WCNDT-EC Wrokshop

Download as pptx, pdf, or txt
Download as pptx, pdf, or txt
You are on page 1of 106

Eddy Current Techniques in NDE

Lalita Udpa
Nondestructive Evaluation Laboratory
Department of Electrical and
Computer Engineering Michigan State University
East Lansing, MI 48824

WCNDT Preconference Workshop,


Durban, April 15, 2012
Outline

 Part I - Physical Principles


 Part II – Probes
 Part III Inspection Modes
 Part IV - Forward Problem in EC-NDE -
Finite Element Modeling
 Part V - Inverse Problem in EC-NDE –
Defect Classification/Characterization
 Part VI – Case Study- Eddy current
inspection of SG tubes
Application of Eddy Current NDE
• Eddy current NDE is commonly used in the inspection
of conducting samples
– Measurement of impedance changes in coils in the presence of an
anomaly in a conducting specimen
• Typical applications
– Steam generator tubing in nuclear power plants
– Aircraft components
• Nuclear Power Industry
–Inspection of Steam Generator Tubing in
Nuclear Power Plants
Aircraft components

• Wheels
• Rivet Holes
• Engines – Impeller bores
• Fuselage skin
Part I - Physical Principles
Part I - Phhysical
• Principles
Alternating magnetic fields are generated by alternating current
excitation – Maxwell Ampere Law
• Magnetic field induces currents (eddy currents) in test specimen –
Maxwell Faraday Law
• Eddy currents establish secondary fields which oppose the primary
fields
• Changes net flux linkage and hence the impedance of the coil
• Anomalies in the test specimen affect the induced field, changing the
net impedance of the coil

Changes in the impedance of the Bprimary


probe coil constitutes the
eddy current signal

Bsecondary
Physical Principles
• Test coil characteristics
 Inductive reactance L  2 f L
 Ohmic resistance R
f : excitation frequency
L : self-inductance of the coil
Coil

Eddy Currents

Test Specimen
Eddy Current Inspection

Significant properties of test specimen


 Electrical conductivity ()
 Dimensions (such as depth of conducting plate)
 Magnetic permeability (µ)
 Material discontinuities (such as cracks or corrosions)

Significant instrument
characteristics
 Frequency of current excitation in the probe coil
 Size and shape of the probe coil
 Distance of test coil to the test specimen (lift off)
Transformer Analog
• Transformer
– Two circuits coupled inductively, in which a change of current in
one winding induces an electromotive force (or voltage) in a
second winding

Magnetic Coupling Between Two Coils


Equivalent Electrical Circuit for air core transformer
Eddy Current Inspection - Transformer Analog

When two coils are coupled by a magnetic field, they are subject to
the effects of a mutual inductance.

M 12  12  N 2 12 M 21  21  N1 21
I1 1 I2 I2
I
where

 M 21 mutual inductance of two coils (henrys)


M 12
 12 ( or  21 ) flux linkage created by flux of coil-1 passing through
the interior of coil-2 (reverse vise)
N1 ( or N 2 ) number of turns in coil-1 (coil-2)

I1 ( or I1 ) current flow in coil-1 (coil-2)

12 ( or mutual flux created by current in coil-1 and passing


21 )
though interior of coil-2 (reverse vise)
Transformer Analog (Cont’d.)

Some of magnetic flux lines associated with each coil do not couple
with both coils, those flux lines are called leakage magnetic flux
lines, and they contribute to self-inductance of the coil.

 
L1  11 N111 L2  22 N222
I1  I1 I2  I2
Transformer Analog (Cont’d.)
Faraday’s law of electromagnetic induction
An electromotive force (emf) is induced within an electric
circuit whenever the magnetic flux linking with the circuit
changes, and it is proportional to the time rate of change of flux
linkage.

d dI d
emf  L N
dt dt dt

emf  induced electromotive force (volts);


  total flux linkage (weber - turns);
t  time (seconds);
I  current (amperes);
L  self - inductance (henrys);
  coupled magnetic flux (webers); and
N  number of turns coupled with magnetic flux.
Transformer Analog
• If coil 1 serves as the primary winding, the voltage
induced in coil 2 is
e2  12 dI1  N 2 d12
M dt dt
• Likewise, if coil 2 is the primary, voltage induced in coil 1
is dI 2 d21
e1  21  N1
M
• Since  M 21 ,
M 12
N1  Turns Ratio TR
1 
e2
• When the secondary is short-circuited,
N2
N1 I1  N 2 I 2 I2 N1 
I1  N 2 TR
Transformer Action In EC Test Systems
• A practical eddy current test system can be considered
analogous to a transformer
– Excitation coil: Primary
– Test object: One-turn secondary
– Properties and dimensions of test object can be reflected in primary
or pickup coil voltages as a consequence of transformer action
• Let I 2 be the current flowing in the test material. Then,
component of primary voltage corresponding to
reaction is
secondary
2
V1  N 1
I 2 R2   N 1 
I1 R2
N2 N 
• Total voltage across test/probecoil2 
 2  dI 1
N 1
V1   R1    R2  I 1  1 dt
Coil Winding Resistance N2  Self inductance term
 L

Skin Depth

For a plane wave incident on an conducting half plane,


The variation of induced current with depth is given by -

J x  Jo exp(x  f where δ
1 is called
πfμσ
skin depth
 )
Factors Affecting Eddy Current
Measurements
• Factors affecting eddy current transducers
– Lift-off: Separation between the coil and the specimen surface
• Impedance of the coil changes as the probe is moved from air till it
touches the material surface – liftoff curve
• Minimized by the use of surface-riding probes or multifrequency
measurements
• Can be used to determine the thickness of non-conducting coatings on
conducting surfaces

– Skin effect
• Eddy currents decay exponentially with depth in the material
• Standard depth of penetration: (depth at which eddy currents become
1/e the surface value)
• This limits the sensitivity of eddy current method to the surface of the
conducting specimen
Part II - Eddy Current Sensors
Part II - Eddy Current Sensors

 Absolute probe
 Differential Bobbin probe
 Plus Point & Array Probe
 Meandering coil
 Eddy Current – Magneto-optic (MOI) sensor
 Eddy Current – Magneto-resistive (MR) sensor
Absolute Probe
• Absolute probe: a single coil is used for
impedance measurements

• Disadvantage
– Detects small changes in impedance, factors such as changes in
coil parameters or lift-off may mask this small signal.
Absolute Probe
Impedance plane trajectory of a coil over a ferromagnetic (or
nonferromagnetic ) specimen with and without a
discontinuity

1. Coil in air 1. Coil in air


2. Coil over a nonferromagnetic specimen 2. Coil over a ferromagnetic specimen
containing a discontinuity containing a discontinuity
3. Coil over a nonferromagnetic specimen 3. Coil over a ferromagnetic specimen
containing no discontinuities containing no discontinuities
Differential Bobbin Probe
• Differential bobbin probe
– Two bobbin coils with current in opposite direction
– Resulting signal is a difference signal from two coils
• Flaw signal is more distinguishable from
relatively constant background signal.
Differential Bobbin Probe
Impedance-plane trajectory of a coil over a
conducting nonferromagnetic test specimen

Impedance plane trajectory from a differential bobbin probe over a defect


Bobbin Probe
• Diff. mode: two identical bobbin coils, with fixed space
– Very sensitive to abrupt anomalies, such as pitting corrosion, and fretting
wear
– resistant to probe wobble, temperature variations, and gradual variations in
tube’s conductivity, diameter, and ovality

• Abs. mode: one testing bobbin coil, one shielded ref. coil
– gradually varying wall thinning, which could not be detected by diff. mode
Differential coil
Diff. sig.

FBH, OD Differential mode


20%-100%
Absolute coil Ref. coil
Abs. sig.

Axial
ZETEC Bobbin Probes
Absolute mode Metallic Shield
Bobbin Probe Disadvantages
• Merits:
– inexpensive, fast scanning( typ. up to 1m/s )
– Reliably detect and size volumetric flaws, such
as fretting wear and pitting corrosion
• Disadvantage:
– insensitive to circ. oriented flaws, because induced
eddy current parallel to flaws and not perturbed by
the flaws
– limited sensitivity at expansions, U-bend, and support
plates
– low resolution for flaw location and characterization
Rotating Probe - RPC
• Components of RPC:
– Typ. 4~8 surface-riding pancake coils placed around the circumference
– Driven motor, rotating circumferentially and moving forward (helical
pattern)

• Characteristics:
– Capable of both axial and circ. oriented flaws
– Very sensitive to cracking in transition zone

- 3 coils spaced 120 deg. Apart


- Each scan the inner surface of
the tube in a helical path
- multi-freq. : 400, 300, 200
kHz
- C-Scan impedance plot

C
ir
Rotating Probe - Plus Point

• Two orthogonal coils connected in diff.


mode crossing at a point
• Affect simultaneously by material and
geometric distortion, such as defects
• sensitive with circ. and axial flaws
• Rotate and move forward same as
RPC
• Slow scan speed
Array Probe - X-Probe

- 40 times faster than rotating


probes

C01: work both on transmit and receive mode


Make control circuit much complicated

X-Probe Diff. Bobbin


Combined with diff. bobbin coils
Planar Coils for Plates

Exciting coil: meander coil


Perpendicular to each
Receiving coil: mesh coil

S. Yamada, etc. 1995, Japan and Canada


Planar Coils for Plates

Lock-in Voltmeter output

Scan line perpendicular to crack


Scan line parallel to crack
Eddy Current – MOI Sensors

Aloha Airlines B-737-200 lost


part of its front fuselage during
a flight in Hawaii, 1985.

Eddy Current Excitation


Magneto-Optic Sensing
Imaging
Operational Principles of MOI

• Faraday rotation effect (1845)

K : Kundt's constant
 KMd M : Magnetization vector
d : path length of light
Magneto-Optic Sensors

Thin film of Bismuth-doped iron garnet grown on 3” diameter, 0.02” thick


substrate of gadolinium gallium garnet

1.Uniaxial magnetic anisotropy, i.e. they have an ‘easy’


axis of magnetization normal to the sensor surface and a
‘hard’ axis of magnetization in the plane of the sensor.

2. Memory, i.e. if the magnetization along the easy axis is


removed, the film will retain most of the established
magnetization.
3. Large Faraday Rotation, i.e. ~ ± 20,000~30,000
degrees/cm
Eddy Current Induction

Conventional methods rely on coils

Surface to be inspected

Magneto-Optic methods rely on


sheet current induction

Eddy current foil

Surface to be inspected
Magneto-Optic Imaging
Schematic of MOI Instrument

LIGHT SOURCE
ANALYZER
POLARIZER

BIAS COIL

SENSOR
LAP JOINT INDUCTION FOIL

• Eddy Current Excitation - f ~ 100Hz-100kHz


• Magneto-Optic Sensing
• Imaging
Multi-Directional (Rotating) Eddy Current
Excitation

Cracks
Linear null
region

MOI image using linear MOI image using rotating


excitation excitation
Advantages of MOI

• NDT systems are evolving towards imaging capability


• MOI systems produce real time analog images of inspected part
• Images both surface breaking and subsurface cracks
• Easy to interpret ( reduces operator fatigue)
• Eliminates need for removing paint or decal
• Can be used on conducting samples as well as composites by
“tagging” with ferromagnetic particles
• Allows rapid inspection of large areas for surface and subsurface
defects
• Use is straightforward requiring minimal training
• Output can be readily videotaped
Eddy Current – GMR Sensors

• A GMR device consists of two or more layers of ferromagnetic material (typically


NiFe, CoFe or related transition metal alloy) separated by ultra-thin non-
magnetic metal spacer layers (Cu, Au or Ru).
• In the absence of an externally applied field, the magnetic layers alternate in
magnetization, resulting in a high resistance. When a magnetic field is applied, this can
overcome the interlayer coupling and force all of the layers to align with the field and
reduce the resistance.
• Since a magnetic field in either direction will cause alignment of the magnetization, the
resulting R vs. H curve is an even function, symmetric about zero
GMR Sensor
• Giant Magneto-Resistive (GMR)
– Decrease in resistance in presence
of a magnetic field
– Uni-polar sensor
• Needs to be biased to measure
sinusoidal (+,-) fields
• GMR sensor kept in a coil
through which a constant DC
current is passed.
• DC field keeps the GMR sensor
biased during operation

Normal

Biased
Operating point

GMR Sensor Calibration and Bias


Source and Sensor Configuration
Uniform linear currents with a single line of symmetry
Alternate Sensors (GMR, MOI)

•Eddy Current Excitation


•Image the magnetic fields associated with the induced
currents

Anomaly free region Defect free rivet region Rivet with defect

Measurement
Region

Local Normal flux Local Normal flux


Tangential flux (Symmetric) (Asymmetric)
GMR Signals

Line of symmetry for placing detector


Results
• Schematic and parameters

Excitation current in y direction


Results
• Finite Element Geometry

Top View Side


30% Depth with 3mm squaVreiew
Results-GMR and EC (z component)

GMR Voltage EC Induced Voltage


Results-GMR and EC (x component)

GMR Voltage EC Induced Voltage


Results-GMR and EC (y component)

GMR Voltage EC Induced Voltage


Part III - Eddy Current Inspection Modes
Part III - Inspection Methods
• Single frequency inspection
– Single frequency sinusoidal excitation
– Magnitude and phase changes can be used to detect flaws
• Multifrequency inspection
– Multiple excitation frequencies multiplexed in time
– Used to detect (and locate) flaws throughout the depth of a specimen
– Can be used to suppress interfering signals that may mask defects
• Pulsed eddy current inspection
– Driving coil excited with a repetitive broadband pulse, such as a square
wave.
– Broad frequency spectrum is produced in one pulse: reflected signal
contains depth information
– Pulse is broadened and delayed as it travels deep into the conducting
material
• Flaws or other anomalies close to the surface will affect the eddy current
response earlier in time than deep flaws
• Remote Field Eddy Current inspection
– Low frequency excitation
Pulsed Eddy Current

1. Driving probe coil excited with a repetitive broadband pulse,


such as a square wave.
2. Broad frequency spectrum is produced in one
pulse: transient/reflected signal contains depth
information
3. Pulse is broadened and delayed as it travels deep into
the •conducting
Flaws or other anomalies close to the surface will affect the eddy
material
current response earlier in time than deep flaws

Motivation
Multilayer aircraft structure needs accurate characterization
of hidden corrosions and cracks
Pulsed Eddy Current

Pulsed Excitation

Top Layer

Bottom Layer
Typical Signals
A-scan : a transient signal measured under typical test conditions.
Amplitude and the time for zero-crossing is most important
features.

Amplitude of voltage measured Difference voltage between measured


by pick-up coil signal and reference signal
Typical Signals
C-scan : at a time instance, measured signal from a observation
plane is displayed as image.

Amplitude of voltage

Time for zero-crossing


Typical Signals
B-scan : Image of a linear cross section of C-scan within a time
period. X-axis represents the probe location, Y-axis represents the
time, amplitude of signal is displayed in gray value.
Remote Field Eddy Current Testing

• Originally developed for inspecting tubular structures


• Is sensitive to both ID and OD defects
• The signal is sensed by a pick-up coil located in the
remote field region
Indirect Path

Pipe Wall

Excitation
Direct Path
Coil

Remote Indirect Path


Field
Sensor
Remote Field Eddy Current Testing
• Direct and indirect
fields interaction
– Direct field energy
dominant in the
near field zone
– Indirect field
dominant in the
remote field zone
– In the transitional
zone DF and IF
may cancel out
Remote Field Probe

Receiver
Active or Passive ShiTerladnsmitter

Test Specimen
IMTT Probe Response Y. Sun, at al

3.0”  3.0”  0.040” Flat-bottom corrosion 0.5”  0.5”  0.040” Flat-bottom corrosion
0.603” Below Surface 0.603” Below Surface

Corrosion Size Estimate Corrosion Size Estimate

.
Part IV - Forward Modeling
What Simulation Models Can Do
• Simulation Models are useful in
– Solution of forward problem – Predict EC probe signals
– Effect of probe wobble, frequency, sludge characteristics on probe
measurements (POD)
– Visualization of field/flaw interaction
– Optimization of sensor/system design
– Test bed for generating defect signatures
– Useful in Probability of Detection (POD) Models at low cost
– Inverse problem solution (Reverse engineering models for finding root
cause)
• Key Advantages of Simulation Model
– Provides an inexpensive and fast method to simulate realistic test and
defect geometries
Part IV- Forward Modeling
• Maxwell's equations: Constitutive relations for linear and isotropic media:

 E   B
t D  E J
B H
E

B0
is the permittivity
D is the permeability
H, E are the magnetic & electric is the
field strengths conductivity
B, D are the magnetic & electric
flux densities
J, are the current & charge
densities
Finite Element Modeling -Technical Details
The governing equations for the eddy current excitation in
terms of the magnetic vector potential A and electrical
scalar potential V

1  
   ( A )  j ( A  V )  0 1

in   in  1
  ( j ( A  V ))
0 1  in  2
   ( A )  S

J

, are the permeability and conductivity of the media.


Finite Element Formulation
Step 1. Mesh generation - Discretize the solution region
into finite elements –
 Nodes are numbered globally and locally.
 A connectivity array is constructed to describe the relationship
between the elements and nodes.

Step 2. Choose shape functions. (hexahedral element)


Ni ( ,, )  (1  i )(1  i)(1  i=1,2, . . ., 8
 i )
Finite Element Formulation (continued)

Step 3. Compute stiffness matrix and load vector for each element.
 Notation
8 24
A   N j A xxj Nj A yyj  Nj A zzj  N k Ae
k j 1
k 1

8
V   Nj V e
j j
1

 Axje k3j
N jx k3j  e
 2
N k  2N j y A ke   yj
 A e k3j
k  3 j 1
 Azj 1
N j z k3j
k3j
Finite Element Formulation (continued)
24


j e
1
  i  N   N )(  N
j 1 i j dV }Aje 
1
24
 N
 8

 jNi  Nj dV }A j  { e N N dV }V 
e e
j i j j

j { e
i  1,2,,24
1 1

N i  (n   A)dS  
1 1 N i  J SdV
N i  (  A  n)dS
 e

 
 e


e

 
24 8

 jNk  Nj dV }A    Nk N j dV }V


e e


j j
j { e
j { e

1 1  k  1,2,,8
Nk( jA  V )  ndS 
 e
0

[G]e [ A]e  [Q]e
3232 321
321
Finite Element Formulation (continued)

Step 4. Assemble element matrices to global matrix.


[G][ A]  [Q]

Step 5. Boundary conditions.


 Dirichlet boundary (
A , V)
 Neumann boundary ( , H)
B
Finite Element Formulation (continued)

Step 6. Solve the matrix equation.


 Direct solver
 Iterative solver (Transpose Free Quasi-Minimal-Residual method)

Step 7. Calculate other measured quantities.

B   - magnetic flux density


A
J   jA  - electric current density
V
Model Validation
(Palanisamy, 1980)

Resistive

Reactive Impedance Plane

Comparison of Experimental and Predicted Results for An OD Defect


Model Validation
(Palanisamy, 1980)

Resistive

Impedance Plane
Reactive

Comparison of Experimental and Predicted Results for a Tube Support


Model Validation
(Palanisamy, 1980)

Modeling

Geometry

Experimental
Comparison of Experimental and Predicted Results for Defects in a support
SGTSIM v1.0 Features - 2008

Predefined geometries
Support plate
Tube sheet
Free span

Probe s: Bobbin ( absolute, differential, air core, ferrite core),


Pancake coil
Defects: Rectangular : ID, OD
Experimentally Validated
SGTSIM v 2.0 Features –2009

• Model Enhancements
Probes: + Point Probe
• GUI Enhancements
– New graphical interface:
– 2D and 3D surface plot and Lissajous plots
– Manual Calibration
– Data Writing formats

vertical
horizontal vertical horizontal

Simulation
Measured
Real Crack Model - Quantitative Validation –
+ point probe ; 300KHz ; ETSS Data file – Farley-1_25_51

Axial Notch Profile from Experimental signals


MET data - mesh

Average error - 5.31% Simulated signals Average error – 2.17%

Vertical Channel Horizontal Chan


Real Crack Model - Experimental Validation –
+ point probe ; 300KHz ; ETSS Data file - TMI-1_91_55

Axial Notch Profile


from MET data
Experimental signals

Average error – 6.28% Average error – 1.19%

Simulated signals

Vertical Channel Horizontal Cha


Part V - Inverse Problems in EC NDE
Overall Analysis Procedure
Raw EC data

Preprocessing

Features

Signal Classification

No Degradation Degradation

Compensation

Defect Characterization
Automated Signal Classification (ASC)

Raw UT Weld Inspection Signal

Preprocessing
• Noise Filtering
• Invariance
• Feature Extraction

Feature
Vector
(DWT
Coefficie
Training • Clustering Algorithms
nts)
• Neural Networks
Classifier
Signal Class
Model-based Inversion for Defect
Characterization

Initial Measured
Defect Profile Probe Signal

Forward Compare with


Predicted Defect
Model desired
Response probe
response Profile

Update the defect


profile
Characterization

Defect Parameterization

y r

d1 d3 d2

x z
Characterization

Typical Results

(i) (ii)
11.4
11.4 0.25

11.2
11.2
0.2
11

Objective Function
10.8
0.15
)
(mm )

(mm 10.6
10.6
r
r

0.1
10.4

10.2
10.2 : desi red profile
: desired profile
initial guess
guess 0.05
:: initial
10 : reconstructed profile
10 : reconstructed
profile
9.8 0
0 0 1 1 2 2 3 3 4 4 5 5 6 0 5 10 15
z (mm) 6 Iteration Number
z (mm)

d1  d2  55% d 3  60%
50%
Characterization

Typical Results

(i) (ii)
11.4 0.25

11.2
0.2
11

Objective Function
10.8
0.15
r (mm)

10.6

0.1
10.4

10.2 : desired profile


: initial guess 0.05
10 : reconstructed profile

9.8 0
0 1 2 3 4 5 6 0 5 10 15
z (mm) Iteration Number

d1  d2  55% d 3  45%
50%
Characterization

Test Configuration

d1 d 2 … d8

z
Characterization

Typical Results

N=100
Characterization

Typical Results

N=100
Case study: Steam Generator Tube Inspection
Rotating probe coil EC Data
Free Span Drilled Support Broached
Support

Freespan

U bend Tube Sheet Tube Expansion Transition

** Plus Point coil, 300kHz


channel
Filtering

Noisy Data

After Filtering

Volumetric Indications

8
CWT Analysis – RPC Data

The continuous wavelet transform of a signal u(t) is given by




 ( , s)



u(t)
1
 * ((t   ) /
s)dt s
where τ and s represents translation and scale of the mother wavelet ψ(t)
Property of wavelets to perform multi-scale analysis is exploited for flaw detection

Flaw Structur Noise


e

8
CWT – Flaw Detection

• A simple threshold based on statistics of CWT coefficients easily separates out flaw signal
from noise

1D RPC Data CWT Potential Flaw Signals

Flaw

Flaw

8
Compensation: Deconvolution
• Objective
• Development of deconvolution algorithms for removing the response
of the eddy current probe area from the measurements
– Compensation for the finite dimensions of eddy current
probes
• Approach
• Deconvolution algorithm using Wiener filter
• System model (time domain)
• yi, j   d i, j  hi, j ni, j 
• where y(i,j) – measured signal
• d(i,j) – defect footprint
h(i,j) – system impulse response (probe footprint)
n(i,j) – noise
* – convolution
Compensation

Typical Results
2 2 2

4 4 4

6 6 6

8 8 8

10 10 10

12 12 12

14 14 14
16 16
16

18 18
18

20 20
20

5 10 15 20 30 35 5 10 15 20 25 30 35
5 10 15 20 25 35
25
30

True Defect image


Measured data Result of deconvolution

The defect diameter is


• True defect image: 0.06”
• Before deconvolution: 0.1”
• After deconvolution: 0.061”

Transfer function (kernel)


Defect Characterization
• Determine defect parameters from measurement
signal
– Calibration Methods, Neural Networks

6000

4000

2000

-2000
30
Mapping
25
20
25
15 20
10 15
10
5
axial 5
0 0 circum

Eddy Current Signal Defect Profile


Overall Profiling Procedure

CALIBRATED DATA

ROI SELECTION

LENGTH ESTIMATION

METHOD 1: METHOD 2: METHOD 3:


PROFILING USING PROFILING USING PROFILING USING
CLASSICAL ENHANCED RADIAL BASIS
CALIBRATION CALIBRATION NEURAL
CURVE CURVE NETWORK
Length estimation
Depth Profiling
radial basis neural network method

.
.
.
fi, 300kHz
Axial direction .
.
.

.
.
.
fi, 200kHz Trained RBF depthi
. Feature Matrix
Network
.
.

.
.
.
fi, 100kHz
.
. Features from line scans in the
.
background region were also
used to map to zero depth
Radial Basis Function (rbf) Approach

The input-output transformation equation for the RBFNN can be expressed as


 p  
f ( y)  w (|| y  t ||, )

i1 i i i
where = input vector of dimension N,
f( ) = output vector of dimension
M,
= ith basis center,

wi = weight vector of dimension corresponding to the ith center and

= radially symmetric basis function with spread σ . total

of p centers (or nodes in hidden layer) are used in basis function expansion
Feature vector
• For 300kHz, 200kHz and 100kHz data, the maximum magnitude and its corresponding phase angle
is computed for each gradation in axial direction of the ROI.
• The phase spread. which is an additional feature, defined as the range for as three phase angles
computed per gradation
• Seven features are thus obtained per gradation in axial direction

Polar plot of raster in 100kHz, 200kHz, 300kHz channel


90
25
120 60
20

15
150 30
10
f1
5
f2
f  100 , 200 , , m m m 
. 180 0 300 100 200 300 100,300
.
.
 
.
f 210 330
N

240 300
300
270
kHz
200
kHz
100 kHz

9
Radial basis neural network method

1
2
TRAINING MODULE
Length
.
. Estimation Nx7
. (N Axial scans) Feature Matrix
.
N

Training data MET Result

R 1
2
.
B .
.
.
F
N .
1
E .
2 Length Test Data .
.
Estimation Feature T .
.
. (M Axial Matrix W .
M scans) O .
.
R .
K M
TEST MODULE
Variation of basic rbf algorithm (rbf2)

 Three feature vectors corresponding to three spatially contiguous line scans are mapped to depth at center
scan
Training data MET Result TRAINING MODULE

Axial direction fn-1


fn
fn+
} NX21
Feature
Matrix
depthn

F
N
fn-1
Feature vector E
fn Predicted depth at nth scan
T
fn+
W
1
O
TEST MODULE R
K
Results 1
Defect Length & Depth profiles

Max % TW PDA (%) Flaw Length (in)

Log- Mag 26.00 21.69 0.56


Mag 28.53 22.89 0.52
NN (RBF1) 49.23 32.74 0.52
NN2 (RBF2) 56.80 44.52 0.52

MET 60.00 40.97 0.53

RESULTS - profiling

100

90

80
log-mag
70
mag
60 NN
MET
%TW

50

40 NN2
30

20

10

0
0 0.1 0.2 0.3 0.4 0.5 0.6 0.7

Location
Results 2
Defect Length & Depth profiles

Max % TW PDA (%) Flaw Length (in)


Log- Mag 62.00 38.91 0.67
Mag 60.14 39.26 0.67
NN (RBF1) 57.39 33.49 0.69
NN2 (RBF2) 64.17 42.25 0.69

MET 77.00 42.01 0.73


RESULTS - profiling

100

90

80
log-mag
70
mag
60 NN2
NN
%TW

50
MET
40

30

20

10

0
0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8

Location
Results 3
Defect Length & Depth profiles
Max % TW PDA (%) Flaw Length (in)

Log- Mag 44.00 43.00 0.13


Mag 46.25 43.07 0.13
NN (RBF1) 34.47 26.68 0.14
NN2 (RBF2) 31.15 26.73 0.14
MET 35.80 24.22 0.11

RESULTS - profiling

100

90

80
log-
70
mag
60 mag
NN2
NN
%TW

50
ME
40 T

30

20

10

0
0 0.02 0.04 0.06 0.08 0.1 0.12 0.14 0.16

Locatio
n
Results 4
Defect Length & Depth profiles
Max % TW PDA (%) Flaw Length (in)
Log- Mag 22.53 10.35 0.94
Mag 9.00 6.79 0.94
NN (RBF1) 76.85 43.11 0.96
NN2 (RBF2) 64.17 42.25 0.88
MET 77.00 42.01 0.82

RESULTS - profiling

100

90

80
log-mag
70
mag
60 ME
T
%TW

50
NN2
40 NN

30

20

10

0
0 0.2 0.4 0.6 0.8 1 1.2

Location
Summary

Eddy Current - Physical Principles


- Transformer Analogy
- Probe Coil geometry
- Continuous, Pulsed &
Remote excitation
- Simulation models
- Data Analysis
- Defect Classification
Defect Profiling
- Application (SG tube
Inspection)
Questions?

You might also like