaluation, experimental and theoretically
assessed “â vs. a” models must be resolved. Here, an evaluation of the full noise and
signal distributions as a function of crack length is performed and the call criteria is
applied to evaluation POD and probability of a false call (POFC) or false call rate.
Uncertainty (or credibility) bounds on the POD evaluation must be calculated given
limited empirical samples, factor variability, uncertainty propagation, and model
error. Bayesian methods are ideally suited to incorporate empirical data with NDE
models that include prior information/distributions. A typical example of applying
Bayesian methods is through the addition of new empirical data to evaluate the
posterior distribution, a refinement to the original prior distribution is achieved.
Numerical methods such as Markov Chain Monte Carlo (MCMC) methods and
Bayes Factors can be applied to perform this evaluation [125, 132]. Note, care must
be taken to ensure that all assumptions applied in a Bayesian analysis are valid.
Although this process provides significant complexity to address the most complex
inspection problems, one must keep in mind the goal to simplify.
The MAPOD case study problem presented here is the detection of cracks under in-
stalled countersunk fasteners in an airframe structures (see Fig. 21.2) [124,133,134].
For this problem, the key factors to include in the evaluation were identified from
experience with the inspection problem and prior experimental studies. The probe
characteristic response, lift-off, and scan resolution are important factors concerning
the NDE measurement. The surface condition of the samples, thickness of the
layers, type of fastener, fastener fit, and distance between adjacent fasteners and
edges are identified as significant factors related to part geometry and material.
The dimensions, aspect ratio, location around fastener site, and morphology are
Fig. 21.2 Diagram of a fastener site with (a) a first layer corner crack, (b) a second layer corner
crack, and (c) a second layer through crack
21.3 Case Study for MAPOD Evaluation of NDI of Fastener Sites... 425
some error for select experimental data points. However, the transition region for
the crack response just exceeding the noise threshold is not in perfect agreement.
This difference is the primary source of differences in the two POD calculations.
The presence of coherent noise related to surface features (poor paint quality) was
observed for select experimental data. In particular, three of the four outliers noted
in Fig. 21.5a and b were found to be correlated with the presence of surface related
noise. The other outlier was found to be in close proximity to a steel fastener
and experienced greater noise as well. It is possible to actually shift the crack
measurement either out or in from the optimal location to minimize the effect of
the surface noise.
Fig. 21.3 Calibration fit of model with respect to experimental data for titanium fastener site (in
Volts). Comparison plots include (a) the measurement plane, (b) the horizontal component and (c)
the vertical component responses
â = β0 + β1 f (a1 , β2 ) + ε , (21.1)
where ε ∼ N(0, σε2 ), f () is a function call for a physics-based model, β0 and β1 are
model calibration parameters, and β2 is a random variable associated with crack
aspect ratio (b/a). An initial study was performed where the random variables
428 21 Model-Assisted Probability of Detection
Fig. 21.4 Monte Carlo simulation results for varying length of (a) first layer (corner) cracks and
(b) second layer (through and corner) cracks
Fig. 21.5 POD evaluation results for experimental and full-model-assisted POD studies for
(a) first layer corner cracks and (b) second layer cracks
that include aspect ratio as a random variable as shown in Fig. 21.6 can reasonably
address both the mean response and nonconstant variance trends observed in
experimental results. The evaluation of these stochastic model parameters can
be achieved through evaluation of hierarchical Bayesian models. Gelman et al
introduced hierarchical Bayesian models for these classes of problems [146] and
430 21 Model-Assisted Probability of Detection
Fig. 21.6 Simulated results from the surrogate eddy-current model response as a function of crack
length based on VIC-3D R
numerical simulations for first and second layer cracks at a fastener site.
The aspect ratio is considered a Gaussian random variable with the mean and standard deviation
prescribed
where β0 and β1 are offset and mean slope terms of the model, respectively, η is a
random variable associated with varying-slope in the model, and ση2 is the variance
in slope parameter. Here, the random variable, η , is used to simulate the increasing
variance with increasing flaw size present in the physics-based model shown in
Fig. 21.6. The goal of this study is to assess how accurately these four parameters,
β0 , β1 , ση2 and, σâ2 , can be simultaneously estimated with respect to known values.
Two example estimation problems are presented in Fig. 21.7. Test case values
were selected to closely represent examples in prior experimental and simulated
21.4 Bayesian Methods for Estimating Uncertain Parameters in POD/MAPOD Evaluation 431
Fig. 21.7 Hierarchical model test cases estimating calibration parameters and variation in mea-
surement noise and stochastic model slope. Test cases shown (a) for strong model slope variation
and (b) for varying both slope and measurement noise significant factors
results in Fig. 21.6. The first case (a) investigated the condition where variance as a
function of flaw size dominates measurement noise (i.e., the variance independent
of flaw size). Using only 100 samples, the estimates for the two variance terms,
and, were found to be 0.2580 and 0.00139, respectively, close to the exact values of
0.300 and 0.0010. However, the 95 % credible bounds for both estimates just missed
containing the true values. This error may have been due to the specific random
sample used with the limited number of points or could be something systematic in
the estimation problem. Note, the estimates for the calibration parameters, β0 and
β1 , were found to be in good agreement with the true values for these case studies.
The variance terms appear to be the more challenging parameters to estimate in the
hierarchical model.
A second case in Fig. 21.7b investigated the condition where variance as a
function of flaw size is on a similar order as the measurement noise (variance
independent of flaw size). Using only 100 samples, the estimates for the two
variance terms, ση2 and σâ2 , were again found to be quite close to the exact
values of 0.100 and 0.0050, respectively, and the credible bounds included the
true values of the parameters. Both variance parameters are slightly overestimated,
while the calibration parameter β1 was slightly underestimated. All in all, these case
studies demonstrate the potential of simultaneously estimating the model calibration
parameters, model random variables, and measurement error.
432 21 Model-Assisted Probability of Detection
Fig. 21.8 Evaluation of (a) characterization error (ê j ) with respect to damage conditions (ak ) is
analogous to (b) â-vs-a POD analysis
Fig. 21.9 Model-assisted process for NDE characterization error (CE) evaluation
with k = 1 · · · N key factors, for example the critical flaw size. Gaussian process (GP)
models are ideally suited to evaluate this complex multidimensional relationship and
provide a useful model to support CBM+. A Gaussian Process model is a principled
probabilistic approach to representing data as a collection of random variables with
the property that the joint distribution of any finite subset is a Gaussian function.
This task will be to develop an empirical-based evaluation procedure following
MIL-HDBK-1823A but leveraging Gaussian process models.
Although experimental data are necessary in any measurement system evalua-
tion, due to the large number of factors that often must be addressed and the need
for a statistically significant number of test specimens, a model-assisted approach is
proposed as an alternative to an empirical-based evaluation of characterization error.
This process of model-assisted POD evaluation [121] may be adapted to address
a model-assisted approach to characterization error evaluation. A modified process
diagram for model-assisted NDE characterization error (CE) evaluation is presented
in Fig. 21.9. Throughout this process, proper error evaluation and uncertainty
434 21 Model-Assisted Probability of Detection
1. Burrows, M.L.: A theory of eddy-current flaw detection. PhD Thesis, University of Michigan,
University Microfilms, Inc., Ann Arbor, Michigan (1964)
2. Förster, F., Breitfeld, H.: Theoretische und Experimentelle Grundlagen der zerstörungsfreien
Werkstoffprüfung mit Wirbelstromverfahren, (Parts I and II). Z. Metallkd. 43(5), 163–180
(1952)
3. Förster, F., Breitfeld, H., Stambke, K.: Theoretishe und Experimentelle Grundlagen der
zerstörungsfreien Werkstoffprüfung mit Wirbelstromverfahren, (Parts III–VII). Z. Metallkd.
45(5), 166–199 and 221–226 (1954)
4. Dodd, C.V., Deeds, W.E.: Analytical solutions to eddy-current probe coil problems. J. Appl.
Phys. 39(6), 2829–2838 (1968)
5. Luquire, J.W., Deeds, W.E., Dodd, C.V.: Alternating current distribution between planar
conductors. J. Appl. Phys. 41(10), 3983–3991 (1970)
6. Cheng, C.C., Dodd, C.V., Deeds, W.E.: General analysis of probe coils near stratified
conductors. Int. J. Nondestr. Test. 3, 109–130 (1971)
7. Nestor, C.W., Jr., Dodd, C.V., Deeds, W.E.: Analysis and computer programs for eddy current
coils concentric with multiple cylindrical conductors. Report No. ORNL-5220, Oak Ridge
National Laboratory, Oak Ridge, TN 37830, July 1979
8. Deeds, W.E., Dodd, C.V., Scott, G.W.: Computer-aided design of multifrequency eddy-
current tests for layered conductors with multiple property variations. Report No. ORNL/TM-
6858, Oak Ridge National Laboratory, Oak Ridge, TN 37830, October 1979
9. Miller, E.K.: Model-based parameter estimation in electromagnetics: III–applications to EM
integral equations. Appl. Comput. Electrom. 10(3), 9–29 (1995)
10. Murphy, K., Sabbagh, H.A.: A boundary-integral code for electromagnetic nondestructive
evaluation. In: Conference Proceedings: 12th Annual Review of Progress in Applied Compu-
tational Electromagnetics, Applied Computational Electromagnetics Society, 18–22 March
1996, pp. 171–178
11. Xie, H., Song, J., Yang, M., Nakagawa, N.: A novel boundary integral equation for surface
crack model. In: Thompson, D.O., Chimenti, D.E. (eds.) Review of Progress in Quantitative
Nondestructive Evaluation, vol. 29, pp. 329–336. American Institute of Physics, Melville
(2010)
12. Harrington, R.F.: Field Computation by Moment Methods. The Macmillan Company,
New York (1968)
13. Berreman, D.W.: Optics in Stratified and Anisotropic Media: 4 × 4-Matrix Formulation.
J. Opt. Soc. Am. 62(4), 502–510 (April 1972)
14. Altman, C., Schatzberg, A.: Appl. Phys. B 28, 327–333 (1982)
15. Altman, C., Schatzberg, A.: Appl. Phys. B 26, 147–153 (1981)
16. Altman, C., Schatzberg, A., Suchy, K.: IEEE Trans. Antenn. Propag. AP-32(11) (November
1984)
17. Schatzberg, A., Altman, C.: J. Plasma Phys. 26(Part 2), 333–344 (1981)
18. Suchy, K., Altman, C.: J. Plasma Phys. 13(Part 3), 437–449 (1975)
19. Krowne, C.M.: IEEE Antennas and Propagation Symposium Digest, Boston, MA, 25–29 June
1984, pp. 569–572
20. Krowne, C.M.: IEEE Trans. Microw. Theor. Tech. MTT-32(12), 1617–1625 (December
1984)
21. Krowne, C.M.: IEEE Trans. Antenn. Propag. AP-32(11), 1224–1230 (November 1984).
22. Roberts, T.M., Sabbagh, H.A., Sabbagh, L.D.: Electromagnetic interactions with an
anisotropic slab. IEEE Trans. Magn. 24(6), 3193–3200 (November 1988)
23. Roberts, T.M., Sabbagh, H.A., Sabbagh, L.D.: Electromagnetic scattering for a class of
anisotropic layered media. J. Math. Phys. 29, 2675–2681 (December 1988)
24. Roberts, T.M.: Explicit eigenmodes for anisotropic media. IEEE Trans. Magn. 26(6),
3064–3071 (November 1990)
25. Bowler, J.R., Sabbagh, L.D., Sabbagh, H.A.: A theoretical and computational model of eddy-
current probes incorporating volume integral and conjugate gradient methods. IEEE Trans.
Magn. 25(3), pp. 2650–2664 (May 1989)
26. Sabbagh, H.A., Bowler, J.R., Sabbagh, L.D.: A model of eddy-current probes with ferrite
cores. Nondestr. Test. Eval. 5(1), 67–79 (1989)
27. Sabbagh, H.A.: A model of eddy-current probes with ferrite cores. IEEE Trans. Magn. MAG-
23(3), 1888–1904 (May 1987)
28. Sabbagh, H.A., Sabbagh, L.D., Bowler, J.R.: A volume-integral code for eddy-current
nondestructive evaluation. Int. J. Comput. Math. Elec. Electron. Eng. 9(Suppl. A), 67–70
(1990)
29. Bowler, J.R., Sabbagh, L.D., Sabbagh, H.A.: Eddy-current probe impedance due to a surface
slot in a conductor. IEEE Trans. Magn. 26(2), 889–892 (March 1990)
30. Rao, S.M., Wilton, D.R., Glisson, A.W.: Electromagnetic scattering by surfaces of arbitrary
shape. IEEE Trans. Antenn. Propag. AP-30(3), 409–418 (May 1982)
31. Aubin, J-P.: Approximation of Elliptic Boundary-Value Problems. Wiley-Interscience,
New York (1972)
32. Glisson, A.W., Wilton, D.R.: Simple and efficient numerical methods for problems of
electromagnetic radiation and scattering from surfaces. IEEE Trans. Antenn. Propag. AP-29,
593–603 (1980)
33. Wertgen, W., Jansen, R.H.: Efficient direct and iterative electrodynamic analysis of geomet-
rically complex MIC and MMIC structures. Int. J. Numer. Model. Electron. Network. Dev.
Field. 2(3), 153–186 (September 1989)
34. Yaghjian, A.D.: Electric dyadic Green’s functions in the source region. Proc. IEEE. 68,
248–263 (February 1980)
35. Burke, G.J., Dease, C.G., Didwall, E.M., Lytle, R.J.: Numerical modeling of subsurface
communication. UCID-20439 Rev. 1, Lawrence Livermore National Laboratory, August 1985
36. Catedra, M.F., Gago, E., Nuño, L.: A numerical scheme to obtain the RCS of three-
dimensional bodies of resonant size using the conjugate gradient method and the fast Fourier
transform. IEEE Trans. Antenn. Propag. 37(5), 528–537 (May 1989)
37. Peter, A., Zwamborn, M., van den Berg, P.M., Mooibroek, J., Koenis, F.T.C.: Computation of
three-dimensional electromagnetic-field distributions in a human body using the weak form
of the CGFFT method. Appl. Comput. Electrom. 7(2), 26–42 (Winter 1992)
38. Zwamborn, P., van den Berg, P.M.: The three-dimensional weak form of the conjugate
gradient FFT method for solving scattering problems. IEEE Trans. Microw. Theor. Tech.
40(9), 1757–1766 (September 1992)
39. Sabbagh, H.A.: Splines and their reciprocal-bases in volume-integral equations. IEEE Trans.
Magn. 29(6), 4142–4152 (November 1993)
40. http://www.sabbagh.com
41. Dongarra, J.J., Moler, C.B., Bunch, J.R., Stewart, G.W.: LINPACK Users’ Guide. Society for
Industrial and Applied Mathematics, Philadelphia (1979)
References 437
42. Hestenes, M.: Conjugate Direction Methods in Optimization. Springer, New York (1980)
43. Sarkar, T.P.: Application of the Conjugate Gradient Method in Electromagnetics and Signal
Processing. Elsevier, New York (1991)
44. Catedra, M.F., Torres, R.P., Basterrechea, J., Gago, E.: The CG-FFT Method: Application of
Signal Processing Techniques to Electromagnetics. Artech House, Boston (1995)
45. Peterson, A., Ray, S., Mittra, R.: Computational Methods for Electromagnetics. IEEE,
New York (1998)
46. Chew, W.C., Jin, J.M., Michielsssen, E., Song, J.M.: Fast and Efficient Algorithms in
Computational Electromagnetics. Artech House, Boston (2001)
47. Harrington, R.F.: Time-Harmonic Electromagnetic Fields. McGraw-Hill, New York (1961)
48. Collin, R.E.: Field Theory of Guided Waves. McGraw-Hill, New York (1960)
49. Sabbagh, H.A., Sabbagh, E.H., Murphy, R.K.: Recent advances in modeling eddy-current
probes. In: Thompson, D.O., Chimenti, D.E. (eds.) Review of Progress in Quantitative
Nondestructive Evaluation, vol. 21, pp. 423–429. American Institute of Physics, Melville
(2002)
50. Wozencraft, J.M., Jacobs, I.M.: Principles of Communication Engineering. Wiley, New York
(1965)
51. Umashankar, K.R., Nimmagadda, S., Taflove, A.: Numerical analysis of electromagnetic
scattering by electrically large objects using spatial decomposition technique. IEEE Trans.
Antenn. Propag. 40(8), 867–877 (August 1992)
52. Murphy, K., Sabbagh, H.A., Treece, J.C.: Thickness measurements with eddy-current probes:
a simple inversion problem. In: Thompson, D.O., Chimenti, D.E. (eds.) Review of Progress
in Quantitative Nondestructive Evaluation, vol. 13, pp. 927–934. Plenum Press, New York
(1994)
53. Lawson, C.L., Hanson, R.J.: Solving Least Squares Problems. Prentice-Hall, Inc., Englewood
Cliffs (1974)
54. Murphy, K., Sabbagh, H.A., Treece, J.C.: Some inversion problems in nondestructive
evaluation. In: Thompson, D.O., Chimenti, D.E. (eds.) Review of Progress in Quantitative
Nondestructive Evaluation, vol. 14, pp. 857–861. Plenum Press, New York (1995)
55. Baltzersen, O.: Model-based inversion of plate thickness and liftoff from eddy current probe
coil measurements. Mat. Eval. 51, 72–76 (1993)
56. Förster, F. Libby, H.: Electromagnetic Testing, 2nd edn, pp. 178–179. The American Society
or Nondestructive Testing, Columbus (1986)
57. Sabbagh, H.A., Murphy, R.K. Sabbagh, E.H.: Advances in modeling eddy-current NDE of
ferromagnetic bodies. In: Thompson, D.O., Chimenti, D.E. (eds.) Review of Progress in
Quantitative Nondestructive Evaluation, vol. 22, pp. 383–389. American Institute of Physics,
Melville (2003)
58. Rinnooy Kan, A.H.G., Timmer, G.T.: Stochastic global optimization methods. Part I:
clustering methods. Math. Program. 39, 27–56 (1987)
59. Rinnooy Kan, A.H.G., Timmer, G.T.: Stochastic global optimization methods Part II: multi
level methods. Math. Program. 39, 57–78 (1987)
60. Byrd, R.H., Dert, C.L., Rinnooy Kan, A.H.G., Schnabel, R.B.: Concurrent stochastic methods
for global optimization. Math. Program. 46, 1–29 (1990)
61. Nakhkash, M., Huang, Y., Fang, M.T.C.: Application of the multilevel single-linkage method
to one-dimensional electromagnetic inverse scattering problem. IEEE Trans. Antenn. Propag.
47(11), 1658–1668 (November 1999)
62. Golub, G.H., Van Loan, C.F.: Matrix Computations. The Johns Hopkins University Press,
Baltimore (1983)
63. Rousseeuw, P., Yohai, V.: Robust regression by means of s-estimators. In: Robust and
Nonlinear Time Series Analysis: Proceedings of a Workshop, pp. 256–272, 1984
64. Van Trees, H.L.: Detection, Estimation, and Modulation Theory: Part I. Wiley, New York
(1968)
65. Kay, S.M.: Fundamentals of Statistical Signal Processing: Estimation Theory. Prentice-Hall,
Inc., Upper Saddle River (1993)
438 References
66. Devaney, A.J., Tsihrintzis, G.A.: Maximum likelihood estimation of object location in
diffraction tomography. IEEE Trans. Signal Process. 39(3), 672–682 ( March 1991)
67. Herman, G.T., Lent, A., Hurwitz, H.: A storage-efficient algorithm for finding the regularized
solution of a large, inconsistent system of equations. J. Inst. Math. Appl. 25, 361–366 (1980)
68. Censor, Y.: Row-action methods for huge and sparse systems and their applications. SIAM
Rev. 23, 444–446 (October 1981)
69. Herman, G.T., Meyer, L.B.: Algebraic reconstruction techniques can be made computation-
ally efficient. IEEE Trans. Med. Imag. 12(3), 600–609 (September 1993)
70. Sabbagh, H.A., Sabbagh, L.D., Vernon, S.N.: Verification of an eddy-current flaw inversion
algorithm. IEEE Trans. Magn. 22(6), 1881–1886 (November 1986)
71. Papoulis, A.: A new algorithm in spectral analysis and band-limited extrapolation. IEEE
Trans. Circ. Syst. CAS-22, 735–742 (1975)
72. Youla, D.C.: Generalized image restoration by the method of alternating orthogonal projec-
tions. IEEE Trans. Circ. Syst. CAS-25, 694–702 (1978)
73. Youla, D.C., Webb, H.: Image restoration by the method of convex projections: Part 1-Theory.
IEEE Trans. Med. Imag. MI-1, 81–94 (October 1982)
74. Sezan, M.I., Stark, H.: Image restoration by the method of convex projections:
Part 2-applications and numerical results. IEEE Trans. Med. Imag. MI-1, 95–101 (October
1982)
75. Oskoui-Fard, P., Stark, H.: Tomographic image reconstruction using the theory of convex
projections. IEEE Trans. Med. Imag. 7(1), 45–58 (March 1988)
76. Bucci, O.M., D’Elia, G., Mazzarella, G., Panariello, G.: Antenna pattern synthesis: a new
general approach. Proc. IEEE. 82(3), 358–371 (March 1994)
77. Oh, S., Marks II, R.J., Atlas, L.E.: Kernel synthesis for generalized time-frequency distri-
butions using the method of alternating projections onto convex sets. IEEE Trans. Signal
Process. 42(7), 1653–1661 (July 1994)
78. Lent, A., Tuy, H.: An iterative method for the extrapolation of band limited functions. J. Math.
Anal. Appl. 83, 554–565 (1981)
79. Kaczmarz, S.: Angenaherte auflosung von systemen linearer gleichungen. Bull. Pol. Acad.
Sci. Lett. A, 6–8A, 355–357 (1937)
80. Tanabe, K.: Projection method for solving a singular system. Numer. Math. 17, 203–214
(1971)
81. Hounsfield, G.N.: A method of and apparatus for examination of a body by radiation such as
x-ray or gamma radiation. Patent Specification 1283915, The Patent Office (1972)
82. Kak, A.C., Slaney, M.: Principles of Computerized Tomographic Imaging. IEEE Press,
New York (1988)
83. Goodman, J.W.: Introduction to Fourier Optics. McGraw-Hill, San Francisco (1968)
84. Pratt, W.K.: Digital Image Processing. Wiley, New York (1978).
85. Sabbagh, L.D., Sabbagh, H.A., Klopfenstein, J.S.: Image enhancement via extrapolation
techniques: a two dimensional iterative scheme and a direct matrix inversion scheme. In:
Thompson, D.O., Chimenti, D.E. (eds.) Review of Progress in Quantitative Nondestructive
Evaluation, vol. 5, pp. 473–483. Plenum Press, New York (1986)
86. Sabbagh, L.D., Sabbagh, H.A.: Inversion of eddy current data and the reconstruction of flaws
Part 2: inversion of data. In: Thompson, D.O., Chimenti, D.E. (eds.) Review of Progress in
Quantitative Nondestructive Evaluation, vol. 6, pp. 619–626. Plenum Press, New York (1987)
87. Sabbagh, L.D., Sabbagh, H.A.: Eddy-current modeling and flaw reconstruction. J. Nondestr.
Eval. 7(1/2), 95–110 (1988)
88. Sabbagh, H.A., Sabbagh, L.D.: An eddy-current model for three-dimensional inversion. IEEE
Trans. Magn. MAG-22(4), 282–291 (July 1986)
89. Sabbagh, L.D., Sabbagh, H.A.: Eddy current modeling and signal processing in NDE. In:
Chen, C.H. (ed.) Signal Processing and Pattern Recognition in Nondestructive Evaluation of
Materials: NATO ASI Series, vol. F44, pp. 145–154. Springer, Berlin (1988)
90. Theodoulidis, T.P., Poulakis, N., Bowler, J.R.: Developments in modeling eddy current coil
interactions with a right-angled conductive wedge. In: Takahashi, S., Kikuchi, H. (eds.)
Electromagnetic Nondestructive Evaluation X, pp 41–48. IOS Press, Amsterdam (2007)
References 439
91. Sabbagh, H.A., Sabbagh, L.D., Roberts, T.M.: An eddy-current model and algorithm for
three-dimensional nondestructive evaluation of advanced composites. IEEE Trans. Magn.
24(6), 3201–3212 (November 1988)
92. Murphy, K., Sabbagh, H.A. Treece, J.C.: Verification of a model of eddy-current probes with
ferrite cores. In: Thompson, D.O., Chimenti, D.E. (eds) Review of Progress in Quantitative
Nondestructive Evaluation, vol. 13, pp. 1089–1093. Plenum Press, New York (1994)
93. Mittleman, D.M., Jacobsen, R.H., Buss, M.C.: T-ray imaging. IEEE J. Sel. Top. Quant.
Electron. 2(3), 679–692 (September 1996)
94. Sze, S.M.: Physics of Semiconductor Devices. Wiley, New York (1969)
95. Sabbagh, H.A., Murphy, R.K., Woo, L.W., Sabbagh, E.H., Krzywosz, K.: Recent advances
in modeling eddy-current probe-flaw interactions. In: Thompson, D.O., Chimenti, D.E. (eds.)
Review of Progress in Quantitative Nondestructive Evaluation, vol. 15, pp. 331–338. Plenum
Press, New York (1996)
96. Burke, S.K., Ditchburn, R.J.: Mutual impedance of planar eddy-current driver-pickup spiral
coils. Res. Nondestr. Eval. 19, 1–19 (2008)
97. Moré, J.J., Garbow, B.S., Hillstrom, K.E.: USER GUIDE FOR MINPACK-1. ANL-80-74,
Argonne National Laboratory, August 1980
98. Collin, R.E.: Foundations for Microwave Engineering, Chap. 4. McGraw-Hill Book Com-
pany, New York (1966)
99. Sabbagh, H.A., Sabbagh, E.H., Murphy, R.K., Nyenhuis, J.: Assessing thermal barrier
coatings by eddy current inversion. Mater. Eval. 59(11), 1307–1312 (November 2001)
100. Sabbagh, H.A., Sabbagh, E.H., Murphy, R.K., Nyenhuis, J.: Assessing thermal barrier
coatings by eddy current inversion. In: Thompson, D.O., Chimenti, D.E. (eds.) Review of
Progress in Quantitative Nondestructive Evaluation, vol. 21, pp. 722–727. American Institute
of Physics, Melville (2002)
101. Wang, S. Solid-State Electronics. McGraw-Hill Book Company, New York (1966)
102. Knopp, J.S., Aldrin, J.C., Misra, P.: Considerations in the validation and application of models
for eddy current inspection of cracks around fastener holes. J. Nondestr. Eval. 25(3), 123–138
(2006)
103. Carpenter, D.C.: Use of the finite element method in simulation and visualization of
electromagnetic nondestructive testing applications. Mater. Eval. 58(7), 877–881 (2000)
104. Palanisamy, R., Lord, W.: Prediction of eddy current probe signal trajectories. IEEE Trans.
Magn. 16(5), 1083–1085 (1980)
105. Knopp, J.S., Aldrin, J.C., Sabbagh, H.A., Jata, K.V.: Estimation theory metrics in electromag-
netic NDE, electromagnetic nondestructive evaluation workshop proceedings, Seoul, Korea,
10–12 June 2008
106. Trefethen, L.N. Bau, D.: Numerical Linear Algebra. SIAM, Philadelphia (1997)
107. Sabbagh, H.A., Murphy, R.K., Sabbagh, E.H. Aldrin, J.C., Knopp, J., Blodgett, M.: Com-
putational electromagnetics and model-based inversion: a modern paradigm for eddy-current
nondestructive evaluation. Appl. Comput. Electrom. 24(6), 533–540 (December 2009)
108. Sabbagh, E.M.: Circuit Analysis. Ronald Press Company, New York (1961)
109. Ramo, S., Whinnery, J.R., Van Duzer, T.: Fields and Waves in Communication Electronics.
Wiley, New York (1965)
110. Vernon, S.N.: The universal impedance diagram of the ferrite pot core eddy current transducer.
IEEE Trans. Magn. 25(3), 2639–2645 (May 1989)
111. Bowler, J.R., Sabbagh, L.D., Sabbagh, H.A.: The reduced impedance function for cup-core
eddy-current probes. IEEE Trans. Magn. 25(3), 2646–2649 (May 1989)
112. ’Validation of Direct and Inverse Models of SEACURE Ferritic Tubes With Benchmark Data,’
PID069961, Technical Update, October 2010, Electrical Power Research Institute. Prepared
by Victor Technologies, LLC
113. Todorov, E., Levesque, S., Ames, N., Krzywosz, K.: Measurement of magnetic properites of
ferromagnetic tubes for heat exchangers. Trans. Am. Nucl. Soc. 104, 297–298 (26–30 June
2011, Hollywood, Florida)
440 References
114. Scully, J.R.: Hidden corrosion, what should be measured to improve emerging anticipate
and manage strategies. In: 32nd Annual Review of Progress in Quantitative Non-destructive
Evaluation, QNDE-Brunswick, Maine (Key Note Lecture for Symposium Kick-off General
Session) (2005)
115. Tian, Y., Tamburrino, A., Udpa, S.S., Udpa, L.: Time-of-flight measurements from eddy
current tests. In: Thompson, D.O., Chimenti, D.E. (eds.) Review of Progress in Quantitative
Nondestructive Evaluation, vol. 22, pp. 593–600. American Institute of Physics, Melville
(2003)
116. Liu, Z., Safizadeh, M.-S. Forsyth, D.S., Lepine, B.A.: Data fusion method for the optimal
mixing of multi-frequency eddy current signals. In: Thompson, D.O., Chimenti, D.E.
(eds.) Review of Progress in Quantitative Nondestructive Evaluation, vol. 22, pp. 577–584.
American Institute of Physics, Melville (2003)
117. Liu, X., Deng, Y., Zeng, Z., Udpa, L., Knopp, J.: Model based inversion technique of
GMR signals using element-free Galerkin method. In: Conference Proceedings: 24th Annual
Review of Progress in Applied Computational Electromagnetics, Applied Computational
Electromagnetics Society, pp. 221–226 (March 2008)
118. Gray, J.N., Gray, T.A., Nakagawa, N., Thompson, R.B.: Models for predicting NDE reliabil-
ity, nondestructive evaluation and quality control, Metals Handbook 17, pp 702–715. ASM
International, Ohio, 1989
119. Thompson, R.B.: Using Physical Models of the Testing Process in Determination of
Probability of Detection. Mater. Eval. 69(7), 861–865 (2001)
120. Thompson, R. B.: A unified approach to the model-assisted determination of probability of
detection. Mater. Eval. 66, 667–673 (2008)
121. U.S. Department of Defense, Handbook, Nondestructive Evaluation System Reliability
Assessment, MIL-HDBK-1823A, 7 April 2009
122. Smith, K.D., Thompson, R.B., Brasche, L.: Model-Based POD: Successes and Opportunities,
1st Meeting of the MAPOD Working Group, Albuquerque, New Mexico, 23–24 September
2004. Web site: http://www.cnde.iastate.edu/MAPOD/
123. Harding, C., Hugo, G., Bowles, S.: Model-assisted Probability of Detection Validation of
Automated Ultrasonic Scanning for Crack Detection at Fastener Holes, Proceedings of the
10th Joint FAA/DoD/NASA Conference on Aging Aircraft, Palm Springs, CA, 16–19 April
2007
124. Aldrin, J.C., Knopp, J.S., Lindgren E.A., Jata, K.V.: Model-assisted probability of detection
(MAPOD) evaluation for eddy current inspection of fastener sites. Rev. progr. Quant.
Nondestr. Eval. 28, AIP, 1784–1791 (2009)
125. Aldrin, J.C., Medina, E.A., Lindgren, E.A., Buynak, C., Knopp, J.: Case studies for model-
assisted probabilistic reliability assessment for structural health monitoring systems. Rev.
progr. Quant. Nondestr. Eval. 30, AIP, 1589–1596 (2011)
126. Aldrin, J.C., Medina, E.A., Lindgren, E.A., Buynak, C.F., Knopp, J.S.: Protocol for reliability
assessment of structural health monitoring systems incorporating model-assisted probability
of detection (MAPOD) approach. In: Chang, F.-K. (ed.) Proceedings of the 8th International
Workshop on Structural Health Monitoring, Stanford, 13–15 September 2011
127. Aldrin, J.C., Medina, E.A., Santiago, J., Lindgren, E.A., Buynak, C.F., Knopp, J.S.:
Demonstration study for reliability assessment of SHM systems incorporating model-assisted
probability of detection approach. In: Thompson, D.O., Chimenti, D.E. (eds.) Review of
Progress in Quantitative Nondestructive Evaluation, vol. 31, pp. 1543–1550. American
Institute of Physics, Melville (2012)
128. Aldrin, J.C., Sabbagh, H.A., Murphy, R.K., Sabbagh, E.H., Knopp, J.S., Lindgren, E.A.,
Cherry, M.R.: Demonstration of model-assisted probability of detection evaluation method-
ology for eddy-current nondestructive evaluation. In: Thompson, D.O., Chimenti, D.E. (eds.)
Review of Progress in Quantitative Nondestructive Evaluation, vol. 31, pp. 1733–1740.
American Institute of Physics, Melville (2012)
129. Oberkampf, W.L., Roy, C.J.: Verification and Validation in Scientific Computing. Cambridge
University Press, New York (2010)
References 441
130. Knopp, J.S., Aldrin, J.C., Blodgett, M.P.: Efficient propagation of uncertainty in simula-
tions via the probabilistic collocation method. In: Chady, T., Gratkowski, S., Takagi, T.,
Udpa, S.S. (eds.) Electromagnetic Nondestructive Evaluation (XIV), pp. 141–148. IOS Press,
Amsterdam (2011)
131. Frey, H.C.: Quantitative analysis of uncertainty and variability in environmental policy mak-
ing. Fellowship Program for Environmental Science and Engineering, American Association
for the Advancement of Science, Washington, DC (1992)
132. Mahadevan, S., Rebba, R.: Validation of reliability computational models using Bayes
networks. Reliab. Eng. Syst. Saf. 87, 223–232 (2005)
133. Knopp, J.S., Aldrin, J.C., Lindgren, E., Annis, C.: Investigation of a model-assisted approach
to probability of detection evaluation. Rev. Prog. Quant. Nondestr. Eval. 26, 1775–1782
(2007)
134. Aldrin, J.C., Knopp, J.S.: Modeling and simulation for nondestructive testing with applica-
tions to aerospace structures. Mater. Eval. 66(1), 53–59 (2008)
135. Knopp, J.S., Aldrin, J.C., Misra, P.: Considerations in the validation and application of models
for eddy current inspection of cracks around fastener holes. J. Nondestr. Eval. 25(3), 123–138
(2006)
136. Dominguez, N., Feuillard, V., Jenson, F., Willaume, P.: Simulation assisted POD of a phased
array ultrasonic inspection in manufacturing. Rev. Progr. Quant. Nondestr. Eval. 31, AIP,
1765–1772 (2012)
137. Kennedy, M.C., O’Hagan, A.: Bayesian calibration of computer models. J. R. Stat. Soc. B 63,
425–464 (2001)
138. Gelman, A., Carlin, J.B., Stern, H.S., Rubin, D.B.: Bayesian Data Analysis. CRC, London
(2003)
139. Christensen, R., Johnson, W., Branscum, A.: Bayesian Ideas and Data Analysis: An Introduc-
tion for Scientists and Statisticians. CRC, Boca Raton (2010)
140. Meeker, W.Q., Escobar, L.A.: Introduction to the use of Bayesian methods for reliability data.
In: Statistical Methods for Reliability Data, pp. 343–368. Wiley, New York (1998)
141. Thompson, R.B.: A Bayesian approach to the inversion of NDE and SHM data. Rev. Progr.
Quant. Nondestr. Eval. 29, AIP, 679–686 (2010)
142. Leemans, D.V., Forsyth, D.: Bayesian approaches to using field test data in determining the
probability of detection. Mater. Eval. 62, 855–859 (2004)
143. Knopp, J.S., Zeng, L.: Statistical analysis of hit/miss data. Mater. Eval. 71(3), 323–329 (2013)
144. Li, M., Meeker, W.Q., Hovey, P.: Joint estimation of NDE inspection capability and flaw-size
distribution for in-service aircraft inspections. Res. Nondestr. Eval. 23, 104–123 (2012)
145. Aldrin, J.C., Knopp, J.S., Sabbagh, H.A.: Bayesian methods in probability of detection
estimation and model-assisted probability of detection evaluation. In: Thompson, D.O.,
Chimenti, D.E. (eds.) Review of Progress in Quantitative Nondestructive Evaluation, vol. 32,
pp. 1733–1740. American Institute of Physics, Melville (2013)
146. Gelman, A., Hill, J.: Data Analysis Using Regression and Multilevel/Hierarchical Models.
Cambridge University Press, Cambridge (2007)
147. ASTM Standard E2782, 2011: Standard guide for measurement systems analysis (MSA).
ASTM International, West Conshohocken (2011). doi:10.1520/E2782-11, www.astm.org
Index