Multifrequency Eddy Current Signal Analysis

Iowa State University Capstones, Theses and
Retrospective Theses and Dissertations

Dissertations
1997
Multifrequency eddy current signal analysis

Avanindra
Iowa State University
Follow this and additional works at: https://lib.dr.iastate.edu/rtd

Part of the Electromagnetics and Photonics Commons, Power and Energy Commons, and the
Signal Processing Commons
Recommended Citation
Avanindra, "Multifrequency eddy current signal analysis" (1997). Retrospective Theses and Dissertations. 16995.
https://lib.dr.iastate.edu/rtd/16995
This Thesis is brought to you for free and open access by the Iowa State University Capstones, Theses and Dissertations at Iowa State University Digital
Repository. It has been accepted for inclusion in Retrospective Theses and Dissertations by an authorized administrator of Iowa State University Digital
Repository. For more information, please contact [email protected].
Multifrequency eddy current signal analysis
by
Avanindra
A thesis submitted to the graduate faculty
in partial fulfillment of the requirements for the degree of
MASTER OF SCIENCE
Major: Electrical Engineering
Major Professor: Satish Udpa
Ames, Iowa
1997
Copyright © Avanindra, 1997. All rights reserved.

11
Graduate College
This is to certify that the Master's thesis of
Avanindra
has met the thesis requirements of Iowa State University
Signatures have been redacted for privacy

III
TABLE OF CONTENTS
ACKNOWLEDGEMENTS
ABSTRACT . . . . . . VI
1 INTRODUCTION 1
1.1 Motivation . . . . . . . . 2
1.2 Summary of contributions and outline of thesis. 6
2 PRINCIPLES OF NONDESTRUCTIVE EDDY CURRENT EVAL-

UATION . . . . . . . . . . . . . . . . . . . . . . . . . . . . 8
2.1 Basics of single frequency eddy current testing 8
2.1.1 Eddy current probes 12
2.1.2 Frequency selection . 15
2.2 Multifrequency eddy current techniques. 18
2.2.1 Need for multifrequency testing . 18
2.2.2 Fundamentals of multifrequency eddy current inspection 20

2.2.3 Operating principles of multiparameter analysis systems 24
3 ALGORITHMS FOR MULTIPARAMETER SIGNAL ANALYSIS 32
3.1 Introduction........... 32
3.2 Affine transformation equations 33
3.3 Conjugate gradient method .. 35

3.3.1 Heuristics to improve conjugate gradient minimization 41
3.4 Fourier series method . . . . . . . . . . . . . . . . . . . . . . . 45

IV
3.5 Discrete cosine transform method . . . . 50

3.5.1 Invariant parameters using DCT . 57
4 RESULTS AND CONCLUSIONS . . . . . . . . . . . 60

4.1 Conclusions and recommendations for future work . . . . . 63
BIBLIOGRAPHY . . . . . . . . . . . . . . . . . . . . . . . . 70
v
ACKNOWLEDGEMENTS
I wish to express my most sincere gratitude to my advisor Dr. Satish Udpa for his
guidance and invaluable help for this research. His insight and knowledge has motivated
me throughout my graduate study. The professional example set by him will always be
a source of inspiration. I would also like to thank Dr. Lalita Udpa for her willingness
to help whenever I needed to use the MIZ-40 eddy current instrument. I wish to thank
my committee members Dr. William Lord, Dr. Greg R. Luecke, Dr. Julie Dickerson
and Dr. Lalita Udpa for their support and encouragement. I am also thankful to Kiran
Kumar Dasoju and Pradeep Ramuhalli for their help in preparing the slides for my oral
examination. I would like to thank all the members of MCRG group at Iowa State
University who made working here an enjoyable experience.
I am grateful to my parents, family members and friends for their constant support
and encouragement. I wish to dedicate this work to them.

Vl
ABSTRACT
This thesis presents a novel procedure for representing and processing multifrequency
eddy current signals. Multifrequency eddy current NDE methods are used extensively
for the inspection of steam generator tubes in nuclear power plants. Existing methods
utilize computationally expensive time domain procedures to process the data. The
procedure outlined in the thesis uses frequency domain methods to minimize the com-
putational effort significantly. Two different approaches are evaluated. The first method
uses the Fourier descriptor to represent the signal. The Fourier coefficients are utilized
to obtain the rotation, scaling and translation parameters required for mixing. The
second approach uses the cosine transform. The mixing parameters are derived from
the transform coefficients. Fast algorithms can be used to compute the transform. A
spin-off of the approach is the ability to obtain rotation, translation and scale invariant
parameters from the coefficients directly. Experimental results supporting the validity
of the approach are presented.

1
1 INTRODUCTION
Nondestructive Evaluation (NDE) by definition involves the evaluation of a test
specimen for defects without affecting its usefulness or serviceability. Nondestructive
evaluation is used widely in various industries for the characterization of materials and
the detection of flaws. This is due to the high cost and time requirements associated with
destructive methods of testing. Indeed, in many situations destructive testing is not even
possible such as in the case of civil infrastructure. In recent years, with advancements in
computer technology, a lot of emphasis has been placed on NDE techniques for quality
control and maintenance testing, to increase the reliability of manufactured products.
NDE also plays a crucial role in areas such as nuclear power and transportation industries
in ensuring the integrity of operational components. Safety and a desire for avoiding
disruption in service are prime considerations in these cases.
Over the years a multitude of NDE methods have been developed as a result of
increased research efforts. These methods address a variety of needs and applications.
Prominent among them are ultrasonic, radiographic and electromagnetic methods (151.
Nondestructive evaluation methods share a common strategy. An excitation energy is
applied to the test specimen. The interaction between the specimen and the applied
energy produces a response signal which contains useful information about the material
characteristics or defects. This information is extracted by employing proper signal
processing techniques and interpreting the response signal. Computers are used for
signal processing and interpretation as well as the storage of the response signal.
Eddy current NDE methods rely on the principles of electromagnetic induction, to

2
identify or differentiate between a wide variety of physical, structural and metallurgical
conditions in electrically conductive materials. Eddy currents offer high sensitivity to
surface breaking defects as well as other anomalies that are close to the surface of the
specimen. Also eddy current testing techniques are "contactless" and do not require any
coupling to the test specimen as is the case with ultrasonic methods.
1.1 Motivation
Eddy current testing is used extensively for testing heat exchanger tubes. A heat
exch~nger is a device that is used to transfer heat from a fluid flowing on one side of
a barrier to another fluid flowing on the other side of the barrier [5]. Heat exchangers
are used in a variety of industries, including, power stations, petrochemical plants, oil
refineries and air conditioning and refrigeration units. The barrier between the fluids is
usually a metal wall such as that of a tube or pipe. Heat exchangers typically utilize a
bundle of tubes through which one of the fluids flows. The other fluid is directed in its
flow in the space outside the tubes through various arrangements of passes. This fluid
is contained by the heat exchanger shell. The most common materials used for tubes is
carbon steel and alloy steel because of the strength they offer. Because of excellent heat
conductance, different types of brass and copper alloys also find wide use in exchanger
manufacture. Corrosion plays a key role in the selection of construction materials. Where
corrosion is likely to be a major problem, more expensive corrosion resistant materials
such as stainless steels, nickel alloys and titanium may be used. Nuclear power plants,
for example use heat exchangers called steam generators that employ Inconel tubes.
Steam generator transfers heat from hot pressurized water in the tubes to the sur-
rounding water which boils and produces steam. The tubing is the boundary between
the radioactive water from the reactor, and the steam generated outside, which is used
to run turbines. Typically these U-tubes are made from Inconel and are approximately
3
7.5 m high with an internal diameter of 15.5 mm and 1 mm wall thickness [2]. The
array of tubes is forced through holes in the ferromagnetic support plates, which are
distributed along their length [19]. The tubing usually becomes radioactive with use
making the inspection environment hostile to humans.
Historically steam generator inspection has been a difficult problem. There are nu-
merous examples of unscheduled plant shutdowns. Since a plant outage in a utility can
typically cost $500,000 a day, there are strong economic incentives to develop reliable
NDE methods. Visual examination and ultrasonic techniques have limited use as they
are very slow and only a small percentage of the tubes can be inspected.
These problems have led to a widespread use of eddy current techniques for the
inspection of non-ferrous tubing, particularly in the nuclear power industry. In such
cases, eddy current inspection offers the following advantages:
• Relatively fast (0.5-1.0 mls testing speed).
• Can be carried out from inside the tube.
• Can detect both internal and external defects.
• Can detect both gradual and sharp defects.
• Provides quantitative flaw measurements (size, depth etc.).
• Repeatable.
• Lends itself to automated results analysis.
• Can be used to inspect a wide range of tubes.
Disadvantages of eddy current testing are that defects under known support struc-
tures (such as support plates,tube sheets, antivibration bars, roll transitions etc.) are
4
difficult to detect. Figure 1.1 shows some of the support structures that appear as "dis-
continuities" to the eddy current probe. To examine the issue further, consider a typical
situation in a steam generator, some 3600 tubes have to be examined. There are around
10 support plates and a tube sheet anchoring a single tube. Since a nuclear power plant
employs four steam generators, this results in over 158,000 sites where there is a high
probability that damage in the form of dents, cracks and corrosion pits is present. When
the defects are close to support structures the eddy current signal is masked by the re-
sponse from these structures making it difficult to characterize the defect. The detection
of such flaws is critical since failure to identify tubes with flaws in advanced stages of
deterioration can lead to catastrophic failure. Proper inspection can help determine
which tubes are likely to deteriorate unacceptably before the next overhaul. Such tubes
may then be plugged or replaced, depending on the condition of the steam generator.
In addition to the detection of such defects, eddy current testing can also be used to
monitor other conditions, such as build up of external sludge and to verify the degree of
expansion at tube sheets during manufacture [3].
One method of overcoming the difficulty of isolating the defect signal from the artifact
signals is to collect eddy current data at more than one excitation frequency. When
data at many frequencies are available, it is possible to eliminate the effect of unwanted
variables. In general, at least as many independent eddy current readings must be
taken as there are properties whose variation may affect the readings [28). If sinusoidal
eddy currents are used, only two independent quantities, such as the magnitude and
phase, can be measured at each frequency. Therefore, if more than two properties
need to be determined, multiple frequencies should be used. The increased amount of
information gained using multifrequency methods, may be combined to nullify unwanted
signal components.
This thesis presents some new techniques for processing multi frequency data to elim-
inate unwanted signals. Examples of unwanted signals include support plates, dents and
5
Dent
Deposit
Zipper
crack
Internal
cold
Pilgering
noise
Sleev
Expansion
Anomalies
Figure 1.1 Heat Exchanger Design and Manufacturing Discontinuities [14]
magnetite deposits on the outside of tubes. These benign signals sometimes mask harm-
ful defect signals. Multifrequency processing techniques can be employed to suppress
the unwanted signals without degrading the quality of defect signals. Other desirable
features that are sought also include low computational complexity in order to ensure
that the algorithm can be implemented in real time.
Both time and frequency domain approaches for processing the multifrequency data
are evaluated. The time domain method eliminates the unwanted signal by minimizing
a cost function in the least squares sense. A drawback is that the method is computa-
tionally intensive. The problem is minimized by employing fast algorithms and taking
advantage of the underlying heuristics. This allows the approach to be implemented in

6
real time.
1.2 Summary of contributions and outline of thesis
This thesis explores the feasibility of using frequency domain approaches for reduc-
ing the computational burden. Methods that are evaluated include Fourier series and
discrete cosine transform based approaches where the frequency domain coefficients of
the signals containing information are used for suppressing the benign signals. The de-
sirable feature of frequency domain techniques is that they require a fixed amount of
computation. This makes these approaches superior to the time domain techniques. The
validity of the approach is established using simulated as well as experimental data.
An additional outcome of the research is that cosine transform based features that
are insensitive to rotation, scaling and translation of the original signal can be obtained.
The invariant features offer a benefit in that the features allow independent scaling of the
real and imaginary components of the eddy current signal. These invariant descriptors
can be used for the classification of defect signals in automated nondestructive evaluation
systems.
Chapter 2 presents an introduction to the principles of eddy current nondestructive
evaluation. Both single and multifrequency eddy current methods are described. A
brief description of various multifrequency analysis systems that are currently used in
industry is also included.
A new approach for analyzing multifrequency eddy current signals is presented in
chapter 3. The chapter describes novel time and frequency domain techniques developed
for processing multifrequency eddy current signals. A section is devoted to describe
the cosine transform based features that are invariant to signal rotation, scaling and
translation.
Chapter 4 discusses the results obtained by applying the multifrequency techniques

7
to suppress support plate signals. The results are compared with those obtained using
a commercially available multifrequency instrument. This is followed by a section that
presents a few concluding remarks and recommendations for future work.

8
2 PRINCIPLES OF NONDESTRUCTIVE EDDY

CURRENT EVALUATION
2.1 Basics of single frequency eddy current testing
When a coil carrying an alternating current is brought in close proximity to a ma-
terial, an electromotive force proportional to the rate of change of the magnetic field is
induced in accordance with Faraday's law of electromagnetic induction. If the material
is conducting, closed current loops are established in the material. These are called eddy
currents. In conformity with Lenz's law, the direction of the induced eddy currents, and
consequently the secondary field generated by these currents, is such as to oppose the
change in the primary field [6]. This is shown in Figure 2.1.
Current carrying
Coil H primary
~
Direction of primary
/ current
~
current
) I
==--
H Induced
Figure 2.1 Alternating current coil over a conducting specimen showing di-
rection of primary and induced current
9
Assuming that the specimen is nonferromagnetic, the flux linkages associated with
the coil decreases because of the opposing nature of the primary and secondary fields.
The inductance of the coil, which is defined as the flux linkage per ampere, consequently
decreases as the coil is brought close to the specimen [30]. Accompanying the decrease
in inductance is an increase in resistance, to account for the eddy current losses in the
specimen. The presence of discontinuity or inhomogeneity in the test specimen causes
a reduction as well as a redistribution of the eddy currents as shown in the Figure 2.2.
Consequently, the changes in the coil impedance are reduced.
Conducting slab
Current carrying coil
--I
Eddy current Path
Crack
Figure 2.2 The disturbance of the flow of eddy current in the material due
to the presence of a defect
Eddy current responses of a single coil may be conveniently described by reference
to the "impedance plane". This is a graphical representation of the complex probe
impedance where the abscissa (X value) represents the resistance and the ordinate (Y
value) represents the inductive reactance [10]. Figure 2.3 shows the change in impedance
for a nonferromagnetic conducting specimen under different conditions. The underlying
process is more complicated when the test specimen is ferromagnetic. Due to the higher
permeability of the material the inductance of the coil increases [6]. This effect, very
often, dominates the decrease in inductance due to eddy currents. The resistance
10
(3)
•
R
1. Coil in air
2. Coil over a nonferromagnetic specimen
containing a discontinuity
3. Coil over a nonferromagnetic specimen
containing no discontinuities
Figure 2.3 Impedance plane trajectory of a coil over a nonferromagnetic

specimen with and without a discontinuity; changes are greatly
exaggerated for clarity
also increases due to eddy current and hysteresis losses.
The variations in coil impedance caused by discontinuities in the test specimen are
often very small in comparison with the quiescent value of the coil impedance. The
detection and measurement of these small changes is often accomplished using bridge
circuits [17]. Factors which influence the eddy current field, [16] [5] and therefore the
coil impedance, are:
• the separation between the coil and surface, called lift-off.
• the electrical conductivity of the specimen.
• the magnetic permeability of the specimen.

11
• the frequency of the AC inducing the eddy current field.
• the design of the eddy current probe
• geometric factors,
• discontinuities such as cracks, corrosion, pitting.
Successful detection and characterization of flaws requires a careful design of signal
processing procedures to negate or compensate for these effects. It is this elimination
of undesired response that forms the basis of much of the technology of eddy current
inspection. The following paragraphs explain the effect of each of the factors.
As mentioned earlier, lift-off is the separation or distance between the probe and
specimen surface. The closer a probe coil is to the surface the greater will be the effect
on coil impedance. This has two main effects: The "lift-off" signal is generated as the
probe is moved on and off the surface and there is a reduction in sensitivity as the coil to
material spacing increases. This fact is used to design probes to give a better response
signal as described in the next section.
The conductivity of test specimen has a very direct effect on the eddy current flow:
the greater the conductivity of the test material, the greater the flow of eddy currents
on the surface. Conductivity is often measured by an eddy current technique. Based
on the conductivity measurement inferences can be drawn about the different factors
which affect conductivity, such as material composition, heat treatment, surface coating
thickness, etc.
The magnetic permeability of a material may be described as the ease with which a
material can be magnetized. For non-ferrous metals such as copper, brass etc., and for
austenitic stainless steels the permeability is the same as that of 'free space'. For ferrous
metals however, the relative permeability J-lr may be several hundred, and this has a very
12
significant influence on the eddy current response. In addition the permeability varies
greatly within a metal part due to localized stresses and heating effects.
Geometric factors such as finite dimensions, curvature, edges, grooves etc. affect the
eddy current distribution and hence the response also. Test techniques must recognize
this, for example to test for a crack near an edge, the probe should be moved parallel
to the edge so that any change in response is due to a discontinuity and not due to the
fact that eddy current redistributes near an edge. The effect of frequency and probe
configuration are also factors which affect eddy current response. These are factors
which can be controlled by the designer. These are explained in greater detail in the
following sections.
The change in the impedance of the coil, which is known as the eddy current signal is
measured by most instruments. The amplitude and phase of the signals can be displayed
on a cathode ray oscilloscope. Impedance plane diagrams are also used for displaying
eddy current signals. The analysis of these signals and interpretation give information
about a variety of material properties including conductivity, permeability, specimen
thickness and lift-off. In the case of tube inspection, the phase of the signal can give
important information regarding the nature of the defect and depth.
2.1.1 Eddy current probes
One obvious factor which influences the impedance of the coil is its configuration.
It is necessary to use a coil which is most sensitive to the kind of defects expected
during the test. Hence the choice of probes is dictated by the specimen to be evaluated.
Several different probe types are available. Sometimes special probes are designed to
meet the needs of a specific application. The probe configurations described below are
for inspecting tubes.
An absolute probe which is a single coil wound circumferentially, is sensitive to both
inner diameter (ID) and outer diameter (OD) defects [7]. As the name suggests, an
13
absolute probe can be used to evaluate conditions in absolute terms. It is useful for
measuring conditions which change slowly along the tube, such as gradual metal loss
(thinning). However due to sensitivity to slowly changing conditions, it is often difficult
to accurately identify smaller defects. In addition, changes in the coil parameters. due
to environmental factors and lift-off, can often mask changes due to discontinuities.
making signal interpretation very difficult. Figure 2.4 shows absolute mode response
for different defects.
A variation of the absolute eddy current probe is the differential eddy current probe.
Figure 2.5 shows a differential eddy current probe designed for inspecting tubes and
the response signal to various defects. The probe consists of two identical coils mounted
on the same axis as the tube but spaced apart by a small distance. The two coils form
the two arms of a bridge circuit. The bridge imbalance signal is the voltage difference
across the impedance of the two coils. As the probe is scanned past a discontinuity,
the change in impedance of the leading coil when it scans a discontinuity results in an
imbalance voltage. The differential voltage traces a trajectory bl-a2-b2 in the impedance
plane as shown in Figure 2.5. Similarly when the trailing coil scans the discontinuity,
the differential impedance traces the trajectory b2-c2-b3 in the opposite direction. The
shape of the impedance plane trajectory is a function of the nature of the discontinuity.
Figure 2.5 shows the eddy current signals for different types of defects.
Differential probes give extremely accurate evaluation of small defects such as inter-
nal pitting. Using a differential test, it is possible to evaluate both the size and depth
of defects, so that a small hole can be distinguished from a large area of shallow pitting.
Some other probes that are used are cross axis and driver-pickup probes [7] [14]. A cross
axis probe is a special type of differential probe in which one coil is wound circumferen-
tially, as in a standard coil, with the second coil in line with the tube. The second coil is
sensitive to circumferential cracks, to which a standard differential coil gives very little
response.
14
Short Long with Sharp Edges Long and Gradual
~ . • • : . I
'. '''\<'''''';~~ ~<~',);, ",' >.::,~ ........ " «;,..- '" ':< .... " ........ '" '<~>." ~~""""~"~,''''~' '", ", ~~" '" '''''~", '
y y y
1 3 3 7
x x x
2
5 9
Figure 2.4 Absolute mode response to discontinuities in a tube [14]
Driver-pickup coils have three coils. A driver coil is connected to the oscillator circuit
of the instrument, and the resulting magnetic field is sensed by separate pickup coils.
This has the advantage that the coil characteristics can be better optimized for the job
they are to perform. A comparison of Figures 2.5 and 2.4 demonstrates the sensitivity
of absolute and differential probes to gradual and sharp defects respectively.
Ocassionally a combination of absolute and differential coils is used for complete
defect characterization. For the eddy current test to be sensitive, the probe should fit
the tube as closely as possible. If the probe is a sloppy fit there will be two problems [7].
There is excessive noise due to wobble or lift-off, which makes it difficult to measure signal
phase, hence the defect cannot be characterized. Secondly the sensitivity is reduced:
particularly to defects at one end of the tube, due to the probe lying on the diametrically
15
Short Long with Sharp Edges Long and Gradual
,;; ...
"
, '
, '
," "
,
,"
. '
, "
:.,
", ' , ' - ~,,', ' ",'" .:. ~,
~
,
a2 y
blO
x x b8 x
b7
b9
blJ
c2 b3
b4
Figure 2.5 Differential mode response to discontinuities in a tube [14]
opposite end. If the probe is too tight, there is risk of it sticking in the tube, especially
when there are dents in the tube. Therefore it is necessary to allow some clearance.
The 'goodness of fit' is commonly expressed in terms of fill-factor - the ratio of probe
cross-sectional area to tube ID squared. Usually a fill factor of 0.7 or more is considered
a good fit.
2.1.2 Frequency selection
The selection of probe excitation frequency is the primary eddy current test param-
eter under operator control. Frequency selection affects both the magnitude of response
from different flaws and the phase relationship. Thus selection of operating frequency
16
is very important in obtaining good resolution of flaw signals in the presence of other
variables which may affect the test. The frequency selected for an eddy current test
is influenced by several factors. The frequency selected should be sensitive to all the
flaws and should be able to accurately size them. It should be sensitive to all relevant
extraneous discontinuities, and should be able to discriminate them from the damage.
An important criterion that is used in frequency selection is the "depth of penetration".
The eddy current density, and thus the strength of the response from a flaw, is greatest
on the surface of the material being tested [13] and declines in some special cases expo-
nentially with depth as shown in Figure 2.6. Hence higher frequencies enable inspection
of areas near or on the surface of the metal whereas lower frequencies allow inspection
of defects that are deeper within the specimen. The excitation frequency of the eddy
current probe is typically between 10 Hz and 10 MHz [18].
The "standard depth of penetration or skin depth" 8 is defined by the formula:
5 _ 1
(2.1 )
- ~7rp.(JI
where
f is the excitation frequency,
p. is the magnetic permeability,
a is the electrical conductivity
This is the depth at which the eddy current is lie (37%) of its surface value when
an infinite half sheet of current is induced in a metallic half plane. In the case of tube
inspection the skin d~pth merely serves as a useful measure to consider in choosing the
excitation frequency for a given material (p. and a ).
The frequency selected is thus influenced by the tubing wall thickness and conduc-
tivity of the material. Thicker walls require a lower frequency so that all the defects
17
c Eddy Current Strength
37% 100%
,Ie------~ '.
,
I
c:::>
- - - - - - - - - - - - - - -\
I I
)-
.'
/ 0
I
\
\
,
\
I \
/
.'
Figure 2.6 Depth of penetration for an infinite conductive half space
within the tube can be characterized. Other factors being equal, the frequency chosen is
inversely proportional to the material conductivity, since higher the conductivity lower
is the depth of penetration. The frequency selected is also dictated by the required form
of response; most commonly a test is set up in accordance with a specification such as
ASME boiler and pressure vessel code. For differential testing this means that the probe
wobble signal is set horizontally on the screen and the displayed phase of a through wall
hole is approximately 45 degrees clockwise from the horizontal [7] . The final selection
of the frequency is done using an appropriate calibration tube.
In order to provide accurate defect analysis, it is necessary to obtain a piece of tube of
the same material and dimensions as the tubes to be inspected and use it for calibration
of the instrument. Defects similar to those that are likely to occur during inspection, are
introduced into this specimen piece. These defects have known size and origin and can
be used for calibration. This is essential to optimize the frequency and sensitivity setting
required in order that actual defects can be classified relative to the reference defects.
Some calibration tubes attempt to simulate the types of defects encountered in service
18
[5]. For example, flat bottomed holes can be used to simulate pitting corrosion, spark
eroded notches can be used to simulate cracks and machined circumferential grooves can
be used to simulate general thinning. The main advantage is that more accurate depth
assessment of defects can be given.
As an aside, for ferromagnetic material the relative permeability is typically 500-2000
[5]. This means that for ferromagnetic materials the eddy currents are concentrated
on the surface and defects that are buried deep in the tube wall are not detectable.
Also, small variations in permeability gives rise to relatively high noise levels. Hence
eddy current inspection techniques have limited applicability to ferromagnetic tubing
inspection.
2.2 Multifrequency eddy current techniques
2.2.1 Need for multifrequency testing
Single frequency eddy current tests offer excellent sensitivity to a number of different
types of steam generator tubing under normal conditions. However conditions are often
complicated by a number of factors and consequently inspection needs cannot be effec-
tively solved by single frequency examination. Some extraneous discontinuities (such
as tube support plate, internal noise due to sludge build up, probe wobble, dents etc.)
distort or mask defect signals which are located near them [4]. This creates mistaken
interpretation of the eddy current signal resulting in unnecessary tube plugging. Lack
of detection may also lead to unexpected leaks and costly shutdowns. The detection of
other discontinuities such as wall thinning, sludge height, dents etc. needs several suc-
cessive probe passes at different frequencies and measurement mode [23]. This increases
the inspection time, which from safety as well as economic reasons should be kept to a
mlmmum.
State of the art multifrequency eddy current testing overcomes most of the single
19
frequency limitations. The multifrequency technique consists of collecting data simulta-
neously using several excitation frequencies from just one probe pulling. This provides
data which are analyzed using multifrequency mixing or multiparameter techniques.
This technique allows the effect of extraneous discontinuities to be "nullfied". Alter-
nating currents of different frequencies are either summed and sent simultaneously to
the test coil, or multiplexed and sent successively. After frequency separation, using
bandpass filters or the timing information in multiplexed method, the coil impedance is
estimated and displayed for each frequency separately. Multiparameter or mixing tech-
niques are then used to analyze the data to classify and characterize the defects. Several
superimposed test frequencies are sent to the probe. The frequency which would be
used normally when conducting a single frequency examination is the basic frequency
or primary frequency. The others are auxiliary frequencies. As the probe passes un-
der a discontinuity, the signatures obtained using different excitation frequencies can be
compared.
As mentioned earlier, each frequency is sensitive to a certain type of discontinuity.
A typical inspection uses three or four frequencies in both differential and absolute
modes. Low frequencies have a large skin depth and hence give clear signals from support
structures which are located away from the coil. They are sometimes used to determine
location along the tube. They can also be used to detect loose parts on the outside of
the tubes such as magnetite deposits. High frequencies have a much smaller skin depth
and for defects on the outside of the tube, the depth can be estimated from the phase of
the eddy current signal [2] [5]. Higher excitation frequency also gives information about
the probe wobble when the fill factor of the probe is low. Because of the different skin
depths at different frequencies, the relationship between signals from defects and support
features changes with frequency. Consequently, it is possible to combine the signals from
two different frequencies so as to subtract out a support feature but leave the signal from
defect. In effect, this means that multifrequency response signals have more information
20
which can be analyzed to extract relevant features. Thus multifrequency techniques
have the following advantages [14]:
• Collects data at several test frequencies simultaneously. This decreases inspection
time by preventing retesting of the tube with another frequency.
• Decreases the in-service inspection time and human exposure to radiation in nu-
clear plant inspections.
• Allows separation of discontinuities which give similar signals at one frequency.
• Improves sensitivity to different types of discontinuities.
• Easy interpretation of complex signals.
• Improves the detection, interpretation and sizing of defects even in the presence
of artifacts that complicate the analysis procedure.
• Can also be used in conjunction with multimode (differential or absolute) technique
for complete analysis.
2.2.2 Fundamentals of multifrequency eddy current inspection
The multifrequency method is useful for solving multiple signal problems, i.e. when
the useful defect signal occurs with background noise or in the presence of a benign sig-
nal from a support structure. The basic assumption, confirmed by practice, is the linear
superposition of the signals which states "the signal resulting from two discontinuities,
or from one discontinuity and background noise, is equal to the linear combination (
vectorial sum) of two signals considered separately". The whole concept of multifre-
quency analysis is based on this important "linearity principle". Figure 2.7 and 2.8
show the result of adding different discontinuities in a tube, from a differential probe. In
Figure 2.7 both the external and the internal defects are less than 10% of the tube wall
21
thickness. The presence of the two defects at the same location, leads to the resultant
signal which has a phase similar to a through hole defect signal. Note that in Figure
2.7 if the defect signatures are associated with vectors as shown, the resultant signal can
be interpreted as vector addition of the ID and OD defect signals. This interpretation
is possible since the two defect signals have similar shape. For Figure 2.8 also. the
principles of vector operation apply, but since the two signals have completely different
shapes it is not as apparent. The assumption of linearity is valid under very restrictive
conditions [28], A more detailed discussion of issue as well as methods for nonlinear
combination of signals have been proposed by Horne et al. [8] and others [27].
The multifrequency analysis uses a composite signal and subtracts the undesirable
signal to leave only the useful defect signal, as if the useful signal had been detected
alone. Thus the problem is to produce the signal used for subtraction.
Internal Discontinuity External Discontinuity Resultant Signal
Figure 2.7 Impedance plane trajectories observed when using a differential

probe on internal and external discontinuities in a tube along
with their vectorial representation
The result of subtraction would be perfect if the undesirable signal alone were avail-
able. The signal can then be directly subtracted from the composite signal to obtain
the defect signature. This is sometimes possible, when the signal is perfectly repeat-
able each time an inspection is carried out. In such cases there is no need for using
multiple excitation frequencies. However, the multi frequency process is far more flexible
22
o
Support Plate Edge
++ Defect Addition
.0
Signal with Noise Noise Pure Defect

Signal
Figure 2.8 Demonstration of linearity principle
and versatile because the signal need not be stored; it simply has to be measured at
more than one frequency. The measurement is done using the same probe excited with
multiple frequencies, with the auxiliary excitation frequency selected such that it results
in a preferential detection of the signal to be eliminated. Unfortunately the test carried
out at the auxiliary frequency is sensitive to both the undesirable signal and the defect
signal. Subtraction of auxiliary signal from the primary frequency signal does not cancel
the extraneous discontinuity signal; changing the frequency also changes the sensitivity
and the nature of the response to discontinuities. The net effect is that the desired signal
is also altered in amplitude and the phase is changed. Thus the signal interpretation
procedure has to be adjusted to account for these factors. Figure 2.9 explains this
phenomenon in terms of vectors.
The auxiliary or mixing frequency is chosen such that the amplitude and phase sep-
aration between the defect and the unwanted discontinuity signals is very different from
that obtained at the basic excitation frequency. Also the mixing frequency should not
be too different, otherwise the signal shapes would not be retained, which is required for
mixing. In reality when two undesirable responses to the two frequencies are adjusted in
23
,,
A'
UI', ,'-U2 Ul-U2=O
,V
Basic Frequency
FI Z
,
UI'
' Ul+OI
Vectorial
FI-F2
[]
Mixing Frequency "P2 Subtraction Mixer Output
F2
~
U2'
, U2+02
Y--~P; 01-02=0
01
O=OEFECf
U = UNWANTED DISCONTINUITY
Figure 2.9 Two frequency mix principle to suppress one unwanted discon-
tinuity [14]
amplitude and phase, their waveforms are slightly different, resulting in a residual. The
aim is to minimize this residual which looks like noise in the defect signal after mixing.
With the gain and phase controls of the mixing frequency, the unwanted discontinuity
(U2) signal is shaped in phase and amplitude to match its basic frequency signal (Ul).
Then subtraction in the mixer eliminates the unwanted signal. Before subtraction, the
defect signal has a different amplitude and phase at the basic (Dl) and mixing (D2)
frequencies; subtraction does not cause suppression of the defect signal. The resulting
defect (D) signal is displayed at the mixer output. This is treated as response of a system
that is sensitive to defects, but insensitive to the suppressed unwanted discontinuity.
A significant advantage of mixing is that it is purely static; the result of the signal
combination is absolutely independent of the probe movement speed so that the result
is the same at rest, at low or high speed [17].

24
2.2.3 Operating principles of multiparameter analysis systems
Multifrequency eddy current testing uses the amplitude and phase parameters at
different frequencies to separate the discontinuity signals. These parameters vary dif-
ferently under different frequencies and hence when data is collected at more than one
frequency, there is enough information to isolate the effect of extraneous or benign dis-
continuities. Manufacturing and design discontinuities are referred to as benign signals.
Various analysis systems are used for

, suppression of a given parameter. These sys-
terns have signal outputs from two channels corresponding to the test frequencies .f1
and .f2. The signals represent the inphase and quadrature components of the complex
impedance of the probe. For test frequency /1, let the inphase or real component be
xl and imaginary component be yl. Similarly for test frequency .f2, let the real and
imaginary components be x2 and y2. The signal processing methods that are commonly
used are the algebraic method, coordinate transform method and phasor rotation and
subtraction or combinations of these method [27]. Each of these methods are described
next.
2.2.3.1 Algebraic method
The algebraic method assumes that the multifrequency data are linear functions of n
parameters. These parameters represent n discontinuities or changes such as conductiv-
ity, permeability, support plates, wall thickness or defects present in the specimen. This
assumption is valid if there is a linear relationship between variations in the parameters
and the probe impedance. It has been verified experimentally and shown to be valid
under small signal conditions (121 [131. The component values aij and Pj are measured
from some zero, null or nominal values. This is also consistent with the general practice
adopted by commercial eddy current test equipment where the output signals represent
deviations from some input null, reference or electrical bridge balance conditions.
25
In the case of a two frequency system, the signals are expressed as follows:
(2.2)
Yl a21 Pl + a22 P 2 + a23 P3 + a24 P 4

X2 a31H + a32 P 2 + a33 P3 + a34P4
Y2 a4l P I + a42 P 2 + a43 P3 + a44 P 4
where
PI, P2 , P3 and P4 are the parameters

and aij is the coefficient of parameter Pj
The solution of these equations is given by
CllXl + Cl2YI + Cl3 X 2 + Cl4Y2 {2.3}
C21 Xl + C22YI + C23 X 2 + C24Y2

P3 C3l X I + C32YI + C33 X 2 + C34Y2
P4 C41 X l + C42YI + C43 X 2 + C44Y2
Coefficient Cij may be obtained either by calculation after the signals have been digitized
or by the successive approximation method. This is done in the calibration stage where
typical values of parameter Pj are measured using standards appropriate for the kind of
specimen to be evaluated.
Figure 2.10 shows the basic system for a two frequency instrument, which directly
outputs the desired parameter. Additional circuits are needed for more parameters.
A variation of the algebraic method is used by Prince et. al. for the measurement of
Zircaloy cladding thickness on uranium [24]. Slightly enriched uranium is extruded into
a tubular shape and coated with zircaloy. As part of the quality assurance program,
26
the fuel elements must be 100 percent surface inspected for zircaloy clad thickness. It
is observed that the cladding thickness bears an exponential relation with the signal
amplitude and can, therefore be estimated using a polynomial of the form:
t = b1 + b2 ln VSOk + b3 ln VSM + b4 (ln VSOk)2 +

b5(1n VSM)2 + b6 (1n VSOk)(1n V5M ) (2.4)
where
t is the clad thickness
b1 ••• b6 are coefficients
V50k and VSM are signal amplitudes at 50 kHz and 5 MHz respectively.
Two different frequencies are used to remove the effect of lift-off variations. The 5
MHz signal is sensitive to lift-off and 50 kHz signal is sensitive to both lift-off and clad
thickness. The coefficients are calculated using fuel element clad thickness standards
at different lift-off and thickness conditions. The data obtained is fitted to the above
polynomial by minimizing the square error. The authors obtain an accuracy of 0.025
mm for the clad thickness with the lift-off varying from 0.102 to 0.305 mm.
2.2.3.2 Coordinate transformation method
This method also defines a set of parameters. It is assumed that each parameter
is represented by a vector in a space of n dimensions. The curves displayed on the
instrument represent the evolution of the end of these vectors on a projection plane cor-
responding to the frequency used. If the test frequency is changed, the projection plane
rotates around its own axis, thus modifying the phase shift between the discontinuity
signals [17].
The method suppresses the effects of undesired artifacts by applying successive rota-
tion and projection transforms to the probe coil impedance data. The discrimination is
27
• I (PI.Pl,P3, P4 .--_~
C-)
~---,
~~-YI
YI(PI, Pl, Pl, P4) ----.-----
"----I
PI
Y2 (P1'Pl'Pl'P4)--r-_-i~
INVERTERS WEIGIfTING ADDER
Figure 2.10 Analog multiparameter system using algebraic method for pa-
rameter PI [17]
achieved by projecting the undesired data on to a hyperplane and the defect data on to
a plane which is normal to the hyperplane. For a two frequency instrument, there are
two projection planes corresponding to each frequency. As shown in Figure 2.11 using
various components one of the planes is rotated within the space and a plane is recon-
stituted so that only parameter P4 has a vertical component and the other parameters
coincide with the horizontal axis.
The complex impedance plane data can be rotated by applying the two dimensional
coordinate transformation.
x' x cos </> + y sin </> (2.5)
y' -x sin </>+ ycos</>
where
x and yare the initial coordinates of the signal
x' and y' are the coordinates obtained after rotation </>
The goal of successive rotations is to transform the unwanted parameters onto one
plane. Figure 2.12 shows how this is achieved for a two frequency instrument. Each
28
rotator is a ganged sine-cosine potentiometer and is configured as shown. When the
signals are available as digital data, the rotation transformations may be carried out
using a digital signal processor. The three phase shifters, which are identical and can
rotate the input signal from 0 - 360 degrees independent of each other, are connected
in cascade. The phase shifters are adjusted until the desired parameter is normal to
the unwanted ones. The four parameters PI, P2, P3, P4 are shown in Figure 2.11(a).
The phase shifters <PI and <P2 are adjusted successively to obtain the separation of the
required parameter PI (shown in Figure 2.11(b) and 2.11(c)). Parameters P2, P3 and
P4 coincide in 2.11(e). Phase shifter <P3 adjusts the residual signals P2, P3 and P4
horizontally so that only the parameter PI has a vertical component. The resulting
signal is displayed on an oscilloscope. The separation of parameters needs a careful
choice of frequencies [26].
(e)
I
Figure 2.11 Adjustment steps for coordinate transformation [17]
29
,-----------------------------------,
I
XI (PI,P2,P3,P4) - - - . - - - {
YI (PI, P2, PJ, P4)
"2 (PI, P2, PJ, P4) ---._ _ _ _ _ _----'.
Y (PI,P2,PJ,P4)
2
Figure 2.12 General coordinate transformation system
2.2.3.3 Combination method
This is one of the more commonly used methods. One frequency is used for elimina-
tion of each parameter although it might sometimes be possible to eliminate more than
one parameter using the auxiliary frequency. The primary frequency is chosen such that
it is more sensitive to the defect while the auxiliary frequencies are used to remove the
effect of benign artifacts or noise as explained in detail earlier.
Figure 2.13 shows various stages in parameter elimination. It is assumed that only
two parameters exist. PI is the parameter to be eliminated and P2 is the discontinuity
signal. The continuous curve represents the signals obtained using the primary excitation
frequency and the dashed curve represents the signals corresponding to the auxiliary
frequency.
Processing involves (1) modifying the curve PI' by weighting to make it similar to
PI as shown (PI" in Figure 2.I3(b)). (2) executing a phase rotation to superimpose
the signal PI" and PI. (3) subtracting the components of the signals [17] as follows:
(2.6)
This is the same as vectorial subtraction explained earlier. The components of the
disturbing signal PI are canceled out, while the discontinuity signal components retain
30
a significant value.
The system can be extended for mixing and eliminating more than one parameter.
For example, to simultaneously eliminate the effect of support plate as well as dent
another auxiliary frequency is added. The auxiliary frequency 11 is sensitive to the
support plate while 12 is sensitive to dents. The combination of 11 and 12 with the basic
excitation frequency signal removes the effect of the support plate. The combination of
the signals thus obtained further removes the effect of the dent, so that the final mix
shows only the defect signal.
An alternate approach requires only two mixers. Mixing 11 and the basic frequency
signal gives an output with partial subtraction of support plate. This signal is mixed
with h to completely suppress the support plate and dent. Both methods need two
auxiliary frequencies. These have been tested extensively in the field. The results
obtained from both methods are same although the two mixer method is more difficult
to optimize. However it has the advantage that it is less operator dependent than the
three mixer method. The three mixer method has the advantage that three mixing
outputs are available for analysis. These mixers can be optimized to suppress various
combinations of extraneous discontinuities such as: support plate and dent, expansion
transition and tube sheet plate, internal pilgering noise and support plate, antivibration
bar and V-bend transition, probe wobble and copper deposits, and so on.
The next chapter presents some novel algorithms that have been developed for mixing
the basic and auxiliary frequency signals efficiently. These algorithms implement steps
(1) and (2) of the combination method.

31
(a) P2
PI'
PI
(b)
P2
PI"
PI
P2"
P2
(c) PI
PI'
(d)
Figure 2.13 Adjustment stages in the combination method: (a) signals be-
fore analysis; (b) signals after weighting; (c )signals after phase
rotation; and (d) signals after subtraction
32
3 ALGORITHMS FOR MULTIPARAMETER SIGNAL
ANALYSIS
3.1 Introduction
The objective of a mixing algorithm is to subtract the undesirable signal component
generated by an artifact from a composite signal to produce the defect signal. This is
accomplished by conducting the eddy current test at two excitation frequencies. The
first excitation frequency is chosen so as to ensure the highest level of sensitivity to the
artifact. The auxiliary frequency data is then rotated, translated and scaled such that it
resembles the composite signal. The transformed auxiliary signal is subtracted from the
composite signal to obtain the defect signal. The purpose of the procedures presented
in this chapter is to shape and rotate the auxiliary frequency signals appropriately. The
success of the multiparameter method is dependent on the efficiency with which the
transformation procedure is implemented. Extensive survey of the available literature
showed a lack of such methods. Most systems use a calibration stage in which the mixing
parameters (the weights for shaping the signal and the rotation angle) are calculated
off-line. These parameters are then used to transform the auxiliary frequency signal for
subtraction from the basic frequency signal as the probe is scans the tube. This is an
inefficient approach, as the mixing parameters are optimized for suppression of a par-
ticular kind of discontinuity. In addition it is likely that the nature of the discontinuity
being suppressed may differ from the artifact standard used during the calibration stage.
Hence the mixing may not be optimum. An alternative approach is to mix signals using
33
the procedure described earlier. This has the advantage that mixing is optimized for
each discontinuity and is not dependent on a reference standard. For such a method
to be effective, the mixing algorithms have to be computationally efficient so that the
results of mixing can be seen in real time. In this thesis the signals are shaped and
rotated using affine transformations [11].
3.2 Affine transformation equations
An affine transformation is employed to rotate, scale and translate an eddy current
signal such that it resembles a second eddy current signal [29].
Translation is represented by the following equations.
x' x+Tx (3.1 )
y' y+Ty
or
1 0 0
[x' y' 1] [x y 1] 0 1 0 (3.2)
Tx Ty 1
where
x, yare the original coordinates of a point in the plane
x', y' are the transformed coordinates

Tx is the X-axis translation
Ty is the Y-axis translation
The operation of rotation in the counter-clockwise direction about the origin is ob-
tained by:
x' x cos () - y sin ()

34
y' x sin () + y cos ()
or
cos () sin () 0
[x' y' 1] [x y 1] - sin () cos () 0 (3.3)
0 0 1
where
x', y' are the transformed coordinates.

() is the angle of rotation
Scaling is achieved by:
x' (3.4)
or
1 0 0
[x' y' 1] [xy1] 0 1 0 (3.5)
Sx Sy 1
where
x' ,y' are the transformed coordinates
Sx is the X-axis scaling coefficient

Sy is the Y-axis scaling coefficient
Thus the transformations can be combined and written using a compact notation as:
35
[x', y/] [x, y, l]A (3.6)
where
Sx cos 0 Sy sinO
A -Sx sin 0 Sy cos 0 (3.7)
Sx(Tx cos 0 - Ty sin 0) Sy(Tx sin 0 + Ty cos 0

A is referred to as the affine transform matrix.
The objective of the mixing algorithm is to estimate the parameters Tx, Ty, 0, S:r and
Sy such that the transformed auxiliary signal data is as close to the primary signal data
as possible.
3.3 Conjugate gradient method
The optimum affine transform parameters can be estimated using a least squares
estimation procedure in the time domain [29]. Let Sb be the basic frequency signal and
Sa be the auxiliary frequency signal. Then S~ the transformed signal is given by:
S~ = SaA(Sx' Sy, T x , Ty, 0)
The error function E is defined as the square of the error between the transformed
version of the auxiliary frequency data and the primary frequency data.
(3.8)
The optimum values of the affine transform parameters are estimated by minimiz-
ing the error function E. This involves solving the following five nonlinear equations
simultaneously.
36
8E
8Tx -
0 (3.9)
8E
0 (3.10)
8Ty
8E
0 (3.11)
80
8E
0 (3.12)
8Sx
8E
0 (3.13)
8Sy
Solving this system of equations yields the affine transform parameters, that is op-
timum in the minimum error squares sense. Each of the sequences Sa and Sb are rep-
resented by n samples in the impedance plane. The fully expanded form of these five
equations is as follows:
n
2 'L {[Xdi
i=l
* [-Sy sin On = 0 (3.14)

n
2 'L{[Xdi
i=l
* [-Sy cos On = 0 (3.15)

n
2'L{[Xdi
i=l
37
(3.16)
(Sx cos O)Xi + (Sx sin O)Yi - SxTx cos 0
+ Ty sin 0] = 0 (3.17)
n
2 "L[Ydi
i=l
SyTy cos 0][(- sin O)Xi - (cos O)Yi - Tx sin 0
Ty cos 0] = 0 (3.18)
where
Xdi, Ydi is the primary frequency data, the elements of sequence Sb
Xi, Yi is the auxiliary frequency data, the elements of sequence Sa
Since the gradient function is available, one of several descent methods can be uti-
lized for minimizing the error E. Many of these are iterative methods and require an
initial choice of the parameters. In each successive step, new values of the parameters are
generated. The cost function (in this case, the error E) is evaluated for the new param-
eters and tested for convergence. The iterative procedure is repeated if the convergence
criterion is not satisfied, otherwise the current values of the parameters are taken to be
optimum. Descent methods that can be employed include the steepest gradient, con-
jugate gradient or Fletcher-Reeve's method, Newton's method and the variable metric
method [25]. The steepest descent method and conjugate gradient method were eval-
uated since they are intuitively satisfying and allow heuristics to be included to speed
up convergence. The Newton and variable metric method require extra storage, and
consequently were not evaluated.
The flow chart for Fletcher-Reeve's method is shown in Figure 3.1. The vector Xi
38
shown in the Figure is defined as:
Txi
TYi
Xi ()i (3.19)
Sxi
Syi
where
i stands for ith iteration
V E is the gradient of the error function defined as:
8E
8T%
8E
8Ty
VE 8E
8if
(3.20)
8E
88%
8E
88y
The algorithm starts with an initial estimate of the vector Xi denoted Xl' The
gradient of the cost function E is evaluated at Xl. Vector Sj is the search direction used
to update Xi
(3.21)
where
Ai is the step length.

In general at the ith iteration, the vector Si is defined as:
(3.22)
39
Start with Xl
Find 'i7E 1 = 'i7E(X 1)

Set St = -'VEl
Find A.*1 to minimize E(}f + A. 1 S 1 )
Set i = i+l
Take "opt = X i+l
Figure 3.1 Flow chart for the Fletcher-Reeves method

40
The steepest descent method is a special case of the more general Fletcher-Reeve's
method where Sj is always defined as
It has been suggested for the conjugate gradient method that after every m iterations
Si should be chosen as
to reduce the cumulative effect of round off errors. In the actual implementation, m
was chosen by experimentation to be 5 or 6. This is close to the recommended value
[rao] of m = k + 1 where k is the number of design variables. For the affine transform
minimization case, the number of variables k = 5.
The choice of A to minimize E(Xi + AiSi) requires another minimization process
within each iterative step. A heuristic method, explained later, is used for choosing the
optimal value of Ai. Another step in the flow chart that needs explanation is the choice
of convergence criterion. As the gain of the impedance plane data and also the number
of samples or data points in the impedance plane is not fixed, an absolute value for the
error E cannot be used. Instead the convergence criterion is based on the derivative, i.e.
vj = 0,1,2,3,4 (3.23)
where
Xj is the ph component of vector Xi.
This is true, since all the partial derivatives are close to zero when the parameter
estimates are close to the optimal value. The value of f is chosen to be 10- 3 . A lower
value only tends to increase the number of iterations without affecting the quality of the
parameter estimates significantly.

41
It was observed that by using the same initial estimates for the parameters, the con-
jugate gradient method requires fewer iterations to converge than the steepest gradient
method. However the quality of the estimates were superior to the steepest descent
method.
3.3.1 Heuristics to improve conjugate gradient minimization
The conjugate gradient method can be improved by choosing good initial estimates
for the transformation parameters, namely Tx, Ty, (), SXl Sy. (}1 is obtained by setting it
equal to the phase difference between the two impedance plane trajectories Sa and Sb.
Sxl is chosen as the ratio of the inphase phase component of the amplitude of the two
signals. Similarly Syl is the ratio of the quadrature components of the amplitude of the
signals. Tx and Ty are initially set to a value representing the difference between the
averages of the two signals. This significantly reduces the number of iterations. For
the example presented in Figure 3.3, the number of iterations is reduced by a factor
of ten. The approach has been tested using real data and gives good initial estimates
of the parameters. As an example, Figure 3.2 shows a through-hole defect signal in
the presence of a support plate at auxiliary excitation frequency, that is translated to
match the primary excitation frequency data using the initial estimates of the translation
parameters.
A reduction in the computational complexity at each iterative step can be achieved by
considering the equations used for calculating the partial derivatives. If the summation
is taken inside in equations (3.14) through (3.18), the sum need not be performed for
i = 0 to i = n at each iterative step. Thus equation (3.14) becomes:
n n
-2[2:Xdi - (Sxcos()) LXi + (Sx sin (})
;=1 ;=1
42
n
* L: Yi - n * SxTx cos () + n * SxTy sin 0)
i=l
n
* [-SxcosOJ + [LYdi - (SysinO)
i=l
n n
* L:Xi + (SycosO) L:Yi - n * SyTxsinO
i=l i=l
n * SyTy cos OJ * [-Sy sin OJ (3.24)
If we define the running and product sums as,
n
X di = L:i=l Xdi Ydi = L:Ydi (3.25)
i=l
n
Xi = L:i=l Xi Y£= LYi
i=l
n
Sum xx = L:i=l Xi Xi Sum XdXd = L: XdiXdi
i=l
n
Sum yy = L:i=l YiYi Sum YdYd = L: YdiYdi
i=l
n
Sum xy = I:i=l XiYi Sum XdYd = L: XdiYdi
i=l
n
Sum XdX = I:i=l XdiXi Sum ydy = LYdiYi
i=l
n
Sum XYd = L:i=l XiYdi Sum xdy = LXdiYi
i=l
the computational effort in calculating V' E at each iterative step is minimized. Another
advantage is that the performance of the algorithm is independent of the sampling rate
of the impedance plane data since the number of samples in the impedance plane data
does not affect convergence speed. This is evident from the fact that the data is used
only once to compute various terms defined in equations (3.25).
The method implemented for estimating the optimum value of the step size A has a
significant effect on the performance of the algorithm due to the fact that it is repeated
in each iteration. Various alternatives are available such as choosing a small constant
43
1
I
]
j
-1
-1
-2
-3
-3,L--_3~---"_2--_~'_~_-,-----,~~----1
--4 -3 -2 _1
(a) Primary frequency signal (b) Auxiliary frequency sig-

at 400 kHz nal at 200 kHz
1.5
0.5
.... 5
_1
-1.5
-2
_2.5,L-~ _____ -,-----"_~_~---,~_
-.( -3 -2 -1
(c) signal in (b) translated

to (a)
Figure 3.2 Signal transformation for a through hole defect in the presence
of support plate using initial estimates
44
step size, or using an appropriate optimization technique for minimizing E(Xi + AiSi)
as a function of Ai. Different methods were used to choose an optimal value of A. The
drawback associated with choosing constant A is that it is difficult to select a value which
works well under all conditions. An inappropriate value of A may result in divergence of
the iterative procedure. In contrast, a small value of A results in poor performance. The
Newton-Raphson method was used to estimate optimal values of A. Though Newton-
Raphson method reduces the number of iterations, the overall effect is to slow down the
algorithm since the method is a computationally intensive procedure.
A heuristic method is used in the final implementation to choose the values of A at
each iterative step. Initially A is chosen to be a sufficiently small number which ensures
that the cost function does not increase. As the iterative procedure progresses, A is
chosen as follows:
IF E,~est < E·1-1 THEN
A = AO:, 0: > 1.0

ELSE
A = A{3, 0.0 < {3 < 1.0
Go back to the beginning of IF statement
The parameters controlling the adaptation of A are experimentally found to be:
Astart = 10- 5 , 0: = 1.2 and (3 = 0.9
The validity of the approach was evaluated using finite element model (FEM) impedance
plane data. The FEM model [21] was used to simulate an inner diameter tube defect at
frequencies of 25 kHz and 50 kHz. The impedance plane trajectories obtained using the
FEM are presented in Figure 3.3(a) and 3.3(b). The estimated values of the transfor-
mation parameters are: Tx = 0.0805, Ty = -0.0486 , e = -155.7377 , Sx = -2.6682 and

45
Sy = 1.8923. The auxiliary frequency data was then mapped using these transformation
parameters. The results are shown in Figure 3.3( c). As can be seen, the auxiliary
frequency data is closely mapped to the primary frequency data. Additional results are
presented in the next chapter.
3.4 Fourier series method
The algorithm presented in the last section is based on a time domain approach.
Alternatively, the transformation parameters can be estimated using frequency domain
methods. The strategy involves estimating the transformation parameters based on
information contained in frequency components of the primary and auxiliary impedance
plane data. We use the Fourier series expansion of the signals to obtain the frequency
domain representation.
Consider a simple closed impedance plane curve III the two dimensional coordi-
nate system as shown in Figure 3.3(a). We obtain a parametric representation of
the impedance plane curve by representing each point as a function of the arc length
I along the curve, measured from an arbitrary starting point Po. Thus (x(l), y(l)) is a
point on the curve located I arc length units away from the starting point. Next we
define a function u( I) as:
u( l) = x( l) + jy(l), j=v=r (3.26)
If L is the total length of the curve then,
u(l + nL) = u(l), n = ... -1,0,1, ...
The periodicity of the function u( l) allows its expansion in a Fourier series. Thus
u(l) = n~oo en exp -L-

00 [j27rnlj (3.27)
46
'5
'0
-'0
-15 _'5
~~~--~--~--~--~--~~~~
-20 -15 -10 -5 20 ~~0---~'5---_~'0---_~5--~--~--~'0---'~5--~~

at 50kHz nal at 25kHz

to (a)
Figure 3.3 Signal transformation for FEM generated data using the conju-
gate gradient method
47
where
en = ~
L Jo
r u (I) exp (-j27rnl)dl
L
L (3.28)
If the curve is translated by Tx and Ty, rotated by () and scaled by S the Fourier
coefficient c~ of the translated curve is given by [31]:
1
L
rL 5exp(j()) (u(l) + Tx + jTy) exp( -j27rnl
Jo L )dl
5exf(j()) Io L
u(l) exp( -j~7rnl )dl
5 exp(j()) rL(T °T ) ( - j27rnl )dl

+ L Jo x + J y exp L
5exf(j()) Io L
u(l)exp(-j~7rnl)dl n =I 0 (3.29)
Since
r exp( -j27rnl
Jo
L
L )dl = 0 , (3.30)
for n =0
or
c~ = 5exp(j())[eo + Tx + jTy] (3.31 )
From the above equations it is clear that the ratio SL

en
can be used to estimate the
scaling (5) as well as the rotation (()) parameters. Once 5 and () are known, equating the
real and imaginary parts of equation (3.31) yield estimates of the translation parameters
Tx and Ty respectively.
In practice the impedance plane data is sampled. Persoon and Fu [22] provide a
procedure for estimating the Fourier coefficients by approximating the curve by a polygon
48
of m sides with vertices at Vo, Vb'" ,Vm-l. The vertices correspond to the data points
of the impedance plane trajectory. Vertex Vm = Vo to satisfy the assumption of a closed

curve. The expression for estimating Cn is as follows:
L ~ -j27rntk
Cn - -2-2 Ljbk-1-bk)exp( L ), n -1= 0 (3.32)
47r n k=l
1 m
-I:Vk (3.33)
m k=l
where
k
lk = I: IVi - vi-d, k>O and to = 0 (3.34)
i=l
and
(3.35)
The estimates of Cn can be used to calculate the scaling and rotation parameters.
Although technically, any coefficient pair can be used to calculate these parameters, it
is preferable to use coefficients that are large valued. This usually implies the use of
lower order coefficients. Coefficients ranging from C-lO to CIO were calculated and used
to estimate Sand (J. All of them gave similar values for the parameters. The translation
parameters were estimated using eo. The computational burden of associated with the
use of Fourier series based method is significantly lower compared to the time domain
conjugate gradient method. The procedure is O( N) where N is the number of samples.
Figure 3.4( a) shows support plate signal obtained at the primary frequency, 3.4(b)
shows the auxiliary frequency signal and 3.4( c) shows the original and transformed
signals.
49
1.5
0.5
-0.5
-1
-1.5
-2
-2.5
-3
-4 -3 -2 -1

1.5
0.5
-0.5
-1
-1.5
-2
-2.5
-3
-4 -3 -2 -1

to (a)
Figure 3.4 Support plate signal translation using Fourier senes method:
dotted line shows the transformed signal
50
It is obvious from the results shown in Figure 3.4 that the procedure is not satisfac-
tory. This is due to the fact that the algorithm assumes
The reduction in the degrees of freedom clearly affects the performance. The algorithm
is not able to match the FEM impedance plane data pairs that was used for evaluating
algorithm using the conjugate gradient method. This is understandable since the optimal
values of Sx = -2.6681 and Sy = 1.8923 are different. It must be mentioned though
that the frequency domain methods are not iterative and involve a fixed amount of
computation, which is a significant advantage relative to the time domain methods.
This limitation led to the search for other frequency domain methods which would allow
independent scaling of the inphase and quadrature components of the impedance plane
data. One such method that was investigated uses the cosine transform for obtaining
the frequency domain representation.
3.5 Discrete cosine transform method
The discrete cosine transform (nCT) of a data sequence,
X(m), m = 0,1,···, (n - 1) is defined as
V2
n
I:
m=O
X(m) (3.36)
2 ~ X() (2m + 1)k7r ,

Gx ( k ) = - ~ m cos k = 1,2, ... ,(n -1) (3.37)
n m=O 2n
where Gx(k) is the kth nCT coefficient.
As can be seen, the DCT is a real transform [9]. This means that the nCT can be
evaluated for expanding the inphase and the quadrature components of the impedance
plane data independent of each other.

51
Let Sa and Sb be the sequences as defined earlier. The sequence Sa is given by
where n is the length of the sequence
The points represent complex impedance of the probe coil in the complex impedance
plane. Hence the sequence may also be written as
(Xo + jyo), (Xl + jyt},.", (Xn-l + jYn-l)
Also the sequence Re(Sa) is the real part of sequence Sa given by
Xo, Xl, X2 ... ,Xn-l
Similarly the sequence Im(Sa) is the imaginary part of the sequence and is given by
Yo, yl, Y2 ... ,Yn-l
The sequence S~ is the transformed signal so that it maps to the basic frequency
signal Sb.
The sequence S~ is given by
(X~, y~), (x~, yD, (x;, y~), ... , (x~_l' Y~-l)
The sequence S~ is obtained using the transformation
where the matrix A is given by:
Sx cos () Sy sin ()
A -Sx sin () Sy cos ()
Sx(Tx cos () - Ty sin ()) Sy{Tx sin () + Ty cos ()
Alternatively the relation between the sequences Sa and S~ can be written as:
52
[x', y'] = [x, y, l]A
The above affine transform relation between data sequences Sa and S~ in the ex-
panded form is given by:
x' - xSx cos () - ySx sin () + S:r;{Tx cos () - Ty sin ())
- xSx cos () - ySx sin () + a31 (3.38)
y' - xSy sin () + ySy cos 0 + Sy(Tx sin 0 + Ty cos 0)
- xSy sin () + ySy cos () + a32 (3.39)
where
a31 and a32 represent appropriate elements of the matrix A.
The DCT of Re(Sa) denoted by Gx is

v'2 n-l
- -Lx
n m=O
m (3.40)
2 ~ (2m + l)k7r
- L-J Xm cos 2 ' k = 1,2, ... ,(n -1) (3.41)
n m=O n
The DCT of Re( S~) denoted by G~ is

J2 n-l
G~(O) - -Lx~
n m=O
(3.42)
2 ~, (2m+1)k7r
G~(k) = - L-J xm cos 2 ' k = 1,2, ... , (n - 1) (3.43)
n m=O n
Substituting for x~ in the above relation using equation (3.37) gives,
G~(O)
53
-J2 n-l .J2 n-l -J2 n-l

Sx cos {}- L Xm - Sx sin (}---=: L Ym +- L a31
n m=O n m=O n m=O
Sx cos (}Gx(O) - Sx sin (}Gy(O) + a31 V2 (3.44)
Similarly,
2 ~ I (2m+l)k7r
G~(k) - L.J xm cos 2
n m=O n
2 n-l . (2m + l)krr
-L (xmSx cos {} - YmSx sm {} + a31) cos -'-------'-
n m=O 2n
2 ~ (2m + l)krr
S x cos {} - L.J Xm cos -'-------'--
n m=O 2n
. {}2 ~ (2m + l)krr
S x sm - L.J Ym cos -'-------'--
n m=O 2n
2 ~ (2m + l)krr
+ a13- L.J cos 2
n m=O n
Sx cos (}Gx(k) - Sxsin{}Gy(k) (3.45)
The constant term in G~(k) vanishes since
2 n-l (2m + 1 )krr

-n m=O
L cos 2
n
= 0, k = 1"2 ... , (n -1) (3.46)
Similar computations for G~(O) and G~(k) give the following relations
G~(O) Sy sin (}Gx(O) + Sy cos (}Gy(O) + a32V2 (3.47)
G~(k) - Sysin{}Gx(k) + Sycos{}Gy(k) (3.48)
For k = 1,2" .. , (n - 1), the above relations can also be written as
G~(k) SxRe[G(k)eiO ] (3.49)
G~(k) Sylm[G(k)eiO ] (3.50)
where
54
The above equations clearly show the relation between the DCT coefficients of se-
quences Sa and S~. Given the DCT coefficients of sequences Sa and Sb, it is possible
using the above equations to calculate the values of the translation parameters. The
advantage over the Fourier series method described earlier is that the equations allow
determination of the parameters Sx and Sy. Another advantage of the DCT is that fast
FFT based algorithms can be used to calculate the DCT coefficients. This has major
implications with regard to the computational burden associated with these algorithms.
The affine transform parameters Tx, Ty, (), Sx, Sy can be calculated as follows.
Let
(3.51 )
Hence
G~(kl) Re[G(kl )ei8 ]

Rklk2 = G~(k2) = Re[G(k2)ei8] (3.52)
Expanding the above equation and rearranging the terms gives:
(3.53)
This gives two values of () between 0 - 360 which are 180 degrees apart (say Ol and
After 0 is estimated, Sx can be calculated using
G~(kl)
(3.54)
Sx = Re[G(kt}ei8 ]
Similarly,
G~(kd
(3.55)
Sy = Im[G(kt}ei8 ]
(3.56)
55
As can be verified from the equations (3.54) and (3.55) if the solution using ()l is
SxI and Syl the other solution obtained using () = ()l + 7l" is, - SxI and - Syl. The two
solutions result in same values for the parameters Tx and Ty. Both sets of solutions map
the auxiliary frequency signal to the primary frequency signal in an identical manner.
This can also be verified by replacing () in the matrix A. with ()+7l" and Sx, Sy by -Sx, -Sy
respectively.
Tx and Ty are calculated from equations (3.43) and (3.46) relating the zeroth neT
coefficients of the two signals. All other terms in the equations are known, consequently
the two equations can be solved simultaneously for Tx and Ty.
The algorithm for finding the optimal affine transform parameters is implemented
using the above equations. The neT coefficients are calculated for the primary and
the auxiliary frequency signals. These coefficients can be evaluated efficiently using fast
algorithms [1] [20]. Another noteworthy aspect of this procedure is that only the first
few neT coefficients need to be calculated. This reduces the computational burden even
further. Once the neT coefficients are calculated, the optimal transform parameters
can be estimated using the above equations.
It can be seen that the parameters can be estimated using any arbitrary value of kl
and k2 • However it has been observed, that not all coefficients provide identical estimates
of the parameters. Hence the optimal parameters are those that result in a minimum
value for the cost function E. Figure 3.5 shows the results obtained using the neT
based algorithm for calculating the optimal affine transform parameters for the FEM
modeled data. The results are the same as those obtained using the conjugate gradient
method. Additional results are presented in the next chapter.
The closed form relations between the neT coefficients of the primary and auxiliary
frequency signals suggests that invariant feature vectors can be found which would be
insensitive to translation, rotation and scaling of the original signal. These invariant
features are dependent only on the shape of the impedance plane trajectory. This result
56
15 15
10 10
-5 -5
-10 -10
-15 -15
_ro~~--~--~--~--~--~~---J
-20 -15 -10 -5 10 1S 20 ~~~-'5~--~'0----5~~~~--~'0---'~5--~ro


to (a)
Figure 3.5 Signal transformation for FEM generated data using DCT coef-
ficients
57
is of great significance especially in the area of NDT since the invariant descriptors
can be embedded in feature vectors for the classification of defect signals. This is an
important outcome of the above analysis and is presented next.
3.5.1 Invariant parameters using neT

The neT terms vary if the signal is rotated, scaled and translated. In order to
obtain descriptors that are invariant to these transformations, we begin by defining the
function,
(3.57)
It can be shown that A(kb k2' k3, k4 ) is invariant to these transformations. This can
be proved as follows:
Hence
B(kb k2)
(3.58)
B(k3, k4 )
Now
G~(kdG~(k2) - G~(kdG~(k2)
SXSy[(Gx(kd cos () - Gy(kd sin (})(G x(k 2) sin () + Gy(k2) cos (})
(G x(k2) cos () - Gy(k2) sin 0)( Gx(kd sin () + Gy(kd cos (})]
SXSy[Gx(k l ) cos (}Gx(k 2) sin () + Gx(k l ) cos (}G y(k2) cos ()
Gy(kd sin (}G x(k 2) sin () - Gy(kd sin (}G y(k2) cos ()
Gx(k 2) cos (}Gx(kd sin () - Gx(k2) cos (}Gy(kd cos ()
+ Gy( k2) sin (}G x( kl ) sin () + Gy(k2) sin (}G y( k l ) cos 0]
which on rearranging gives

58
(3.59)
Consequently,
B'(k l , k2)
B'(k3, k 4)
SxSyB(k}, k 2)
SxSyB(k3, k4)
- A(k}, k2, k3, k 4) (3.60)
Hence the function defined by A(k}, k2' k3, k4) is invariant under rotation, indepen-
dent scaling of the X and Y components, and translation along the X and Y axis.
The above function defines a family of invariant features. By experimenting with vari-
ous combinations of the coefficient kll k2' k3 and k4' appropriate feature vectors can be
constructed and used for training a neural network for defect classification.
Figure 3.6 plots the function A(kl' k2' k3, k 4) for the FEM derived primary frequency
data and the same data after being translated, scaled and rotated by arbitrarily chosen
amount. The dotted lines show the invariant features for the primary frequency signal
and the points marked by 'x' shows for the translated signal. The index along the X
axis stands for k l . The variables k2' k3 and k4 are set equal to kl + 1, kl + 2 and kl +3
respectively. As can be seen the feature vectors match perfectly for the FEM modeled
primary frequency impedance plane trajectory and its transformed version.

59
15
¥
10 I'
.\
1 .
\
x.
5
" ~\ )\
/
/
"X ,, 1 \ / ./\
_.)j:
1
0 >I X-'
\ 1
\
\ X
/
/
/
\
1
\ .- \ \ \
1 X' X. . 1
\ \ . \
1
-5 \ 1 *
X
\ /
-10 ~/
\
-15 \
\
I
-20 \
\
-25
-30
0 2 4 6 8 10 12 14 16
Figure 3.6 Invariant features for FEM modeled signal

60
4 RESULTS AND CONCLUSIONS
The multi frequency eddy current analysis techniques developed in the previous chap-
ters are evaluated using experimental data. The results obtained using the new methods
were compared with those obtained using a commercial eddy current multifrequency in-
strument (MIZ-40, Zetec Inc.) The MIZ-40 was used for collecting eddy current defect
signatures at multiple excitation frequencies in the presence of a support plate.
The experimental setup consisted of a calibration tube with machined defects such
as through-wall hole defect, deep fiat bottom holes, 60%, 40% and 20% wall thickness
deep and 2 mm wide OD defects. A ferromagnetic support ring was used to simulate a
support plate. A differential eddy current probe excited simultaneously at two different
frequencies is used to the collect eddy current signals. Data obtained at two frequencies
are normally sufficient to nullify the effect of a support plate. The primary and auxiliary
excitation frequencies are selected experimentally by following the procedure outlined
in chapter 2. It was observed that using a primary frequency of 400 kHz and an aux-
iliary frequency of 200 kHz achieves the best phase and amplitude separation between
the defect and the unwanted support plate signal, for the given setup. The complex
impedance of the differential probe is displayed on the instrument for each excitation
frequency, after separating the test frequencies and digitizing the signal. A sampling
rate of 400 Hz is used for digitization which is considered sufficient. A multiplexed
version of the eddy current signals, is available through the analog output ports of the
instrument. These signals are sampled at 1 kHz but since the signal is multiplexed,
the effective sampling rate for each frequency is 100 Hz. This is the signal that is used
61
for multiparameter analysis using the approaches that have been developed. Since the
maximum sampling rate available is 100 Hz, the sampling rate on MIZ-40 is also set to
the same value for comparison.
Figure 4.1 presents the step by step method used by MIZ-40 for support plate
suppression. First, the support ring is placed on a defect free region of the tube. The
tube is scanned and pure support plate signatures are obtained. These are displayed in
Figure 4.1 (a) and 4.1 (c). The mixing parameters are calculated from these signals and
stored. The tube is scanned again, this time to locate and characterize flaws in the tube.
As the tube is scanned, the auxiliary excitation frequency data are transformed using
the mixing parameters, subtracted from the basic frequency data, and then displayed
on the instrument screen. Figure 4.1 shows the results obtained using this approach
on data collected from the instrument. The affine transform based conjugate gradient
method was used to calculate the transformation parameters Tx, Ty, B, Sx and Sy. Figure
4.1 (e) shows the auxiliary support signal (dotted lines) translated to map the primary
frequency support signal which is also displayed. The result of subtraction is shown in
Figure 4.1(g). As can be observed, the residual signal after subtraction is small. Figure
4.1(b) and Figure 4.1(d) show the composite signal when the support ring is placed
above the through-hole defect. Figure 4.1(f) shows the auxiliary frequency composite
signal mapped onto the primary frequency composite signal using the transformation
parameters calculated above. Figure 4.1(h) shows the final result of subtraction of the
translated composite signal from the primary composite signal. The support plate signal
is completely suppressed (except for the residue in the center which does not affect the
phase analysis). Figure 4.2 shows the same steps for the MIZ-40 instrument. The
algorithm used by MIZ-40 for calculating the transformation parameters is not known.
A comparison of Figure 4.1(h) and 4.2(f) shows that similar results are obtained for
MIZ-40 and the time-domain conjugate gradient method. Figure 4.3 presents the results
obtained from the Zetec instrument and the conjugate gradient method for comparison,
62
for a 60 % wall thickness and 2 mm wide OD defect with the support ring aligned with
the defect. The poor quality of the defect signal is a direct consequence of the low
sampling rate. This is clear from 4.3(g) which shows the defect signal after support
plate suppression (done on MIZ-40) at 400 Hz sampling rate.
The method used by MIZ 40 for support plate suppression is not very efficient since in
reality the nature of the support plate signal may be different from the artifact standard
that is used during the calibration stage. It is also likely to be different for support plates
located along the tubes due to the varying conditions that exist in nuclear power plants.
The alternative approach that is proposed in this thesis uses the composite signals
for calculating the transformation parameters. The parameters are used to translate
the auxiliary frequency signal to the primary signal which is then subtracted from the
primary frequency signal to nullify the effect of the support plate. Similar results are
obtained using the same approach for a through-hole defect as shown in Figure 4.4. The
conjugate gradient based approach is used for calculating the translation parameters.
The results obtained using cosine transform based method are similar to the results
obtained using the time-domain conjugate gradient method. This is demonstrated in
Figure 4.5 for the through-wall hole defect. The results are compared with those ob-
tained using the MIZ-40 instrument for a 40 % OD defect in Figure 4.6. The cosine
transform based method was used to estimate the translation parameters from the com-
posite signal in this case. In effect, the results obtained using the conjugate gradient
and the cosine transform based techniques are similar to those obtained using the MIZ-
40 instrument both when the support plate signal in the defect free region is used for
mixing or when the composite signals are directly mixed.

63
4.1 Conclusions and recommendations for future work
A novel approach for eddy current signal analysis has been described in this thesis.
The validity of the approach has been demonstrated using experimental data. The theo-
retical basis for the cosine transform based approach is presented in chapter 3. The time
domain approach provides an optimal result as it minimizes a cost function in the least
squares error sense. The cosine transform based approach achieves similar results with
the additional benefit that it involves a fixed computational effort. The computational
effort associated with the conjugate gradient method is variable. A comparison of the
results shows that the proposed approaches offer performance levels that are similar to
those using MIZ-40. The DCT based method is superior to the Fourier series method.
However the proposed techniques are computationally fast and can be implemented in
real time. The proposed methods also do not require the mixing to be optimized for
each extraneous discontinuity that is encountered as the tube is inspected. Additionally
the new methods do not require the user to undergo a calibration procedure.
These techniques will be incorporated in the multifrequency eddy current instrument
that is being developed by the MCRG group at Iowa State University. The function
A(kt, k2' k3, k4 ) described in chapter 3 defines a family of invariant features that are
independent of scaling, rotation and translation of the signal. A more detailed study
of these invariant features should be conducted so that appropriate features can be se-
lected and used for defect signal classification in automated NDE systems. An alternate
Fourier series based approach can be investigated by encapsulating the real and imag-
inary components of the impedance in a two dimensional vector of real and imaginary
components. This might allow independent scaling of the real and imaginary components
of the impedance plane data, so that the signals can be mapped more accurately. The
method does not work for ali pairs of DCT coefficients. This needs further investigation.
64
(a) Sup- (b) Support

port plate plate and de-
(400kHz) fect (400kHz)
(c) Support (d) Support

(200kHz) and defect
(200kHz)
.
.
(e) trans- (f) transformed

formed sup- defect signal
port signal
(g) mixed (h )mixed de-

support signal fect signal
Figure 4.1 Mixing results using conjugate gradient technique for a through
hole defect in the presence of a support plate
65
,
., ---, - T
, , ,
..; - ,- ,
-,- ... - .. -,-
,,
,
"T
,
-.- -,- ..., - ,- ,
-,-
(a) Sup- (b) Support

port plate plate and de-
(400kHz) fect (400kHz)
(c) Support (d) Support

(200kHz) and defect
(200kHz)
(e) mixed (f) mixed

support signal defect signal
Figure 4.2 Mixing results for MIZ-40 instrument for a through hole defect
in the presence of a support plate
66
., - .- -,-
,
T ', -
, ,
T
I
-r- -,-
(a) Primary (b) Primary
frequency frequency
signal signal
(c) Auxiliary (d) Auxiliary

frequency frequency
signal signal
(e) Zetec Mix (f) Mix us-

mg affine
transform
(g) Zetec Mix,

400 Hz sam-
pling rate
Figure 4.3 60% OD defect in the presence of support plate: a,c,e are ob-
tained using the MIZ-40 instrument, b,d,f are obtained using the
conjugate gradient technique
67
.
,. - .. -.-·
.... - .- ·-
·
-.-
frequency frequency
signal signal
(c) Auxiliary ( d) Auxiliary

frequency frequency
signal signal
(e)
Zetec Mix (f) Mix us-
mg affine
transform
Figure 4.4 Through hole defect in the presence of support plate: a ,c,e are
obtained using the MIZ-40 instrument, b,d,f are obtained using
the conjugate gradient method on the composite signal
68
·
...
·
...
...·
-
(a) Primary
"
- ,,-
-
,
-,-
-
t
(b) Primary
frequency frequency
signal signal
( c) Auxiliary (d) Auxiliary

frequency frequency
signal signal
J
(e)
Zetec Mix (f) Mix us-
. .
mg cosme
transform
approach
Figure 4.5 Through hole defect in the presence of support plate: a ,c,e are
obtained using the MIZ-40 instrument , b,d,f are obtained using
the cosine transform based approach on the composite signal
69

frequency frequency
signal signal
( c) Auxiliary (d) Auxiliary

frequency frequency
signal signal
(e) Zetec Mix (f) Mix us-

mg affine
transform
Figure 4.6 40% OD defect in the presence of support plate: a,c,e are ob-
tained using the MIZ-40 instrument, b,d,f are obtained using the
cosine transform based approach on the composite signal
70
BIBLIOGRAPHY
[1] N. Ahmed, T. Natarajan, and K.R. Rao, "Discrete cosine transform," IEEE trans-
actions on computers, Jan. 1994, pp. 90-93.
[2] K.J. Bowker, M.C. Warnes and M.D. Ashworth, "Some developments in eddy cur-
rent techniques for in-service inspection," Insight Vol. 37, No.3, March 1995. pp.
163-168.
[3] S.D. Brown, "Multifrequency IMultiparameter eddy current steam generator NDE,"
Quantitative NDE in the nuclear industry, Edited by R.B. Clough, American Soci-
ety For Metals, Metals Park, Ohio, 1983. pp 99-105.
[4] "Field experiences with multifrequency-multiparameter eddy current technology,"
EPRI NP-2299. Batelle, Columbus Laboratories and EPRI NDE Center. Electric
Power Research Institute, Palo Alto, Calofornia March 1982.
[5] R.K. Granville, "In service eddy current examination of non-ferrous industrial heat
exchanger tubing," British journal of NDT, Vol. 33, No.8, Aug. 1991, pp. 403-408
[6] D.J. Hagemaier, Fundamentals of eddy current testing, American Society for Non-
destructive Testing, Columbus, Ohio, 1990.
[7] "Eddy current inspection of non-ferrous heat-exchanger tubing," Insight, Vol. 36,
No.4, April 1994, pp. 218-220.
71
[8] D. Horne, S. Udpa and W. Lord, "Superposition of eddy current probe signals,"
Materials Evaluation, Vol. 42, June 1984, pp. 930-933.
[9] A.K. Jain, Fundamentals of digital image processing, Prentice hall of Private limited,
New Delhi, 1995. pp. 150-154.
[10] M. Junger, C. Brook, "Beginner's guide to sensor selection and evaluation tech-
niques for eddy current testing," British journal of NDT, Vol. 32, No.9, Sept.
1990, pp. 463-466.
[11] D. Kahaner, C. Moler and S. Nash, Numerical methods and software, Englewood
Cliffs, Prentice-Hall, New Jersey, 1989.
[12] H.L. Libby, "Eddy current test for tubing flaws in support regions," Research tech-
niques in nondestructive testing, Vol. 2 Edited by R.S. Sharpe, London, Academic
Press,1973, pp. 151-184.
[13] H.L. Libby, Introduction to electromagnetic nondestructive test methods, John Wiley
& Sons, New York, 1971.
[14] Electromagnetic methods of nondestructive testing, Edited by W. Lord, Gordon and
Breach Science Publishers, London, 1985, pp. 175-304.
[15] R.C. McMaster, Nondestructive testing handbook, Vol. 1, American Society for Non-
[16] R.C. McMaster, Nondestructive testing handbook, Vol. 2, American Society for Non-
[17] R.C. McMaster, P. McIntire and M.L. Mester, Nondestructive testing handbook, Vol.
4, 2nd ed., American Society for Nondestructive Testing, Columbus, Ohio, 1986.
72
[18] R.C. McMaster, "The present and future of eddy current testing," Materials Eval-
uation , Vol. 43, Nov. 1985, pp. 1512-1521.
[19] J.A. Mundis and C.W. DeYoung, "Solutions to NDE problems in PWR steam
generators-an overview," Quantitative NDE in the nuclear industry, Edited by R.B.
Clough, American Society for Metals, Metals Park, Ohio, 1983, pp. 93-98.
[20] M.J. Narasimha, A.M. Peterson, "On the computation of the discrete cosine trans-
form," IEEE transactions on communications, Vol. COM-26, No.6, June 1978, pp.
934-936.
[21] R. Palanisamy and W. Lord, "Finite element analysis of eddy current phenomena,"
Materials Evaluation, Vol. 38, No. 10, Oct. 1980, pp. 39-43.
[22] E. Persoon, and K.S. Fu, "Shape discrimination using Fourier descriptors," IEEE
transactions on systems, man and cybernetics, Vol. SMC-7, No.3, March 1977.
[23] L. de la Pintiere, "Back to basics," Materials Evaluation, Vol. 43, Feb. 1985, pp.
152-153.
[24] J.M. Prince, L.D. Reid, D.L. Lessor, "Two-frequency eddy current instrument for
measuring the thickness of Zircaloy cladding on Uranium under conditions of vary-
ing lift-off," Materials Evaluation, Vol. 43, Nov. 1985, pp. 1562-1565.
[25] S.S. Rao, "Nonlinear programming II: Unconstrained optimization techniques,"
chapter 6, Optimization theory and applications, Wiley Eastern Limited, New Delhi,
India, 1978.
[26] R. Saglio and M. Pigeon, "Concepts of multifrequency eddy current testing", Elec-
tromagnetic testing, Vol. 4, 2nd Ed., Edited by P. McIntire, R. McMuster, American
Society for Nondestructive testing, Columbus, Ohio, 1986, pp. 592-605.

Multifrequency Eddy Current Signal Analysis

Uploaded by

Copyright:

Available Formats

Multifrequency Eddy Current Signal Analysis

Uploaded by

Document Information

Original Title

Copyright

Available Formats

Share this document

Share or Embed Document

Sharing Options

Did you find this document useful?

Is this content inappropriate?

Copyright:

Available Formats

Multifrequency Eddy Current Signal Analysis

Uploaded by

Copyright:

Available Formats

Iowa State University Capstones, Theses and

Retrospective Theses and Dissertations

Multifrequency eddy current signal analysis

Follow this and additional works at: https://lib.dr.iastate.edu/rtd

A thesis submitted to the graduate faculty

in partial fulfillment of the requirements for the degree of

Major: Electrical Engineering

Major Professor: Satish Udpa

Iowa State University

Copyright © Avanindra, 1997. All rights reserved.

This is to certify that the Master's thesis of

has met the thesis requirements of Iowa State University

Signatures have been redacted for privacy

1.2 Summary of contributions and outline of thesis. 6

2 PRINCIPLES OF NONDESTRUCTIVE EDDY CURRENT EVAL-

2.1 Basics of single frequency eddy current testing 8

2.1.1 Eddy current probes 12

2.1.2 Frequency selection . 15

2.2 Multifrequency eddy current techniques. 18

2.2.1 Need for multifrequency testing . 18

2.2.2 Fundamentals of multifrequency eddy current inspection 20

3 ALGORITHMS FOR MULTIPARAMETER SIGNAL ANALYSIS 32

3.2 Affine transformation equations 33

3.3 Conjugate gradient method .. 35

3.4 Fourier series method . . . . . . . . . . . . . . . . . . . . . . . 45

3.5 Discrete cosine transform method . . . . 50

4 RESULTS AND CONCLUSIONS . . . . . . . . . . . 60

University who made working here an enjoyable experience.

and encouragement. I wish to dedicate this work to them.

of the approach are presented.

Nondestructive Evaluation (NDE) by definition involves the evaluation of a test

specimen for defects without affecting its usefulness or serviceability. Nondestructive

control and maintenance testing, to increase the reliability of manufactured products.

disruption in service are prime considerations in these cases.

Nondestructive evaluation methods share a common strategy. An excitation energy is

characteristics or defects. This information is extracted by employing proper signal

Eddy current NDE methods rely on the principles of electromagnetic induction, to

identify or differentiate between a wide variety of physical, structural and metallurgical

conditions in electrically conductive materials. Eddy currents offer high sensitivity to

coupling to the test specimen as is the case with ultrasonic methods.

corrosion is likely to be a major problem, more expensive corrosion resistant materials

making the inspection environment hostile to humans.

inspection of non-ferrous tubing, particularly in the nuclear power industry. In such

cases, eddy current inspection offers the following advantages:

• Relatively fast (0.5-1.0 mls testing speed).

• Can be carried out from inside the tube.

• Can detect both internal and external defects.

• Can detect both gradual and sharp defects.

• Provides quantitative flaw measurements (size, depth etc.).

• Lends itself to automated results analysis.

• Can be used to inspect a wide range of tubes.

expansion at tube sheets during manufacture [3].

variables. In general, at least as many independent eddy current readings must be

need to be determined, multiple frequencies should be used. The increased amount of

information gained using multifrequency methods, may be combined to nullify unwanted

Figure 1.1 Heat Exchanger Design and Manufacturing Discontinuities [14]

ful defect signals. Multifrequency processing techniques can be employed to suppress

that the algorithm can be implemented in real time.

advantage of the underlying heuristics. This allows the approach to be implemented in