SlideShare une entreprise Scribd logo
1  sur  14
PULSED EDDY CURRENT
DETECTION OF LOOSE
PARTS IN CANDU®
NUCLEAR REACTOR
STEAM GENERATORS
Dylan Johnston - 10063536
D.johnston@queensu.ca
Co-Supervisors:Prof. Morelli & Prof. Krause
Physics 590 Honours Research Thesis
Final Report
Submitted: September 4th, 2015
1
Pulsed Eddy Current Detection of Loose Parts in CANDU® Nuclear
Reactor Steam Generators
Abstract – Degradation of support structures and other components in nuclear steam generators (SGs) can
leadto the accumulationof debrisonthe SG tube sheet.Flowinducedvibrationof loose parts can resultin SG
tube damage and eventual lossofSG efficiency.Regularinspectionofsupportstructurescan extend SG life by
detecting and monitoring flaws and cracks in SG tubes as well as identifying accumulation of loose parts.
Currently,conventional eddycurrenttechnologiesare usedtocharacterize andassessthe conditionofsupport
structures. However, these technologies do not perform well in the presence of magnetite fouling or when
loose parts are resting directly on the SG tube sheet. Pulsed eddy current (PEC) combined with a modified
principal components analysis (MPCA) was investigated as a method of detecting a variety of loose part
materials in close proximity to 16 mm outer diameter Alloy-800 SG tubes and the surrounding support
structures. Loose parts were found to be detectable for varying axial positions along the SG tube, including
when in contact with the tube sheet. When offset radially from the SG tube, 1.6 mm diameter carbon steel
parts could be detectedupto 3.2 mm away, while stainlesssteel partsofthe same size couldonly be detected
up to 1.5 mm away. Signal variation in the presence of loose parts was attributed to differences in the
amplitude ofthe transientresponsesand in relaxationtimesfordiffusionof transientmagneticfieldsthrough
the variousmaterials. The abilityto accurately detectand identifyloose partslocated in close proximityto SG
tubes establishes PEC combined with MPCA as a potential candidate for SG inspection.
I. INTRODUCTION
Steam generators (SGs), which are a standard
component in many types of thermal power
generators including nuclear reactors, are used to
convert thermal energy into electrical energy via
steam powered rotary motion of turbine
generators. Steam generators undergo regular
inspectionsinordertoextendthe lifetimeof nuclear
reactors, while providing evidence of their
continued safe operation [1]. Therefore, it is
desirable todevelopinspectionstrategiescapableof
identifying flaws and other forms of degradation
efficiently, while reducing reactor down time. Eddy
current testing (ET) and ultrasonic testing (UT) are
two such methods that are used for a wide variety
of non-destructive testing procedures, including
flaw detection, condition assessment, and flaw
growth prediction within a SG [2].
A variety of degradation modes can be present
withinthe SGsof CANDUnuclearreactors,including
stresscorrosioncracking andfrettingwearontheSG
tubing, support structure degradation and
corrosion, and the accumulation of magnetite
sludge on the SG support structures. Magnetite
sludge isa corrosionproductcomposedprimarilyof
iron oxides,howevercopperand copperoxides can
alsobe present,dependingonthe materialsusedto
construct the SG [3]. Additionally, soluble species
such as sulphates and silicates can become
concentrated in magnetite sludge, and eventually
precipitate within the porous structure of the
deposits [3].
Small metallicparts that are not attached to the
supportstructures within a SG can compromise the
integrityof aSG.Loose partsmayappearinSGsafter
maintenance is performed or can detach from
support structures due to corrosion. Accumulation
occurs primarily on top of the tube sheet of the SG
[4],eitherdirectlyonthe tube sheetitself,orresting
on top of magnetite sludge, that has already
accumulated onthe tube sheet. Loose parts thatare
in physical contact withSG tubes,whenexposedto
flow induced vibration, can cause fretting wear,
produce more metallic debris and, eventually,
compromise the integrity of SG tube structures, if
not removed [2].
Detection of these loose parts is imperative to
maintaining plant efficiency, since SG tubes are
‘plugged’ once 40% through-wall fretsare detected
[5]. Currently, ET is the primary non-destructive
testing method used to detect loose parts that lie
within 3 millimetres of SG tubes [2] and not in
contact with the tube sheet. However, the
inspection performance of ET is reduced in the
presence of the magnetite fouling [2]. Additionally,
2
ET encountersdifficultiesindetectingloose partsin
the presence of ferromagnetic materials, such as a
tube sheetor othersupportstructureswithinanSG
[2].
PulsedEddyCurrenttesting(PEC) isanemerging
non-destructive testing technique that has been
developed for the inspection of nuclear and
aerospace structures [6] [7]. Instead of the
continuous sinusoidal signal utilized by ET, PEC
employsasquare wave voltagepulse toinduce eddy
currentsin conductive media. The transientvoltage
responsesinducedinthe pick-upcoilsbythese eddy
currents can be consideredas consisting of a series
of discrete frequencies [8].Thus,PEChasaspectrum
of frequencies available for interaction with the
inspected component, as opposed to the discrete
frequenciesthat are generatedby the ETdrivesignal
[9]. In addition, the approach to direct current (DC)
excitation in the pulse provides magnetization of
ferromagnetic materials, which enhances the pick-
upcoil response asopposedtohinderingit, asisthe
case in conventional ET [2]. PEC pick-up coil
responses maybe analyzedusingmodifiedPrincipal
Components Analysis (MPCA) [10] [11], which
reconstructs the signal using orthogonal basis
vectors and associated scores, reducing the
dimensionality of the data. MPCA differs from PCA
in that the data mean is not subtracted from the
original data [11].
When combined with PCA, PEC has been found
to be a robust analysistechnique [12].Inaerospace
NDT, PEC was able to detect stress-corrosion
cracking defects, which were undetectable using
conventional eddy current due to large lift-off
distance [6]. In nuclear reactor non-destructive
testing, PEC has been shown to characterize wall
loss, tube fretting wear, and corrosion of support
structuresmore successfullythanET [12],whichhas
limited performance when multiple modes of
degradation or magnetite fouling are present [2].
This report examines seven types of loose parts
made out of different metals placed in two basic
orientations. Modified PCA and the theory of
magnetic field diffusion are described in Section II,
followed by the experimental setup and a
description of the measurement technique usedin
Section III. Section IV discusses the results of using
MPCA to analyze PEC signals for detecting loose
parts, while Section V provides a conclusion and
summary of this report, as well as outlining the
future research suggested for PEC.
II. THEORY
A. Electromagnetic Diffusion
The square wave voltage pulse employed in PEC
inducestransienteddycurrentsinconductivemedia
that lie in close proximity to the drive coil. These
eddy currents are induced in a target material via
Faraday’s law of electromagnetic induction [13].
Changes in the eddy currents and associated
magnetic field can in turn be sensed using pick-up
coils. The approach to DC excitation in the pulse
provides a secondary magnetization effect in the
presence of ferromagnetic materials. The diffusion
of eddy currents is described by the diffusion
equation (7) below. We can derive this equation
beginningwithone of Maxwell’sequations, namely
Ampère’s circuital law [13](1):
𝛁⃗⃗ × 𝑩⃗⃗ = 𝜇0(𝑱+ 𝜀0
𝜕𝑬⃗⃗
𝜕𝑡
) (1)
The diffusionequationisobtainedbytakingthe curl
of both sides of the equation:
−𝛁⃗⃗ 2 𝑩⃗⃗ + 𝛁⃗⃗ (𝛁⃗⃗ ∙ 𝑩⃗⃗ ) = 𝜇0( 𝛁⃗⃗ × 𝑱) + 𝜇0 𝜀0(𝛁⃗⃗ ×
𝜕𝑬⃗⃗
𝜕𝑡
) (2)
By substituting in Ohm’s law (3), Faraday’s law of
electromagnetic induction (4) and Gauss’s law of
magnetism (5):
𝑱 = 𝜎𝑬⃗⃗ (3)
𝛁⃗⃗ × 𝑬⃗⃗ = −
𝜕𝑩⃗⃗
𝜕𝑡
(4)
𝛁⃗⃗ ∙ 𝑩⃗⃗ = 0 (5)
and using a vector identity,the diffusionequation
can be written in terms of just the magnetic field:
𝛁⃗⃗ 2 𝑩⃗⃗ = 𝜇0 𝜎
𝜕𝑩⃗⃗
𝜕𝑡
+ 𝜇0 𝜀0
𝜕2 𝑩⃗⃗
𝜕𝑡2
(6)
A goodconductorwill have alarge 𝜎 value.Thus,we
can approximate the second order term as
negligible, and the diffusion of magnetic flux, 𝐵⃗ ,at
low frequencies (<108 𝐻𝑧) can be written as [14]:
𝛁2 𝑩⃗⃗ = 𝜇𝜎
𝜕𝑩⃗⃗
𝜕𝑡
(7)
3
For conductive media, the general solution to the
diffusion equation of magnetic fields (7) has the
form [14]:
𝑩⃗⃗ = 𝑓(𝑒
−
𝑡
𝜏 𝐷) (8)
Solutions of this form can oftenbe expressedas
a series of relaxation times that are dependent on
the conductivity, 𝜎, and permeability, 𝜇, of a given
material,aswellassome characteristiclengthof the
system, ℓ. The characteristic diffusion time of the
material, 𝜏 𝐷,can be estimated as [15] [14]:
𝜏 𝐷~ 𝜇𝜎ℓ2 (9)
With longer times providing increased depth of
penetration by the induced eddy currents. To
understand the complete transient response of the
system,a seriesof relaxationtimes,asdescribedby
(9), is sufficient. The eddy current diffusion time is
dominated primarily by the 𝜇𝜎 product; the
characteristiclength, ℓ,dependsonseveralphysical
parameters of the situation including the size,
proximity and orientation of the loose part.
The difference in relaxation times between
different materials and different physical situations
(eg.size,orientation,distance) isexpectedtocause
variation in the transient signal response exhibited
by the probe’s pick-up coil array. Several other
variables affect the response of the pick-up coils,
including the magnetization of the ferromagnetic
conductive materials, the temperature of the
system, the resistance of the pick-up coils, and the
back EMF induced onthe drivingpulse.However,as
these factors are not drastically different between
non-ferromagnetic materials throughout the
investigation, the variation in pick-up coil response
is expected to be dominated by the 𝜏 𝐷 term. In the
case of ferromagnetic materials,the magnetization
component is expected to dominate [16].
The measured conductivityvaluesforthe various
loose part materials are given in Table 1.
B. Modified PCA
Principal Components Analysis (PCA) is an
analysis technique used to reduce the
dimensionality of data by separating large, highly
correlated data sets into a combination of linearly
uncorrelated principal components and associated
scores [10].
The principal component of adatasetisfoundby
creatingthe covariance matrix for an n-dimensional
data set that has had the mean subtracted from
everypoint. The eigenvectors,or‘components’,and
eigenvalues of this (𝑛 × 𝑛)-matrix are then
determined. ModifiedPCA differs in that the mean
is not subtracted from the data points before the
covariance matrix iscalculated [11].Byretainingthe
data mean in this analysis, results become less
prone tosystematicoffsetscausedbya differencein
the average signals.
Due to the nature of eigenvectors, these
componentswill be orthogonal toone another.The
principal components must also be of unit length,
and therefore must be normalized [10]. The
eigenvector with the largest eigenvalue is the
principal component, and represents the greatest
variation within the data set. The second largest
component describes the next largest variation
within the data set, and so on. By ordering the
eigenvectorsintermsof theireigenvalue,highestto
lowest,arankedlistof the componentsisobtained.
The components are ranked from most significant,
to least significant, in terms of the relationships
between dimensionsof the dataset [10].Discarding
components with small eigenvalues reduces the
dimensionality of the data at the cost of removing
information from the reconstructed signal.
Generally, discarding all but the first 3-5 principal
components allows for 99% signal reproduction [6]
[7] [12].
Reconstruction of the signalisquitesimpleatthis
point:the eigenvectorsthatwere keptare organized
as the rows of a matrix, with the most significant
eigenvector at the top. Multiplying this matrix on
the leftof the transposedoriginal dataset (i.e each
row contains a separate dimension) yields the
original data solely in terms of the retained
components. The original signal, 𝒀, minus the
information discarded in higher order components,
can be represented by a sum of the retained
components, 𝑽 𝒊, multiplied by their associated
scores, 𝑠𝑖 [10].
𝒀 = ∑𝑠𝑖 𝑽𝒊 (10)
This is similar to recreating a function using a
Fourier series of harmonic basis functions and
associated Fourier coefficients, however, fewer
4
terms are generally required compared to Fourier
analysis.
Using MPCA, the dimensionality of the data is
effectively reduced to the number of principal
components used to reconstruct the signal, while
signal reconstruction is accomplished through the
determination of several scores 𝑠𝑖. This method of
analysis represents a minimization of the sum
square residualsinterpretationof the result,instead
of a variance minimum interpretation, which is
obtainedinconventionalPCA [6].The programused
to perform the modified principal components
analysis was developed at RMC [11].
III. EXPERIMENTAL SETUP
A. Apparatus
Six Alloy-800 tubes, 45.7 cm in length, and one
Alloy-800 tube, 30.5 cm in length, were used to
recreate the SGtube structure forthisinvestigation.
Each tube had a wall thickness of 1.2 mm and an
outer diameter of 15.9 mm. A tube sheet support
plate of thickness 25 mm and a broach support
structure of thickness 26 mm were usedtoholdthe
Alloy-800 tubes in place (Figure 1). The 305 mm
Alloy-800tube was installed asmanufactured,while
the othersix tubes all possessedvarious degreesof
frettingwear.The unfrettedtube was usedfordata
acquisition asit representsnominal tube condition,
andwill be the SGtube referredtohereafter.The SG
tube wasmarkedwithmaskingtape forthe purpose
of offsetting the loose parts away from the tube
sheet, defined asthe axial direction (see Figure 1).
B. Loose Parts
The loose parts used in this investigation are
listedin Table1alongwiththeirrelevantdimensions
and measured conductivities. These values were
obtained using a 4-point measurement, and, aside
from lead, aluminum and tin, agree with literature
values within their uncertainty.
The brass, carbon steel and stainless steel
samples were loaned by the Canadian Nuclear
Laboratories(CNL) and are shownin Figure 2 (Top).
The tin, copper, aluminum and lead samples were
acquired and machined at Queens University and
are shown in Figure 2 (Bottom).
Material Label Diameter
(mm)
Length
(mm)
Conductivity (𝝁𝐒/𝒎)
(T=𝟐𝟐℃± 𝟐℃)
Carbon
Steel
1/16” 1.6 27.0 4 ± 0.5
1/8” 3.2 25.0
1/4” 6.3 25.9
1/2” 12.7 25.4
1” 25.4 25.6
Brass 1/8” 3.2 25.9 15 ± 3.7
1/4” 6.3 26.7
1/2” 12.7 25.9
1” 25.4 24.0
Stainless
Steel
(300
Series)
1/16” 1.6 26.2 1 ± 0.2
1/8” 3.2 25.3
1/4” 6.3 26.1
1/2” 12.7 25.7
1” 25.4 25.9
Copper 1/8” 3.2 26.3 55 ± 9.3
Aluminum 1/8” 3.2 25.7 28 ± 3.8
Tin 1/8” 3.3 29.3 8 ± 1.3
Lead 1/6” - 41.8 4 ± 0.2
Table 1 – Loose parts included in investigation (Lead part was a
rectangular prism approximately 4 mm in height and width)
Four 3.2 mm thick plastic offsetting shims,
loaned by CNL, were used to offset the loose parts
away from the SG tube wall, defined as the radial
Figure 2 - (Top) Brass, carbon steel and stainless steel 1” long
loose parts ranging from 1.6 mm to 25.4 mm in diameter.
(Bottom) 1” long copper, aluminum and tin loose parts with3.2
mm diameter, lead loose part.
Figure 1 - Steam generator tube structure withloose part in
parallel orientation; broach support not shown in figure.
Radial Offset
Axial Offset
5
direction (see Figure 1). Two magnetite collars, of
thickness 2.5 mm and 3.2 mm, were used to
simulate magnetitesludgesurroundingthe SGtube.
Both collarshad a relative magneticpermeabilityof
1.9 [17]. Materialsusedtooffsetthe loose parts are
shown in Figure 3.
C. Probe
The probe used in this investigation was
developed by RMC and manufactured by CNL. The
probe head, as described elsewhere [18], consisted
of a drive coil and 8 pick-up coils: 4 located above
(front array) and 4 located below (back array) the
drive coil,spacedazimuthally at90 degree intervals,
as shown in Figure 4. Pick-up coils that were 180°
apart were paired for analysis. These pairs were
matched as closely as possible in terms of signal
response inanattemptto increase the sensitivityof
the probe to induced eddy currents.
The probe body itself has a length of 77.6 mm
andan outerdiameterof 13.5mm.The lengthof the
probe body was encased in heat shrink material,
then labeled in 2 mm increments originating at the
drive coil.The drivecoil waswoundcoaxiallyaround
the probe 127 times using 36 AWG copper wire.
Each of the eight pick-up coils was wound with 360
turns using 42 AWG copper wire. The axis of each
pick-upcoil wasorientedperpendicularlytothe axis
of the drive coil, with shielded twisted pair wire
carrying the signals from the pick-up coils to a
purpose-built amplification circuit that was
developed at RMC.
The signals from each pick-up coil were
individually amplified with a gain of 100 before
being digitized at 1MHz using a 16-bit resolution
NI6356 USB DAQ. The DAQ was connected to a
laptopcomputerthatwasusedtosave the data and
perform MPCA on the acquired data. The NI6356
USB DAQ was also used to generate the excitation
pulse, which was carried to the drive coil using a
coaxial cable.The pulse, generatedbythe USB DAQ
at 1000 Hz and 50% duty cycle, hadan amplitudeof
1.5 V. A block diagram of the equipment used for
data acquisition is shown in Figure 5.
D. Method
The probe was initially insertedintothe SG tube
a distance of 90 mm, so that the drive coil was 10
mm from the edge of the tube sheet, before
measurementsbegan.Measurementsweretakenat
2 mm intervals within the Alloy-800 SG tube as the
probe was axially translated a distance of 130 mm
through the SG tube, for a total of 65 data points.
The loose partswereaffixedtotheSGtube suchthat
they were aligned with the probe, in order to
maximize the pick-up coil response. Temperature
was recorded for each data acquisition. A timed
LabVIEW program was used to generate excitation
pulses and to collect voltage responses from the
pick-up coil array. Many data sets, spanning a wide
variety of possible loose part sizes, compositions,
andboth axial andradial offset positions,wereused
in the construction of the principal components.
3.2 mm
thick plastic
shims
2.5 mm thick
magnetite collar
3.2 mm thick
magnetite collar
1.5 mm thick
plastic shim
0.5 mm thick
plastic shim
collar
Figure 3 - Materials used to offset loose parts radially from the
SG tube.
Figure 4 – PEC Probe Schematic
Figure 5 – Block diagram of experimental setup used
for investigation.
6
This ensuredthatthe sensitivity of the components
was notrestrictedtoa specificsituationof interest.
IV. RESULTS & DISCUSSION
A. Probe Response
The transient response of the probe in the
presence of just the SG tube (null trial), a 3.2 mm
diametercarbonsteel part, and a 3.2 mm diameter
copperpart, is showninFigure 6. The response was
found to exhibit a unique variation in the presence
of carbon steel, compared to the variations
observedforthe othermaterialssampled.Thiswas
attributed to carbon steel being the most
ferromagnetic material included in the
investigation.Stainlesssteel was found to have the
weakest transient response, and could not be
detected without more advanced analysis. Figure 6
clearly shows the variation in transient response
associated with differences in characteristic
diffusion times in the presence of different
materials, as discussed in Sec. II. A.
B. MPCA
The modified PCA described in Sec. II. B. was
performed ona collectionof datasets that covered
a wide range of part sizes andcompositions, inboth
radial and axial offset positions. One data set from
each situation of interest was used to generate the
principal components, as it was found that the
transient response of the probe was highly
reproducible. Additionally, the average mean
square error associated with signal reproduction
was found to be negligibly different when multiple
trials of each situation of interest were used to
generate the principal components (see Table 2 of
the Appendix).
The first four principal components were used
for signal reproduction, resulting in a reproduced
signal with an average mean square error of 1.04%
from the as-measuredsignal.Higherorderprincipal
componentswere foundto containnoise,andwere
discarded. The normalizedeigenvectorscanbe seen
in Figure 7.
The first principal component (𝑉1) accounts for
the average response of the data, and thus,
unsurprisingly, resembles the raw transient signal
shown in Figure 6. The scores associated with the
first principal component were foundto be at least
an orderof magnitude largerthanothercomponent
scores.The secondprincipal component(𝑉2) canbe
seen peaking earlier than the first principal
component, and possesses a second peak of
opposite sign. The third principal component (𝑉3)
shares a similar shape as compared with the first
principal component(𝑉1),resemblingthe rawsignal,
but peakslater in time;thissuggestsan association
with materials further from the probe, as observed
by J.Buck [12]. The final principal component(𝑉4) is
noisierthanthe previousthree,andpossessesthree
peaks.
C. Score Comparison
The most direct method of determining the
effectthe presenceof variousloosepartshadonthe
principal components scores was to compare the
-0.2
-0.15
-0.1
-0.05
0
0.05
0.1
0.15
0.2
0.25
0 50 100 150 200
Eigenvectors[V]
Time [𝜇s]
V1 V2
V3 V4
Figure 7 - Eigenvectors constructed by performing MPCA on a
global data set
-0.5
0.5
1.5
2.5
3.5
4.5
5.5
6.5
0.025 0.075 0.125 0.175 0.225
Voltage(V)
Time (ms)
Null Trial Transient Response
Carbon Steel Transient Response
Copper Transient Response
4.5
5
5.5
6
6.5
0.055 0.06 0.065 0.07 0.075
Voltage(V)
Time (ms)
Figure 6 - Transient voltage response of a pick-up coil in the
presence of an Alloy-800 tube, an Alloy-800 tube and a 3.2 mm
diameter carbon steel part, and an Alloy-800 tube and a 3.2 mm
copper part. Neither part possessed a radial offset.
7
scores of a trial containing a loose part with the
scores of a null trial as a function of axial
displacement. It was expected that, in the absence
of a loose part, scores should be approximately
equal between data sets. Therefore, if a score
variation was observed for a collection of data
points, it could be inferred that a loose part was
present at that axial position on the SG tube.
It wasobservedthatavertical offsetwaspresent
betweenindividual datapointseveninthe absence
of loose parts.Thismade the detectionof significant
score variationdifficultaseverydata pointin a trial
showedvariationfromthe correspondingdatapoint
in the null trial. This effect was determined to be a
resultof the temperature at whichthe data set was
taken, which was found to vary by as much as 4
degrees Celsius (see Appendix Figure 1). A
‘normalizing function’ was implemented, which
aligned data points corresponding to Alloy-800
tubing,allowingthevariationdue tothe presence of
a loose part to be more readily identified.
This method of loose part identification proved
to be useful for the majority of data sets that
involved aloose partin physical contactwiththe SG
tube. Typically, data sets possessedseveral distinct
traits:A large score deflectioninthe presenceof the
tube sheet, followed by a smaller magnitude
variationfromthe null setinthe presence of aloose
part, and, finally, little to no variation from the null
set for the remainder of the data set. Figure 8
demonstrates this type of comparison for two
different principal component scores.
D. Cluster Analysis
For loose partssmallerthan6.4 mm indiameter,
as well as for parts which contained a radial offset
from the SG tube, the score comparison method of
identification could not consistently be used to
determine the presence of loose parts in individual
principal componentscores.Instead,scores of both
a null trial, and a trial of interest, were plotted in
three dimensional PCA space, in order to identify
variations as linear combinations of individual
principal component scores. Figure 9 demonstrates
that, with proper viewing angle, variations in PCA
space are more readilydetectable thaninindividual
principal componentscores,especiallywithsmaller
parts and parts with higher resistivity values.
The second, third, and fourth principal
componentscoreswere mostcommonly plottedfor
this type of analysis, since the more difficult to
detectparts demonstratedthe leastvariationinthe
firstprincipal componentscore.Thisclusteranalysis
improved the recognition of loose parts that were
already detectable using score comparison, and
allowed for smaller loose parts, with larger radial
offsets, to be detected more easily.
Cluster analysis was only applied to parts with
diameterof 3.2 mmor less,asparts largerthan this
were foundto be quite readilydetectable. Itshould
be notedthatthe temperature normalizingfunction
described in Sec. IV. C. was also implemented for
cluster analysis.
-2.30
-1.30
-0.30
0.70
1.70
2.70
3.70
-43.00
-42.00
-41.00
-40.00
-39.00
-38.00
-37.00
6.8 9.8 12.8 15.8 18.8
V2Score
V1Score
Distance(cm)
Null V1
Brass V1
Null V2
Brass V2
Figure 8 - 3.2 mm diameter brass loose part oriented parallel to
the SG tube with no axial or radial offset. Variation of both 𝑉1
and 𝑉2 from the null set at approximately 128 mm indicates the
presence of the loose part.
-0.15
-0.1
-0.05
0
-0.2
-0.15
-0.1
-0.05
0
0.05
0.1
0
0.05
0.1
0.15
V3 ScoreV2 Score
V4Score
Stainless Steel (D = 1.6mm)
Null Trial
Figure 9 - 1.6 mm stainless steel loose part oriented parallel to
the SG tube with no axial or radial offset. Clustering of data
points indicates presence of loose part; score variations using
the method described in Sec. IV. C only occurred for a 25.4 mm
diameter stainless steel part in the same orientation.
8
With the exception of carbon steel, Copper was
the most detectable material. Parts with higher
conductivity values were found to be more easily
detectable (Figure 10). This is attributed to the
dependence of transient response variation on
differences in characteristic diffusion times, as
discussed in Sec. II. A. The most difficult to detect
material, stainless steel, not only possessed the
largest resistivity of the materials tested, but had
the most similar resistivity value compared to the
Alloy-800 tubing used in this investigation.
As expected from Eqn. 9, loose parts that were
largerinsize were easiertodetectforboththistype
of analysis (Figure 11), as well as for the score
comparison analysis mentioned in Sec. IV. C.
Detectability was also found to decrease rapidly
whenpartswere offsetradiallyfromthe SGtube, as
seen in Figure 12. Both Figure 11 and Figure 12
demonstrate score variations attributed to
differences in the characteristic length of the
system, as described by Eqn. 9.
Score variations were found to be more distinct
when a loose part was offset radially using a
magnetite collar, as compared to a part offset
radiallybythe same amountusinganon-conducting
plastic shim, as shown in Figure 13. This result can
be attributed to the increase in flux caused by the
magnetization of the ferromagnetic magnetite
collar.
The magnitude of the score variation when the
loose part was placed in the perpendicular
orientation was smaller than when it was placed in
the parallel orientation.Furthermore,loose partsin
the perpendicularorientationwere more difficultto
detect due to the score variation from the null set
Figure 11 – Brass loose parts ranging from 3.2 mm in diameter
to 25.4 mm in diameter with no radial or axial offset. Clearly,
larger loose parts result in more pronounced score variations
-0.5
0
0.5
-1.2-1-0.8-0.6-0.4-0.20
-1.2
-1
-0.8
-0.6
-0.4
-0.2
0
V3 Score
V2 Score
V4Score
Brass (D = 25.4mm)
Brass (D = 12.7mm)
Brass (D = 6.4mm)
Brass (D = 3.2mm)
Null Trial
Figure 12 – 3.2 mm diameter carbon steel loose part in contact
with the tube sheet at various (non-magnetite) radial offsets.
Score variation magnitude drops off quickly with radial offset.
-0.4
-0.2
0
0.2
0.4
0.6
-2-1.5-1-0.50
0
0.1
0.2
V2 Score
V3 Score
V4Score
Carbon Steel (OS = 0 mm)
Carbon Steel (OS = 3.2 mm)
Carbon Steel (OS = 6.4 mm)
Carbon Steel (OS = 9.6 mm)
Null Trial
Figure 13 – 3.2 mm diameter brass loose part in contact with
tube sheet offset 3.2 mm radially by plastic shim and by 3.2 mm
thick magnetite collar (P = 1.9). Clustering in PCA space is more
obvious when offset is due to magnetite.
0.1
0.2
0.3
0.4
0.5
-0.1
0
0.1
-0.1
-0.05
0
0.05
0.1
0.15
V2 Score
V3 Score
V4Score
Brass (OS = 3.2 mm Magnetite Collar)
Brass (OS = 3.2 mm)
Null Trial (3.2 mm Magnetite Collar
Null Trial
0.1
0.2
0.3
0.4
0.5
-0.0500.050.10.150.20.25
0
0.1
0.2
0.3
0.4
V2 Score
V3 Score
V4Score
Copper (D = 3.2 mm)
Aluminum (D = 3.2 mm)
Brass (D = 3.2 mm)
Stainless Steel (D = 3.2 mm)
Null Trial
Figure 10 –3.2 mmdiameter loose parts ofvarying compositionwith no
radial or axialoffset. Materials with higher conductivities possess larger
PCA space variation. Stainless steel is slightly ferromagnetic, and thus
shows variation in the opposite direction.
9
occurringfora fewernumberof datapoints,asseen
in Figure 14, due to a shorter length of interaction
with the probe
The score variation observed from a loose part
that was offset axially from the tube sheet was
found to be of the same, or of slightly lower
magnitude compared to that of a loose part in
contact with the tube sheet (Appendix Figure 2).
This effect is likely caused by the voltage
amplification due to the production of a secondary
magnetic field in the ferromagnetic tube sheet.
A summaryof the limitationsondetectabilityfor
a variety of loose parts can be found in Table 2.
E. Investigationof Multinomial Logistic
Regression (MLR)
Before PEC becomes a practical method for
identifying loose parts, a more efficient means of
detecting loose parts, and identifying the
characteristics it possesses, is required. A
multinomial logistic regression (MLR) function that
is included in MatLab was investigated for this
purpose.
MLR is a method used to generalize logistic
regression problems with more than two discrete
outcomes [19].Givenasetof independentvariables,
the model is used to classify groups of data by
calculating the probabilities of the various possible
outcomes of a categorically distributed dependent
variable [19].Inessence,datapointsare sortedinto
binsbasedontheirsimilaritytodatapointsonwhich
the model was trained on, in a logistic regression
sense.
Loose Part
(diameter)
Non-magnetite
radial offset
detectable
Magnetite radial
offset detectable
Carbon Steel
(3.2 mm)
9.6 mm 3.2 mm
Carbon Steel
(1.6 mm)
3.2 mm 3.2 mm
Brass (3.2 mm)
3.2 mm 3.2 mm
Stainless Steel
(3.2 mm)
3.2 mm -
Stainless Steel
(1.6 mm)
1.5 mm -
Aluminum
(3.2 mm)
0.5 mm 2.5 mm
Copper
(3.2 mm)
3.2 mm 2.5 mm
Tin (3.2 mm)
1.5 mm 2.5 mm
Lead
0.5 mm -
Table 2 - Magnetite and non-magnetite radial offset limitations
on detectability for loose parts in contact with the tube sheet. ‘-
‘ indicates a part could not be detected for the smallest offset
available.
The model was initiallytrainedusing entire data
setscontaining3.2 mm diameterloose parts,which
had been offset axially from the tube sheet by 30
mm. These data sets were used due to the spatial
separation between the tube sheet and the loose
part; data points could be easily categorized as
either loose part material, Alloy-800 Tube, or tube
sheet.
While the MLR model was successful at
identifying when an axially offset part was present,
it was found to have difficulties discriminating
between materialswith similar conductivity values.
Specifically, composition probabilities were split
betweencopperandaluminum, andbetweenbrass
and tin for a number of validation sets (Figure 15),
suggestingthatthe response of the pick-upcoil toa
loose part isstronglydependenton conductivity,as
expected from Eqn. 9.
The model also encountered difficulties in the
presence of anomalousdatapointswithinvalidation
sets;small score variationsinthe absence of aloose
part were identifiedaseithertin,aluminumorbrass
loose partswith greaterthan80% confidence.While
these results could be readily discounted in the
context of this investigation, in reactor this might
not be the case. Loose parts used in this
0 10 20 30 40 50 60 70
-0.5
0
0.5
1
1.5
2
Data Point
V3score
Null Trial
Carbon Steel Parallel
Carbon Steel Perpendicular
Figure 14 – 3.2 mm diameter carbon steel loose part placed in
the perpendicular and parallel orientation. The loose part has a
smaller variation when in the perpendicular orientation.
Tube Sheet
10
investigationhadaknownlength. Thus, ifthe model
detected a part for less data points than the length
of a loose part, it could be discounted. However, in
reactor,debrisare of unknownsize,andthus cannot
be counted as false positives.
The model was found to perform poorly when
the validation set contained a loose part size or
radial offset that had not been included in the
training sets. In situations such as these, the
confidence of the model either decreased by as
much as 60%, or yielded several additional false
positives, depending on the material used. This
resultisnotideal,asitimpliesthatonlyorientations
and sample sizes used to train the model can be
detectedconfidently,andincludingall possibleparts
and configurations commonly found in situ in the
training set would not be a feasible task.
V. CONCLUSION & FUTURE WORK
This report investigated the use of pulsed eddy
current and principal components analysis as a
method for detecting loose parts in close proximity
to the tube sheetsof CANDU nuclearreactor steam
generators. Several materialsspanningawide range
of conductivitieswereincludedinthe investigation.
In the presence of magnetite fouling, the
identification of data point clustering in PCA space
was found to improve for several materials,
includingbrass,aluminumandtin.Incloseproximity
to ferromagneticstructures,suchasthe tube sheet,
PEC could detect a variety of loose part
compositions as small as 3.2 mm in diameter at
varying radial offsets.
PCA isa robustanalysistechniquethatiscapable
of identifyingvariationsin PECpick-upcoil response
due to the presence of conductive materials. This
result is attributed to a difference in
electromagnetic field diffusion times when
conductive materials are presentin close proximity
to the SG tube.
PEC combinedwithPCA has beendemonstrated
as being a viable analysis technique for the
detection of loose part accumulation in close
proximity to the tube sheet of SGs.
A multinomial logistic regression model was
used to calculate the probability that a loose part
was present.The model was foundto be successful
at detecting the presence of loose parts, but could
only categorize scored data that it had been
specifically trained on. Additionally, the model had
difficulty differentiating between non-
ferromagnetic materials with similar conductivity
values due to a similarity in characteristic diffusion
times.
Further investigation in the study of loose parts
detection would include identifying the effects of
azimuthal offset from a pick-up coil (by as much as
45°). The capabilities of PEC when operating at
temperatures more closely resembling those found
in-situ must also be considered. Additionally, a
model capable of more accurately identifying loose
part properties should be investigated. Artificial
neural networks may be another avenue of
investigation to achieve this goal [20].
Figure 15 – Probability that a loose part is present, that the tube
sheet is present, and the probabilities that the part is composed
of tin or brass, as determined by the MLR model.
11
Appendix
Numberof RepeatedTrials Used Average Mean Square Error (V) MaximumMean Square Error (V)
1 𝟑. 𝟗𝟒 × 𝟏𝟎−𝟓
𝟏. 𝟑𝟒 × 𝟏𝟎−𝟑
2 𝟑. 𝟗𝟏 × 𝟏𝟎−𝟓
𝟏. 𝟐𝟕 × 𝟏𝟎−𝟑
5 𝟑. 𝟗𝟏 × 𝟏𝟎−𝟓
𝟏. 𝟐𝟒 × 𝟏𝟎−𝟑
10 𝟒. 𝟎𝟗 × 𝟏𝟎−𝟓
𝟏. 𝟑𝟏 × 𝟏𝟎−𝟑
Table 1 – Average and maximum mean square error associated with signal reproduction when 1, 2, 5 and 10 trials of the situations of
interest are used to generate the eigenvectors
Figure 1 – Effect of temperature on principal component scores;data points which should have the same score are offset for different
data sets, relationship between offset and temperature is linear.
-37.4
-37.3
-37.2
-37.1
-37
-36.9
-36.8
21.2 21.7 22.2 22.7 23.2 23.7 24.2 24.7 25.2
V1Score(ArbitraryUnits)
Temperature (°C)
V1 ScoreVersus Temperature
5th V1 Score Value
60th V1 Score Value
Figure 2 – 3.2 mm diameter copper loose part, parallel to and in contact with the SG tube for varying axial offsets.
Score variation magnitude is found to remain fairly constant for different axial offsets.
20 25 30 35 40 45 50
-0.8
-0.6
-0.4
-0.2
0
0.2
0.4
0.6
0.8
Data Point
V2score
Offset Plotting
Null Trial
OS=0mm, H=0mm
OS=0mm, H=10mm
OS=0mm, H=20mm
OS=0mm, H=30mm
12
References
[1] D. R. Diercks,W. J.Shack and J.Muscara, "Overviewof SteamGeneratorTube DegradationandIntegrity
Issues,"NuclearEngineering and Design, vol.194, no. 1, pp. 19-30, 1999.
[2] M. H. Attia,"FrettingFatigue andWearDamage of Structural ComponentsinNuclearPowerStations-
FitnessforService andLife ManagementPerspective," Tribology International, vol.39,no. 10, pp.1294-
1304, 2006.
[3] M. Vepsalainen,"DepositformationinPWRsteamgenerators,"FinnishNational ResearchProgramon
NPPSafety,2010.
[4] B. L. Jarman,I. M. Grant, R. Garg and AECB, "Regulationof AgeingSteamGenerators,"in Canadian
NuclearSociety,Toronto, 1998.
[5] K.-W.Ryu,C.-Y.Park, H.-N.KimandH. Rhee,"Predictionof FrettingWearDepthforSteamGenerator
TubesBasedOn VariousTypesof WearScars," Journalof NuclearScience and Technology, vol.47, no.5,
pp.449-456, 2010.
[6] P. F.Horan, P. R. Underhill andT.W. Krause,"Real Time PulsedEddyCurrentDetectionof CracksinF/A-
18 InnerWingSpar Using DiscriminantSeperationof ModifiedPrincipalComponentsAnalysisScores,"
IEEE SensorsJournal, vol.14,no. 1, pp.171-177, 2014.
[7] C. A.Stott, P.R. Underhill,V.K.Babbarand T. W. Krause,"PulsedEddyCurrentDetectionof Cracksin
MultilayerAluminumLapJoints," IEEESensorsJournal, vol.15,pp.956-962, 2014.
[8] P. Hammond,"The Calculationof the MagneticFieldof RotatingMachines.Part1: The Fieldof a Tubular
Current,"Proc.IEE - PartC: Monograph, vol.106, no.10, pp.158-164, 1959.
[9] T. W. Krause,C. Mandache and J. H. V.Lefebvre,"Diffusionof PulsedEddyCurrentsinThinConducting
Plates,"Reviewof QuantitativeNondestructiveEvaluation, vol.27,pp. 368-375, 2008.
[10] J. Lattin,D. Carroll andP. Green,in Analyzing MultivariateData,PacificGroove,CA,Brooks/Cole,2003,
pp.264-275.
[11] P. Horan,P. R. Underhill andT.W. Krause,"PulsedEddyCurrentDetectionof CracksinF/A-18InnerWing
Spar WithoutWingSkinRemoval UsingModifiedPrincipal ComponentAnalysis," NDT&EInternational,
vol.55, no.1, pp. 21-27, 2013.
[12] J. A.Buck, P.R. Underhill,S.G.Mokros, J. Morelli,V.K.Babbar,B. Lepine,J.RenaudandT. W. Krause,
"PulsedEddyCurrentInspectionof SupportStructuresinSteamGenerators," IEEESensorsJournal, vol.
15, no.8, pp.4305-4312, 2015.
[13] D. J. Griffiths,IntroductiontoElectrodynamics(4thEdition),Essex:PearsonEducation,Inc.,2012.
[14] H. C. Ohanian,"Onthe Approachto Electro- and Magento-staticEquilibrium,"American Journalof
Physics, vol.51, pp. 1020-1022, 1983.
13
[15] J. D. Jackson,"Quasi-StaticMagneticFieldsinConductorslEddyCurrents;MagneticDiffusion,"in Classical
Electrodynamics3rd Edition,Hoboken,NJ,Wiley&Sons,1999, pp. 219-238.
[16] D. R. Desjardins,G.Vallieres,P.P.WhalenandT. W. Krause,"Advancesintransient(pulsed) eddycurrent
for inspectionof milti-layeredaluminumstructuresinthe presence of ferrousfasteners," Reviewof
Progressin QualitativeNondestructiveEvaluation, vol.31,pp.400-407, 2012.
[17] S. G. Mokros, "PulsedEddyCurrentInspectionof BroachSupportPlatesinSteamGenerators,"Queen's
University,Kingston,2015.
[18] T. W. Krause,V.K. Babbar andP. R. Underhill,"A PulsedEddyCurrentProbe forInspectionof Support
PlatesFromWithinAlloy-800SteamGeneratorTubes,"Review of Progressin QuantitativeNondestructive
Evaluation, vol.33, pp.1352-1358, 2014.
[19] Mathworks,"Documentation - Multinomial LogisticRegression,"Mathworks,2015. [Online].Available:
http://www.mathworks.com/help/stats/mnrfit.html.[Accessed1st-28thAugust2015].
[20] J. A.Buck, P.R. Underhill,J.Morelli andT.W. Krause,"Simultaneousmulti-parametermeasurementin
pulsededdycurrentsteamgeneratordatausingartificial neural networks," IEEETrans.Instrumentation
and Measurements, ToBe Published.

Contenu connexe

En vedette (14)

Skolem’s “paradox” as logic of ground: The mutual foundation of both proper a...
Skolem’s “paradox” as logic of ground: The mutual foundation of both proper a...Skolem’s “paradox” as logic of ground: The mutual foundation of both proper a...
Skolem’s “paradox” as logic of ground: The mutual foundation of both proper a...
 
Viatges3r a
Viatges3r aViatges3r a
Viatges3r a
 
Zs prakovce
Zs prakovceZs prakovce
Zs prakovce
 
стеклодув 2
стеклодув 2стеклодув 2
стеклодув 2
 
адлет Buduschee reshenie
адлет Buduschee reshenieадлет Buduschee reshenie
адлет Buduschee reshenie
 
Betka
BetkaBetka
Betka
 
Protipoziarne
ProtipoziarneProtipoziarne
Protipoziarne
 
Yet Another PaaS
Yet Another PaaSYet Another PaaS
Yet Another PaaS
 
CML-Studiendatenbank: Patienteninformationen für Studien
CML-Studiendatenbank: Patienteninformationen für StudienCML-Studiendatenbank: Patienteninformationen für Studien
CML-Studiendatenbank: Patienteninformationen für Studien
 
20160310 els contes inf
20160310 els contes inf20160310 els contes inf
20160310 els contes inf
 
Relógio Eletrônico de Ponto Henry Prisma Super Fácil R04 - Aplicativo Web Em...
Relógio Eletrônico de Ponto Henry Prisma Super Fácil  R04 - Aplicativo Web Em...Relógio Eletrônico de Ponto Henry Prisma Super Fácil  R04 - Aplicativo Web Em...
Relógio Eletrônico de Ponto Henry Prisma Super Fácil R04 - Aplicativo Web Em...
 
Primers dies del curs 2015-16 dels nens i nenes d’EI3
Primers dies del curs 2015-16 dels nens i nenes d’EI3Primers dies del curs 2015-16 dels nens i nenes d’EI3
Primers dies del curs 2015-16 dels nens i nenes d’EI3
 
3.4 fryzjer
3.4 fryzjer3.4 fryzjer
3.4 fryzjer
 
5.3 fryzjer
5.3 fryzjer5.3 fryzjer
5.3 fryzjer
 

Similaire à Undergraduate_Thesis_Dylan_Johnston_10063536

Set and seu analysis of cntfet based designs in harsh environments
Set and seu analysis of cntfet based designs in harsh environmentsSet and seu analysis of cntfet based designs in harsh environments
Set and seu analysis of cntfet based designs in harsh environmentseSAT Publishing House
 
PERFORMANCE OF HYBRID ELECTROMAGNETIC DAMPER FOR VEHICLE SUSPENSION
PERFORMANCE OF HYBRID ELECTROMAGNETIC DAMPER FOR VEHICLE SUSPENSIONPERFORMANCE OF HYBRID ELECTROMAGNETIC DAMPER FOR VEHICLE SUSPENSION
PERFORMANCE OF HYBRID ELECTROMAGNETIC DAMPER FOR VEHICLE SUSPENSIONijiert bestjournal
 
Resonant-tunneling-diode effect in Si-based double-barrier structure sputtere...
Resonant-tunneling-diode effect in Si-based double-barrier structure sputtere...Resonant-tunneling-diode effect in Si-based double-barrier structure sputtere...
Resonant-tunneling-diode effect in Si-based double-barrier structure sputtere...IJRES Journal
 
Effect of voltage on multiple particles and collisions in a single Phase Gas ...
Effect of voltage on multiple particles and collisions in a single Phase Gas ...Effect of voltage on multiple particles and collisions in a single Phase Gas ...
Effect of voltage on multiple particles and collisions in a single Phase Gas ...IJERA Editor
 
Faults detection in metallic tubes using eddy current.
Faults detection in metallic tubes using eddy current.Faults detection in metallic tubes using eddy current.
Faults detection in metallic tubes using eddy current.ministry of electricity
 
Molecular dynamics simulation studies of mechanical properties of different c...
Molecular dynamics simulation studies of mechanical properties of different c...Molecular dynamics simulation studies of mechanical properties of different c...
Molecular dynamics simulation studies of mechanical properties of different c...Johan Zhi Kang Kok
 
Electrostatic Edge Plasma Turbulence in the Uragan-3M torsatron
Electrostatic Edge Plasma Turbulence in the Uragan-3M torsatronElectrostatic Edge Plasma Turbulence in the Uragan-3M torsatron
Electrostatic Edge Plasma Turbulence in the Uragan-3M torsatronAleksey Beletskii
 
Performance analysis of a monopole antenna with fluorescent tubes at 4.9 g hz...
Performance analysis of a monopole antenna with fluorescent tubes at 4.9 g hz...Performance analysis of a monopole antenna with fluorescent tubes at 4.9 g hz...
Performance analysis of a monopole antenna with fluorescent tubes at 4.9 g hz...Alexander Decker
 
Modeling the Dependence of Power Diode on Temperature and Radiation
Modeling the Dependence of Power Diode on Temperature and RadiationModeling the Dependence of Power Diode on Temperature and Radiation
Modeling the Dependence of Power Diode on Temperature and RadiationIJPEDS-IAES
 
Structural and Optical properties of Multiwalled Carbon Nanotubes Modified by...
Structural and Optical properties of Multiwalled Carbon Nanotubes Modified by...Structural and Optical properties of Multiwalled Carbon Nanotubes Modified by...
Structural and Optical properties of Multiwalled Carbon Nanotubes Modified by...TELKOMNIKA JOURNAL
 
Effect of mesh grid structure in reducing hot carrier effect of nmos device s...
Effect of mesh grid structure in reducing hot carrier effect of nmos device s...Effect of mesh grid structure in reducing hot carrier effect of nmos device s...
Effect of mesh grid structure in reducing hot carrier effect of nmos device s...ijcsa
 
Time-Frequency Representation For Estimating Young’s Modulus And Poisson’s Ra...
Time-Frequency Representation For Estimating Young’s Modulus And Poisson’s Ra...Time-Frequency Representation For Estimating Young’s Modulus And Poisson’s Ra...
Time-Frequency Representation For Estimating Young’s Modulus And Poisson’s Ra...IJERA Editor
 
EFFECT OF ULTRAVIOLET RADIATION ON STRUCTURAL PROPERTIES OF NANOWIRES
EFFECT OF ULTRAVIOLET RADIATION ON STRUCTURAL PROPERTIES OF NANOWIRESEFFECT OF ULTRAVIOLET RADIATION ON STRUCTURAL PROPERTIES OF NANOWIRES
EFFECT OF ULTRAVIOLET RADIATION ON STRUCTURAL PROPERTIES OF NANOWIRESijoejournal
 
Physical Models of LENR Processes Using the BSM-SG Atomic Models
Physical Models of LENR Processes Using the BSM-SG Atomic ModelsPhysical Models of LENR Processes Using the BSM-SG Atomic Models
Physical Models of LENR Processes Using the BSM-SG Atomic ModelsStoyan Sarg Sargoytchev
 
Geometric and process design of ultra-thin junctionless double gate vertical ...
Geometric and process design of ultra-thin junctionless double gate vertical ...Geometric and process design of ultra-thin junctionless double gate vertical ...
Geometric and process design of ultra-thin junctionless double gate vertical ...IJECEIAES
 
Magnetic Flux Controllers in Induction Heating and Melting by Robert Goldstei...
Magnetic Flux Controllers in Induction Heating and Melting by Robert Goldstei...Magnetic Flux Controllers in Induction Heating and Melting by Robert Goldstei...
Magnetic Flux Controllers in Induction Heating and Melting by Robert Goldstei...Fluxtrol Inc.
 
Coilgun: Simulation and Analysis
Coilgun: Simulation and AnalysisCoilgun: Simulation and Analysis
Coilgun: Simulation and Analysispaperpublications3
 
COMBINED SKEWED CMOS RING OSCILLATOR
COMBINED SKEWED CMOS RING OSCILLATORCOMBINED SKEWED CMOS RING OSCILLATOR
COMBINED SKEWED CMOS RING OSCILLATORecij
 

Similaire à Undergraduate_Thesis_Dylan_Johnston_10063536 (20)

Set and seu analysis of cntfet based designs in harsh environments
Set and seu analysis of cntfet based designs in harsh environmentsSet and seu analysis of cntfet based designs in harsh environments
Set and seu analysis of cntfet based designs in harsh environments
 
PERFORMANCE OF HYBRID ELECTROMAGNETIC DAMPER FOR VEHICLE SUSPENSION
PERFORMANCE OF HYBRID ELECTROMAGNETIC DAMPER FOR VEHICLE SUSPENSIONPERFORMANCE OF HYBRID ELECTROMAGNETIC DAMPER FOR VEHICLE SUSPENSION
PERFORMANCE OF HYBRID ELECTROMAGNETIC DAMPER FOR VEHICLE SUSPENSION
 
Resonant-tunneling-diode effect in Si-based double-barrier structure sputtere...
Resonant-tunneling-diode effect in Si-based double-barrier structure sputtere...Resonant-tunneling-diode effect in Si-based double-barrier structure sputtere...
Resonant-tunneling-diode effect in Si-based double-barrier structure sputtere...
 
Effect of voltage on multiple particles and collisions in a single Phase Gas ...
Effect of voltage on multiple particles and collisions in a single Phase Gas ...Effect of voltage on multiple particles and collisions in a single Phase Gas ...
Effect of voltage on multiple particles and collisions in a single Phase Gas ...
 
Faults detection in metallic tubes using eddy current.
Faults detection in metallic tubes using eddy current.Faults detection in metallic tubes using eddy current.
Faults detection in metallic tubes using eddy current.
 
Molecular dynamics simulation studies of mechanical properties of different c...
Molecular dynamics simulation studies of mechanical properties of different c...Molecular dynamics simulation studies of mechanical properties of different c...
Molecular dynamics simulation studies of mechanical properties of different c...
 
Electrostatic Edge Plasma Turbulence in the Uragan-3M torsatron
Electrostatic Edge Plasma Turbulence in the Uragan-3M torsatronElectrostatic Edge Plasma Turbulence in the Uragan-3M torsatron
Electrostatic Edge Plasma Turbulence in the Uragan-3M torsatron
 
Performance analysis of a monopole antenna with fluorescent tubes at 4.9 g hz...
Performance analysis of a monopole antenna with fluorescent tubes at 4.9 g hz...Performance analysis of a monopole antenna with fluorescent tubes at 4.9 g hz...
Performance analysis of a monopole antenna with fluorescent tubes at 4.9 g hz...
 
Modeling the Dependence of Power Diode on Temperature and Radiation
Modeling the Dependence of Power Diode on Temperature and RadiationModeling the Dependence of Power Diode on Temperature and Radiation
Modeling the Dependence of Power Diode on Temperature and Radiation
 
Structural and Optical properties of Multiwalled Carbon Nanotubes Modified by...
Structural and Optical properties of Multiwalled Carbon Nanotubes Modified by...Structural and Optical properties of Multiwalled Carbon Nanotubes Modified by...
Structural and Optical properties of Multiwalled Carbon Nanotubes Modified by...
 
Effect of mesh grid structure in reducing hot carrier effect of nmos device s...
Effect of mesh grid structure in reducing hot carrier effect of nmos device s...Effect of mesh grid structure in reducing hot carrier effect of nmos device s...
Effect of mesh grid structure in reducing hot carrier effect of nmos device s...
 
Time-Frequency Representation For Estimating Young’s Modulus And Poisson’s Ra...
Time-Frequency Representation For Estimating Young’s Modulus And Poisson’s Ra...Time-Frequency Representation For Estimating Young’s Modulus And Poisson’s Ra...
Time-Frequency Representation For Estimating Young’s Modulus And Poisson’s Ra...
 
EFFECT OF ULTRAVIOLET RADIATION ON STRUCTURAL PROPERTIES OF NANOWIRES
EFFECT OF ULTRAVIOLET RADIATION ON STRUCTURAL PROPERTIES OF NANOWIRESEFFECT OF ULTRAVIOLET RADIATION ON STRUCTURAL PROPERTIES OF NANOWIRES
EFFECT OF ULTRAVIOLET RADIATION ON STRUCTURAL PROPERTIES OF NANOWIRES
 
LifeTimeReport_v7
LifeTimeReport_v7LifeTimeReport_v7
LifeTimeReport_v7
 
Q04605101105
Q04605101105Q04605101105
Q04605101105
 
Physical Models of LENR Processes Using the BSM-SG Atomic Models
Physical Models of LENR Processes Using the BSM-SG Atomic ModelsPhysical Models of LENR Processes Using the BSM-SG Atomic Models
Physical Models of LENR Processes Using the BSM-SG Atomic Models
 
Geometric and process design of ultra-thin junctionless double gate vertical ...
Geometric and process design of ultra-thin junctionless double gate vertical ...Geometric and process design of ultra-thin junctionless double gate vertical ...
Geometric and process design of ultra-thin junctionless double gate vertical ...
 
Magnetic Flux Controllers in Induction Heating and Melting by Robert Goldstei...
Magnetic Flux Controllers in Induction Heating and Melting by Robert Goldstei...Magnetic Flux Controllers in Induction Heating and Melting by Robert Goldstei...
Magnetic Flux Controllers in Induction Heating and Melting by Robert Goldstei...
 
Coilgun: Simulation and Analysis
Coilgun: Simulation and AnalysisCoilgun: Simulation and Analysis
Coilgun: Simulation and Analysis
 
COMBINED SKEWED CMOS RING OSCILLATOR
COMBINED SKEWED CMOS RING OSCILLATORCOMBINED SKEWED CMOS RING OSCILLATOR
COMBINED SKEWED CMOS RING OSCILLATOR
 

Undergraduate_Thesis_Dylan_Johnston_10063536

  • 1. PULSED EDDY CURRENT DETECTION OF LOOSE PARTS IN CANDU® NUCLEAR REACTOR STEAM GENERATORS Dylan Johnston - 10063536 D.johnston@queensu.ca Co-Supervisors:Prof. Morelli & Prof. Krause Physics 590 Honours Research Thesis Final Report Submitted: September 4th, 2015
  • 2. 1 Pulsed Eddy Current Detection of Loose Parts in CANDU® Nuclear Reactor Steam Generators Abstract – Degradation of support structures and other components in nuclear steam generators (SGs) can leadto the accumulationof debrisonthe SG tube sheet.Flowinducedvibrationof loose parts can resultin SG tube damage and eventual lossofSG efficiency.Regularinspectionofsupportstructurescan extend SG life by detecting and monitoring flaws and cracks in SG tubes as well as identifying accumulation of loose parts. Currently,conventional eddycurrenttechnologiesare usedtocharacterize andassessthe conditionofsupport structures. However, these technologies do not perform well in the presence of magnetite fouling or when loose parts are resting directly on the SG tube sheet. Pulsed eddy current (PEC) combined with a modified principal components analysis (MPCA) was investigated as a method of detecting a variety of loose part materials in close proximity to 16 mm outer diameter Alloy-800 SG tubes and the surrounding support structures. Loose parts were found to be detectable for varying axial positions along the SG tube, including when in contact with the tube sheet. When offset radially from the SG tube, 1.6 mm diameter carbon steel parts could be detectedupto 3.2 mm away, while stainlesssteel partsofthe same size couldonly be detected up to 1.5 mm away. Signal variation in the presence of loose parts was attributed to differences in the amplitude ofthe transientresponsesand in relaxationtimesfordiffusionof transientmagneticfieldsthrough the variousmaterials. The abilityto accurately detectand identifyloose partslocated in close proximityto SG tubes establishes PEC combined with MPCA as a potential candidate for SG inspection. I. INTRODUCTION Steam generators (SGs), which are a standard component in many types of thermal power generators including nuclear reactors, are used to convert thermal energy into electrical energy via steam powered rotary motion of turbine generators. Steam generators undergo regular inspectionsinordertoextendthe lifetimeof nuclear reactors, while providing evidence of their continued safe operation [1]. Therefore, it is desirable todevelopinspectionstrategiescapableof identifying flaws and other forms of degradation efficiently, while reducing reactor down time. Eddy current testing (ET) and ultrasonic testing (UT) are two such methods that are used for a wide variety of non-destructive testing procedures, including flaw detection, condition assessment, and flaw growth prediction within a SG [2]. A variety of degradation modes can be present withinthe SGsof CANDUnuclearreactors,including stresscorrosioncracking andfrettingwearontheSG tubing, support structure degradation and corrosion, and the accumulation of magnetite sludge on the SG support structures. Magnetite sludge isa corrosionproductcomposedprimarilyof iron oxides,howevercopperand copperoxides can alsobe present,dependingonthe materialsusedto construct the SG [3]. Additionally, soluble species such as sulphates and silicates can become concentrated in magnetite sludge, and eventually precipitate within the porous structure of the deposits [3]. Small metallicparts that are not attached to the supportstructures within a SG can compromise the integrityof aSG.Loose partsmayappearinSGsafter maintenance is performed or can detach from support structures due to corrosion. Accumulation occurs primarily on top of the tube sheet of the SG [4],eitherdirectlyonthe tube sheetitself,orresting on top of magnetite sludge, that has already accumulated onthe tube sheet. Loose parts thatare in physical contact withSG tubes,whenexposedto flow induced vibration, can cause fretting wear, produce more metallic debris and, eventually, compromise the integrity of SG tube structures, if not removed [2]. Detection of these loose parts is imperative to maintaining plant efficiency, since SG tubes are ‘plugged’ once 40% through-wall fretsare detected [5]. Currently, ET is the primary non-destructive testing method used to detect loose parts that lie within 3 millimetres of SG tubes [2] and not in contact with the tube sheet. However, the inspection performance of ET is reduced in the presence of the magnetite fouling [2]. Additionally,
  • 3. 2 ET encountersdifficultiesindetectingloose partsin the presence of ferromagnetic materials, such as a tube sheetor othersupportstructureswithinanSG [2]. PulsedEddyCurrenttesting(PEC) isanemerging non-destructive testing technique that has been developed for the inspection of nuclear and aerospace structures [6] [7]. Instead of the continuous sinusoidal signal utilized by ET, PEC employsasquare wave voltagepulse toinduce eddy currentsin conductive media. The transientvoltage responsesinducedinthe pick-upcoilsbythese eddy currents can be consideredas consisting of a series of discrete frequencies [8].Thus,PEChasaspectrum of frequencies available for interaction with the inspected component, as opposed to the discrete frequenciesthat are generatedby the ETdrivesignal [9]. In addition, the approach to direct current (DC) excitation in the pulse provides magnetization of ferromagnetic materials, which enhances the pick- upcoil response asopposedtohinderingit, asisthe case in conventional ET [2]. PEC pick-up coil responses maybe analyzedusingmodifiedPrincipal Components Analysis (MPCA) [10] [11], which reconstructs the signal using orthogonal basis vectors and associated scores, reducing the dimensionality of the data. MPCA differs from PCA in that the data mean is not subtracted from the original data [11]. When combined with PCA, PEC has been found to be a robust analysistechnique [12].Inaerospace NDT, PEC was able to detect stress-corrosion cracking defects, which were undetectable using conventional eddy current due to large lift-off distance [6]. In nuclear reactor non-destructive testing, PEC has been shown to characterize wall loss, tube fretting wear, and corrosion of support structuresmore successfullythanET [12],whichhas limited performance when multiple modes of degradation or magnetite fouling are present [2]. This report examines seven types of loose parts made out of different metals placed in two basic orientations. Modified PCA and the theory of magnetic field diffusion are described in Section II, followed by the experimental setup and a description of the measurement technique usedin Section III. Section IV discusses the results of using MPCA to analyze PEC signals for detecting loose parts, while Section V provides a conclusion and summary of this report, as well as outlining the future research suggested for PEC. II. THEORY A. Electromagnetic Diffusion The square wave voltage pulse employed in PEC inducestransienteddycurrentsinconductivemedia that lie in close proximity to the drive coil. These eddy currents are induced in a target material via Faraday’s law of electromagnetic induction [13]. Changes in the eddy currents and associated magnetic field can in turn be sensed using pick-up coils. The approach to DC excitation in the pulse provides a secondary magnetization effect in the presence of ferromagnetic materials. The diffusion of eddy currents is described by the diffusion equation (7) below. We can derive this equation beginningwithone of Maxwell’sequations, namely Ampère’s circuital law [13](1): 𝛁⃗⃗ × 𝑩⃗⃗ = 𝜇0(𝑱+ 𝜀0 𝜕𝑬⃗⃗ 𝜕𝑡 ) (1) The diffusionequationisobtainedbytakingthe curl of both sides of the equation: −𝛁⃗⃗ 2 𝑩⃗⃗ + 𝛁⃗⃗ (𝛁⃗⃗ ∙ 𝑩⃗⃗ ) = 𝜇0( 𝛁⃗⃗ × 𝑱) + 𝜇0 𝜀0(𝛁⃗⃗ × 𝜕𝑬⃗⃗ 𝜕𝑡 ) (2) By substituting in Ohm’s law (3), Faraday’s law of electromagnetic induction (4) and Gauss’s law of magnetism (5): 𝑱 = 𝜎𝑬⃗⃗ (3) 𝛁⃗⃗ × 𝑬⃗⃗ = − 𝜕𝑩⃗⃗ 𝜕𝑡 (4) 𝛁⃗⃗ ∙ 𝑩⃗⃗ = 0 (5) and using a vector identity,the diffusionequation can be written in terms of just the magnetic field: 𝛁⃗⃗ 2 𝑩⃗⃗ = 𝜇0 𝜎 𝜕𝑩⃗⃗ 𝜕𝑡 + 𝜇0 𝜀0 𝜕2 𝑩⃗⃗ 𝜕𝑡2 (6) A goodconductorwill have alarge 𝜎 value.Thus,we can approximate the second order term as negligible, and the diffusion of magnetic flux, 𝐵⃗ ,at low frequencies (<108 𝐻𝑧) can be written as [14]: 𝛁2 𝑩⃗⃗ = 𝜇𝜎 𝜕𝑩⃗⃗ 𝜕𝑡 (7)
  • 4. 3 For conductive media, the general solution to the diffusion equation of magnetic fields (7) has the form [14]: 𝑩⃗⃗ = 𝑓(𝑒 − 𝑡 𝜏 𝐷) (8) Solutions of this form can oftenbe expressedas a series of relaxation times that are dependent on the conductivity, 𝜎, and permeability, 𝜇, of a given material,aswellassome characteristiclengthof the system, ℓ. The characteristic diffusion time of the material, 𝜏 𝐷,can be estimated as [15] [14]: 𝜏 𝐷~ 𝜇𝜎ℓ2 (9) With longer times providing increased depth of penetration by the induced eddy currents. To understand the complete transient response of the system,a seriesof relaxationtimes,asdescribedby (9), is sufficient. The eddy current diffusion time is dominated primarily by the 𝜇𝜎 product; the characteristiclength, ℓ,dependsonseveralphysical parameters of the situation including the size, proximity and orientation of the loose part. The difference in relaxation times between different materials and different physical situations (eg.size,orientation,distance) isexpectedtocause variation in the transient signal response exhibited by the probe’s pick-up coil array. Several other variables affect the response of the pick-up coils, including the magnetization of the ferromagnetic conductive materials, the temperature of the system, the resistance of the pick-up coils, and the back EMF induced onthe drivingpulse.However,as these factors are not drastically different between non-ferromagnetic materials throughout the investigation, the variation in pick-up coil response is expected to be dominated by the 𝜏 𝐷 term. In the case of ferromagnetic materials,the magnetization component is expected to dominate [16]. The measured conductivityvaluesforthe various loose part materials are given in Table 1. B. Modified PCA Principal Components Analysis (PCA) is an analysis technique used to reduce the dimensionality of data by separating large, highly correlated data sets into a combination of linearly uncorrelated principal components and associated scores [10]. The principal component of adatasetisfoundby creatingthe covariance matrix for an n-dimensional data set that has had the mean subtracted from everypoint. The eigenvectors,or‘components’,and eigenvalues of this (𝑛 × 𝑛)-matrix are then determined. ModifiedPCA differs in that the mean is not subtracted from the data points before the covariance matrix iscalculated [11].Byretainingthe data mean in this analysis, results become less prone tosystematicoffsetscausedbya differencein the average signals. Due to the nature of eigenvectors, these componentswill be orthogonal toone another.The principal components must also be of unit length, and therefore must be normalized [10]. The eigenvector with the largest eigenvalue is the principal component, and represents the greatest variation within the data set. The second largest component describes the next largest variation within the data set, and so on. By ordering the eigenvectorsintermsof theireigenvalue,highestto lowest,arankedlistof the componentsisobtained. The components are ranked from most significant, to least significant, in terms of the relationships between dimensionsof the dataset [10].Discarding components with small eigenvalues reduces the dimensionality of the data at the cost of removing information from the reconstructed signal. Generally, discarding all but the first 3-5 principal components allows for 99% signal reproduction [6] [7] [12]. Reconstruction of the signalisquitesimpleatthis point:the eigenvectorsthatwere keptare organized as the rows of a matrix, with the most significant eigenvector at the top. Multiplying this matrix on the leftof the transposedoriginal dataset (i.e each row contains a separate dimension) yields the original data solely in terms of the retained components. The original signal, 𝒀, minus the information discarded in higher order components, can be represented by a sum of the retained components, 𝑽 𝒊, multiplied by their associated scores, 𝑠𝑖 [10]. 𝒀 = ∑𝑠𝑖 𝑽𝒊 (10) This is similar to recreating a function using a Fourier series of harmonic basis functions and associated Fourier coefficients, however, fewer
  • 5. 4 terms are generally required compared to Fourier analysis. Using MPCA, the dimensionality of the data is effectively reduced to the number of principal components used to reconstruct the signal, while signal reconstruction is accomplished through the determination of several scores 𝑠𝑖. This method of analysis represents a minimization of the sum square residualsinterpretationof the result,instead of a variance minimum interpretation, which is obtainedinconventionalPCA [6].The programused to perform the modified principal components analysis was developed at RMC [11]. III. EXPERIMENTAL SETUP A. Apparatus Six Alloy-800 tubes, 45.7 cm in length, and one Alloy-800 tube, 30.5 cm in length, were used to recreate the SGtube structure forthisinvestigation. Each tube had a wall thickness of 1.2 mm and an outer diameter of 15.9 mm. A tube sheet support plate of thickness 25 mm and a broach support structure of thickness 26 mm were usedtoholdthe Alloy-800 tubes in place (Figure 1). The 305 mm Alloy-800tube was installed asmanufactured,while the othersix tubes all possessedvarious degreesof frettingwear.The unfrettedtube was usedfordata acquisition asit representsnominal tube condition, andwill be the SGtube referredtohereafter.The SG tube wasmarkedwithmaskingtape forthe purpose of offsetting the loose parts away from the tube sheet, defined asthe axial direction (see Figure 1). B. Loose Parts The loose parts used in this investigation are listedin Table1alongwiththeirrelevantdimensions and measured conductivities. These values were obtained using a 4-point measurement, and, aside from lead, aluminum and tin, agree with literature values within their uncertainty. The brass, carbon steel and stainless steel samples were loaned by the Canadian Nuclear Laboratories(CNL) and are shownin Figure 2 (Top). The tin, copper, aluminum and lead samples were acquired and machined at Queens University and are shown in Figure 2 (Bottom). Material Label Diameter (mm) Length (mm) Conductivity (𝝁𝐒/𝒎) (T=𝟐𝟐℃± 𝟐℃) Carbon Steel 1/16” 1.6 27.0 4 ± 0.5 1/8” 3.2 25.0 1/4” 6.3 25.9 1/2” 12.7 25.4 1” 25.4 25.6 Brass 1/8” 3.2 25.9 15 ± 3.7 1/4” 6.3 26.7 1/2” 12.7 25.9 1” 25.4 24.0 Stainless Steel (300 Series) 1/16” 1.6 26.2 1 ± 0.2 1/8” 3.2 25.3 1/4” 6.3 26.1 1/2” 12.7 25.7 1” 25.4 25.9 Copper 1/8” 3.2 26.3 55 ± 9.3 Aluminum 1/8” 3.2 25.7 28 ± 3.8 Tin 1/8” 3.3 29.3 8 ± 1.3 Lead 1/6” - 41.8 4 ± 0.2 Table 1 – Loose parts included in investigation (Lead part was a rectangular prism approximately 4 mm in height and width) Four 3.2 mm thick plastic offsetting shims, loaned by CNL, were used to offset the loose parts away from the SG tube wall, defined as the radial Figure 2 - (Top) Brass, carbon steel and stainless steel 1” long loose parts ranging from 1.6 mm to 25.4 mm in diameter. (Bottom) 1” long copper, aluminum and tin loose parts with3.2 mm diameter, lead loose part. Figure 1 - Steam generator tube structure withloose part in parallel orientation; broach support not shown in figure. Radial Offset Axial Offset
  • 6. 5 direction (see Figure 1). Two magnetite collars, of thickness 2.5 mm and 3.2 mm, were used to simulate magnetitesludgesurroundingthe SGtube. Both collarshad a relative magneticpermeabilityof 1.9 [17]. Materialsusedtooffsetthe loose parts are shown in Figure 3. C. Probe The probe used in this investigation was developed by RMC and manufactured by CNL. The probe head, as described elsewhere [18], consisted of a drive coil and 8 pick-up coils: 4 located above (front array) and 4 located below (back array) the drive coil,spacedazimuthally at90 degree intervals, as shown in Figure 4. Pick-up coils that were 180° apart were paired for analysis. These pairs were matched as closely as possible in terms of signal response inanattemptto increase the sensitivityof the probe to induced eddy currents. The probe body itself has a length of 77.6 mm andan outerdiameterof 13.5mm.The lengthof the probe body was encased in heat shrink material, then labeled in 2 mm increments originating at the drive coil.The drivecoil waswoundcoaxiallyaround the probe 127 times using 36 AWG copper wire. Each of the eight pick-up coils was wound with 360 turns using 42 AWG copper wire. The axis of each pick-upcoil wasorientedperpendicularlytothe axis of the drive coil, with shielded twisted pair wire carrying the signals from the pick-up coils to a purpose-built amplification circuit that was developed at RMC. The signals from each pick-up coil were individually amplified with a gain of 100 before being digitized at 1MHz using a 16-bit resolution NI6356 USB DAQ. The DAQ was connected to a laptopcomputerthatwasusedtosave the data and perform MPCA on the acquired data. The NI6356 USB DAQ was also used to generate the excitation pulse, which was carried to the drive coil using a coaxial cable.The pulse, generatedbythe USB DAQ at 1000 Hz and 50% duty cycle, hadan amplitudeof 1.5 V. A block diagram of the equipment used for data acquisition is shown in Figure 5. D. Method The probe was initially insertedintothe SG tube a distance of 90 mm, so that the drive coil was 10 mm from the edge of the tube sheet, before measurementsbegan.Measurementsweretakenat 2 mm intervals within the Alloy-800 SG tube as the probe was axially translated a distance of 130 mm through the SG tube, for a total of 65 data points. The loose partswereaffixedtotheSGtube suchthat they were aligned with the probe, in order to maximize the pick-up coil response. Temperature was recorded for each data acquisition. A timed LabVIEW program was used to generate excitation pulses and to collect voltage responses from the pick-up coil array. Many data sets, spanning a wide variety of possible loose part sizes, compositions, andboth axial andradial offset positions,wereused in the construction of the principal components. 3.2 mm thick plastic shims 2.5 mm thick magnetite collar 3.2 mm thick magnetite collar 1.5 mm thick plastic shim 0.5 mm thick plastic shim collar Figure 3 - Materials used to offset loose parts radially from the SG tube. Figure 4 – PEC Probe Schematic Figure 5 – Block diagram of experimental setup used for investigation.
  • 7. 6 This ensuredthatthe sensitivity of the components was notrestrictedtoa specificsituationof interest. IV. RESULTS & DISCUSSION A. Probe Response The transient response of the probe in the presence of just the SG tube (null trial), a 3.2 mm diametercarbonsteel part, and a 3.2 mm diameter copperpart, is showninFigure 6. The response was found to exhibit a unique variation in the presence of carbon steel, compared to the variations observedforthe othermaterialssampled.Thiswas attributed to carbon steel being the most ferromagnetic material included in the investigation.Stainlesssteel was found to have the weakest transient response, and could not be detected without more advanced analysis. Figure 6 clearly shows the variation in transient response associated with differences in characteristic diffusion times in the presence of different materials, as discussed in Sec. II. A. B. MPCA The modified PCA described in Sec. II. B. was performed ona collectionof datasets that covered a wide range of part sizes andcompositions, inboth radial and axial offset positions. One data set from each situation of interest was used to generate the principal components, as it was found that the transient response of the probe was highly reproducible. Additionally, the average mean square error associated with signal reproduction was found to be negligibly different when multiple trials of each situation of interest were used to generate the principal components (see Table 2 of the Appendix). The first four principal components were used for signal reproduction, resulting in a reproduced signal with an average mean square error of 1.04% from the as-measuredsignal.Higherorderprincipal componentswere foundto containnoise,andwere discarded. The normalizedeigenvectorscanbe seen in Figure 7. The first principal component (𝑉1) accounts for the average response of the data, and thus, unsurprisingly, resembles the raw transient signal shown in Figure 6. The scores associated with the first principal component were foundto be at least an orderof magnitude largerthanothercomponent scores.The secondprincipal component(𝑉2) canbe seen peaking earlier than the first principal component, and possesses a second peak of opposite sign. The third principal component (𝑉3) shares a similar shape as compared with the first principal component(𝑉1),resemblingthe rawsignal, but peakslater in time;thissuggestsan association with materials further from the probe, as observed by J.Buck [12]. The final principal component(𝑉4) is noisierthanthe previousthree,andpossessesthree peaks. C. Score Comparison The most direct method of determining the effectthe presenceof variousloosepartshadonthe principal components scores was to compare the -0.2 -0.15 -0.1 -0.05 0 0.05 0.1 0.15 0.2 0.25 0 50 100 150 200 Eigenvectors[V] Time [𝜇s] V1 V2 V3 V4 Figure 7 - Eigenvectors constructed by performing MPCA on a global data set -0.5 0.5 1.5 2.5 3.5 4.5 5.5 6.5 0.025 0.075 0.125 0.175 0.225 Voltage(V) Time (ms) Null Trial Transient Response Carbon Steel Transient Response Copper Transient Response 4.5 5 5.5 6 6.5 0.055 0.06 0.065 0.07 0.075 Voltage(V) Time (ms) Figure 6 - Transient voltage response of a pick-up coil in the presence of an Alloy-800 tube, an Alloy-800 tube and a 3.2 mm diameter carbon steel part, and an Alloy-800 tube and a 3.2 mm copper part. Neither part possessed a radial offset.
  • 8. 7 scores of a trial containing a loose part with the scores of a null trial as a function of axial displacement. It was expected that, in the absence of a loose part, scores should be approximately equal between data sets. Therefore, if a score variation was observed for a collection of data points, it could be inferred that a loose part was present at that axial position on the SG tube. It wasobservedthatavertical offsetwaspresent betweenindividual datapointseveninthe absence of loose parts.Thismade the detectionof significant score variationdifficultaseverydata pointin a trial showedvariationfromthe correspondingdatapoint in the null trial. This effect was determined to be a resultof the temperature at whichthe data set was taken, which was found to vary by as much as 4 degrees Celsius (see Appendix Figure 1). A ‘normalizing function’ was implemented, which aligned data points corresponding to Alloy-800 tubing,allowingthevariationdue tothe presence of a loose part to be more readily identified. This method of loose part identification proved to be useful for the majority of data sets that involved aloose partin physical contactwiththe SG tube. Typically, data sets possessedseveral distinct traits:A large score deflectioninthe presenceof the tube sheet, followed by a smaller magnitude variationfromthe null setinthe presence of aloose part, and, finally, little to no variation from the null set for the remainder of the data set. Figure 8 demonstrates this type of comparison for two different principal component scores. D. Cluster Analysis For loose partssmallerthan6.4 mm indiameter, as well as for parts which contained a radial offset from the SG tube, the score comparison method of identification could not consistently be used to determine the presence of loose parts in individual principal componentscores.Instead,scores of both a null trial, and a trial of interest, were plotted in three dimensional PCA space, in order to identify variations as linear combinations of individual principal component scores. Figure 9 demonstrates that, with proper viewing angle, variations in PCA space are more readilydetectable thaninindividual principal componentscores,especiallywithsmaller parts and parts with higher resistivity values. The second, third, and fourth principal componentscoreswere mostcommonly plottedfor this type of analysis, since the more difficult to detectparts demonstratedthe leastvariationinthe firstprincipal componentscore.Thisclusteranalysis improved the recognition of loose parts that were already detectable using score comparison, and allowed for smaller loose parts, with larger radial offsets, to be detected more easily. Cluster analysis was only applied to parts with diameterof 3.2 mmor less,asparts largerthan this were foundto be quite readilydetectable. Itshould be notedthatthe temperature normalizingfunction described in Sec. IV. C. was also implemented for cluster analysis. -2.30 -1.30 -0.30 0.70 1.70 2.70 3.70 -43.00 -42.00 -41.00 -40.00 -39.00 -38.00 -37.00 6.8 9.8 12.8 15.8 18.8 V2Score V1Score Distance(cm) Null V1 Brass V1 Null V2 Brass V2 Figure 8 - 3.2 mm diameter brass loose part oriented parallel to the SG tube with no axial or radial offset. Variation of both 𝑉1 and 𝑉2 from the null set at approximately 128 mm indicates the presence of the loose part. -0.15 -0.1 -0.05 0 -0.2 -0.15 -0.1 -0.05 0 0.05 0.1 0 0.05 0.1 0.15 V3 ScoreV2 Score V4Score Stainless Steel (D = 1.6mm) Null Trial Figure 9 - 1.6 mm stainless steel loose part oriented parallel to the SG tube with no axial or radial offset. Clustering of data points indicates presence of loose part; score variations using the method described in Sec. IV. C only occurred for a 25.4 mm diameter stainless steel part in the same orientation.
  • 9. 8 With the exception of carbon steel, Copper was the most detectable material. Parts with higher conductivity values were found to be more easily detectable (Figure 10). This is attributed to the dependence of transient response variation on differences in characteristic diffusion times, as discussed in Sec. II. A. The most difficult to detect material, stainless steel, not only possessed the largest resistivity of the materials tested, but had the most similar resistivity value compared to the Alloy-800 tubing used in this investigation. As expected from Eqn. 9, loose parts that were largerinsize were easiertodetectforboththistype of analysis (Figure 11), as well as for the score comparison analysis mentioned in Sec. IV. C. Detectability was also found to decrease rapidly whenpartswere offsetradiallyfromthe SGtube, as seen in Figure 12. Both Figure 11 and Figure 12 demonstrate score variations attributed to differences in the characteristic length of the system, as described by Eqn. 9. Score variations were found to be more distinct when a loose part was offset radially using a magnetite collar, as compared to a part offset radiallybythe same amountusinganon-conducting plastic shim, as shown in Figure 13. This result can be attributed to the increase in flux caused by the magnetization of the ferromagnetic magnetite collar. The magnitude of the score variation when the loose part was placed in the perpendicular orientation was smaller than when it was placed in the parallel orientation.Furthermore,loose partsin the perpendicularorientationwere more difficultto detect due to the score variation from the null set Figure 11 – Brass loose parts ranging from 3.2 mm in diameter to 25.4 mm in diameter with no radial or axial offset. Clearly, larger loose parts result in more pronounced score variations -0.5 0 0.5 -1.2-1-0.8-0.6-0.4-0.20 -1.2 -1 -0.8 -0.6 -0.4 -0.2 0 V3 Score V2 Score V4Score Brass (D = 25.4mm) Brass (D = 12.7mm) Brass (D = 6.4mm) Brass (D = 3.2mm) Null Trial Figure 12 – 3.2 mm diameter carbon steel loose part in contact with the tube sheet at various (non-magnetite) radial offsets. Score variation magnitude drops off quickly with radial offset. -0.4 -0.2 0 0.2 0.4 0.6 -2-1.5-1-0.50 0 0.1 0.2 V2 Score V3 Score V4Score Carbon Steel (OS = 0 mm) Carbon Steel (OS = 3.2 mm) Carbon Steel (OS = 6.4 mm) Carbon Steel (OS = 9.6 mm) Null Trial Figure 13 – 3.2 mm diameter brass loose part in contact with tube sheet offset 3.2 mm radially by plastic shim and by 3.2 mm thick magnetite collar (P = 1.9). Clustering in PCA space is more obvious when offset is due to magnetite. 0.1 0.2 0.3 0.4 0.5 -0.1 0 0.1 -0.1 -0.05 0 0.05 0.1 0.15 V2 Score V3 Score V4Score Brass (OS = 3.2 mm Magnetite Collar) Brass (OS = 3.2 mm) Null Trial (3.2 mm Magnetite Collar Null Trial 0.1 0.2 0.3 0.4 0.5 -0.0500.050.10.150.20.25 0 0.1 0.2 0.3 0.4 V2 Score V3 Score V4Score Copper (D = 3.2 mm) Aluminum (D = 3.2 mm) Brass (D = 3.2 mm) Stainless Steel (D = 3.2 mm) Null Trial Figure 10 –3.2 mmdiameter loose parts ofvarying compositionwith no radial or axialoffset. Materials with higher conductivities possess larger PCA space variation. Stainless steel is slightly ferromagnetic, and thus shows variation in the opposite direction.
  • 10. 9 occurringfora fewernumberof datapoints,asseen in Figure 14, due to a shorter length of interaction with the probe The score variation observed from a loose part that was offset axially from the tube sheet was found to be of the same, or of slightly lower magnitude compared to that of a loose part in contact with the tube sheet (Appendix Figure 2). This effect is likely caused by the voltage amplification due to the production of a secondary magnetic field in the ferromagnetic tube sheet. A summaryof the limitationsondetectabilityfor a variety of loose parts can be found in Table 2. E. Investigationof Multinomial Logistic Regression (MLR) Before PEC becomes a practical method for identifying loose parts, a more efficient means of detecting loose parts, and identifying the characteristics it possesses, is required. A multinomial logistic regression (MLR) function that is included in MatLab was investigated for this purpose. MLR is a method used to generalize logistic regression problems with more than two discrete outcomes [19].Givenasetof independentvariables, the model is used to classify groups of data by calculating the probabilities of the various possible outcomes of a categorically distributed dependent variable [19].Inessence,datapointsare sortedinto binsbasedontheirsimilaritytodatapointsonwhich the model was trained on, in a logistic regression sense. Loose Part (diameter) Non-magnetite radial offset detectable Magnetite radial offset detectable Carbon Steel (3.2 mm) 9.6 mm 3.2 mm Carbon Steel (1.6 mm) 3.2 mm 3.2 mm Brass (3.2 mm) 3.2 mm 3.2 mm Stainless Steel (3.2 mm) 3.2 mm - Stainless Steel (1.6 mm) 1.5 mm - Aluminum (3.2 mm) 0.5 mm 2.5 mm Copper (3.2 mm) 3.2 mm 2.5 mm Tin (3.2 mm) 1.5 mm 2.5 mm Lead 0.5 mm - Table 2 - Magnetite and non-magnetite radial offset limitations on detectability for loose parts in contact with the tube sheet. ‘- ‘ indicates a part could not be detected for the smallest offset available. The model was initiallytrainedusing entire data setscontaining3.2 mm diameterloose parts,which had been offset axially from the tube sheet by 30 mm. These data sets were used due to the spatial separation between the tube sheet and the loose part; data points could be easily categorized as either loose part material, Alloy-800 Tube, or tube sheet. While the MLR model was successful at identifying when an axially offset part was present, it was found to have difficulties discriminating between materialswith similar conductivity values. Specifically, composition probabilities were split betweencopperandaluminum, andbetweenbrass and tin for a number of validation sets (Figure 15), suggestingthatthe response of the pick-upcoil toa loose part isstronglydependenton conductivity,as expected from Eqn. 9. The model also encountered difficulties in the presence of anomalousdatapointswithinvalidation sets;small score variationsinthe absence of aloose part were identifiedaseithertin,aluminumorbrass loose partswith greaterthan80% confidence.While these results could be readily discounted in the context of this investigation, in reactor this might not be the case. Loose parts used in this 0 10 20 30 40 50 60 70 -0.5 0 0.5 1 1.5 2 Data Point V3score Null Trial Carbon Steel Parallel Carbon Steel Perpendicular Figure 14 – 3.2 mm diameter carbon steel loose part placed in the perpendicular and parallel orientation. The loose part has a smaller variation when in the perpendicular orientation. Tube Sheet
  • 11. 10 investigationhadaknownlength. Thus, ifthe model detected a part for less data points than the length of a loose part, it could be discounted. However, in reactor,debrisare of unknownsize,andthus cannot be counted as false positives. The model was found to perform poorly when the validation set contained a loose part size or radial offset that had not been included in the training sets. In situations such as these, the confidence of the model either decreased by as much as 60%, or yielded several additional false positives, depending on the material used. This resultisnotideal,asitimpliesthatonlyorientations and sample sizes used to train the model can be detectedconfidently,andincludingall possibleparts and configurations commonly found in situ in the training set would not be a feasible task. V. CONCLUSION & FUTURE WORK This report investigated the use of pulsed eddy current and principal components analysis as a method for detecting loose parts in close proximity to the tube sheetsof CANDU nuclearreactor steam generators. Several materialsspanningawide range of conductivitieswereincludedinthe investigation. In the presence of magnetite fouling, the identification of data point clustering in PCA space was found to improve for several materials, includingbrass,aluminumandtin.Incloseproximity to ferromagneticstructures,suchasthe tube sheet, PEC could detect a variety of loose part compositions as small as 3.2 mm in diameter at varying radial offsets. PCA isa robustanalysistechniquethatiscapable of identifyingvariationsin PECpick-upcoil response due to the presence of conductive materials. This result is attributed to a difference in electromagnetic field diffusion times when conductive materials are presentin close proximity to the SG tube. PEC combinedwithPCA has beendemonstrated as being a viable analysis technique for the detection of loose part accumulation in close proximity to the tube sheet of SGs. A multinomial logistic regression model was used to calculate the probability that a loose part was present.The model was foundto be successful at detecting the presence of loose parts, but could only categorize scored data that it had been specifically trained on. Additionally, the model had difficulty differentiating between non- ferromagnetic materials with similar conductivity values due to a similarity in characteristic diffusion times. Further investigation in the study of loose parts detection would include identifying the effects of azimuthal offset from a pick-up coil (by as much as 45°). The capabilities of PEC when operating at temperatures more closely resembling those found in-situ must also be considered. Additionally, a model capable of more accurately identifying loose part properties should be investigated. Artificial neural networks may be another avenue of investigation to achieve this goal [20]. Figure 15 – Probability that a loose part is present, that the tube sheet is present, and the probabilities that the part is composed of tin or brass, as determined by the MLR model.
  • 12. 11 Appendix Numberof RepeatedTrials Used Average Mean Square Error (V) MaximumMean Square Error (V) 1 𝟑. 𝟗𝟒 × 𝟏𝟎−𝟓 𝟏. 𝟑𝟒 × 𝟏𝟎−𝟑 2 𝟑. 𝟗𝟏 × 𝟏𝟎−𝟓 𝟏. 𝟐𝟕 × 𝟏𝟎−𝟑 5 𝟑. 𝟗𝟏 × 𝟏𝟎−𝟓 𝟏. 𝟐𝟒 × 𝟏𝟎−𝟑 10 𝟒. 𝟎𝟗 × 𝟏𝟎−𝟓 𝟏. 𝟑𝟏 × 𝟏𝟎−𝟑 Table 1 – Average and maximum mean square error associated with signal reproduction when 1, 2, 5 and 10 trials of the situations of interest are used to generate the eigenvectors Figure 1 – Effect of temperature on principal component scores;data points which should have the same score are offset for different data sets, relationship between offset and temperature is linear. -37.4 -37.3 -37.2 -37.1 -37 -36.9 -36.8 21.2 21.7 22.2 22.7 23.2 23.7 24.2 24.7 25.2 V1Score(ArbitraryUnits) Temperature (°C) V1 ScoreVersus Temperature 5th V1 Score Value 60th V1 Score Value Figure 2 – 3.2 mm diameter copper loose part, parallel to and in contact with the SG tube for varying axial offsets. Score variation magnitude is found to remain fairly constant for different axial offsets. 20 25 30 35 40 45 50 -0.8 -0.6 -0.4 -0.2 0 0.2 0.4 0.6 0.8 Data Point V2score Offset Plotting Null Trial OS=0mm, H=0mm OS=0mm, H=10mm OS=0mm, H=20mm OS=0mm, H=30mm
  • 13. 12 References [1] D. R. Diercks,W. J.Shack and J.Muscara, "Overviewof SteamGeneratorTube DegradationandIntegrity Issues,"NuclearEngineering and Design, vol.194, no. 1, pp. 19-30, 1999. [2] M. H. Attia,"FrettingFatigue andWearDamage of Structural ComponentsinNuclearPowerStations- FitnessforService andLife ManagementPerspective," Tribology International, vol.39,no. 10, pp.1294- 1304, 2006. [3] M. Vepsalainen,"DepositformationinPWRsteamgenerators,"FinnishNational ResearchProgramon NPPSafety,2010. [4] B. L. Jarman,I. M. Grant, R. Garg and AECB, "Regulationof AgeingSteamGenerators,"in Canadian NuclearSociety,Toronto, 1998. [5] K.-W.Ryu,C.-Y.Park, H.-N.KimandH. Rhee,"Predictionof FrettingWearDepthforSteamGenerator TubesBasedOn VariousTypesof WearScars," Journalof NuclearScience and Technology, vol.47, no.5, pp.449-456, 2010. [6] P. F.Horan, P. R. Underhill andT.W. Krause,"Real Time PulsedEddyCurrentDetectionof CracksinF/A- 18 InnerWingSpar Using DiscriminantSeperationof ModifiedPrincipalComponentsAnalysisScores," IEEE SensorsJournal, vol.14,no. 1, pp.171-177, 2014. [7] C. A.Stott, P.R. Underhill,V.K.Babbarand T. W. Krause,"PulsedEddyCurrentDetectionof Cracksin MultilayerAluminumLapJoints," IEEESensorsJournal, vol.15,pp.956-962, 2014. [8] P. Hammond,"The Calculationof the MagneticFieldof RotatingMachines.Part1: The Fieldof a Tubular Current,"Proc.IEE - PartC: Monograph, vol.106, no.10, pp.158-164, 1959. [9] T. W. Krause,C. Mandache and J. H. V.Lefebvre,"Diffusionof PulsedEddyCurrentsinThinConducting Plates,"Reviewof QuantitativeNondestructiveEvaluation, vol.27,pp. 368-375, 2008. [10] J. Lattin,D. Carroll andP. Green,in Analyzing MultivariateData,PacificGroove,CA,Brooks/Cole,2003, pp.264-275. [11] P. Horan,P. R. Underhill andT.W. Krause,"PulsedEddyCurrentDetectionof CracksinF/A-18InnerWing Spar WithoutWingSkinRemoval UsingModifiedPrincipal ComponentAnalysis," NDT&EInternational, vol.55, no.1, pp. 21-27, 2013. [12] J. A.Buck, P.R. Underhill,S.G.Mokros, J. Morelli,V.K.Babbar,B. Lepine,J.RenaudandT. W. Krause, "PulsedEddyCurrentInspectionof SupportStructuresinSteamGenerators," IEEESensorsJournal, vol. 15, no.8, pp.4305-4312, 2015. [13] D. J. Griffiths,IntroductiontoElectrodynamics(4thEdition),Essex:PearsonEducation,Inc.,2012. [14] H. C. Ohanian,"Onthe Approachto Electro- and Magento-staticEquilibrium,"American Journalof Physics, vol.51, pp. 1020-1022, 1983.
  • 14. 13 [15] J. D. Jackson,"Quasi-StaticMagneticFieldsinConductorslEddyCurrents;MagneticDiffusion,"in Classical Electrodynamics3rd Edition,Hoboken,NJ,Wiley&Sons,1999, pp. 219-238. [16] D. R. Desjardins,G.Vallieres,P.P.WhalenandT. W. Krause,"Advancesintransient(pulsed) eddycurrent for inspectionof milti-layeredaluminumstructuresinthe presence of ferrousfasteners," Reviewof Progressin QualitativeNondestructiveEvaluation, vol.31,pp.400-407, 2012. [17] S. G. Mokros, "PulsedEddyCurrentInspectionof BroachSupportPlatesinSteamGenerators,"Queen's University,Kingston,2015. [18] T. W. Krause,V.K. Babbar andP. R. Underhill,"A PulsedEddyCurrentProbe forInspectionof Support PlatesFromWithinAlloy-800SteamGeneratorTubes,"Review of Progressin QuantitativeNondestructive Evaluation, vol.33, pp.1352-1358, 2014. [19] Mathworks,"Documentation - Multinomial LogisticRegression,"Mathworks,2015. [Online].Available: http://www.mathworks.com/help/stats/mnrfit.html.[Accessed1st-28thAugust2015]. [20] J. A.Buck, P.R. Underhill,J.Morelli andT.W. Krause,"Simultaneousmulti-parametermeasurementin pulsededdycurrentsteamgeneratordatausingartificial neural networks," IEEETrans.Instrumentation and Measurements, ToBe Published.