Escolar Documentos
Profissional Documentos
Cultura Documentos
EURASIP Journal on
Advances in Signal Processing
Editor-in-Chief
Phillip Regalia, Institut National des Telecommunications, France
Associate Editors
Adel M. Alimi, Tunisia
Yasar Becerikli, Turkey
Kostas Berberidis, Greece
Enrico Capobianco, Italy
A. Enis Cetin, Turkey
Jonathon Chambers, UK
Mei-Juan Chen, Taiwan
Liang-Gee Chen, Taiwan
Satya Dharanipragada, USA
Kutluyil Dogancay, Australia
Florent Dupont, France
Frank Ehlers, Italy
Sharon Gannot, Israel
Samanwoy Ghosh-Dastidar, USA
Norbert Goertz, Austria
M. Greco, Italy
Irene Y. H. Gu, Sweden
Fredrik Gustafsson, Sweden
Sangjin Hong, USA
Jiri Jan, Czech Republic
Magnus Jansson, Sweden
S. Jayaweera, USA
Soren Holdt Jensen, Denmark
Contents
Signal Processing in Advanced Nondestructive Materials Inspection, Joao Manuel R. S. Tavares and
Joao Marcos A. Rebello
Volume 2010, Article ID 954623, 2 pages
Geometrical Feature Extraction from Ultrasonic Time Frequency Responses: An Application to
A Study of Concrete Hydration and Dielectric Relaxation Mechanism Using Ground Penetrating Radar
and Short-Time Fourier Transform, W. L. Lai, T. Kind, and H. Wiggenhauser
Volume 2010, Article ID 317216, 14 pages
Strain and Cracking Surveillance in Engineered Cementitious Composites by Piezoresistive Properties,
Jia Huan Yu and Tsung Chan Hou
Volume 2010, Article ID 402597, 6 pages
Heuristic Enhancement of Magneto-Optical Images for NDE, Matteo Cacciola, Giuseppe Megali,
` Salvatore Calcagno, Mario Versaci, and Francesco Carlo Morabito
Diego Pellicano,
Volume 2010, Article ID 485695, 11 pages
A Machine Learning Approach for Locating Acoustic Emission, N. F. Ince, Chu-Shu Kao, M. Kaveh,
A. Tewfik, and J. F. Labuz
Volume 2010, Article ID 895486, 14 pages
Editorial
Signal Processing in Advanced Nondestructive Materials
Inspection
Joao Manuel R. S. Tavares1 and Joao Marcos A. Rebello2
1 Department
of Mechanical Engineering, Faculty of Engineering, University of Porto, Rua Dr. Roberto Frias, 4200-465 Porto, Portugal
of Metallurgy and Materials, Faculty of Engineering, Federal University of Rio de Janeiro, Technological Center, Room
F-210, Ilha do Fundao, RJ, Brazil
2 Department
Acknowledgments
The guest editors would like to express their deep gratitude to
the Editor-in-Chief and Associate Editors of EURASIP Journal on Advances in Signal Processing for this opportunity, to
all authors who shared their excellent works with us, and to
all members of the Scientific Committee of this special issue,
helped us in the review process.
Joao Manuel R. S. Tavares
Joao Marcos A. Rebello
Research Article
Geometrical Feature Extraction from Ultrasonic Time Frequency
Responses: An Application to Nondestructive Testing of Materials
1 Ramon
1. Introduction
Signal processing is an essential tool in nondestructive material characterization. Modern technologies can take benefit
of more sophisticated algorithms allowing to classify and
characterize materials precisely. One of the techniques that
takes advantage of all these advances is the nondestructive
testing (NDT) using ultrasounds. Thanks to the advances in
signal processing it is now easy to find applications of NDT
using ultrasonics in materials, that some years ago was very
hard to find [13].
The Signal Processing Group (GTS) of the Universidad
Politecnica de Valencia published a technique [2] that allows
to characterize dispersive materials by means of pulseecho inspection with ultrasonic energy. The aforementioned
technique was based on extracting time of flight-dependent
parameters from the ultrasonic A-scan. This technique
S() = A e(c ) /B
(1)
S(, z) = A e(c ) /B e0
2
y z
/B 2 (z)
(2)
.
c
,
1 + 0 z B2
B (z) =
B
1 + 0 z
B2
(3)
.
(4)
(5)
2
(z)
eS0 D
2 z
(6)
/B 2 (z)
The new parameters A (z), c (z), and B (z) that take
into account attenuation due to Stochastic scattering can be
derived by equations (7), (8), and (9)
c
,
1 + S0 D z B 2
c (z) =
B (z) =
B
1 + S0 D z B 2
2
(7)
(8)
(9)
3. Conventional Parameter
Extraction for Material Characterization:
The Ultrasonic Signature (US) Concept
As already mentioned, information about the material is
included in the A-scan. Among some other possibilities to
extract information about the materials [5, 7], the analysis
of the variant impulse response (or equivalently the variant
frequency response) of the LTV is a feasible alternative. The
time-variant characteristic of the model leads naturally to a
nonstationary analysis of the recorded signal.
This technique proposes the use of a short-term frequency analysis of the signal to isolate the evolution of
the dierent frequency components. This can be done
by means of explicit implementation of a bank of filters
or, more usually, by means of some type of linear or
nonlinear time-frequency transformation, including nonconstant bandwidth analysis like wavelet transform. From
the time-frequency signal we obtain the US which is a
one-dimensional signal hopefully encompassing the relevant
106
106
2.5
2.5
Pulse central frequency and bandwidth (Hz)
1.5
0.5
0.5
1.5
0.5
1.5
103
Depth (m)
0.5
1
Depth (m)
(a)
1.5
103
(b)
106
2.5
1.5
0.5
0.5
1
Depth (m)
1.5
103
(c)
Figure 1: Simulations of the proposed model for Stochastic scattering attenuation. Simulation parameters of the ultrasonic pulse were
central frequency fc = 500 KHz, fractional bandwidth B = 75%, and mean scatterer size D = {0.5, 1 and 1.5 mm}. (a) D = 0.5 mm, (b) D =
1 mm, and (c) D = 1.5 mm.
c (z) =
|S(, z)| d
,
1 |S(, z)| d
2
(10)
where |S(, z)| is the magnitude of the timefrequency transformation, and [1 , 2 ] defines the
integration band,
(ii) The fractional bandwidth:
BW3 dB (z)
100%
BW% (z) =
2 c (z)
(11)
(12)
4. An Alternative Technique to
Conventional Parameter Extraction for
Material Characterization
2-D shape analysis can be applied to the TFR of the ultrasonic
A-scans for material characterization. The motivation of this
idea was based on the observation that the mean scatterer
106
2.5
Frequency (Hz)
1.5
I(, z) = rect
1
0.5
10
Depth (m)
15
104
(13)
Area I , z, S0 , D
=
c (z)
.
B (z)
=
=
z=z0
z=0
rect
dw dz
B (z)
(14)
B (z)dz
2
S0 D B
1 + S0 D z0 B2 1 .
z p q I(, z),
(15)
106
106
1.8
1.6
Centroid
1.4
1.2
Pulsation ()
Pulsation ()
5
6
Depth (z)
9
103
Centroids
1.8
1.6
1.4
1.2
5
6
Depth (z)
103
Figure 4: Centroid estimation using decomposition of the binarized TFR into small rectangles.
m10
,
m00
c =
m01
.
m00
(16)
1
(z cz ) p ( c )q I(, z).
m00
(17)
211
1
.
= arctan
2
20 02
(18)
b
a
(19)
4.1.5. Boundary Signature (BS). The BS is a 1-D representation of an object boundary. One of the most simple ways
to generate the BS of a region is to plot the distance from
the center of gravity of the region to the boundary as a
function of the angle, . Figure 3 illustrates this concept.
The changes in size of binarized TFR result in changes in
the amplitude values of the corresponding BS. It is expected
that the higher the value of D is, the lower the amplitude of
the corresponding BS. Moreover, the BS not only provides
information about area changes but also provides the angular
direction of such changes. To compute the BS we need to
compute for each angle, , the Euclidean distance between
the center of gravity and the boundary of the region. As will
be demonstrated, it is expected that dierent values of D,
for the model or material under test, will correspond with
dierent BSs for the binarized TFR.
4.1.6. Frequency-Derived Parameters. Some frequencyderived parameters have been also tested such as centroid
frequency, central frequency, or bandwidth.
To compute the central frequency we work with the TFR
in gray scale (not binarized). We divide the TFR diagram into
small rectangles (see Figure 4) along the z-axis (horizontal),
and then we compute the maximum of each rectangle.
The final result is the evolution of central frequency along
depth.
To compute the centroid frequency evolution with depth
we divide the binzarized TFR into small rectangles along the
z-axis and then we compute the center of gravity (described
above) for every rectangle, the final result is the evolution of
centroid frequency along horizontal direction.
The process to compute the bandwidth evolution is
similar to centroid frequency computation. In the case of
bandwidth the width of each rectangle is computed, thus
obtaining the evolution of bandwidth with depth.
100
200
300
400
500
Grain microstructure of the material
K K type distribution of M parameter
ARMA (4,4) pulse model
c (z)
B (z)
A (z)
Observation noise
n AWGN
n} = 0
n2 , E{
5. Experiments
5.1. Application of Conventional and Geometrical Descriptors
to Simulated Signals. Simulated signals have been generated
according to the model presented in Section 2.
Transducer response (ARMA order) was estimated using
phantom data based on the final prediction error and
residual time series methods [10]. The best results for
the employed transducer that will be later used in this
section, were achieved with an ARMA(4, 4) model. The LTV
system was modeled according to (20) with the expressions
for A (z), c (z) and B (z) given by (7), (8) and (9),
respectively.
y(n) = b0 (z) x(n) + b1 (z)x(n 1) + b2 (z)x(n 2)
+ b3 (z)x(n 3) + b4 (z)x(n 4) a1 (z)y(n 1)
a2 (z)y(n 2) a3 (z)y(n 3) a4 (z)y(n 4).
(20)
The block diagram of the simulated signal generator
is represented in Figure 5. Several signals (A-scans) were
generated using this model. The purpose of these simulations
was to compare conventional signal-processing estimators
with the geometrical estimators described in Section 4.1. The
results are presented in the form of bias/variance graphs
for each estimator as the amount of observation noise
increases (AWGN). The variance is presented in the graphs
with vertical color bars whereas the bias is presented with
a convenient marker to distinguish between conventional or
geometrical estimators.
Figure 6 was generated using D = 0.5 mm and AWGN
variance varying from 0.05 to 0.5. The figure shows the
variance (red bars) and bias (marker ) of conventional
estimators: central frequency, centroid frequency, and fractional bandwidth, as described in Section 3. Superimposed,
Figure 6, also represents the variance (green bars) and bias
(marker o) of shape analysis operators: central frequency,
105
Central frequency
4
3
2.5
1
0
1.5
1
0.5
0.1
0.2
0.3
AWGN
Centroid frequency
Bias/variance
Bias/variance
0.4
0.5
0.5
0.1
0.2
(a)
0.3
AWGN
0.4
0.5
(b)
Fractional BW
15
Bias/variance
10
10
0.1
0.2
0.3
AWGN
0.4
0.5
Conventional estimator
Geometrical estimator
(c)
Figure 6: Central Frequency, Centroid Frequency, and Fractional BandWidth operators variance and deviation from the theoretical result
(bias). Simulation parameters for LTV were mean scatterer diameter D = 0.5 mm, AWGN of variance varying from 0.05 to 0.5. Bias (marker
or o), and variance (vertical bars).
Agar-Agar
concentration
Ner of scatterers
mean D
2% in distilled
water
2% in distilled
water
2% in distilled
water
2% in distilled
water
1000 molecular
sieves
1000 molecular
sieves
1000 molecular
sieves
1000 molecular
sieves
0.5 mm
0.7 mm
1.3 mm
1.8 mm
105
Central frequency
Centroid frequency
4
3.5
3
2.5
Bias/variance
Bias/variance
2
1
0
1.5
1
0.5
2
3
0.5
0.1
0.2
0.3
AWGN
0.4
0.5
0.1
0.2
(a)
0.3
AWGN
0.4
0.5
(b)
Fractional BW
15
Bias/variance
10
10
0.1
0.2
0.3
AWGN
0.4
0.5
Conventional estimator
Geometrical estimator
(c)
Figure 7: Central Frequency, Centroid Frequency, and Fractional BandWidth operators variance and deviation from the theoretical result.
Simulation parameters for LTV were mean scatterer diameter D = 1.5 mm, AWGN of variance varying from 0.05 to 0.5. Bias (marker or
o), and variance (vertical bars).
The measurement equipment was a PC with an ultrasonic board IPR-100 (Physical Acoustics) working in pulseecho mode with 400 V of attack voltage, 40 dB in the
receiver amplifier, and damping impedance of 2000 Ohms.
The transducer frequency was chosen to be 1 MHz (K1SC
transducer probe from Krautkramer and Branson). Received
signal was acquired with the Tektronix 3000 oscilloscope ( fs
= 50 MSamples/s).
The set of 8 test pieces was separated in two subsets: the
odd subset (composed by test pieces 1, 3, 5, and 7) and the
even subset (composed by test pieces 2, 4, 6, and 8). Both subsets were measured separately and individual estimators were
computed and compared between subsets. The measurement
9
106
2.5
1.5
1
1000
1500
2000
Test Piece
1 and 2
3 and 4
5 and 6
7 and 8
Area
651
383
258
279
Orientation
0.0953
0.0344
0.3297
0.1041
2500
3000
Time (samples)
3500
4000
4500
Eccentricity
0.8762
0.8329
0.6965
0.6423
(a)
106
Table 2 shows area, orientation, and eccentricity parameters obtained from the test pieces created in the experiment.
The area values obtained in Table 2 agree with the
expected behavior described in Section 4. It can be noticed
that higher scatterer sizes get lower value of the area
descriptor. This trend is coarsely maintained among all
scatterer diameter sizes (D).
The orientation parameter values presented in Table 2
also agree with the expected behavior described by (7) since it
predicts a downshift in the TFR shape (see Figure 1). Physical
explanation is based on the fact that the higher the value
of D, the higher the attenuation of the ultrasonic energy at
high frequencies. As a result of that, higher D values get
higher negative slope (with respect to horizontal axis). The
orientation parameter allows to distinguish coarsely between
small scatterer test pieces (D = 0.5 and 0.7 mm) and large
(D = 1.3 and 1.8 mm).
However, there are geometrical parameters that allow a
precise classification of test pieces according to D: eccentricity, centroid frequency, and BS are the main ones.
The eccentricity parameter values presented in Table 2
show that the higher D the lower the eccentricity value is.
This behavior agrees with theoretical equations and allows to
classify all the test pieces.
Figure 9 represents the centroid frequency evolution with
depth (time of flight). The parameter has been estimated
using both techniques presented. The left figure was obtained
using the conventional estimator (see (10)) and the right figure was obtained using the geometrical estimator (Figure 4).
Results were averaged for both subsets (even and odd). Both
estimators should give results of the same order of magnitude
as it can be verified. However, as the ultrasonic pulse travels
deep into the agar-agar matrix (increasing time axes) it
2.5
1.5
1
1000
1500
2000
2500
3000
Time (samples)
3500
4000
4500
10
Signature
80
70
Amplitude
60
50
40
30
20
10
0
50
100
150
200
Angle
250
300
350
6. Conclusions
In this paper we show that parameters extracted from
the TFR of ultrasonic A-scans can be used for material
characterization/classification. The novelty of this work is
based on the use of TFRs as input information in 2D-shape
analysis algorithms, specifically geometrical descriptors. This
technique compliments traditional classification parameters
(attenuation, longitudinal ultrasonic velocity, etc.) with
shape-related parameters. Additionally, for some parameters,
the new technique allows to obtain lower variance estimators.
When binarized TFRs are processed and 2-D geometrical
modeling, inherent in our approach, is used, a new set
of estimators can be derived. The proposed geometrical
estimators can provide better estimates and moreover, they
are less sensitive to noise than conventional estimators.
Thanks to this superior performance, in terms of bias
Acknowledgment
This work was supported by the national R + D program
under Grant TEC2008-02975 (Spain), FEDER programme
and Generalitat Valenciana PROMETEO 2010/040.
References
[1] M. Edwards, Ed., Detecting Foreign Bodies in Food, Woodhead,
Cambridge, UK; CRC Press, Boca Raton, Fla, USA, 2004.
[2] L. Vergara, J. Gosalbez, J. V. Fuente, R. Miralles, and I. Bosch,
Measurement of cement porosity by centroid frequency
profiles of ultrasonic grain noise, Signal Processing, vol. 84,
no. 12, pp. 23152324, 2004.
[3] J. Gosalbez, A. Salazar, I. Bosch, R. Miralles, and L. Vergara, Application of ultrasonic nondestructive testing to the
diagnosis of consolidation of a restored dome, Materials
Evaluation, vol. 64, no. 5, pp. 492497, 2006.
[4] P. He, Simulation of ultrasound pulse propagation in lossy
media obeying a frequency power law, IEEE Transactions on
Ultrasonics, Ferroelectrics, and Frequency Control, vol. 45, no.
1, pp. 114125, 1998.
[5] R. Demirli and J. Saniie, Model-based estimation of ultrasonic echoespart I: analysis and algorithms, IEEE Transactions on Ultrasonics, Ferroelectrics, and Frequency Control, vol.
48, no. 3, pp. 787802, 2001.
[6] M. Karaoguz, N. Bilgutay, and B. Onaral, Modeling of
scattering dominated ultrasonic attenuation using power-law
function, in Proceedings of the IEEE Ultrasonics Symposium,
vol. 1, pp. 793796, October 2000.
[7] L. Vergara, J. Gosalbez, J. V. Fuente, et al., Ultrasonic
nondestructive testing on marble rock blocks, Materials
Evaluation, vol. 62, no. 1, pp. 7378, 2004.
[8] I. Pitas, Digital Image Processing Algorithms and Applications,
Wiley-Interscience, New York, NY, USA, 1st edition, 2000.
[9] R. C. Gonzalez and R. E. Woods, Digital Image Processing,
Prentice-Hall, Englewood Clis, NJ, USA, 2007.
[10] A. K. Nandi, Blind Estimation Using Higher-Order Statistics,
Kluwer Academic Publishers, Boston, Mass, USA, 1999.
Research Article
Cyclic Biaxial Stress Measurement Method Using the Grain
Growth Direction in Electrodeposited Copper Foil
Yuichi Ono, Cheng Li, and Daisuke Hino
Department of Mechanical and Aerospace Engineering, Tottori University, 4-101 Koyama-cho minami, Tottori-shi,
Tottori 680-8552, Japan
Correspondence should be addressed to Yuichi Ono, ono@mech.tottori-u.ac.jp
Received 28 December 2009; Accepted 11 April 2010
Academic Editor: Joao Marcos A. Rebello
Copyright 2010 Yuichi Ono et al. This is an open access article distributed under the Creative Commons Attribution License,
which permits unrestricted use, distribution, and reproduction in any medium, provided the original work is properly cited.
A method that uses grain growth direction in electrodeposited copper foil to measure cyclic biaxial stress is examined in this paper.
The grain growth direction is measured by image processing software after a cyclic loading test for various biaxial stress ratios is
carried out. Since the grain growth occurs in two directions and its directions correspond closely with the direction of maximum
shearing stress when the biaxial stress ratio is negative, the principal stress can be measured using Mohrs stress circle. On the
other hand, when the biaxial stress ratio is positive, above-mentioned feature does not occur. Therefore, the first principal stress
can be measured based on the grain growth density. The number of grains necessary to measure the biaxial stress is estimated by a
statistical approach.
1. Introduction
The copper electroplating method is used to measure cyclic
stress that causes metal fatigue [13]. If copper foil adhered
to a machine element is subjected to repeated loads, grain
growth occurs in the copper foil. Since the grain growth
density is controlled by the maximum shearing stress and
the number of cycles, the maximum shearing stress can
be measured based on the grain growth density in the
prescribed number of cycles [4]. This method has the
advantage of detecting stress in microscopic regions like the
stress concentration region. Moreover, this method can be
easily applied to rotating machines and machine elements in
sealed casings, since it does not need an output line like an
electrical resistance strain gauge.
Since the principal stresses that are important for
evaluating metal fatigue cannot be detected by this method,
a new method using copper foil with circular holes has been
developed [4, 5]. However, this new method is somewhat
complex, because the grain growth length at hole edges as
well as the grain growth density in the copper foil must be
measured. This also means that two kinds of copper foils
(foil with and without circular holes) are necessary for the
principal stress measurement.
max
max
3 = 0
max
max
(a) C 0
(b) C 0
2
max ,
1C
1 = 2max ,
2 = C1
2 = C1
(C 0),
(C 0).
(1)
(2)
Elongation [%]
15.2
(3)
(4)
3. Experimental Procedures
3.1. Test Specimen and Testing Machine. A copper foil was
obtained as follows. A stainless steel plate (200 mm
100 mm 1 mm) was electroplated with copper sulfate
solution [1]. Since the stainless steel plate is polished by
bung before plating, the deposited layer can easily strip
from the stainless steel plate. This deposited layer is called
a copper foil. All subsequent experiments were carried out
by cutting this single foil to small pieces. The copper foil was
about 20 m thick and the initial grain size was about 1 m
[4]. This grain size is considerably smaller than the grown
grain size.
A titanium alloy (Ti-6Al-4V) was used as the specimen
material. The mechanical properties are shown in Table 1.
3 (z)
3 (z)
max
2 (y)
1
(x)
max
2 (y)
1
(x)
max
max
90
Grown grain
90
Copper foil
(a) C 0
(b) C 0
MB
max
xy max
MT
13
xy
MB
94
110
x
MT
max xy
max
xy
30
Copper foil
4-6
t = 4 mm
(a) Plate-type (C < 0)
MB : bending moment
MT : torsional moment
8-6.5
max
Copper foil
xy
P.C.D 100
2
x
2
xy
120
t = 4 mm
(b) Disk-type (C > 0)
Figure 5 shows the geometry and dimensions of test specimen. A plate-type specimen was used for C < 0, and a
disk-type specimen was used for C > 0. A Schenck type
fatigue testing machine was used, and tests were performed
at a frequency of 60 Hz by mounting the bending-torsion
Specimen
MB
Machine type
Schenck
Servo-hydraulic
[ ]
30
45
60
75
90
120
Specimen fixing block
R722.5
Copper foil
(5)
Specimen
Ix , I y , and Ixy are the moment of inertia of area for the x axis,
the moment of inertia of area for the y axis, and the product
of inertia of area, respectively. Image processing software can
automatically calculate the grain growth direction based on
(5) using a binary image of the grown grains. The calculation
was carried out for grains that did not coalesce with each
other.
Punch block
5
Table 3: Biaxial stress ratio in each experimental condition.
max
C
j [%]
k [ ]
l [rad]
f0 [%]
max
0.16
0.33
0.52
0.72
1.0
3.11
99.3
2.99
2.98
4.13
91.9
2.69
2.82
4.50
94.6
2.48
2.96
4.78
96.1
2.22
3.09
4.85
94.7
1.88
3.03
1.0
1.99
53.1
100 m
(a) C = 1.0
max
max
peak
100 m
100 m
(c) C = 1.0
2
+ l + f0 ,
fsin = j sin
k
(n = 1, 2, 3, . . .).
(7)
is obtained
by substituting (3) for (4). In this figure, peak
(b) C = 0.52
k 4n + 1
=
l
2
2
(6)
1
1
fnormal = exp 2 ( )2 ,
2s
s 2
(8)
20
=0
= 74.7
15
90
15
10
= 164.7
10
0
0
15
30
45
60
75
15
30
45
60
75
[ ]
m = 50
m = 100
m = 200
m = 50
m = 100
m = 200
m = 50
m = 100
m = 200
(a) C = 1.0
(b) C = 0.33
20
15
10
15
30
45
60
75
m = 50
m = 80
(c) C = 1.0
100
110
100
Principal stress 1 [MPa]
peak
and
90
80
70
90
80
70
60
60
50
50
0.8
0.6
0.4
0.2
40
0.8
0.6
0.4
0.2
peak
Theoretical curve
( )
99.99
99.9
[%]
99
90
3
5
7
0.16
0.33
0.52
0.72
1.0
330
117
61
182
66
34
71
26
13
48
18
9
44
16
8
5. Conclusions
70
50
30
10
1
0.1
0.01
20
C = 0.52
C = 0.72
C = 1
40
60
80
Grain growth direction [ ]
100
120
C = 0.16
C = 0.33
References
[1] H. Ohkubo, Copper electroplating method of stress analysis,
Memoirs of the School of Engineering Nagoya University, vol. 2021, p. 1, 1968.
[2] A. Kato and T. Mizuno, Stress concentration factors of grooved
shaft in torsion, Journal of Strain Analysis for Engineering
Design, vol. 20, no. 3, pp. 173177, 1985.
[3] Y. Nagase and T. Yoshizaki, Fatigue gage utilizing slipinitiation phenomenon in electroplated copper foil, Experimental Mechanics, vol. 33, no. 1, pp. 4954, 1993.
[4] S. Kitaoka and Y. Ono, Cyclic biaxial stress measurement by
electrodeposited copper foil with circular holes, Strain, vol. 42,
pp. 4956, 2006.
[5] S. Kitaoka, J.-Q. Chen, N. Egami, and J. Hasegawa, Measurement of biaxial stress using electrodeposited copper foil with a
microcircular hole, JSME International Journal, Series A, vol.
39, no. 4, pp. 533539, 1996.
[6] A. Kato, Stress measurement by copper electroplating aided by
a personal computer, Experimental Mechanics, vol. 27, no. 2,
pp. 132137, 1987.
[7] S. Timoshenko and S. Woinowsky-krieger, Theory of Plates and
Shells, McGraw-Hill, New York, NY, USA, 2nd edition, 1959.
[8] D. C. Montgomery and G. C. Runger, Applied Statistics and
Probability for Engineers, John Wiley & Sons, New York, NY,
USA, 3rd edition, 2003.
[9] D. C. Montgomery and G. C. Runger, Applied Statistics and
Probability for Engineers, John Wiley & Sons, New York, NY,
USA, 4th edition, 2003.
Research Article
Automatic Determination of Fiber-Length Distribution in
Composite Material Using 3D CT Data
Matthias Temann,1 Stephan Mohr,2 Svitlana Gayetskyy,2 Ulf Haler,2
Randolf Hanke,2 and Gunther Greiner1
1 Computer
2 Fraunhofer
1. Introduction
Fiber reinforced polymers (FRPs) are used increasingly in
the aerospace and automotive industry, since those components facilitate the cost-eective building of lightweight
but rigid components. One manufacturing method for
the construction of fiber reinforced polymers is the long
fiber reinforced thermoplastics (LFRP-D) process, where a
matrix material, for example consisting of polypropylene
and additives, is heated and mixed with fibers, for example
carbon or glass fibers. This process happens directly, that
is, without the usage of an intermediate semifinished part.
Thereby, components can be manufactured that are capable
of acting as supporting elements with respect to rigidity and
stability. Due to these properties, many parts made of LFRP
are already used in the automotive industry, for example for
frontends, underbody casing, supporting elements, or parts
of an engine compartment.
The length of the fibers has a strong influence on the
strength, stiness, and impact resistance of the component
(a)
(b)
V3
V2
2. Previous Work
The established method for the determination of a fiber
length distribution within a component is to pyrolyze its
matrix [4, 5]. Therefore, the component is put into an oven
which is heated to about 450 C. After about 90 minutes,
the matrix component is reduced to ashes and the skeleton
of the fibers remain. In order to determine their length
distribution, several methods are possible and in use. One
mechanical process is to sieve the fibers with dierent sieve
sizes. However, this method does not work satisfactory with
long fibers that are strongly felted. Another common method
V1
+Min vector
Min vector
Figure 8: Schematic view of the solution to the gap problem. During the tracing, critical voxels are detected (red). In a subsequent
pass, dierent fibers containing the same critical voxels are merged
into a single fiber.
Figure 6: Schematic view of the fiber crossing problem. The
currently used minimum eigenvector is pointing into a dierent
direction than the previous one. This can lead to a wrong
segmentation result.
3. Fiber Segmentation
A model of an ideal fiber is the foundation of the following
segmentation approach. It is reasonable to assume that a
general fiber is cylindrically shaped. Furthermore, in the
high-resolution scans acquired the grey value profile of the
fibers exhibited a clear maximum at their centers. Moreover,
all fibers of a common class usually have constant and
previously known diameters [5]. As a consequence, these
characteristic features can be exploited by using a modelbased segmentation approach. The segmentation algorithm
itself is modeled as a multistep process. Firstly, the whole
image is filtered and reduced by a closing operation in order
to achieve a good fiber separation in the image data. Then,
(a)
(a)
A
(57 13 61): 0.2625 GV
2 cm
Slice: 61
Timepoint: 0
66, 51, 135, grey, 1
1.000, 1.000, 1.000
User mode
scan:
LUT C/W: 0.2625/0.9275
(b)
(b)
H
(57 13 61): 0.2625 GV
6 cm
Slice: 7
Timepoint: 0
51, 135, 66, grey, 1
1.000, 1.000, 1.000
User mode
scan:
LUT C/W: 0.2625/0.9275
(c)
H
(c)
R
Slice: 13
Timepoint: 0
66, 135,51, grey, 1
1.000, 1.000, 1.000
User mode
scan:
LUT C/W: 0.2625/0.9275
(d)
Figure 10: Real product sample one. Fiber length is high and the
fibers are packed very tightly, including some overlap. However, the
shape of the fibers is quite clear and approximating the ideal case.
Fiber-length distribution
60
50
30
20
450550
410450
310410
250310
170250
110170
User mode
scan:
LUT C/W: 0.425/1
70110
Slice: 63
Timepoint: 0
66, 51, 135, grey, 1
1.000, 1.000, 1.000
030
3070
10
2 cm
40
Fiber-length (m)
(a)
2 I
2 I
2 I
xx xy xz
2 I
2 I
2 I
.
H x, y, z =
yx yy yz
I
2 I
2 I
(b)
(2)
zx zy zz
Since the Hessian is a symmetric matrix, it can be rearranged
to yield
(1)
2 I
2 I
2 I
xx xy xz
2
2
2
I
I
I
.
H x, y, z =
xy
yy
yz
2
2
I
I
I
(3)
xz yz zz
This order facilitates the calculation of the eigenvalues and
eigenvectors of this matrix. We have
1 0 0
H x, y, z = V 0 2 0 VT ,
0 0 3
(4)
10 cm
(a)
(a)
A
(363 69 54): 0.02 GV
20 cm
Slice: 54
Timepoint: 0
511, 511, 129, grey, 1
1.000, 1.000, 1.000
User mode
scan:
Lut C/W: 0.2625/0.9275
(b)
H
(b)
10 cm
Slice: 260
Timepoint: 0
511, 129, 511, grey, 1
1.000, 1.000, 1.000
User mode
scan:
Lut C/W: 0.2625/0.9275
(d)
27
12
0.91
0.8-0.9
0.70.8
0.60.7
0.50.6
0.40.5
0.30.4
(c)
Fiber-length distribution
0.20.3
User mode
scan:
Lut C/W: 0.2625/0.9275
800
700
600
500
400
300
200
100
0
0.10.2
Slice: 300
Timepoint: 0
511, 129, 511, grey, 1
1.000, 1.000, 1.000
Fiber count
10 cm
00.1
Fiber-length (mm)
F(x) =
0,
if 2 > 0 or 3 > 0,
D(x), else,
(5)
with
(6)
where
RA =
|2 |
,
|3 |
|1 |
,
|2 3 |
RB =
S =
(7)
2j .
Correlation in %
97.29%
98.03%
97.87%
(8)
4. Results
For the evaluation of the presented method, two types of
data were used. First of all, artificial test datasets were
created. They contained varying types of fibers with dierent
densities and lengths. The major advantage of this data type
was that the length of each individual fiber was known
beforehand and thus allowed an exact evaluation of the
produced results.
Moreover, the algorithm was also evaluated on CT scans
of real plastic components. Two of them will be presented
in this paper. The first component was built of straight and
long cylindrical fibers, approximating the ideal fiber very
closely. The second dataset, however, was more dicult to
deal with. It contained mostly short, curved, and heavily
overlapping fibers. Nevertheless, good segmentation results
were achieved in all cases.
4.1. Artificial Test Data. Three datasets containing ideal fiber
models were created for evaluation (Figure 9). In order to
simulate a real CT scan, random noise was added to the
image data before processing.
During the detection phase, all fibers from the test
data were found and have been segmented correctly. The
correlation with the known results is very high in all cases, as
shown in Table 1. In summary, the accuracy of the detection
process over the three test cases was about 98%. Due to the
presence of the artificial image noise, the small detection
error is tolerable. Tests without addition of the extra noise
yielded a fiber-length correlation of 100%.
4.2. Real Test Data. From the evaluation on real fiber
reinforced polymer components, two selected examples are
presented. 3D and slice views are shown in Figures 10 and 13,
5. Conclusion
In this paper, a novel, model-based approach for the
automatic detection, segmentation, and length distribution
calculation of fibers in CT data of fiber reinforced polymers
was presented. Since fiber length distribution within the
material is essential for the stability of an assembly, having
a non-destructive evaluation method is highly desirable.
The presented approach uses a segmentation scheme
which was shown to be robust even in the presence of
curved fibers and image noise. The algorithm is also able
to handle tightly packed and crossing fibers, though the
accuracy suers in these cases, as not all fibers may be
detected fully or gaps may occur within single fibers. In order
to estimate a systematic error in this situations, more datasets
have to be investigated and the automatic results have to be
compared to the outcome of a pyrolysis analysis. Moreover,
current CT scanning devices are still restricted to scanning
small sample sizes only, which currently limits the practical
applicability of this method.
However, the results show that the presented algorithm
can achieve a reasonably good segmentation and thus can act
as a basis for further research on this topic. Further research
will include the acquisition of reference data by pyrolizing
Acknowledgments
The authors are thankful to T. Potyra and M. Reif (Fraunhofer ICT) for providing the samples used for evaluation.
This paper was cofinanced by the European Union and the
Free State of Bavaria, Germany.
References
[1] A. M. A. Hug and J. Azaiez, Eects of length distribution on
the steady shear viscosity of semiconcentrated polymer-fiber
suspensions, Polymer Engineering and Science, vol. 45, no. 10,
pp. 13571368, 2005.
[2] M. Neitzel and P. Mitschang, Handbuch Verbundwerkstoe,
Carl Hanser, Munchen, Germany, 2004.
[3] S.-Y. Fu, C.-Y. Yue, X. Hu, and Y.-W. Mai, Characterization
of fiber length distribution of short-fiber reinforced thermoplastics, Journal of Materials Science Letters, vol. 20, no. 1, pp.
3133, 2001.
[4] R. Smallman and R. Bishop, Modern Physical Metallurgy and
Materials Engineering, Butterworth-Heinemann, 1999.
[5] G. Erhard, Designing with Plastics, Hanser Gardner, 2006.
[6] G. Washer and F. Blum Jr., Raman spectroscopy for the
nondestructive testing of carbon fiber, Research Letters in
Materials Science, vol. 2008, Article ID 693207, 3 pages, 2008.
[7] W. Pratt, Digital Image Processing, 2007.
[8] A. F. Frangi, W. J. Niessen, R. M. Hoogeveen, T. van Walsum,
and M. A. Viergever, Modelbased quantitation of 3-d magnetic resonance angiographic images, IEEE Transactions on
Medical Imaging, vol. 18, no. 10, pp. 946956, 1999.
[9] W. H. Press, S. A. Teukolsky, W. T. Vetterling, and B. P.
Flannery, Numerical Recipes: The Art of Scientific Computing,
2007.
[10] M. Fatica, D. Luebke, I. Buck, et al., High-performance
computing with cuda, SUPERCOMPUTING Tutorial, 2007.
Research Article
Digital Radiography Using Digital Detector Arrays Fulfills
Critical Applications for Offshore Pipelines
Edson Vasques Moreira,1, 2 Jose Maurcio Barbosa Rabello,3 Marcelo dos Santos Pereira,4
Ricardo Tadeu Lopes,5 and Uwe Zscherpel6
1 Nondestructive
Testing Laboratory of TenarisConfab, 475 Av. Dr. Gastao Vidigal Neto, Pindamonhangaba, 12414-900, SP, Brazil
Estadual Paulista (UNESP)-FEG, 333 Av. Doutor Ariberto Pereira da Cunha, Guaratingueta, 12516-410, SP, Brazil
3 Petrobras/Engenharia/SL/SEQUI/CI, Km 143 Rodovia Presidente Dutra, S
ao Jose dos Campos, 12220-840, SP, Brazil
4 Department of Materials and Technology, UNESP, FEG, 333 Av. Doutor Ariberto Pereira da Cunha,
Guaratingueta, 12516-410, SP, Brazil
5 Nuclear Instrumentation Laboratory, COPPE, Universidade Federa do Rio de Janeiro, Rio de Janeiro, Brazil
6 BAM, Federal Institute for Materials Research and Testing, Radiological Methods, 87 Unter den Eichen, 12205, Berlin, Germany
2 Universidade
1. Introduction
Industrial radiographic films have been utilized for many
years in the quality control by NDT of a variety of products;
however, the use of digital radiography has recently been
implemented in several sectors, for example, the medical,
aerospace, security, automotive and petrochemical sectors.
In addition to the technological trend it has been demonstrated that digital radiography sometimes oers a series of
benefits in terms of productivity, sensitivity, environmental
aspects, image treatment tools, cost reduction, security, POF
improvement [1], and so forth.
Among the current options, the digital detector array,
DDA, Varian 2520V, 127 m, employed in this paper is
2
Indicator (IQI), Basic Spatial Resolution (BSR), and Signal Noise Ratio (SNR) normalized by the Basic Spatial
Resolution and comparing artificial defects, the digital
method showed better results and advantages compared with
conventional film technique. These experiments were carried
out to support the voting and the development of the first
ISO document applicable to digital radiography using DDA
for weld seam inspection on welded pipes for pressure, the
ISO/DIS 10893-7 specification [3].
2. Digital Radiography
Digital radiography systems oer the possibility of obtaining
images with much less strict exposure requirements than
those of conventional film systems. Exposure imprecision
normally leads to radiographs that are dark, light or show
little contrast, which are easily improved and enhanced using
digital techniques.
Some the advantages of digital radiographic systems
include: image display, reduction of X-ray doses, image
processing, automated acquisition, partially or completely
automated evaluation, image storage, and the retrieval is
significantly reduced.
The entire operation is simplified, from obtaining the
image to the cycle time involved in obtaining, evaluating
and storing each image with ensured traceability [4], as
illustrated in Figure 2.
Dierent to industrial films, a fully integrated environment for digital radiographic images adds even other
advantages [5] to these of the DDAs, for example: productivity and sensitivity are increased resulting in fast decisions
using remote access, meetings, training, Level 3 supervision,
process control monitoring, and so forth.
Ocean
Layer of salt
Pre-salt
Meetings
Video conference
R&D
Supervision
Level 3 inspector
Process control
monitoring
Training
Welding
Repairs
Interpreter
Level 2 inspector
Outside inspector
Costumers
Auditors
Backup
Electronic
comunication
Remote access
Laboratory
4. Results
The basic parameters for evaluation of the image quality are
the following: the normalized Signal-to-Noise Ratio (SNRn)
at base material, the Basic Spatial Resolution (SRb) and the
Contrast Sensitivity (CS) by the wire type IQI. Finally, the
defect visibilities obtained with DDA were compared with
those obtained with digitalized films.
4.1. Normalized Signal-to-Noise Ratio, SNRn. The normalized SNRn (see ASTM E 2597 for details) for the DDA system
TFT
Photo
diode
TFT
Photo
diode
Photo
diode
TFT
Photo
diode
TFT
Photo
diode
TFT
Photo
diode
TFT
Photo
diode
TFT
Photo
diode
TFT
TFT
TFT
TFT
TFT
Photo
diode
TFT
Photo
diode
TFT
Photo
diode
TFT
Photo
diode
TFT
TFT
Photo
diode
Photo
diode
TFT
Photo
diode
TFT
Photo
diode
Photo
diode
TFT
Photo
diode
TFT
Photo
diode
Photo
diode
Photo
diode
TFT
Photo
diode
Photo
diode
TFT
TFT
Photo
diode
TFT
TFT
Photo
diode
Photo
diode
Photo
diode
Photo
diode
Photo
diode
TFT
Photo
diode
TFT
Photo
diode
TFT
Photo
diode
TFT
TFT
Photo
diode
Photo
diode
TFT
TFT
TFT
Photo
diode
TFT
TFT
Photo
diode
Photo
diode
TFT
TFT
TFT
Photo
diode
Photo
diode
TFT
Photo
diode
Photo
diode
TFT
Photo
diode
TFT
Photo
diode
TFT
Photo
diode
TFT
Photo
diode
TFT
Photo
diode
TFT
Photo
diode
SDD
DDA Varian
PaxScan 2520 V
ODD
Sample
Diaphragm
Sample
X-ray tube
Y.TU 225-D04
DDA
X-ray tube
(a)
(b)
4.2. SRb and Contrast IQI. In Figures 7 and 8 the performance of digital radiography is shown in terms of
contrast sensitivity (single wire IQI read-out) as function
5
Digital radiography-ISO 10893-7 (base metal)
SNRn
450
400
350
300
250
200
150
100
50
0
5
10
15
20
25
30
35
13
12
11
10
9
500
400
14
300
13
13
200
100
10
12
12
11
0
0
10
15
20
25
19
(+2 wires)
18
(+1 wire)
17
16
(+1 wire)
15
14
13
4.9 mm
ODD 100
4.9 mm
9.7 mm
9.7 mm
6.4 mm
ODD 55 ODD 103 ODD 108 ODD 57
WT-object detector distance (ODD)
1s
8s
2s
16 s
4s
32 s
Required ISO 10893-7
8s
16 s
32 s
14
14
14
600
SNRn
14
1s
2s
4s
Required ISO 10893-7
8
0
20
15
(a)
4.9 mm-CP04-film D4
(b)
4.9 mm-CP04-1 s
Crack
(c)
4.9 mm-CP04-32 s
Figure 9: WT of 4,9 mm: (a) Film D4, DDA with (b) 1 s and (c) 32 s of integration time.
(a)
25.3 mm-CP03-film D4
(b)
Lack of penetration
25.3 mm-CP03-8 s
(c)
25.3 mm-CP03-32 s
Figure 10: WT of 25,3 mm: (a) Film D4, DDA with (b) 8 s and (c) 32 s of integration time.
18
16
14
KODAK M100
12
10
Film
AGFA D4
8
6
4
ISO 10893-7
Digital
0
4.9
6.4
9.7
19.2
WT (mm)
25.3
32.3
6. Conclusions
Based on the above results, it can be concluded that the direct
digital radiographic technique using DDAs is more sensitive
than the conventional film technique, both in terms of visible
wires of the Image Quality Indicators and in the detection of
small real defects in the welds [13].
Hence, as foreseen in the purposed ISO/DIS 10893-7,
digital radiography using DDAs can be employed directly on
the productionlines of oil and gas pipelines, with important
advantages over the conventional technique.
This digital technique therefore represents an advance
in the quality of radiographic testing currently employed,
in addition to its high degree of automation, which will
allow for improved productivity and greater environmental
friendliness.
Acknowledgments
The authors are indebted to the company XYLON International for carrying out the tests, as well as the sta
responsible for the Post-graduations program at the UNESP,
Universidade Estadual Paulista-FEG. The authors would also
like to thank TenarisConfab for its support in terms of
technical and financial resources, which enabled this paper
to be carried out.
References
[1] L. Pick and O. Kleinberger, Technical highlights of digital
radiography for NDT, Materials Evaluation, vol. 67, no. 10,
pp. 11111116, 2009.
[2] E. V. Moreira, H. R. Simoes, J. M. B. Rabello, J. R. De
Camargo, and M. Dos Santos Pereira, Digital radiography to
inspect weld seams of pipelinesbetter sensitivity, Soldagem
e Inspecao, vol. 13, no. 3, pp. 227236, 2008.
[3] ISO/DIS 10893-7, Non-destructive testing of steel tubes
part 7: digital radiographic testing of the weld seam of
welded steel tubes for the detection of imperfections, Geneva,
Switzerland, 2009.
[4] A. G. Farman, C. M. Levato, D. Gane, and W. C. Scarfe, In
practice: how going digital will aect the dental oce, Journal
of the American Dental Association, vol. 139, 2008.
[5] R. Pincu and O. Kleinberger, Portable X-ray in the service of
art, Materials Evaluation, vol. 68, no. 3, pp. 311318, 2010.
[6] U. Ewert, U. Zscherpel, C. Bellon, G. R. Jaenish, J. Beckmann, and M. Jechow, Flaw size dependent contrast reduction and additional unsharpness by scattered radiation in
radiographyfilm and digital detectors in comparison, in
Proceedings of the 17th World Conference on Non-Destructive
Testing, Shanghai, China, 2008.
[7] API 5L, Specification for Line Pipe, American Petroleum
Institute, Washington, DC, USA, 2007.
[8] ISO 3183, Petroleum and natural gas industriessteel pipes
for pipeline transportation systems, Geneva, Switzerland,
2007.
[9] E. Moreira, M. C. Fritz, H. R. Simoes, J. M. B. Rabello,
and J. R. Camargo, Flat-panel detectors are accepted for
digital radiography in place of conventional radiography in
pipeline weld inspection, in Proceedings of the 4th Conferencia
Panamericana de END, Buenos Aires, Argentina, 2007.
7
[10] E. Moreira, R. Lopes, M. Pereira, J. M. B. Rabello, U. Zscherpel,
and D. Oliveira, Real application stage of DR in weld seam of
pipes for gas and oil linepipes, in 10a COTeq, Conferencia de
Tecnologia, Salvador, Brasil, 2009.
[11] U. Ewert, K. Bavendiek, J. Robbins, et al., New compensation
principles for enhanced image quality in industrial radiology
with digital detector arrays, Materials Evaluation, vol. 68, no.
2, pp. 163168, 2010.
[12] D. F. Oliveira, Analise da Radiografia Computadorizada em
Condicoes de Aguas
Profundas, Dissertacao (Mestrado em
Engenharia Nuclear), COPPEUniversidade Federal do Rio
de Janeiro, Rio de Janeiro, Brazil, 2007.
[13] U. Zscherpel and K. Bavendiek, High quality radiography
with digital detector arrays, in Digital Imaging VIII Conference, Foxwoods, Conn, USA, 2005.
Research Article
Appling a Novel Cost Function to Hopfield Neural Network for
Defects Boundaries Detection of Wood Image
Dawei Qi, Peng Zhang, Xuefei Zhang, Xuejing Jin, and Haijun Wu
College of Science, Northeast Forestry University, Harbin 150040, China
Correspondence should be addressed to Dawei Qi, qidw9806@yahoo.com.cn
Received 31 December 2009; Revised 14 April 2010; Accepted 13 May 2010
Academic Editor: Joao Manuel R. S. Tavares
Copyright 2010 Dawei Qi et al. This is an open access article distributed under the Creative Commons Attribution License,
which permits unrestricted use, distribution, and reproduction in any medium, provided the original work is properly cited.
A modified Hopfield neural network with a novel cost function was presented for detecting wood defects boundary in the image.
Dierent from traditional methods, the boundary detection problem in this paper was formulated as an optimization process
that sought the boundary points to minimize a cost function. An initial boundary was estimated by Canny algorithm first. The
pixel gray value was described as a neuron state of Hopfield neural network. The state updated till the cost function touches the
minimum value. The designed cost function ensured that few neurons were activated except the neurons corresponding to actual
boundary points and ensured that the activated neurons are positioned in the points which had greatest change in gray value. The
tools of Matlab were used to implement the experiment. The results show that the noises of the image are eectively removed, and
our method obtains more noiseless and vivid boundary than those of the traditional methods.
1. Introduction
X-ray wood nondestructive testing is an eective method
for accessing to internal information of wood. Comparing
with other conventional wood nondestructive testing, such as
appearance judgment, acoustic emission testing, ultrasonic
testing, microwave testing, and stress wave testing, this
method can acquire distinct wood internal structure images
by an X-ray imaging system. Through the wood images, the
positions of wood defects can be easily identified; the scales
of the defects can be roughly estimated. Furthermore, we can
make use of computer technology to automatically extract
wood defects information from the images for automatically identifying defects characteristics such as areas, types,
and severity, which can help making the optimal sawing
solution. However, extracting accurate defects information
depends on the accurate boundary detection. There are many
edge detection algorithms. Most previous edge detection
algorithms used first-order derivative operators such as the
Sobel edge operator [1, 2], the Roberts edge operator, and
the Prewitt edge operator [3]. If a pixel point is on the
boundary, its neighborhood will be a zone of transition.
The Laplacian operator [4] is a second-order derivative
2
That method showed the results comparable to those of
standard snakes- based algorithms, but it requires less
computing time.
In this paper, we presented a novel approach to automatically detect wood defects boundaries using a modified
Hopfield neural network with a specific cost function
designed for wood defects image. The boundary detection
problem in this paper was regarded as an optimization
process that sought the boundary points to minimize
a cost function. Hopfield neural network was used as
computational networks for solving optimization problems.
Because of its highly interconnected structure of neurons,
the network was not only very eective in computational
complexity, but also very fault tolerant. In consideration of
the accuracy of the detection, an initial boundary must be
estimated before using the Hopfield neural network. Every
pixel in the image with an initial boundary was represented
by a neuron which was connected to all other neurons but
not to itself. The image was considered as a dynamic system
which was completely depicted by a cost function. The states
of the neurons updated according to the cost function till
the convergence. Then, the result image was given by the
states of the neurons. The tools of Matlab were applied to
implement the experiment in this paper. The results showed
that our method can obtain more continued and more
accurate boundary points than the traditional methods of
boundary detection.
The remainder of this paper is organized as follows. In
Section 2, a basic imaging principle of X-ray and a wood
nondestructive detection imaging system are described. A
Hopfield neural network theory and its application in solving
optimization problems are illustrated in Section 3. Section 4
discusses how to implement the boundary detection algorithm using a Hopfield network. This section is divided into
four phases. We first discuss how to initiate defects boundaries, then how to map the boundary detection problem into
a Hopfield neural network, and then a novel cost function
for wood defects boundaries detection is described. Finally,
we illustrate the summary of the algorithm. In Section 5,
experimental results and a discussion are given. We illustrate
a conclusion and a perspective in Section 6.
I0
X-ray
(1)
Log specimen
Rotation
X-ray source
Image intensifier
Rotating plate
z 1
z 1
3.2. Hopfield Neural Networks for Solving Optimization Problem. Hopfield neural networks have been used successfully
in solving optimization problems such as the traveling
salesman problem (TSP) [1517]. In recent years, take
advantage of their optimization computation capabilities,
Hopfield neural networks were applied in image processing
[1820]. Mapping a practical problem to an energy function
is a key step for Hopfield neural networks to solve optimization problems. The basic form of the energy function was
described in the literature [21] as
z 1
z 1
E=
Neuron
z 1
N N
N
1
Ti j vi v j Ii vi .
2 i=1 j =1
i=1
(4)
Every neuron in a Hopfield neural network has computational capabilities, which can process an input and give
a relevant output. If the ith neuron is described by two
variables: its input ui and its output vi . The output is the
state which is computed by a given activation function f . The
transformation is described as
vi = f (ui ).
(2)
N
j=
/i
Ti j v j + Ii .
(3)
restriction: pi (v1 , v2 , . . . , vN ) 0,
i = 1, 2, . . . , k,
(5)
k
i F pi (v1 , v2 , . . . , vN )
(6)
i=1
ki > 0.
(7)
E(v1 , v2 , . . . , vN )
,
vi
ki > 0.
(8)
N M N M
N
M
1
Ti, jk,l vi, j vk,l
Ii, j vi, j ,
2 i=1 j =1 k=1 l=1
i=1 j =1
M
N
(10)
k=1 l=1
(9)
vi, j = f ui, j ,
(11)
if ui, j >
f ui, j =
0 otherwise.
(12)
A neuron
log
Hopfield
neural network
Crack
Image
m
i=1
n
vi, j
j =1
m
n
i=1 j =1
n
ai, j
j =1
(13)
vi, j
,
Gi, j Gi, j+1 + Gi, j Gi, j 1
n
l=1
ai,l
Gi, j
i, j
,
Gi, j+1 + Gi, j Gi, j 1
(14)
Figure 12: Image after boundary detection using our method with
threshold of 0.006.
7
received a good result. As shown in the Figures 614, the
method based on Hopfield neural network in detecting
boundary of wood defects was eective; the noises were
eectively removed. We can get a more noiseless and vivid
wood defect boundary. Thus, a promising method of wood
boundary detection based on Hopfield neural network with
a novel cost function is provided. All the courses of image
processing and building a Hopfield neural network in this
paper were implemented using the tools of Matlab. The tools
of Matlab are well done in the study of images.
Figure 13: Image after boundary detection using our method with
threshold of 0.005.
Figure 14: Image after boundary detection using our method with
threshold of 0.004.
6. Conclusion
An X-ray imaging technique was applied in wood nondestructive detection. Through wood images acquired by this
technique, the wood defects information such as locations,
scales, and types was visual. The detected defects can be
further processed for recognition of defects types and other
defects characteristics.
Hopfield neural network was applied in the boundary
detection of wood images. We designed a novel cost function
for a Hopfield neural network to detect a defect boundary
as solving an optimization problem. After the boundary
initiation using Canny edge algorithm, a slight adjustment
can be made to seek the actual boundary which will be
implemented by a Hopfield neural network with the cost
function. Those points that decreased the network energy
were detected as boundary points. Taking advantage of
the collective computational ability and energy convergence
capability of the Hopfield neural network, the experiment
References
[1] L. S. Davis, A survey of edge detection techniques, Computer
Graphics Image Processing, vol. 4, pp. 248270, 1975.
[2] D.-S. Lu and C.-C. Chen, Edge detection improvement by ant
colony optimization, Pattern Recognition Letters, vol. 29, no.
4, pp. 416425, 2008.
[3] R. C. Gonzalez, R. E. Woods, and S. L. Eddins, Digital Image
Processing Using MATLAB, Publishing House of Electronics
Industry, Beijing, China, 2004.
[4] R. C. Gonzalez and R. E. Woods, Digital Image Processing,
Publishing House of Electronics Industry, Beijing, China, 2nd
edition, 2002.
[5] J. Canny, Computational approach to edge detection, IEEE
Transactions on Pattern Analysis and Machine Intelligence, vol.
8, no. 6, pp. 679698, 1986.
[6] X. Xu, Z. Yang, and Y. Wang, A method based on rankordered filter to detect edges in cellular image, Pattern
Recognition Letters, vol. 30, no. 6, pp. 634640, 2009.
[7] C. H. Chao and A. P. Dhawan, Edge detection using Hopfield
neural network, in Proceedings of Conference on Applications
of Artificial Neural Networks, vol. 2243, pp. 242251, 1994.
[8] C.-Y. Chang, A contextual-based Hopfield neural network
for medical image edge detection, in Proceedings of the IEEE
International Conference on Multimedia and Expo (ICME 04),
pp. 10111014, June 2004.
[9] M. Kass, A. Witkin, and D. Terzopoulos, Snakes: active
contour models, International Journal of Computer Vision,
vol. 1, no. 4, pp. 321331, 1988.
[10] Y. Zhu and H. Yan, Computerized tumor boundary detection
using a hopfield neural network, IEEE Transactions on
Medical Imaging, vol. 16, no. 1, pp. 5567, 1997.
[11] A. K. Hamou and M. R. El-Sakka, Optical flow active
contours with primitive shape priors for echocardiography,
EURASIP Journal on Advances in Signal Processing, vol. 2010,
Article ID 836753, 10 pages, 2010.
[12] Y. Zheng, G. Li, X. Sun, and X. Zhou, A geometric active
contour model without re-initialization for color images,
Image and Vision Computing, vol. 27, no. 9, pp. 14111417,
2009.
[13] Y. Yang and X. Gao, Remote sensing image registration
via active contour model, AEUInternational Journal of
Electronics and Communications, vol. 63, no. 4, pp. 227234,
2009.
[14] Z. Liu, Modern Ray Detection Technology, China Standard
Press, Beijing, China, 1999.
[15] J. F. Roler and W. H. Gerstacker, On the convergence
of iterative receiver algorithms utilizing hard decisions,
EURASIP Journal on Advances in Signal Processing, vol. 2009,
Article ID 803012, 8 pages, 2009.
8
[16] U.-P. Wen, K.-M. Lan, and H.-S. Shih, A review of Hopfield neural networks for solving mathematical programming
problems, European Journal of Operational Research, vol. 198,
no. 3, pp. 675687, 2009.
[17] J. J. Hopfield and D. W. Tank, Neural computation of
decisions in optimization problems, Biological Cybernetics,
vol. 52, no. 3, pp. 141152, 1985.
[18] G. Pajares, M. Guijarro, and A. Ribeiro, A Hopfield Neural
Network for combining classifiers applied to textured images,
Neural Networks, vol. 23, no. 1, pp. 144153, 2010.
[19] R. Cierniak, A 2D approach to tomographic image reconstruction using a Hopfield-type neural network, Artificial
Intelligence in Medicine, vol. 43, no. 2, pp. 113125, 2008.
[20] R. Sammouda and M. Sammouda, Improving the performance of Hopfield neural network to segment pathological
liver color images, International Congress Series, vol. 1256, pp.
232239, 2003.
[21] J. J. Hopfield, Neurons with graded response have collective
computational properties like those of two-state neurons,
Proceedings of the National Academy of Sciences of the United
States of America, vol. 81, pp. 30883092, 1984.
[22] S. Lu, Z. Wang, and J. Shen, Neuro-fuzzy synergism to
the intelligent system for edge detection and enhancement,
Pattern Recognition, vol. 36, no. 10, pp. 23952409, 2003.
Research Article
Attenuation Analysis of Lamb Waves Using
the Chirplet Transform
Florian Kerber,1 Helge Sprenger,2 Marc Niethammer,3 Kritsakorn Luangvilai,4
and Laurence J. Jacobs4
1 Institute
of Mathematics and Computing Science, University of Groningen, 9700 AV Groningen, The Netherlands
of Applied and Experimental Mechanics, University of Stuttgart, Pfaenwaldring 9, 70569 Stuttgart, Germany
3 Computer Science Department, University of North Carolina, Chapel Hill, NC 27599-3175, USA
4 School of Civil and Environmental Engineering and G.W. Woodru School of Mechanical Engineering, Georgia Institute of Technology,
Atlanta, GA 30332, USA
2 Institute
1. Introduction
Ultrasonic waves are often used in nondestructive testing
to evaluate the integrity of structural components, as
well as to determine material properties of composite or
layered materials. In various disciplines such as civil or
aerospace engineering, multimode, dispersive guided waves
such as Lamb waves have been applied, see Chimenti
[1] for an overview. However, complicated signal analysis
is the trade o for their versatility. In fact, the main
challenges to process Lamb wave signals are due to their
very characteristics. Firstly, dispersion phenomena require
a resolution of the frequency content of a Lamb wave
signal over time which is inherently compromised by the
uncertainty principle. Secondly, Lamb waves are multimodal, which means that interferences between individual
modes complicate the allocation of energy and displacement
t0 , 0 , s, q, p =
=
1
2
(3)
F0 H() = H( 0 ).
(4)
Frequency shift:
F0 h(t) = ei0 t h(t),
Scaling:
1
Ss h(t) = h(t/s),
s
Ss G() = sG(s).
Time shear:
1/2
(6)
p
P p H() = exp i 2 H().
2
Frequency shear:
q
Qq h(t) = exp i t 2 h(t),
2
1/2
Qq H() = (iq)
22
exp i
H() .
q
(7)
Higher order time shear P p1 ,p2 ,... H() can also be applied
resulting in
p1 2 p2 3
+ +
2
3
H(),
(8)
q1 2 q2 3
t + t +
2
3
h(t).
(9)
(10)
stft
C (t0 , 0 ) =
1
=
2
(11)
x(t)gt0 ,0 (t)dt
(12)
X()Gt0 ,0 ()d.
(5)
t2
exp i
h(t),
2p
P p h(t) = ip
to obtain
(1)
Time shift:
ct
(13)
1 t t0
exp
g(t) =
4
2
2
s0
s0
1
2
t() = t0 + p1 ( 0 ) + p2 ( 0 )2 + + p5 ( 0 )5 .
(15)
The group delay g () of a signal H() = A() exp[i()]
in frequency domain is given by
d
().
d
s1
700
(16)
The group delay of the window function Gt0 ,0 ,p1 ,p2 ,...,p5 () =
Tt0 F0 P p1 ,p2 ,...,p5 H() can thus be fitted to (15) for every
mode of propagation by a fifth-order time shear (8) with
parameters p1 , . . . , p5 . Figure 1 depicts the 3-regions of
window functions adjusted by the adaptive algorithm for
the first symmetrical mode s0 . The CT is not calculated
in frequency regions of interference with other modes, for
example, around 2 MHz at the intersection of the a0 - and
s0 -mode. More details about the adaptive algorithm can be
found in [4]. The same Gaussian window function (14) was
also used for the STFT-based analysis.
s4
a1
a2
a4
s2
600
500
a3
a0
400
s0
300
200
s3
100
4
5
6
Frequency (MHz)
10
Figure 1: CT basis functions adjusted to the s0 -mode using 5thorder time shear. The dispersion curves are the solution of the
Rayleigh-Lamb equations (17) for an aluminum 3003 plate of
thickness 0.99 mm and source-receiver distance of 90 mm.
3. Problem Setting
In this paper, Lamb waves traveling in aluminum plate
structures are considered. Due to the relatively simple
geometry of the plate, it is possible to compute dispersion
curves based on the analysis of the Rayleigh-Lamb equations
for stress-free boundaries as derived in Achenbach [9],
4k2 pq
tan qh
=
2 ,
tan ph
q2 k 2
(14)
g () =
800
Time (s)
2
(17)
tan qh
q2 k 2
=
,
tan ph
4k2 pq
where
p2 =
2
k2 ,
cL2
q2 =
2
k2 ,
cT2
(18)
0.8
0.7
0.6
0.5
1.4
1.3
1.2
1.1
0.5
0.4
1.5
2
Frequency (MHz)
2.5
(a)
0.3
0.2
0.9
0.1
0
0
4
5
6
Frequency (MHz)
10
Theoretical
STFT
CT
4. Results
First consider the results obtained for the numerically
simulated signal. The particle displacement associated to a
particular mode is extracted from the modulus |C i (t0 , 0 )|,
i {stft, ct} of each transform. To eliminate the eect of the
excitation source, these values are normalized to a particular
modeTable 1 contains the results for normalization with
respect to a0 and s0 by taking the point-by-point quotient
of the respective moduli at every frequency 0 . Figure 2
shows the ratio s0 /a0 as obtained from the STFT- and CTbased algorithm versus the exact theoretical solution. The
latter are very close to the theoretical solution, while the
amplitude ratio extracted from the STFT deviates especially
at frequencies where individual modes are highly dispersive
such as the s0 -mode for frequencies between 2 and 3 MHz.
Since the STFT does not use window functions adjusted to
2.2
2
1.8
1.6
1.4
1.2
1
0.5
1.5
2
Frequency (MHz)
2.5
(b)
Figure 3: Geometric attenuation for the a0 -mode of the synthetically generated signal determined with the STFT (b) and
CT (a).
Dashed lines represent the theoretically expected solution d2 /d1 ,
dash-dotted lines are the results obtained from the CT- and STFTbased algorithm, respectively, for the distances 40 mm (pink),
50 mm (green), 60 mm (black), 70 mm (blue), and 80 mm (red)
related to 90 mm of propagation distance.
x(t), x(t)
(19)
numerical signal
experimental signal
a1 /a0
9.49
18.60
173.73
115.36
s0 /a0
67.89
686.36
46.96
83.06
s1 /a0
8.46
14.65
343.88
223.07
a0 /s0
10.54
51.33
35.75
52.99
a1 /s0
4.00
66.28
272.29
289.76
s1 /s0
12.19
10.24
484.40
385.70
distance
mode a1
CT
STFT
80/90 mm
70/90 mm
60/90 mm
50/90 mm
40/90 mm
0.31
0.43
0.49
0.24
0.28
3.64
6.40
8.48
10.12
14.04
120/150 mm
90/150 mm
60/150 mm
50/150 mm
1.91
7.26
2.50
3.44
4.16
14.95
16.48
19.21
mode s0
CT
STFT
CT
Synthetically generated signal
1.65
2.54
1.71
2.80
3.62
1.99
4.65
5.20
6.23
1.79
7.80
6.56
3.14
7.28
6.33
Experimentally generated signal
15.03
22.78
17.01
15.83
22.08
9.63
9.71
26.91
13.56
3.12
14.56
10.78
mode s1
STFT
CT
STFT
9.08
11.78
24.37
28.91
28.26
5.61
8.39
13.19
38.22
132.4
19.96
24.37
24.09
32.26
40.03
20.91
20.63
32.68
31.37
32.92
24.63
67.88
76.49
18.7
16.38
14.46
5.29
5. Conclusions
The main goal of this paper is to evaluate the potential of the
chirplet transform for dispersive wave analysis. The problem
of associating displacement or energy related quantities to
individual modes of propagation is of interest in nondestructive evaluation. The theoretical advantage of the proposed
method, that is, tailoring regions of averaging to individual
modes based on the dispersion relation, becomes apparent
when analyzing numerically simulated Lamb wave signals
traveling in an aluminum plate. Extracting displacement
ratios and geometric spreading for individual modes of
propagation succeed with high accuracy in regions with
sucient mode separation. This strongly indicates that
the CT-based algorithm can achieve a better performance
than more conventional approaches like the spectrogram.
The potential to extract displacement and energy-related
quantities associated with a particular mode of a dispersive
wave therefore qualifies it as a versatile tool in NDE
applications. As a model-based approach, the CT based
algorithm uses information about the dispersion relation.
[3]
1.6
1.4
1.2
[4]
1
0.8
0.5
1.5
2
2.5
3
Frequency (MHz)
3.5
[5]
(a)
[6]
[7]
0.5
1.5
2
2.5
3
Frequency (MHz)
3.5
[8]
(b)
[9]
Figure 4: Geometric attenuation for the a0 -mode of the experimentally generated signal determined with the STFT (b) and CT
(a).
Dashed lines represent the theoretically expected solution d2 /d1 ,
dash-dotted lines are the results obtained from the CT- and STFTbased algorithm, respectively, for the distances 50 mm (green),
60 mm (black), 90 mm (blue), and 120 mm (red) related to 150 mm
of propagation distance.
[10]
Acknowledgment
The Deutscher Akademischer Austausch Dienst (DAAD)
provided partial support to F. Kerber.
References
[1] D. E. Chimenti, Guided waves in plates and their use in
materials characterization, Applied Mechanics Reviews, vol.
50, no. 5, pp. 247284, 1997.
[2] M. Niethammer, L. J. Jacobs, J. Qu, and J. Jarzynski,
Time-frequency representations of Lamb waves, Journal of
Research Article
Flexible Riser Monitoring Using Hybrid Magnetic/Optical Strain
Gage Techniques through RLS Adaptive Filtering
Daniel Pipa, Sergio Morikawa, Gustavo Pires, Claudio Camerini, and Joao Marcio Santos
Materials, Equipments and Corrosion Department (TMEC), Petrobras Research and Development Center (CENPES),
Av. Horacio Macedo, 950. Cidade Universitaria, 21941-915 Rio de Janeiro, RJ, Brazil
Correspondence should be addressed to Daniel Pipa, danielpipa@gmail.com
Received 30 November 2009; Revised 5 April 2010; Accepted 7 May 2010
Academic Editor: Joao Manuel R. S. Tavares
Copyright 2010 Daniel Pipa et al. This is an open access article distributed under the Creative Commons Attribution License,
which permits unrestricted use, distribution, and reproduction in any medium, provided the original work is properly cited.
Flexible riser is a class of flexible pipes which is used to connect subsea pipelines to floating oshore installations, such as FPSOs
(floating production/storage/o-loading unit) and SS (semisubmersible) platforms, in oil and gas production. Flexible risers are
multilayered pipes typically comprising an inner flexible metal carcass surrounded by polymer layers and spiral wound steel
ligaments, also referred to as armor wires. Since these armor wires are made of steel, their magnetic properties are sensitive to
the stress they are subjected to. By measuring their magnetic properties in a nonintrusive manner, it is possible to compare the
stress in the armor wires, thus allowing the identification of damaged ones. However, one encounters several sources of noise
when measuring electromagnetic properties contactlessly, such as movement between specimen and probe, and magnetic noise.
This paper describes the development of a new technique for automatic monitoring of armor layers of flexible risers. The proposed
approach aims to minimize these current uncertainties by combining electromagnetic measurements with optical strain gage data
through a recursive least squares (RLSs) adaptive filter.
1. Introduction
Flexible risers are an important component of oshore production systems of oil and gas. They are used to link subsea
pipelines to floating installations, such as FPSOs (floating
production/storage/o-loading unit). Flexible risers have
been one of the preferred deepwater riser solutions in many
regions of the world due to their good dynamic behavior and
reliability [1].
Petrobras is a Brazilian multinational petroleum company whose businesses include oil and gas exploration,
production, transportation, refining, and distribution. Since
most Brazilian oil reserves are located oshore and often
under deepwater, Petrobras oil production is highly dependent on platforms and oshore equipments such as flexible
risers. Integrity management of flexible risers is essential to
ensure the safe operation of a production unit.
The main failure mode of flexible risers, when operating
in deep waters, occurs at the risers top section close to end
fitting due to fatigue in tensile armor wires. It is known,
however, that riser failure only happens after the rupture
2
these current uncertainties by combining electromagnetic
measurements with optical strain gage data through a
recursive least squares (RLSs) adaptive filtering technique.
This paper is organized as follows. Section 2 introduces
the flexible riser and comments its main failure modes.
Section 3 presents the proposed method as well as each
of its components. Finally, Section 4 presents the results
obtained in a laboratory trial, which attests the potential of
the method. A conclusion is drawn in Section 5. Flexible
pipe is a general term and denotes a type of pipe, whereas
flexible riser designates the vertical segment of a pipe
which is usually connected to an oshore production unit.
This article deals with signal processing algorithms,
rather than physical phenomena underlying the correspondence between mechanical load in ferromagnetic materials
and their electromagnetic properties. The idea is to show
that this relation does not need be fully determined and
understood if one uses a global load reference. Additionally,
if there exists an unknown or unstable gap between probe
and sample, this relation can be such complex that a
nonreferenced measurement of stress can be dicult. On
the other hand, some global load estimate can enhance the
results.
2. Flexible Risers
Function
Prevent collapse
2
3
4
5
6
7
8
9
Failure mode
Description
Collapse of carcass and/or pressure armor
Collapse
due to excessive tension, excessive
external pressure or installation overloads
Rupture of tensile or pressure armors due
Burst
to excess internal pressure
Rupture of tensile armors due to excess
Tensile failure
tension
Compressive
Birdcaging of tensile armor wires
failure
Rupture or crack of external or internal
Overbending
sheaths
Torsional
Failure of tensile armor wires
failure
Fatigue
Tensile armor wire fatigue
failure
Erosion
Of internal carcass
Of internal carcass or tensile/pressure
Corrosion
armor exposed to seawater or diused
product
Outer sheath
Outer layer of tensile armor
Anti-wear layer
Inner layer of tensile armor
Anti-wear layer
Back-up pressure armor
Interlocked pressure armor
Internal pressure sheath
Carcass
Pontoon
I-tube
Bellmouth
Bend
stiener
Flexible riser
Figure 2: End-fitting.
2.2. Flexible Riser Inspection and Monitoring. The Recommended Practice for Flexible Pipe [7], also known as API
17B, from the American Petroleum Institute, recommends
some inspection and monitoring methods for in-service
flexible pipes. Table 3 lists the monitoring methods as well
y(n)
X(n)
Display
W(n)
Load reference
Adaptive
filter
e(n)
d(n)
18
16
14
12
10
8
6
4
2
0
2
x(n)
y(n)
Adaptive
filter
Desired signal
Adaptive
algorithm
10
20
30
40
50
60
Output
e(n)
d(n)
Load
MAPS-FR f1
MAPS-FR f2
MAPS-FR f3
MAPS-FR f4
Proposed method
In the second methodology, a thin steel collar instrumented with FBG strain gages was placed around the riser
outer layer, measuring circumferential strains and changes in
its diameter. Wire failures can be detected as they can cause
variation in the external diameter of the polymeric jacket
covering the riser. The disadvantage of this technique is that
the number of broken wires needed to cause a detectable
variation in the external diameter can be significant.
Another scheme using FBG strain gages was proposed
in [1]. It is based on a retrofit clamp that monitors axial
elongation and torsion of a flexible riser. The clamp is
instrumented with FBG strain gages. As the previously
presented methods, it suers from sensibility. That is, one
single broken wire is unlikely to be detected as its eects on
external geometry are minimal.
In [22] a technology that integrates FBG sensors along
grooves in the tensile wires during manufacturing of the pipe
is described. Thus, strain and temperature can be monitored
along several meters of the wires and ruptures are easily
detected. Although new flexible pipes can be manufactured
with this feature, the technology cannot be applied to existing
pipes.
The electromagnetic tool MAPS-FR, on which the proposed method is based, is described in [3]. This equipment
can estimate the stress on armor wires in a noninvasive
manner. Additionally, it is sensitive to a single broken wire
Connector A
Connector B
Figure 9: Trial.
3. Proposed Method
Description
FM
covered
1, 4, 5,
6, 8, 9
Pressure test
1, 2, 5
Destructive
analysis of
removed
samples
8, 9
Load,
deformation
and
environment
monitoring
2, 3, 4,
6, 7
Nondestructive testing
of pipes in
service
Gaging
operations
Spool Piece
Test pipe
Annulus
Monitoring
1, 8, 9
8, 9
8, 9
7, 9
35
MAPS: 74
45
2.5
2
40
1
30
1.5
35
1
25
0.5
20
0
15
0.5
10
30
0.5
25
20
0.5
1
15
1.5
10
1
5
30
40
50
60
2.5
30
70
40
50
(a)
60
70
(b)
MAPS: 274
45
15
35
1
30
40
10
35
0
25
20
15
10
3
30
25
20
15
10
10
15
5
4
230
240
250
260
20
230
270
(a)
240
250
260
270
(b)
comprehending 45 sensing coils. Hence, the current MAPSFR set permits monitoring of approximately 45 wires on the
external armor layer, although this can be altered to suit
requirements.
3.1.2. MAPS-FR Data. The goal achieved by current MAPSFR technology is to compare tensile stress present in
armor wires. Nonetheless, the interpretation of raw data
requires an analysis by MAPS-FR experts. As a result, an
indication of a possible wire rupture is signalized including
its circumferential localization.
During the development of MAPS-FR system, Petrobras
and MAPS Technology jointly performed several controlled
laboratory tests. In these tests, specific wires were induced to
failure by the introduction of notches on their surfaces. Blind
tests were also performed, where only the Petrobras team was
FO
MAPS: 463
45
35
30
25
40
35
30
4
2
0
25
20
15
10
5
4
420
430
440
450
20
15
10
460
420
430
(a)
440
450
460
10
(b)
MAPS: 613
45
35
30
25
20
15
40
35
4
3
30
10
570
580
590
600
25
20
15
10
2
3
570
610
(a)
580
590
600
610
(b)
MAPS: 714
45
35
2
40
30
25
20
35
30
2
1
0
25
20
15
10
5
4
670
680
690
700
15
10
670
710
680
690
(a)
700
710
(b)
FO
2
MAPS: 354
45
10
35
1
30
40
5
35
25
20
15
30
25
20
10
10
15
15
10
4
20
310
320
330
340
310
350
(a)
320
330
340
350
(b)
Let x(n) be the input signal, let y(n) be the output signal,
and let d(n) be the desired signal, with n representing the
time. That is, d(0) is the value of desired signal at time 0. The
input vector is formed by the last N + 1 values of the input
signal and is given by
x(n) = [x(n) x(n 1) x(n N)]T .
(1)
The filter, which transforms the input signal x(n) into the
output y(n), is given by
w(n) = [w0 (n) w1 (n) wN (n)]T ,
(2)
9
MAPS: 547
FO
2
45
35
1
30
40
35
25
20
15
10
30
1
25
20
15
500
510
520
530
10
2
5
500
540
510
520
(a)
540
530
(b)
MAPS: 847
45
35
15
2
40
30
25
30
20
25
15
20
10
10
35
15
5
10
5
10
800
810
820
830
800
840
810
820
(a)
840
830
(b)
(3)
(n) =
(4)
n
n
i=0
i=0
ni 2 (i) =
2
ni d(i) xT (i)w(n) ,
(5)
10
MAPS: 1184
45
35
2
1
30
25
20
15
10
8
40
6
35
30
25
20
15
10
4
1140
1150
1160
1170
1140
1180
1150
1160
(a)
1170
1180
(b)
MAPS: 1282
FO
45
35
30
25
40
35
30
20
15
10
1240
1250
1260
1270
1280
4
3
2
25
20
15
10
1240
(a)
1250
1260
1270
1280
(b)
into realistic load values. These linear systems are continuously recalculated at every iteration to compensate the slow
time drift exhibited by MAPS-FR signals. Although magnetic
properties of metals vary nonlinearly with mechanical load,
linear systems can be used to do this mapping if some
adaptation is permitted. That is, the correspondence holds
(i.e., mapping becomes linear) in a small region surrounding
a given operation point. Once the operation point changes,
the adaptive filter recalculates its coecients. The new filter
coecients are valid within this new region.
The hybrid approach needs an estimate of riser global
load to be used as the desired signal d(n). Indeed, if all wires
are unbroken, the riser global load is approximately equally
divided to each wire and it can be used as an estimate of stress
in each wire. Since only dierences between wire stresses are
11
Initialization
S(1) = I
w(1) = [0 0 0]T
Do for n 0:
e(n) = d(n) xT (n)w(n 1)
(n) = S(n
1)x(n)
(n) T (n)
1
S(n 1)
S(n) =
+ T (n)x(n)
w(n) = w(n 1) + e(n)S(n)x(n)
y(n) = xT (n)w(n)
(n) = d(n) y(n)
(1)
(n)
x2(1) (n)
x(1) (n)
x45
(1)1
x (n 1) x(1) (n 1) x(1) (n 1)
1
2
45
..
..
..
..
.
.
.
.
(1)
(1)
(1) (n
N) x2 (n N) x45 (n N)
x1
x1(2) (n)
(2)
x1 (n 1)
..
(2)
x1 (n N)
(3)
X(n) =
x1 (n)
(3)
x1 (n 1)
..
.
(3)
x (n N)
1
x1(4) (n)
(4)
x (n 1)
1
..
(4)
x1 (n N)
x2(2) (n)
(2)
x2 (n 1)
..
..
.
.
x2(2) (n N)
x2(3) (n)
(3)
x2 (n 1)
..
..
.
.
(3)
x2 (n N)
x2(4) (n)
x2(4) (n 1)
..
..
.
.
(4)
x2 (n N)
1
(2)
x45 (n)
(2)
(n 1)
x45
..
(2)
x45 (n N)
(3)
(n)
x45
, (6)
(3)
x45 (n 1)
..
(3)
x45 (n N)
(4)
x45 (n)
(4)
(n 1)
x45
..
(4)
x45 (n N)
(p)
.
.
.
.
.
.
.
.
.
W(n) = .
.
.
.
(7)
Initialization
Do for 0 k 45
wk (1) = [1 1 1]T
Do for n 0
Do for 0 k 45
yk (n) = xkT (n)wk (n 1)
Update wk (n) for each k as in Algorithm 1
Algorithm 2: Complete Hybrid algorithm.
Meaning
Desired signal vector: global riser
d(n) =
load estimate obtained from
[d(n) d(n) d(n)]T
optical strain gages.
MAPS-FR signals matrix
X(n) = [x1 (n) x45 (n)]
organized as in (6).
y(n) =
Output vector which will be
[y1 (n) y2 (n) y45 (n)]T plotted for wire break detection.
Adaptive filter matrix coecients
W(n) = [w1 (n) w45 (n)] as in (7) that are updated at every
iteration.
(8)
(9)
12
The vanishing behavior is explained next: suppose that
there are as many sensing coils as wires in the external
armor layer. Even if each sensing coil is located exactly above
each wire, respectively, due to the gap between sensing coil
and wire (i.e., external polymer layer thickness), magnetic
field from adjacent wires leaks laterally and aects the
measurements of each other. Therefore, there exists some
portion of magnetic field surrounding a broken wire that
contains signals from unbroken wires. The adaptive filter
parameters wk0 (n) are recalculated based on this portion
of signal which is coherent to the loading d(n). Although
this seems to be a problem, detectability at exact instant of
rupture is unaected, as will be shown in Section 4.
The units of stress and magnetic fields are irrelevant in
this context. The whole system works by comparison; that
is, the goal is to determine whether there is an inactive
wire among active ones. Nevertheless, it is possible to
establish a calibration procedure which would give rise to
consistent measurements, though it is out of the scope of this
document.
4. Results
A trial has been carried out to evaluate the performance of
the proposed method. The trial took place at the riser fatigue
test rig of Physical Metallurgy Laboratory (LAMEF) in Porto
Alegre, Brazil. The facility allows the application of static and
dynamic tensile loads exceeding 220 tons. The test sample
was a section of 6 nominal bore new flexible production
riser, rated for 3000 psi of approximately 5 m length. One end
of the riser was fixed and subjected to axial load (nominally
connector A), whereas the other was free to rotate during
loading (nominally connector B). The loading was cyclic
and sinusoidal, from 160 to 220 ton and at a frequency of
0.0167 Hz.
In the tests, the riser loadings were chosen to simulate as accurately as possible the field conditions, namely,
approximately 30% to 50% of yield point. However, other
field conditions such as internal pressure, arbitrary load
instead of cycling load, and riser orientation (vertical instead
of horizontal) were not considered. The influence of these
circumstances on the results is intended to be studied in
future tests.
Two windows were opened on the external sheath to
access the wires. The first window was near to connector B
and had a circumferential shape, giving access to all wires.
This window (right side of Figure 9) was used to cut the
wires during loading, simulating a real rupture. The second
window was close to connector A and was used to instrument
all wires with optical strain gages. The signals from these
strain gages were used as references (i.e., real stress of wires).
The global riser load needed for the hybrid processing was
estimated averaging all strain gages signals.
The MAPS-FR was installed in the middle of the sample.
This configuration was chosen to ensure that the strain
gages measured similar tension values to those sensed by the
MAPS-FR.
MSE
1.1164
0.9781
0.7956
0.5291
0.2676
2009/06/25
Time
15 h 15
16 h 30
17 h 12
17 h 38
13 h 34
14 h 36
16 h 14
17 h 52
18 h 21
Time stamp
250
439
589
690
330
523
823
1160
1258
Event
Wire 37 break
Wire 35 break
Wire 30 break
Wire 6 break
Wire 5 break
Wire 17 break
Wire 7 break
Wire 13 break
Wire 27 break
13
installed on the external sheath or by the FBG strain gage
collar mentioned in Section 2.2 and described in [21]. The
collar can detect and estimate the riser global load in a
nonintrusive manner.
The presented results showed that the proposed technique produces graphical representations on which visual
detection of wire breaks can be eectively performed in
most cases. The proposed method can be straightforwardly
extended to automatically detect wire ruptures. A simple
fixed threshold or statistically variable threshold could be
employed for this purpose.
Acknowledgments
represent time stamp, whereas the ordinates represent wire
number (a) or sensing coil number (b). Given a wire number
and a time instant, the correspondent color indicates the
stress level. In both graphs, there is a scale bar on the right
which relates the color to a statistically normalized stress
parameter.
Figures 11 to 19 show the rupture moment of several
wires. It is clear on the left graphs which wire was cut. The
color dierence indicates that the damaged wire diverges
from the unbroken ones; that is, cut wires tend to loosen.
Although there is some noise on the right graphs, it is
possible to determine the rupture instant in most cases.
Moreover, the scale range of processed MAPS-FR signals
remains between 3 and +3 during normal operation. When
a rupture occurs, its limits reach from 3.5 to 20. This
could be used as a criterion for automatic detection.
5. Conclusion
Flexible risers are multilayered pipes used in oil and gas
industry. Their complex geometry imposes diculties (i.e.,
unknown wire arrangement) when assessing stress in internal layers through the outer polymeric sheath. Unknown
wire arrangement and in-service wire reaccommodation
introduce uncertainties while measuring internal stress.
This article proposes a new estimation method of
internal stress distribution by combining electromagnetic
measurements with optical strain gage data. Electromagnetic
measurements are converted into load values through adaptive filters. Optical strain gage signal is used as an estimate
of riser global load. This signal is used as desired signal in the
adaptive context. In other words, it is assumed that in normal
conditions riser global load is equally divided between wires.
A set of adaptive linear filters is calculated so as each
of the MAPS-FR signals is converted into load. The filters
inputs are electromagnetic signals, and the filters outputs are
load estimates of correspondent wires. When a wire rupture
occurs, the filter produces indications that the load changed,
and the break can be detected.
The main advantage of the proposed technique is that
it does not need the external sheath to be removed; that
is, it is a nonintrusive technique. Yet, it can detect singlewire ruptures. The riser global load estimate, required by
the proposed method, can be obtained through strain gages
References
[1] H. Corrignan, R. T. Ramos, R. J. Smith, S. Kimminau, and L. El
Hares, New monitoring technology for detection of flexible
armor wire failure, in Proceedings of the Oshore Technology
Conference (OTC 09), 2009.
[2] M. G. Marinho, C. S. Camerini, S. R. Morikawa, D. R.
Pipa, G. P. Pires, and J. M. Santos, New techniques for
integrity management of flexible riser end-fitting connection,
in Proceedings of the 27th International Conference on Oshore
Mechanics and Arctic Engineering, Estoril, Portugal, June 2008.
[3] J. C. McCarthy and D. J. Buttle, Non-invasive magnetic
inspection of flexible riser, in Proceedings of the Oshore
Technology Conference (OTC 09), Houston, Tex, USA, May
2009.
[4] Technip, Flexible pipe brochure, 2008, http://www.technip
.com/pdf/Flexible Pipe.pdf.
[5] Health and S. Executive, Guidelines for integrity monitoring
of unbonded flexible pipe, Tech. Rep., Health and Safety
Executive, 1998.
[6] M. G. Marinho, J. M. dos Santos, and R. D. O. Carneval,
Integrity assessment and repair techniques of flexible risers,
in Proceedings of the 25th International Conference on Oshore
Mechanics and Arctic Engineering (OMAE 06), Hamburg,
Germany, June 2006.
[7] A. P. Institute, API RP 17BRecommended Practice for Flexible
Pipe, API Publishing Services, Washington, DC, USA, 2002.
[8] A. Berg and N. J. Rishj-Nielsen, Integrity monitoring of
flexible risers by optical fibres, in Proceedings of the 21st
International Conference on Oshore Mechanics and Arctic
Engineering (OMAE 02), vol. 3, pp. 4752, 2002.
[9] L. A. Mesquita, J. M. Santos, P. Loureiro, and A. L. Carvalho,
Monitoramento das valvulas de despressurizacao de gas
percolado no espaco anular de risers de producao e exportacao
de oleo
e gas, in Proceedings of the Rio Pipeline Conference and
Exposition, 2005.
[10] A. Felix-Henry and P. Lembeye, Flexible pipes in-service
monitoring, in Proceedings of the 23rd International Conference on Oshore Mechanics and Arctic Engineering (OMAE
04), vol. 3, pp. 149154, 2004.
14
[11] J. Marsh, P. Duncan, and I. MacLeod, Oshore pipeline and
riser integritythe big issues, in Proceedings of the Oshore
Technology Conference (OTC 09), 2009.
[12] C. Saunders and T. OSullivan, Integrity management and
life extension of flexible pipe, in Proceedings of the Oshore
Technology Conference (OTC 07), 2007.
[13] J. W. Picksley, K. Kavanagh, S. Garnham, and D. Turner,
Managing the integrity of flexible pipe field systems: industry
guidelines and their application, in Proceedings of the Annual
Oshore Technology Conference, pp. 609618, 2002.
[14] J. Picksley, State of the art flexible riser integrity issues: study
report, Tech. Rep., MCS International, 2001.
[15] N. Weppenaar, A. Kosterev, L. Dong, D. Tomazy, and F. Tittel,
Fiberoptic gas monitoring of flexible risers, in Proceedings of
the Oshore Technology Conference (OTC 09), 2009.
[16] R. Roberts, S. Garnham, and B. DAll, Fatigue monitoring
of flexible risers using novel shape-sensing technology, in
Proceedings of the Oshore Technology Conference (OTC 07),
2007.
[17] R. Thethi, H. Howells, S. Natarajan, and C. Bridge, A fatigue
monitoring strategy and implementation on a deepwater top
tensioned riser, in Proceedings of the Oshore Technology
Conference (OTC 05), 2005.
[18] E. Binet, P. Tuset, and S. Mjen, Monitoring of oshore
pipes, in Proceedings of the Oshore Technology Conference
(OTC 03), 2003.
[19] M. G. Marinho, C. S. Camerini, J. M. Santos, and G. P. Pires,
Surface monitoring techniques for a continuous flexible riser
integrity assessment, in Proceedings of the Oshore Technology
Conference (OTC 07), Houston, Tex, USA, 2007.
[20] S. D. Soares, C. S. Camerini, and J. M. C. de Santos,
Development of flexible risers monitoring methodology
using acoustic emission technology, in Proceedings of the
Oshore Technology Conference (OTC 09), 2009.
[21] S. R. K. Morikawa, C. S. Camerini, D. R. Pipa et al.,
Monitoring of flexible oil lines using FBG sensors, in
Proceedings of the 19th International Conference on Optical
Fibre Sensors, vol. 7004 of Proceedings of SPIE, pp. 70 046F170 046F-4, April 2008.
[22] M. Andersen, A. Berg, and S. Saevik, Development of an
optical monitoring system for flexible risers, in Proceedings
of the Oshore Technology Conference (OTC 01), 2001.
[23] http://www.maps-technology.com/.
[24] B. DJ, Emerging technologies for in-situ stress surveys, in
Proceedings of the 6th International Conference on Residual
Stresses (ICRS 00), 2000.
[25] B. DJ and S. CB, Residual stresses: measurement using
magnetoelastic eects, in The Encyclopaedia of Materials:
Science and Technology, 2001.
[26] P. S. R. Diniz, E. A. B. da Silva, and S. L. Netto, Digital Signal
Processing: System Analysis and Design, Cambridge University
Press, Cambridge, UK, 2002.
[27] J. B. A. V. Oppenheim and R. W. Schafer, Discrete-Time Signal
Processing, Prentice-Hall, Upper Saddle River, NJ, USA, 2nd
edition, 1997.
[28] P. S. R. Diniz, Adaptive Filtering: Algorithms and Practical
Implementations, Springer, Boston, Mass, USA, 3rd edition,
2008.
[29] S. Haykin, Adaptive Fiter Theory, Prentice-Hall, Upper Saddle
River, NJ, USA, 3rd edition, 1996.
Research Article
Analysis of Approximations and Aperture Distortion for
3D Migration of Bistatic Radar Data with the Two-Step Approach
Luigi Zanzi and Maurizio Lualdi
Dipartimento di Ingegneria Strutturale, Politecnico di Milano, Piazza Leonardo Da Vinci 32, 20133 Milano, Italy
Correspondence should be addressed to Luigi Zanzi, luigi.zanzi@polimi.it
Received 31 December 2009; Accepted 7 June 2010
Academic Editor: Joao Manuel R. S. Tavares
Copyright 2010 L. Zanzi and M. Lualdi. This is an open access article distributed under the Creative Commons Attribution
License, which permits unrestricted use, distribution, and reproduction in any medium, provided the original work is properly
cited.
The two-step approach is a fast algorithm for 3D migration originally introduced to process zero-oset seismic data. Its application
to monostatic GPR (Ground Penetrating Radar) data is straightforward. A direct extension of the algorithm for the application
to bistatic radar data is possible provided that the TX-RX azimuth is constant. As for the zero-oset case, the two-step operator
is exactly equivalent to the one-step 3D operator for a constant velocity medium and is an approximation of the one-step 3D
operator for a medium where the velocity varies vertically. Two methods are explored for handling a heterogeneous medium; both
are suitable for the application of the two-step approach, and they are compared in terms of accuracy of the final 3D operator.
The aperture of the two-step operator is discussed, and a solution is proposed to optimize its shape. The analysis is of interest for
any NDT application where the medium is expected to be heterogeneous, or where the antenna is not in direct contact with the
medium (e.g., NDT of artworks, humanitarian demining, radar with air-launched antennas).
1. Introduction
In 1983, Gibson et al. [1] introduced the fast two-step
migration technique for 3D poststack seismic data. In a
companion paper, Jakubowicz and Levin [2] showed that in
a constant velocity medium the method is equivalent to the
classical one-step 3D migration. In their paper, Gibson et
al. performed a detailed analysis of the dierences between
the two-step approach and the one-step approach when the
velocity varies within the medium. They showed that in
normal conditions these dierences are negligible so that
the method was suggested as a quite attractive solution for
fast 3D migration of poststack seismic data. The extension
of the two-step approach to prestack 3D migration is not
straightforward although achievable as shown by Canning
and Gardner [3]. They proposed a scheme composed of 3D
DMO, cross-line 2D PSI, inline 2D DMO1 , velocity analysis,
and 2D inline depth migration. A variation of this scheme
was proposed by Meinardus et al. [4].
Here, it is shown that under a very restrictive condition,
that is, when the source-receiver azimuth is constant, the
two-step approach can be directly extended to non-zerooset data. In seismics, this would be the case of 3D marine
data collected by a single ship, equipped with a single cable,
and shooting along parallel lines. Of course this is not very
interesting for the seismic community where 3D acquisitions
are designed aiming at a balanced azimuth distribution to
get a good picture of 3D structures. Instead, the result is
interesting for GPR applications where 3D experiments are
normally executed by maintaining a constant orientation of
the antenna box, that is, a constant TX-RX azimuth. With the
present hardware technology, this approach is what is needed
by GPR users to achieve the goal of real-time visualization of
3D migrated volumes. Thus, the following sections discuss
the non-zero-oset extension of the two-step method, the
approximations resulting from the application to vertically
variable velocity fields, and finally the distortion eects on
the aperture of the migration operators. The quantitative
results are derived assuming the usage of an ultra highfrequency radar with air-launched antennas. This type of
hardware is normally preferred to ground-coupled antennas
to speed up the NDT acquisitions on highways and bridges. It
0.3
0.25
0.2
0.15
0.1
0.05
0
10
15
r r0 (cm)
20
25
z0 (cm)
z0 (cm)
2
0
2
4
6
8
10
12
14
16
18
20
0.3
0.25
0.2
0.15
0.1
0.05
0
10
15
r r0 (cm)
(a)
20
25
(b)
50
45
10
15
35
rms
20
25
vsoil = 12 cm/ns
H = 10 cm
Oset = 14 cm
40
LLNL
10
15
20
25 30
r r0 (cm)
vsoil = 8 cm/ns
vsoil = 12 cm/ns
35
40
45
50
vsoil = 16 cm/ns
vsoil = 20 cm/ns
y y0 (cm)
z0 (cm)
Figure 1: Travel time error [ns] for the rms approximation (a) and for the LLNL approximation (b) as a function of z0 (target depth) and
r r0 (lateral distance of the antenna from the vertical above the target). The system is monostatic. The antenna-medium distance is 10 cm,
and the radar wave velocity into the medium is 12 cm/ns.
30
LLNL
25
20
rms
15
10
5
0
10
z0 = 5 cm
z0 = 10 cm
20
30
x x0 (cm)
40
50
z0 = 15 cm
z0 = 20 cm
4(x x0 )2 4(y y0 )2
+
,
v2
v2
(1)
50
50
vsoil = 12 cm/ns
H = 10 cm
Oset = 14 cm
45
40
35
35
30
30
y y0 (cm)
y y0 (cm)
40
25
20
rms
15
20
15
10
5
0
LLNL
25
10
vsoil = 12 cm/ns
H = 10 cm
Oset = 14 cm
45
10
20
30
x x0 (cm)
z0 = 5 cm
z0 = 10 cm
40
50
z0 = 15 cm
z0 = 20 cm
10
20
30
x x0 (cm)
z0 = 5 cm
z0 = 10 cm
40
50
z0 = 15 cm
z0 = 20 cm
(a)
(b)
Figure 4: Migration aperture with the two-step approach for the rms (a) and the LLNL (b) approximations at four dierent target depths.
The antenna-medium distance is 10 cm; the radar wave velocity into the medium is 12 cm/ns; the TX-RX distance is 14 cm with the azimuth
oriented in the y-direction.
t=
t 2 = t02 +
4(y y0 )
,
v2
(2)
2
t0
(y y0 d)2 t0
+
+
v2
2
2
(y y0 + d)2
,
v2
(5)
where
where
2
t02 = T02 +
4(x x0 )
.
v2
(3)
The two-step approach consists of performing a 2D migration in the y-direction according to (2) followed by a 2D
migration in the x direction according to (3). Note that in
the monostatic case any summation order is valid.
Let us consider now a bistatic system where the TX-RX
separation is 2d, and let us assume that the 3D experiment is
executed by keeping a constant TX-RX azimuth. If we rotate
the coordinate system in such a way that the y-direction is
the azimuth direction, the 3D diraction surface will be now
given by
t=
2
T0
2
T0
(x x0 )2 (y y0 d)
+
v2
v2
+
2
(x x0 )2 (y y0 + d)
+
.
v2
v2
(4)
t0
2
2
=
T0
2
2
(x x0 )2
.
v2
(6)
The extension of the two-step approach consists of performing a 2D non-zero-oset migration in the y-direction
according to (5) followed by a 2D zero-oset migration in
the x direction according to (6). Note that in the bistatic
case the summation order is relevant, that is, the first step
must be in the azimuth direction. The conclusion is that
the extension of the two-step approach to a homogeneous
medium investigated with a bistatic radar is possible, and
the algorithm is totally equivalent to an exact one-step 3D
migration.
0.9
0.9
0.8
0.8
0.7
0.7
0.6
0.6
0.5
0.5
0.4
0.4
0.3
0.3
0.2
0.2
0.1
0.1
0
(a)
0
(b)
0.9
0.9
0.8
0.8
0.7
0.7
0.6
0.6
0.5
0.5
0.4
0.4
0.3
0.3
0.2
0.2
0.1
0.1
0
(c)
0
(d)
Figure 5: Dierent forms that can be used to shape the migration operator aperture in order to resemble the real radar footprint; (a) and
(b) are applicable to the one-step 3D operator to simulate a monostatic and a bistatic footprint, respectively; (c) is the aperture of the twostep operator obtained with a trivial 2D aperture limitation applied to both the x and y-direction steps; (d) is the aperture of the two-step
operator that can be obtained with a smooth weighting approach to resemble a shape similar to (b).
z0 (cm)
15
20
25
30
50
60
40
35
20
(cm
20
10
10
30
)
(cm
0
5
10
z (cm)
z (cm)
0
M3B 15
20
M3A 15
15
20
M3B 10
25
M3A 15
M3B 10
10
M3B 15
20
y (c
80
100
40
m)
60
40
y (c
m)
80
60
60
40
80
x (c
20
m)
20
20
(a)
60
40
)
(
x cm
80
100
(b)
90
M3A 15
80
70
M3B 10
0
z (cm)
y (cm)
60
50
M3B 10
40
10
30
M3B 15
20
20
M3A 15
0
20
100
40
y(
cm
60
60
40
80
20
x(
80
cm)
(c)
M3B 15
10
0
10
20
30
40
50
60 70
x (cm)
80
90 100 110
(d)
Figure 7: 3D representations through cross-sections (a, c) and through iso-amplitude plots (b, d) of the focused data over a sector of the
clay box of the JRC mine field. The sector explored with a 1 GHz antenna contains three mines buried at 10 cm and 15 cm. M3B indicates
mines with low metal content, while M3A indicates mines with high metal content.
5. Application Examples
A few examples are shortly presented to illustrate situations
where the vertical heterogeneity of the medium is successfully handled by using the two-step migration approach with
the rms velocity approximation to focus the data collected
with high-frequency bistatic radar systems.
Figure 6 presents the 3D image of a nonmetal dummy
mine (diameter 11 cm, thickness 6 cm) buried 10 cm in a
sand box. The target was explored with a stepped frequency
radar prototyped by RST. The area was manually scanned by
executing parallel profiles with a TX-RX aperture of 14 cm
and keeping the air-launched antennas at an approximate
height of 13 cm above the sand surface. The final scanning
grid was approximately 0.8 cm in the profile direction and
3 cm in the orthogonal direction. The frequency range was
26 GHz with a frequency step of 16 MHz. Despite of the
very low permittivity contrast between the sand and the
mine material, the weak signal scattered by the target was
successfully collapsed by the migration operator producing
0
0
50
x
50
100
100
150
0
20
40
60
80 100
120 140
160 180
y
150
0
50
z
(a)
50
0 0
100
150
(b)
Figure 8: 3D reconstruction of the joints that were used to connect the marble fragments when the monument was rebuilt in the 20th
century. Data collected by sliding a 1 GHz antenna over a cardboard in order to preserve the integrity of the delicate carvings that decorate
the explored wall. The focused data are displayed by showing from two dierent perspectives the 3D contour of an iso-amplitude surface.
Dimensions are indicated in cm.
armored pad plus the weight of the antenna are not expected
to trigger a landmine. The 3D processed data are shown
in Figure 7 using two dierent representation methods. The
quality of the final results is quite good considering that these
data were collected on the unfavorable clay soil box. The
result was also compared with data collected without the
protection pad validating the expectation that the ballistic
material does not introduce any significant degradation. This
is also a demonstration that the two-step migration operator
was properly dealing with the heterogeneity of the medium
consisting of two dierent materials: the ballistic layer and
the clay soil.
The third example is a 3D survey of a wall of a marble
monument built in Rome in 13 BC. The fragments of the
monument were found during archeological excavations,
and the monument was rebuilt in the third decade of the
20th century. Unfortunately, some details about the reconstruction are missing. The data presented in Figure 8 were
collected with a GSSI pulse radar equipment by scanning
the wall with a 1 GHz antenna. A cardboard was interposed
between the antenna and the wall in order to create a flat
surface and preserve the delicate carvings that decorate the
wall. As a result, the antenna is partially detached by the wall.
Again, this situation creates a material heterogeneity that can
be roughly assimilated to a two-layer structure where the first
layer is of a few centimeters and is mainly consisting of voids,
and thus is very fast for the radar signal, while the second
layer consists of marble. The focused data reveal the existence
of few metallic bars that were used to connect the marble
fragments.
The last example is taken from an investigation performed on a historical Palace in Venice in order to map
the position and length of hidden iron connection devices
that have been used until the nineteenth century in Venetian
6. Conclusions
A direct extension of the two-step approach for fast 3D
migration of bistatic GPR data is possible provided that the
90
80
70
y (cm)
60
50
40
30
20
10
0
20
40
60
80
100
120
x (cm)
Figure 9: Depth slice at 12.5 cm below the floor surface extracted from a 3D radar survey performed with a 2-GHz antenna in proximity
of a building corner of a historical palace in Venice. Data were processed with the two-step 3D migration algorithm. Two fiube (one coming
from the building corner) appear in the radar image.
that currently the GPR suppliers are producing multichannel GPR equipment with more and more antennas
mounted in a cart to increase the productivity. These systems
generate huge amounts of data that cannot be migrated
in real time by a single computer unless a very eective
algorithm is used.
Finally, the rms method and the accuracy discussion
are also of interest when radars with ground-coupled
antennas are used to investigate a medium that is vertically
heterogeneous. This is a frequent situation when the GPR
is applied to NDT inspections of layered structures such as
walls, floors, and pavements.
Acknowledgments
The authors are grateful to RST GmbH that developed
the stepped frequency radar prototype for humanitarian
demining, to the Joint Research Center in Ispra that gave free
access to the mine test field, to Dr. G. Lenzi of ISMES S.p.A.
who performed the acquisitions on the marble monument in
Rome, and to IDS S.p.A. that supplied the 2 GHz system for
the experiments in the Venetian Palace.
References
[1] B. Gibson, K. Larner, and S. Levin, Ecient 3-D migration
in two steps, Geophysical Prospecting, vol. 31, no. 1, pp. 133,
1983.
[2] H. Jakubowicz and S. Levin, A simple exact method of 3-D
migrationtheory, Geophysical Prospecting, vol. 31, no. 1, pp.
3456, 1983.
[3] A. Canning and G. H. F. Gardner, A two-pass approximation to
3-D prestack migration, Geophysics, vol. 61, no. 2, pp. 409421,
1996.
[4] H. Meinardus, C. Nieto, A. Chaveste, and J. Castaneda,
Ecient, target-oriented 3-D prestack depth migration in two
steps, Leading Edge, vol. 19, no. 2, pp. 138144, 2000.
[5] E. M. Johansson and J. E. Mast, Three-dimensional groundpenetrating radar imaging using synthetic aperture timedomain focusing, in Advanced Microwave and Millimeter-Wave
Detectors, vol. 2275 of Proceedings of SPIE, pp. 205214, July
1994.
[6] G. Mirabella-Roberti, L. Zanzi, and F. Trovo, Detecting hidden
ties in historic venetian palace by means of GPR, in Proceedings of International RILEM Conference on Site Assessment of
Concrete, Masonry and Timber Structures (SACoMaTiS 08), pp.
965974, Varenna, Italy, September 2008.
Research Article
ICA Mixtures Applied to Ultrasonic Nondestructive Classification
of Archaeological Ceramics
Addisson Salazar and Luis Vergara
Grupo de Tratamiento de Senal, Instituto de Telecomunicaciones y Aplicaciones Multimedia, iTEAM,
Universidad Politecnica de Valencia, Camino de Vera s/n, 46022 Valencia, Spain
Correspondence should be addressed to Luis Vergara, lvergara@dcom.upv.es
Received 23 December 2009; Revised 7 May 2010; Accepted 7 June 2010
Academic Editor: Joao Manuel R. S. Tavares
Copyright 2010 A. Salazar and L. Vergara. This is an open access article distributed under the Creative Commons Attribution
License, which permits unrestricted use, distribution, and reproduction in any medium, provided the original work is properly
cited.
We consider a classifier based on Independent Component Analysis Mixture Modelling (ICAMM) to model the feature jointprobability density. This classifier is applied to a challenging novel application: classification of archaeological ceramics. ICAMM
gathers relevant characteristics that have general interest for material classification. It can deal with arbitrary forms of the
underlying probability densities in the feature vector space as nonparametric methods can do. Mutual dependences among the
features are modelled in a parametric form so that ICAMM can achieve good performance even with a training set of relatively
small size, which is characteristic of parametric methods. Moreover, in the training stage, ICAMM can incorporate probabilistic
semisupervision (PSS): labelling by an expert of a portion of the whole available training set of samples. These properties of
ICAMM are well-suited for the problem considered: classification of ceramic pieces coming from four dierent periods, namely,
Bronze Age, Iberian, Roman, and Middle Ages. A feature set is obtained from the processing of the ultrasonic signal that is
recorded in through-transmission mode using an ad hoc device. A physical explanation of the results is obtained with comparison
with classical methods used in archaeology. The results obtained demonstrate the promising potential of ICAMM for material
classification.
1. Introduction
Determining the historical period of archaeological ceramic
shards is important for many archaeological applications,
particularly to reconstruct human activities of the past. In
fact, the standardization of an ecient and nondestructive testing (NDT) method for ceramic characterization
could become an important contribution for archaeologists.
Chemical, thermoluminescence, and other analyses have
shown to measure the age of ceramics accurately, but they are
expensive, time-consuming and involve some destruction of
the analyzed pieces [1]. Relative dating by comparison with
ceramic collections is nondestructive but very inaccurate
[1].
Ultrasound has been used in archaeological applications
such as ocean exploration to detect wrecks, imaging of
archaeological sites, and cleaning archaeological objects [2
4]. In this paper, we consider a method to sort archaeological
ceramic shards based on ultrasonic nondestructive evaluation. This method aims to be economic, fast, precise,
and innocuous for the ceramic pieces. It consists of three
steps: measuring by the through-transmission technique,
extracting features from the measured ultrasonic signals, and
classifying the feature set in classes corresponding to historic
or protohistoric periods.
The estimation of the chronological period of an archaeological fragment is not a straightforward work, especially
if we consider that the fragment might be moved from
its context of origin due to migrations, wars, or trade
exchange, and so forth. In addition, some external features
used for classification of archaeological objects, such as
particular shapes and decorations, might be not evident in
the fragments, and thus these aspects would not provide
information for a correct classification of the fragments.
Through-transmission was selected because the ceramic
produces large attenuation to the propagation of ultrasound,
2
so the pulse-echo technique cannot be implemented at
the required operating frequency. Time, frequency, and
statistical features (to be described later) were extracted using
standard signal processing techniques. The characteristics
of the classification problem oer a good case study for
testing advanced classifiers, like those based on modelling
the underlying statistical densities of the feature space as
mixtures of independent component analyzers.
In consequence, we dedicate Section 2 to presenting the
ultrasound through-transmission model from a linear system perspective and to defining the selected features. Then,
in Section 3 we present the rationale for these classifiers and
describe them based on mixtures of independent component analyzers. Section 4 presents the experiments and the
results obtained in the sorting of ceramic pieces from four
dierent periods: Bronze Age, Iberian, Roman, and Middle
Ages. Section 5 presents the conclusions and future line of
work.
We reported some preliminary results related to this
archaeological application which was presented in conference [5]. The following significant new contributions are
presented in this paper: rationale and selection of new
ultrasonic features; use of a classifier that is based on
probabilistic semisupervision of independent component
analyzers (ICA) mixture models that are suitable for handling expert uncertainty; implementation of an ad hoc
device designed to avoid the uncontrolled conditions of a
totally manual measurement procedure; and demonstration
of physical interpretation of the results obtained by the
proposed method in comparison with classical methods
used in archaeology. Therefore, this work provides the
foundations to implement a practical method to complement or even replace some of the destructive and
time-consuming techniques that are currently employed in
archaeology.
Centroid frequency fc :
f2
f1
us(t) = fc (t) =
f Px f , t df
f2
f1
Px f , t df
(1)
Bandwidth (BW):
f2
us(t) = BW(t) =
f1
Px f , t df
.
f fc (t)
f2
f1
2
Px f , t df
us(t) = max Px f , t .
These signatures are measures of the spectral content
variations that are aected by the ultrasonic pulse travelling
inside the material. They can be estimated by means of wellknown smoothing techniques of time-frequency spectral
analysis [6].
From us(t), we can obtain features in dierent
forms. For
t
example, the time average value (1/(t1 t0 )) t01 us(t)dt or the
instantaneous value at one particular time instant us(t0 ) can
be elements of the feature vector in the observation space.
Other time-domain features, such as the parameters A
and corresponding to an exponential model of the signal
attenuation x
(t) = Aet or the total signal power received
T
P = 0 |x(t)|2 dt/T, are also possible to complement the
frequency-domain features.
More features can be defined considering special conditions of the through-transmission model. For example,
higher-order statistics can be used to detect the possible
degree of non-Gaussianity of the reflectivity by measuring
higher-order moments of the received signal like HOM =
E[x(nTs ) x((n 1)Ts ) x((n 2)Ts )] [7], where E[] means
statistical expectation and 1/Ts is the sampling frequency.
Departures from the linear model of Figure 1 can be tested
in dierent forms, for example, using the so-called timereversibility [8], which is defined by TR = E[(dx(t)/dt)3 ].
x(t)
H(t, )
LTV
Material reflectivity:
Gaussian or nonGaussian white noise
Ultrasound pulse:
Attenuation and dispersion
eects
Linear time variant system
Recorded signal:
Grain noise (Gaussian or
non-Gaussian coloured noise)
Observation white Gaussian
noise
(i1)
(i1)
det Wk p(i1) skm p(Ck )
= K
,
(i1) (i1) (i1)
sk m p(Ck )
k =1 det Wk p
C(x) = max
p(Ck /x),
Ck
= max
|det Wk | p(sk )p(Ck ),
k = 1 K,
(2)
m = M1 + 1 M,
Ck
b(i)
k
where sk = Wk (x bk ), and Wk ,bk and p(sk ) are estimated
by means of a PSS training. This is achieved using an iterative
algorithm that we briefly describe below (a more detailed
description can be found in [15]). A relevant concept
in ICAMM learning is the embedded ICA algorithm. As
ICAMM is a set of multiple ICA models, learning of the
ICAMM parameters is essentially equivalent to simultaneous
learning of a set of ICA parameters. Thus, any ICA algorithm
could be used as part of the global ICAMM learning
algorithm as we describe below.
Let us consider that the set of training feature vectors
is formed by xm m = 1 M. We divide the set into
two subsets. The first subset is formed by xm m =
1 M1 M1 M vectors such that the expert archaeologist
is capable of assigning some p(0) (Ck /xm )ranging between 0
and 1 for k = 1 K. The second subset is formed by
M2 = M M1 vectors where no knowledge exists about
the possible class they belong to. The learning algorithms
proceed in the following manner.
Initialization. For k = 1 K, compute
M1
(0)
b(0)
m=1 xm p (Ck /xm ) (If M1 = 0, then select
k =
the initial centroids randomly);
W(0)
k (randomly);
p(0) (sk ) (in a form depending on the selected embed(0)
(0)
ded ICA algorithm) using s(0)
km = Wk (xm bk ).
Updating. For i = 1 I and for k = 1 K, compute
p(i)
Ck
xm
= p(0)
Ck
xm
m = 1 M1 .
(3)
M
m=1
xm p
(i)
(4)
Ck
,
xm
(i1)
W(i)
+ Wk(i1) ,
k = Wk
Wk(i1) =
M
m=1
(i1)
Wkm(ICA)
p(i1)
Ck
xm
(i1)
is the updating due to training sample xm
(Wkm(ICA)
in the selected embedded ICA algorithm).
p(i) (sk ) (in a form depending on the selected embed(i)
(i)
ded ICA algorithm) using s(i)
km = Wk (xm bk ).
Bronze Age
Iberian
Roman
Middle Ages
NP-ICA
0.83
0.79
0.72
0.65
JADE
0.81
0.75
0.67
0.64
TDSEP
0.79
0.69
0.66
0.55
fastICA
0.81
0.71
0.60
0.59
RBF
0.64
LVQ
0.59
MLP
0.67
kNN
0.64
Piece
thickness
measure
Coupling
medium
Ultrasound
transducer
Connection
to ultrasound
system
Recevier
case
Pressure
control
Ceramic
shard
Emitter
case
Transducer
carcass
Connection
to ultrasound
system
(a)
(b)
Figure 3: Measurement device employed in ultrasonic signal acquisition. A detail of the ultrasound transducer case is included.
5. Discussion
In order to draw a physical interpretation of the results
obtained by ultrasounds, a diversity of morphological and
physiochemical characterization analyses were carried out
using conventional instrumental techniques. A stratified
random sampling analysis was made using data from the
physical analysis of the pieces: open porosity and apparent
Bronze Age
0.79
0
0.05
0.02
Iberian
0
0.89
0.19
0
Roman
0.07
0.09
0.69
0.05
Middle Ages
0.14
0.02
0.07
0.93
7
Frequency
0
v = 1431.7
P = 28.4
0.1
1000
2000
3000
Statistics
fc = 1529303.7
fmax = 112500 A fmax = 0.4
BW = 15000 17500
fc (t0 ) = 250636.8
0.4
0.3
0.2
1000
Bronze Age
= 94101
A = 0.1
0.1
Bronze Age
Bronze Age
Time
0.1
4000
0.5
1.5
HOM = 0
TR = 0
500
2.5
0.1 0.05
0.05
0.1
(a)
Frequency
v = 1052.9
P = 1.7
1000
2000
3000
fc = 918334.4
fmax = 202500 A fmax = 259.7
BW = 95000 207500
fc (t0 ) = 440885
250
200
150
100
50
4000
Statistics
0.5
1.5
Iberian
= 171490.7
A = 3.2
Iberian
Iberian
Time
4
2
0
2
4
6
1000
500
0
2.5
HOM = 0.4
TR = 0.1
1500
(b)
Frequency
0
v = 832.1
P = 12.4
1000
2000
3000
Statistics
fc = 1488143.1
fmax = 107500 A fmax = 3.1
BW = 15000 110000
fc (t0 ) = 447521.2
Roman
= 222167.7
A = 1.1
Roman
Roman
Time
4000
0.5
1.5
1000
500
0
2.5
HOM = 0
TR = 0.4
1500
(c)
Frequency
0.5
0
0.5
v = 969
P = 16.6
1000
2000
3000
4000
Statistics
fc = 1117217.7
fmax = 115000 A fmax = 39.7
BW = 110000 115000
fc (t0 ) = 315079.8
40
30
Middle Ages
= 221114.5
A = 0.8
Middle Ages
Middle Ages
Time
20
10
0
0.5
1.5
2.5
3000
HOM = 0
TR = 0.3
2000
1000
0
0.5
0.5
(d)
Figure 4: Some examples of time signals, spectra, histograms, and corresponding features extracted from ultrasonic signals for archaeological ceramic pieces from dierent periods. Units are: Time axis: sample number; Frequency axis: MHz; Statistics axis: bins of signal values; P
(dB); v (m/s); fc , fmax , BW, fc (t0 ) (Hz).
Mean
deviation
i
deviation
i
(1) Bronze Age 28.20
1.80
3.7794
0.0676
(2) Iberian
22.70
1.85
3.3320
0.0663
(3) Roman
31.06
1.79
8.3532
0.1607
(4) Middle Ages 22.69
1.84
5.3441
0.0949
Figure 5: Bits taken from the ceramic fragments included in the test
probes prepared for the Scanning Electron Microscope.
Table 5: Statistics of the stratified random sampling for open porosity and apparent density.
Open porosity
Apparent density
Ni
47
155
138
140
47
155
138
140
ni
5
16
35
23
5
17
38
23
i
29,30
22,50
32,00
23,80
1,85
1,77
1,87
1,78
30,90
23,29
33,22
24,82
1,88
1,79
1,89
1,80
10
matrix of Table 3; however, it must be taken into account that
the training process introduced some degree of arbitrariness
because of the probabilistic labelling of the expert. Thus, it
seems that the expert was able to clearly identify the pieces
from Iberian and Middle Ages, but had more diculties
with the Bronze Age and Roman ones. This uncertainty may
have been transmitted to the classifier during the training
stage.
The experiments with classical methods of ceramic
characterization used in archaeology not only show that
correlations between the extracted parameters from the
ultrasound signals and the physical properties of the materials were found. Moreover, they also have demonstrated some
advantages of the proposed ultrasound method. The equipment required for nondestructive evaluation by ultrasound
is, in general, less costly, and the experiments are easier to
perform. The pieces are not damaged in any way during
testing, nor is it necessary to alter or destroy any of the
material that is analyzed. Very significant dierences for the
time required to analyze the pieces were demonstrated: the
ultrasound analysis (measuring, processing, and automatic
classification) for 480 pieces took only 6 hours; the SEM
analysis (tube preparation and electron microscope analysis)
for 80 pieces took 274 hours; the porosity and density
analyses (immersion and weighing of the pieces) for 80 pieces
took 288 hours.
There are limitations to the application of this procedure
due to the fact that the training of the classifier must be
performed from a specific set of data. Thus, the classifier
must be adapted to a specific data model and its eciency
is restricted by the fact that the new data to be classified must
follow the same data model. Nevertheless, the training of
the classifier could progressively be improved by increasing
the number of pieces for each known chronological period.
With proper training, the classifier would be able to provide
a prediction of the chronological period for pieces that
do not have clear chronological markers. In addition, the
semisupervised training mode could be used to model the
uncertainty that expert archaeologists may have about the
chronological period to which the pieces belong.
6. Conclusions
We have presented the results of applying ICAMM to a
challenging application in the area of nondestructive testing
of materials: the classification of archaeological ceramic
pieces into dierent historic periods. We have demonstrated
the interest of using methods that are able to consider nonGaussian models of the underlying probability densities in
the feature vector space. Thus, an ICAMM classifier was
tested using dierent variants depending on the embedded
ICA algorithm. ICAMM has the additional merit of allowing
PSS labelling, which is of practical interest in the considered
application. Note that in any ICAMM variant, the mutual
dependence among features is modelled in a parametric
form; also note that in nonparametric ICAMM, the estimated marginals are nonparametric. This confirms that
nonparametric ICAMM shares the good general modelling
capability of nonparametric classifiers and also can work
Acknowledgments
This paper has been supported by the Spanish Administration and the FEDER Programme of the European
Community under Grant TEC 2008-02975; the Generalitat
Valenciana under Grant GV-ACOMP/2009/340, and Grant
PROMETEO/2010/040.
References
[1] R. E. Taylor and M. J. Aitken, Chronometric Dating in
Archaeology, vol. 2 of Advances in Archaeological and Museum
Science Series, Springer, New York, NY, USA, 1997.
[2] R. Cribbs and F. Saleh, An ultrasonic based system used for
non-destructive imaging of archaeological sites, in Proceedings of Informatica ed Egittologia allInizio Degli Anni 90, pp.
97108, Rome, Italy, 1996.
[3] A. Murray, M. F. Mecklenburg, C. M. Fortenko, and R. E.
Green, Detection of delaminations in art objects using aircoupled ultrasound, in Proceedings of Materials Issues in Art
and Archaeology III Symposium, pp. 371378, San Francisco,
Calif, USA, 1992.
[4] W. I. Sellers and A. T. Chamberlain, Ultrasonic cave mapping, Journal of Archaeological Science, vol. 25, no. 9, pp. 867
873, 1998.
[5] A. Salazar, R. Miralles, A. Parra, L. Vergara, and J. Gosalbez, Ultrasonic signal processing for archaeological ceramic
restoration, in Proceedings of the IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP 06),
pp. III1160III1163, Toulouse, France, May 2006.
[6] L. Cohen and P. Loughlin, Recent Developments in TimeFrequency Analysis, Springer, New York, NY, USA, 1998.
[7] R. Miralles, L. Vergara, and J. Gos`albez, Material grain noise
analysis by using higher-order statistics, Signal Processing, vol.
84, no. 1, pp. 197205, 2004.
[8] R. Miralles, L. Vergara, A. Salazar, and J. Igual, Blind detection of nonlinearities in multiple-echo ultrasonic signals,
IEEE Transactions on Ultrasonics, Ferroelectrics, and Frequency
Control, vol. 55, no. 3, pp. 637647, 2008.
[9] A. K. Jain, R. P. W. Duin, and J. Mao, Statistical pattern
recognition: a review, IEEE Transactions on Pattern Analysis
and Machine Intelligence, vol. 22, no. 1, pp. 437, 2000.
[10] O. Chapelle, B. Scholkopf, and A. Zien, Semi-Supervised
Learning, MIT Press, Cambridge, Mass, USA, 2006.
[11] T.-W. Lee, M. S. Lewicki, and T. J. Sejnowski, ICA mixture
models for unsupervised classification of non-Gaussian classes
and automatic context switching in blind signal separation,
[12]
[13]
[14]
[15]
[16]
[17]
[18]
[19]
[20]
[21]
[22]
[23]
[24]
[25]
[26]
[27]
[28]
[29]
[30]
[31]
[32]
11
Research Article
On the Evaluation of Texture and Color Features for
Nondestructive Corrosion Detection
Fatima N. S. Medeiros,1 Geraldo L. B. Ramalho,2 Mariana P. Bento,1
and Luiz C. L. Medeiros3
1 Teleinformatics
Engineering Department, Federal University of Ceara, Campus of Pici, 6007 Fortaleza, CE, Brazil
Institute of Education, Science, and Technology of Ceara, Campus of Maracanau, Av. Contorno Norte 10,
61925-315 Maracanau, CE, Brazil
3 Petroleo Brasileiro S.A., LUBNOR/IE, Av. Leite Barbosa s/n, 60180-420 Fortaleza, CE, Brazil
2 Federal
1. Introduction
Corrosion is the destructive attack of a metal by chemical
or electrochemical reaction with its environment [1]. The
exposure of metallic surfaces structures to rust degradation
during their operational life is a known problem and it aects
storage tanks, steel bridges, pipelines, and ships [2].
Storage tanks and pipelines are commonly made of
carbon steel and low alloy steel not resistant to corrosion
in natural environments. Corrosion resistant paints and
coatings are used in almost all applications. Despite of this
protection, some steels will rust quite rapidly in humid air
even though condensation is not evident [3].
Atmospheric corrosion causes economic losses usually
due to production interruptions, replacement of expensive
materials, and contamination of products. A good deal of
attention should be paid to safety risks and the environment pollution due to corrosion [4]. Therefore, corrosion
monitoring is a relevant task to detect corroded regions
before failures occur by using inspection methods, so that
2
function [6]. Another issue is the random aspect of texture,
because size, shape, and orientation of pattern elements can
vary over a region.
The potentiality of image processing techniques for
automatic rust steel detection is investigated in [2]. The
methodology introduced an iterative multivariate data analysis to examine the eects of rust steel descriptors, that is,
texture and color distribution on a set of classifier algorithms.
In this analysis, a selector of classifiers indicates the algorithm
that provides good classification results (high sensitivity)
and acceptable time response for the automation of the
system [2].
In 1981, Itzhak et al. [7] employed computer image
processing techniques for statistical evaluation of pitting
corrosion in a plate of AISI 304L stainless steel exposed to a
corrosive water solution containing 10%FeCl3 . The purpose
of this work was to introduce and to evaluate new tools for
analyzing the eects of pitting corrosion process [7]. The
algorithm was capable of estimating the number and area
of pits in the binary image and therefore provided better
evaluation of pitting corrosion damages.
A popular image processing algorithm for texture analysis extracts features from the gray level co-occurrence matrix
(GLCM) [8]. In this paper, we explore the power of these
features to deal with the stochastic pattern of corrosion for
damage detection in metallic surfaces. Parameters extracted
from the GLCM can be used to define similarity properties
for corrosion detection purpose in image segmentation
methods based on region approach. This approach consists
in determining the regions that contain neighbor pixels in
the image that have similar properties, that is, gray level
and spatial relationship [9]. Two GLCM parameters, namely
contrast and energy, are considered to be the most ecient
for discriminating dierent textural patterns [10].
A wide variety of literature works [6, 8, 10, 11] have
reported that texture features are proper to characterize
corroded surfaces. In addition, typical color changes of
metallic surfaces are often related to corrosion. Thus, color
attributes carry out relevant information to design corrosion
detection systems. Moreover, some works have reported that
feature combination carries more discriminant power to
applications designed on small database image samples [12].
Methods based on neural networks and feature selection are
able to handle with high data dimensionality maintaining
good generalization level [13, 14].
This paper proposes a robust feature set for reliable
detection of atmospheric corrosion on metallic surfaces
using optical images acquired by charge-coupled device
(CCD) cameras. A total of 13 attributes per image sample
were computed using color and texture models: HSI (hue,
saturation, and intensity) color histogram statistics and
GLCM probabilities. The GLCM probabilities measure the
roughness, and HSI statistics characterize the color of
metallic surface samples. A sequential bottom-up feature
selection procedure [15] was applied as a result of the small
sample size and the high data dimensionality. We use Fisher
linear discriminant analysis (FLDA) and receiver operating
characteristic (ROC) curve to investigate the performance of
2
S i, j ,
(1)
Correlation =
i jS i, j i j
i j
i, j
Energy =
2
S i, j ,
(2)
(3)
i, j
Homogeneity =
i, j
S i, j
2 .
1+ i j
(4)
R+G+B
,
3
(5)
3
[min(R, G, B)],
R+G+B
I = cos1
[(R G) + (R B)]/2
(R G)2 + (R B)(G B)
(6)
.
(7)
(8)
E(X) =
1
(xn E(X))2 ,
N n=1
(9)
1
(xn E(X))3 ,
N n=1
(10)
1
(xn E(X))4 .
N n=1
(11)
E (X E(X))2 =
E (X E(X))3 =
E (X E(X))4 =
Digital
image
Predicted
ROIs
Corroded
Non corroded
Corroded
0.4167
0.0953
Real
Non corroded
0.1071
0.3809
HSI
statistics
GLCM
Texture attributes
Color attributes
Feature
combination
Corrosion descriptors
1
(n1 1 + n2 2 ),
n2
(12)
T
w (m1 m2 )2
wT SW w
(13)
4. Experimental Results
We performed both qualitative and quantitative discriminant
analysis in order to evaluate the eectiveness of dierent
feature subsets separability.
A set of 13 attributes per image sample were computed
from color and texture models using HSI statistics and gray
level co-occurrence matrices (GLCMs) probabilities.
The feature subsets discriminant performance was compared in terms of classification error and execution time
on a 13-dimensional (two feature subsets), 2-class (1 and
2 ) image database. For sake of simplicity, the two classconditional densities were admittedly Gaussian-like, with
mean vectors m1 and m2 and a pooled covariance matrix SW
defined in (12).
The experimental results are summarized in Tables 14
and Figures 4, 5 and 6. The correct class discrimination for
texture feature is computed as the sum of the highlighted
values in the main diagonal of the confusion matrices. The
values in Tables 1 and 2 reveal that about 79% hits were
obtained for texture features as much as for color features.
The false negatives and false positives rates are of the same
order (20%) for both subsets.
Combining feature sets leads to obtain more discriminant power without loss of generality. We achieved over
90% hits and an appreciably reduction of false positives to
Feature subset
Corroded pattern
Texture
or
Corrosion
descriptors
FLDA
discriminant
analysis
Color
or
Texture
and color
PCA
Non corroded
pattern
0.9
0.8
0.9
0.7
0.8
True positive rate (TPR)
0.6
0.5
0.4
0.3
0.2
0.7
0.6
0.4
0.2
0.1
0.1
0.2
0.3
0.4
0.5
0.6
0.7
0.8
0.9
0
0
AUCCH = 0.8763
0.3
0.1
0
AUC = 0.8486
0.5
9 attributes
13 attributes
0.1
0.2
0.8
0.9
Predicted
Corroded
Non corroded
Corroded
0.3809
0.1309
Real
Non corroded
0.0833
0.4049
Predicted
Corroded
Non corroded
Corroded
0.4118
0.0393
Real
Non corroded
0.0392
0.5099
0.8
0.3
0.7
0.6
AUC = 0.8826
0.5
0.2
0.4
AUCCH = 0.9039
0.3
0.2
0.1
0.1
0
1
0
0.1
0.2
0.3
0.4
0.5
0.6
0.7
0.8
0.9
5 6 7 8 9
Feature subset size
10 11 12 13
0.9
True positive rate (TPR)
0.8
0.7
5. Conclusion
0.6
AUC = 0.9115
0.5
0.4
AUCCH = 0.9311
0.3
0.2
0.1
0
0
0.1
0.2
0.8
0.9
AUC
0.8486
0.8826
0.9115
AUCCH
0.8763
0.9039
0.9311
Acknowledgment
The authors would like to thank ANP PRH31 for providing
financial assistance.
References
[1] J. William and D. Callister, Fundamentals of Materials Science
and Engineering, John Wiley & Sons, New York, NY, USA, 5th
edition, 2001.
[2] M. Trujillo and M. Sadki, Sensitivity analysis for texture
models applied to rust steel classification, in Electronic
Imaging Science and Technology, vol. 5303 of Proceedings of
SPIE, pp. 161169, San Jose, Calif, USA, January 2004.
[3] M. Kutz, Handbook of Environmental Degradation of Materials,
William Andrew, NewYork, NY, USA, 2007.
[4] E. Bardal, Corrosion and Protection, Springer, Berlin, Germany,
2003.
[5] P. R. Roberge, Corrosion Inspection and Monitoring, John Wiley
& Sons, New York, NY, USA, 2007.
[6] A. S. Arko Lucieer and P. Fisher, Multivariate texture-based
segmentationof remotely sensed imagery for extraction of
objects and theiruncertainly, International Journal of Remote
Sensing, vol. 29172936, pp. 610621, 2005.
[7] D. Itzhak, I. Dinstein, and T. Zilberberg, Pitting corrosion
evaluation by computer image processing, Corrosion Science,
vol. 21, no. 1, pp. 1722, 1981.
[8] K. S. Robert, R. M. Haralick, and I. Dinstein, Textural features
for image classification, IEEE Transactions on Systems, Man
and Cybernetics, vol. 3, no. 6, pp. 610621, 1973.
[9] S. Chabrier, B. Emile, C. Rosenberger, and H. Laurent, Unsupervised performance evaluation of image segmentation,
EURASIP Journal on Applied Signal Processing, vol. 2006,
Article ID 96306, pp. 112, 2006.
[10] A. Baraldi and F. Parmiggiani, An investigation of the textural
characteristics associated with gray level cooccurrence matrix
statistical parameters, IEEE Transactions on Geoscience and
Remote Sensing, vol. 33, no. 2, pp. 293304, 1995.
[11] K. Y. Choi and S. S. Kim, Morphological analysis and
classification of types of surface corrosion damage by digital
image processing, Corrosion Science, vol. 47, no. 1, pp. 115,
2005.
[12] D. F. A. Lopes, G. L. B. Ramalho, F. N. S. de Medeiros, R. C.
Combining features to improve
S. Costa, and R. T. S. Araujo,
oil spill classification in SAR images, in Structural, Syntactic,
and Statistical Pattern Recognition, vol. 4109 of Lecture Notes
in Computer Science, pp. 928936, Springer, Berlin, Germany,
2006.
[13] G. L. B. Ramalho and F. N. S. Medeiros, Using boosting to
improve oil spill detection in SAR images, in Proceedings of
the 18th International Conference on Pattern Recognition (ICPR
06), pp. 10661069, August 2006.
[14] G. L. B. Ramalho and F. N. S. de Medeiros, Improving
reliability of oilspill detection systems using boosting for highlevel feature selection, in Proceedings of the International
Conference on Image Analysis and Recognition, Lecture Notes
in Computer Science No. 4633, pp. 11721181, August 2007.
[15] A. Jain and D. Zongker, Feature selection: evaluation, application, and small sample performance, IEEE Transactions on
Pattern Analysis and Machine Intelligence, vol. 19, no. 2, pp.
153158, 1997.
[16] S. Livens, Image analysis for material characterisation, Ph.D.
dissertation, Instelling Antwerpen University, 1988.
[17] A. R. Webb, Statistical Pattern Recognition, John Wiley & Sons,
London, UK, 2nd edition, 2002.
Review Article
Fluctuation Analyses for Pattern Classification in
Nondestructive Materials Inspection
A. P. Vieira,1 E. P. de Moura,2 and L. L. Goncalves2
1 Instituto
2 Departamento
1. Introduction
Many nondestructive materials-inspection tools provide
information about material structure in the form of time
series. This is true for ultrasonic probes, acoustic emission,
magnetic Barkhausen noise, among others. Ideally, signatures of material structure are contained in any of those
time series, and extracting that information is crucial for
building a reliable automated classification system, which is
as independent as possible from the operators expertise.
As in any pattern classification task, finding a set of
relevant features is a key step. Common in the literature
are attempts to classify patterns from time series by directly
feeding the time series into neural networks, by measuring
statistical moments, or by employing Fourier or wavelet
transforms. These last two approaches are hindered by the
presence of noise, and by the nonstationary character of
many time series. Sometimes, however, relevant information
is hidden in the noise itself, as this can reflect memory eects characteristic of underlying physical processes.
Analysis of the statistical properties of the time series can
reveal such eects, although global calculations of statistical
moments miss important local details. Here, we show
that properly defined local fluctuation measures of time
2. Fluctuation Analyses
All techniques of fluctuation analysis employed here start
by dividing the signal into time intervals containing
(1)
(t) = 2K f t
H
(2)
1
zi .
iIk
(3)
Zi,k =
z j z,k ,
(4)
j =k
iIk
(5)
2
1
=
zi z,k .
iIk
(6)
Finally, we calculate the rescaled range R,k /S,k , and take its
average over all nonoverlapping intervals, obtaining
()
1 R,k
,
n k S,k
(7)
in which n = L/ is the (integer) number of nonoverlapping intervals of size than can be fit onto a time series of
length L.
For a purely stochastic curve, with no underlying trends,
the rescaled range should satisfy the scaling form
() H ,
(8)
i
z j z ,
(9)
j =1
z =
(10)
i,k = Zi Zi,k ,
(11)
F,k
(12)
iIk
Xj =
1
F,k .
n k
(13)
(14)
(15)
(16)
max
i0 ii0 + 1
zi
min
i0 ii0 + 1
zi ,
(17)
Hk ,
(18)
(19)
j
Yj =
xi ,
i=1
j
yi .
(20)
i=1
3. Pattern-Classification Tools
Having obtained curves of dierent fluctuation estimates
Q() as functions of the time interval size , we make use
of standard pattern-recognition tools in order to group the
signals according to relevant classes. The first step towards
classification is to build feature vectors from one or more
fluctuation analyses of a given signal. In the simplest case,
a set of d fixed interval sizes { j } is selected, and the values
of the corresponding functions Q(j ) at each j , as calculated
for the ith signal, define the feature (column) vector xi of that
signal,
Q(1 )
Q( )
2
xi =
.. .
.
(23)
Q(d )
In our studies, unless stated otherwise, we select as interval
sizes the nearest integers obtained from powers of 21/4 ,
starting with 1 = 4 and ending with d equal to the length
of the shortest series available.
It is also possible to concatenate vectors obtained from
more than one fluctuation analysis to obtain feature vectors
of larger dimension. This usually leads to better classifiers.
The following subsections discuss dierent methods
designed to group feature vectors into relevant classes. All
methods initially select a subset of the available vectors
as a training group in order to build the classifier, whose
generalizability is then tested with the remaining vectors.
This procedure has to be repeated for many distinct choices
of training and testing vectors, as a way to evaluate the
average eciency of the classifier. One can then study the
resulting confusion matrices, which report the percentage of
vectors of a given class assigned to each of the possible classes.
S=
(24)
m=
(25)
SW =
(26)
(27)
NC
Nk
N
k=1
(28)
(29)
4. Applications
4.1. Cast-Iron Microstructure from Ultrasonic Backscattered
Signals. An early application of the ideas described in this
review aimed at distinguishing microstructure in graphite
cast iron through Hurst and detrended-fluctuation analyses
of backscattered ultrasonic signals.
As detailed in [2], backscattered ultrasonic signals were
captured with a 5 MHz transducer, at a sampling rate
of 40 MHz, from samples of vermicular, lamellar, and
spheroidal graphite cast iron. Double-logarithmic plots of
the resulting R/S and DFA calculations, shown in Figure 1,
reveal that in all cases two regimes can be identified,
reflecting short- and long-time structure of the signals,
respectively. From the discussion in Sections 2.1 and 2.2, this
implies that one can define two sets of exponents, related to
the short- and long-time fractal dimensions of the signals,
as estimated from the corresponding values of the Hurst
exponent H and the DFA exponent . See (16).
Lamellar cast iron is readily identified as having smaller
short- than long-time fractal dimension, contrary to both
vermicular and spheroidal cast irons. These latter types, in
turn, can be identified on the basis of the relative values of H
and on the dierent regimes.
As discussed in the following subsections, this fortunate
clear distinction on the basis of a very small set of exponents
is not possible in more general applications. Nevertheless, a
set of relevant features can still be extracted from fluctuation
or fractal analyses by using tools from the pattern recognition
literature.
DF
2.5
RS
RS
< H >= 0.35
0.5
1
1.5
2
log10
2.5
1.5
(a)
2
log10
2.5
(b)
Spheroidal(S) cast iron
2.5
< >= 0.92
2
DF
1.5
0.5
1
1.5
2
log10
2.5
(c)
Figure 1: Double-logarithmic plots of the curves obtained from Hurst (R/S) and detrended-fluctuation (DF) analyses of backscattered
ultrasonic signals propagating in lamellar (a), vermicular (b), and spheroidal (c) cast iron. The values of and H are obtained by
averaging the slopes of all curves in the corresponding intervals, as shown by the solid lines.
4.2. Welding Defects in Metals from TOFD Ultrasonic Inspection. The TOFD (time-of-flight diraction) technique aims
at estimating the size of a discontinuity in a material by
measuring the dierence in time between ultrasonic signals
scattering o the opposite tips of the discontinuity. For
welding-joint inspection, the conventional setup consists of
one emitter and one receiver transducer, aligned on either
side of the weld bead. (Longitudinal rather than transverse
waves are used, for a number of reasons, among which is
higher propagation speed.)
In the case studied in [13], 240 signals of ultrasound
amplitude versus time were captured, with a TOFD setup,
from twelve test samples of steel plate AISI 1020, welded
by the shielded process. (Details on materials and methods
can be found in [14].) The signals used in the study
were extracted from sections with no visible defects in the
welding, and from sections exhibiting lack of penetration,
300
250
250
200
200
150
150
100
100
50
50
100
200
300
400
500
100
200
(a)
300
250
250
200
200
150
150
100
100
50
50
100
200
400
500
300
400
500
(b)
300
300
300
400
500
(c)
100
200
(d)
Figure 2: Typical examples of signals obtained from samples with (a) lack-of-fusion defects, (b) lack-of-penetration defects, (c) porosities,
and (d) no defects. The horizontal axes correspond to the time direction, in units of the inverse sample rate of the equipment.
LF
LP
PO
ND
LF
91.07 (0.37)
2.61 (0.37)
6.43 (0.32)
1.01 (0.15)
LP
1.69 (0.16)
83.96 (0.45)
13.99 (0.47)
2.55 (0.20)
PO
6.88 (0.33)
12.14 (0.41)
72.66 (0.58)
6.92 (0.32)
ND
0.35 (0.08)
1.28 (0.14)
6.92 (0.34)
89.51 (0.40)
Table 2: The same as in Table 1, but now for a majority vote involving classifications based on each fluctuation analysis separately.
LF
LP
PO
ND
LF
87.11 (0.40)
2.04 (0.18)
7.13 (0.34)
2.26 (0.19)
LP
0.64 (0.10)
90.06 (0.40)
19.16 (0.52)
1.38 (0.17)
PO
6.96 (0.33)
5.88 (0.34)
65.18 (0.61)
7.81 (0.34)
ND
5.28 (0.27)
2.01 (0.18)
8.53 (0.35)
88.54 (0.41)
Table 3: Average percentage of correctly classified testing signals coming from toothless and normal gears working in the absence of load.
rpm
Toothless
Normal
400
69.4 1.9
69.3 1.8
600
86.3 1.5
100
800
96.2 0.7
100
1000
49.2 2.9
64.1 2.4
1200
68.8 2.1
91.5 1.2
1400
48.2 2.5
45.1 2.5
than with the other two lengths. This can be traced to the
fact that, as illustrated by Figure 5, distinguishing features
(such as average slopes and discontinuities) between curves
corresponding to dierent transfer modes tend to happen
at intermediate time scales. For a given length, detrendedcross-correlation analysis of voltage and current signals
yields an intermediate classification eciency as compared
to either voltage or current signals analyzed separately.
The best classifier is obtained with the Hurst analysis of
signals containing L = 2048 points, yielding a negligible
classification error of 0.1%.
In contrast, as shown in the bottom two rows of Table 5,
similar classifiers in which feature vectors are defined by the
full Fourier spectra of the various signals yield much larger
classification errors, and at a much higher computational
cost (since the size of feature vectors scales as L, whereas for
fluctuation analyses it scales as log L).
4.5. Stainless Steel Microstructure from Magnetic Measurements. Barkhausen noise is a magnetic phenomenon produced when a variable magnetic field induces magnetic
domain wall movements in ferromagnetic materials. These
movements are discrete rather than continuous, and are
caused by defects in the material microstructure, generating
magnetic pulses that can be measured by a coil placed on the
material surface.
Magnetic Barkhausen noise (BN) and magnetic flux
(MF) measurements were performed on samples of stainlesssteel steam-pressure vessels, as detailed in [17]. These
presented coarse ferritic-pearlitic phases (named stage A)
before degradation. Owing to temperature eects, two different microstructures were obtained from pearlite that has
partially (stage BC) or completely (stage D) transformed
to spheroidite. Measurements were performed by using a
sinusoidal magnetic wave of frequency 10 Hz, each signal
consisting of 40 000 points, with a sampling rate of 200 kHz.
A total of 144 signals were captured, 40 signals corresponding
to stage A, 88 to stage BC, and 16 to stage D. Typical signals
are shown in Figure 6. Notice that, as regards the magnetic
flux, the dierence between signals from the various stages
seems to lie on the intensity of the peaks and troughs,
Amplitude
0
0
0.2
0.4
500
1000
Time
1500
2000
2
log10
0.2
log10 F()
Amplitude
0.4
2
0.6
4
500
1000
Time
1500
2000
0.1
0.2
0.3
2
log10
log10 F()
Amplitude
0.4
500
1000
Time
1500
2000
2
log10
Figure 3: Representative signals and DFA curves obtained from the three types of gear, working under load at a rotation frequency of
1400 rpm. In the signal plots, time is measured in units of the inverse sampling rate.
Table 4: The same as in Table 3, but now for gears working under load.
rpm
Toothless
Normal
400
100
94.8 0.8
600
100
97.5 0.7
800
100
98.5 0.5
1000
100
95.6 0.7
1200
100
81.3 1.7
1400
100
100
9
400
40
Dip
350
30
Current (A)
Voltage (V)
300
20
250
200
10
150
0
0.1
0.2
Time (s)
0.3
100
0.4
0.1
(a)
0.2
Time (s)
0.3
0.4
0.3
0.4
0.3
0.4
(b)
200
34
Globular
180
Current (A)
Voltage (V)
32
30
160
140
28
120
26
0
0.1
0.2
Time (s)
0.3
100
0.4
0.1
(c)
0.2
Time (s)
(d)
199
25
Spray
198
Current (A)
Voltage (V)
24
23
22
196
195
21
20
197
0.1
0.2
Time (s)
(e)
0.3
0.4
194
0.1
0.2
Time (s)
(f)
Figure 4: Examples of voltage (left) and current (right) time series obtained during the welding process under dip (top), globular (center),
and spray (bottom) metal-transfer modes.
10
4
Hurst I
Hurst V
2.5
log R/S
log R/S
2
2
1.5
1
1
0.5
0
2
log10
(a)
2
log10
(b)
5
DFA I
DFA V
4
3
log F
log F
1
2
0
1
1
2
log
(c)
2
log
(d)
DCC I V
log FDCC
4
2
0
2
4
2
log10
Dip
Globular
Spray
(e)
Figure 5: Examples of curves obtained from Hurst (top), detrended-fluctuation (center), and detrended-cross-correlation (bottom) analyses
to current (I) and voltage (V ) sample signals obtained under dip (top), globular (center), and spray (bottom) metal-transfer modes.
Logarithms are in base 10, and the time window size is measured in tenths of a millisecond.
11
0.3
0.2
0.1
0
0.1
1
0
1
2
0.2
10000
20000
Time (5 s)
30000
40000
Stage A
Stage BC
Stage D
10000
20000
Time (5 s)
30000
40000
Stage A
Stage BC
Stage D
(a)
(b)
Figure 6: Typical signals of (a) magnetic flux and (b) Barkhausen noise obtained from stainless-steel samples at dierent stages of
microstructural degradation. Plots in (b) have been vertically shifted for clarity.
Table 5: Average percentage classification errors of testing voltage (V ) and current (I) signals containing L points, produced by classifiers
based on Hurst, detrended-fluctuation (DF), or detrended-cross-correlation (DCC) analyses. Also shown are results for classifiers based on
Fourier spectra.
L
DF, V
Hurst, V
DF, I
Hurst, I
DCC, V + I
Fourier, V
Fourier, I
512
3.1 0.2
6.5 0.4
2.1 0.2
14.5 0.5
3.2 0.3
23.6 0.9
22.7 2.5
1024
2.2 0.4
3.1 0.5
0.6 0.2
5.4 0.6
1.5 0.3
21.8 0.8
27.5 1.9
2048
3.6 0.7
0.1 0.1
0.5 0.3
4.0 0.9
2.4 0.7
18.7 1.2
8.7 0.9
4096
5.3 1.3
0.7 0.7
1.6 1.1
2.7 1.3
7.7 1.3
36.7 2.3
14.5 1.9
Table 6: Average percentage of correctly classified testing signals coming from stainless-steel samples in dierent degradation stages.
Classifiers employed detrended-fluctuation (DFA), Hurst (RS), or Fourier spectral (FS) analyses on either Barkhausen noise (BN) or
magnetic flux (MF).
Stage A
Stage BC
Stage D
DFA/BN
54.8 1.9
57.6 1.2
68.4 2.9
RS/BN
34.2 1.6
49.5 1.5
31.0 2.7
DFA/MF
83.0 1.3
87.2 0.8
96.4 1.5
RS/MF
90.5 1.0
92.5 0.6
98.0 1.4
FS/MF
67.8 1.7
77.0 1.1
78.6 2.9
12
The best performance is obtained by the Hurst classifier,
with maximum error of about 10%, followed by the DFA
classifier, with a maximum error around 17%. Somewhat
surprisingly, in view of the long-time regularity of the
magnetic flux signals evident in Figure 6, the Fourierspectral classifier shows the worst performance, with an
average classification error of 25%.
5. Conclusions
We have reviewed and supplemented recent work on application of fluctuation analysis as a pattern-classification tool in
nondestructive materials inspection. This approach has been
shown to lead to very ecient classifiers, with a performance
comparable, and usually quite superior, to more traditional
approaches based, for instance, on Fourier transforms. The
present approach also requires less computational eort to
achieve a given eciency, which would be an important issue
when building automated inspection systems for field work.
An extension of the present approach to defect recognition from radiographic or ultrasonic images can be achieved
based on generalizations of the fluctuation analyses to measure surface roughness [18, 19]. Given any two-dimensional
image, a corresponding surface can be built by a color-toheight conversion procedure, and mathematical analyses can
then be performed.
Acknowledgments
The authors acknowledge financial support from the Brazilian agencies FUNCAP, CNPq, CAPES, FINEP (CT-Petro),
and Petrobras (Brazilian oil company).
References
[1] P. Barat, Fractal characterization of ultrasonic signals from
polycrystalline materials, Chaos, Solitons & Fractals, vol. 9, no.
11, pp. 18271834, 1998.
[2] J. M. O. Matos, E. P. de Moura, S. E. Kruger, and J. M. A.
Rebello, Rescaled range analysis and detrended fluctuation
analysis study of cast irons ultrasonic backscattered signals,
Chaos, Solitons & Fractals, vol. 19, no. 1, pp. 5560, 2004.
[3] F. E. Silva, L. L. Goncalves, D. B. B. Fereira, and J. M. A.
Rebello, Characterization of failure mechanism in composite
materials through fractal analysis of acoustic emission signals,
Chaos, Solitons & Fractals, vol. 26, no. 2, pp. 481494, 2005.
[4] H. E. Hurst, Long-term storage capacity of reservoirs,
Transactions of the American Society of Civil Engineers, vol. 116,
pp. 770799, 1951.
[5] J. Feder, Fractals, Plenum Press, New York, NY, USA, 1988.
[6] B. B. Mandelbrot and J. W. van Ness, Fractional brownian
motion, fractional noises and applications, SIAM Review, vol.
10, pp. 422437, 1968.
[7] P. S. Addison, Fractals and Chaos, IOP, London, UK, 1997.
[8] A. R. Webb, Statistical Pattern Recognition, John Wiley & Sons,
West Sussex, UK, 2nd edition, 2002.
[9] C.-K. Peng, S. V. Buldyrev, S. Havlin, M. Simons, H. E.
Stanley, and A. L. Goldberger, Mosaic organization of DNA
nucleotides, Physical Review E, vol. 49, no. 2, pp. 16851689,
1994.
Research Article
A Study of Concrete Hydration and Dielectric Relaxation
Mechanism Using Ground Penetrating Radar and Short-Time
Fourier Transform
W. L. Lai, T. Kind, and H. Wiggenhauser
BAM, Federal Institute for Materials Research and Testing, Unter den Eichen 87, 12205 Berlin, Division VIII.2, Germany
Correspondence should be addressed to W. L. Lai, wai-lok.lai@bam.de
Received 7 January 2010; Revised 6 June 2010; Accepted 5 July 2010
Academic Editor: Joao Marcos A. Rebello
Copyright 2010 W. L. Lai et al. This is an open access article distributed under the Creative Commons Attribution License,
which permits unrestricted use, distribution, and reproduction in any medium, provided the original work is properly cited.
Ground penetrating radar (GPR) was used to characterize the frequency-dependent dielectric relaxation phenomena in ordinary
Portland cement (OPC) hydration in concrete changing from fresh to hardened state. The study was experimented by measuring
the changes of GPR A-scan waveforms over a period of 90 days, and processed the waveforms with short-time Fourier transform
(STFT) in joint time-frequency analysis (JTFA) domain rather than a conventional time or frequency domain alone. The signals
of the direct wave traveled at the concrete surface and the reflected wave from an embedded steel bar were transformed with STFT,
in which the changes of peak frequency over ages were tracked. The peak frequencies were found to increase with ages and the
patterns were found to match closely with primarily the well-known OPC hydration process and secondarily, the evaporation
eect. The close match is contributed to the simultaneous eects converting free to bound water over time, on both conventional
OPC hydration and dielectric relaxation mechanisms.
1. Introduction
1.1. Study of Materials Using GPR. Nondestructive testing
methods attract many attentions in civil engineering applications rapidly in recent years [1, 2]. One of the most
widely used methods is ground penetrating radar (GPR)
which images and sees through material structures [3, 4].
Literatures have been reported extensively in the application
on concrete structures such as bridges [4], railway ballast
[5, 6], dams [3], and brick [7]. Principles of GPR were widely
documented and reviewed [1, 2, 8, 9]. While most published
works and field applications focused on locating internal
characteristics (such as concrete reinforcement, voids and
layer thickness), there is comparatively little eort to study
the material properties which usually refers to the host
materials, that is, concrete. From materials perspective, the
mechanism converting water from free to bound form in
OPC concrete was characterized with GPR by measuring the
change in the real part of dielectric permittivity/velocity in
time domain [10]. In this paper, we attempted to characterize
the same mechanism by measuring the changes of frequency-
2
(C-scan) after alignment of multiple B-scans [4]. B-scans
and C-scans have been frequently used to delineate and
interpret the internal characteristics of the spatial material
structures, provided that the features exhibit contrasting
electric properties compared to the host material. However,
the B- and C-scans operated in time domain are unable
to study the frequency features of the time-varying GPR
signals, which are largely aected by the spatial or temporal changes of the concrete properties. In other words,
when the interpretation of GPR signals is extended from
objects location and spatial characterization to the study
of material properties, the time domain method may not
be adequate because the characteristic of the time-varying
frequency is not taken into account. As such, this paper
adopted a two-dimensional joint time-frequency analysis
(JTFA) to transform and evaluate the signal. This method
is based on and back to the most fundamental A-scan
waveforms.
In the JTFA domain, we migrated the A-scans (obtained
from 1.5 and 2.6 GHz GPR antennae) from one dimension
(time) to two dimensions (time and frequency). In such
domain, the localized frequency changes are analyzed jointly
with the localized amplitude changes registered in the radar
time axis (i.e., A-scan). Short-time fourier transform (STFT)
was used as the computation algorithm because it allows the
frequency spectra to be revealed in a preselected, stepped,
and short-time windows compared to traditional Fourier
transform. The sweeping of this short-time window over
the radar time axis stacks the localized frequency spectra
centered at each radar time points, and ultimately compiles
a 2D time-frequency plot with time and frequency being
the x- and y-axis. In this paper, the usefulness of this
methodology is experimented with the well known cement
hydration mechanism in concrete. It is well-known that
concrete hydrates free water in concrete over the curing
processes, and changes the wave propagation velocity, real
part of permittivity and dielectric contrast [10]. In this
paper, we measured the temporal changes of the A-scans
over a period from an initial fresh state to a hardened
state up to 90 days. Then the A-scans were processed and
transformed to 2D time-frequency plots to observe the lateral
frequency changes for both direct wave and an embedded
steel bar reflector. These signals were further characterized
and quantified by determination of the peak frequency, and
justified by the dielectric relaxation and OPC hydration
theories.
2. Theories
2.1. Ground Penetrating Radar (GPR). A GPR is an instrumentation tool that produces high-frequency electromagnetic waves propagating through the host materials. These
waves are subjected to a combined eect of transmission,
reflection, refraction, scattering, absorption, and attenuation
by the materials which are usually dielectrics, such as
concrete, soils, and rocks. The reflected energy is then
captured, recorded, and digitized by a receiving antenna, and
is registered as an amplitude record over the time windows
in a scale of nanosecond, or simply known as an A-scan.
81
70
Frequency associated
with dielectric dispersion
60
50
Relaxation
frequency
40
30
20
10
0
108
109
1010
1011
1012
Real part
Imaginary part
x(t)(t )e jt dt,
(1)
antenna
1.5 GHz
E-field
2 GHz
E-field
antenna
Concrete
500 mm
500 mm
(a)
GPR antenna
100 mm
dia.25 mm steel bar
500 mm
(b)
Figure 2: Plan view (a) and section (b) of the concrete specimen.
4. Data Processing
6000
10000
4000
5000
0
2000 0
10
20
30
4000
40
50
Time (points)
6000
8000
Amplitude
Amplitude
2000
0
0
10
20
30
40
50
60
5000
1000
10000
Time (points)
25 mm
25 mm
15000
A-scan
1st derivative of the A-scan
2st derivative of the A-scan
Bar @25 mm
Bar @50 mm
Time zero at inflection point
(a)
3000
Amplitude
2500
2000
Time zero at the
inflection point of A-scan
1500
1000
500
0
20
500
Time (points)
22
24
26
28
30
A-scan
1st derivative of the A-scan
2st derivative of the A-scan
(b)
Figure 4: (a) 1.5 GHz A-scan and its first and second derivative. (b)
2.6 GHz A-scan and its first and second derivative (close-up view).
3.5G
3.5E + 09
2.5E + 09
2E + 09
1.5E + 09
3G
Frequency (Hz)
Frequency (Hz)
3E + 09
150k
4G
4E + 09
2.5G
75k
2G
1.5G
1G
1E + 09
500 M
5E + 09
0
500P
0E + 00
1n
2n
(a)
3n
4n
5n
(b)
6000
Amplitude
4000
2000
0
0.5
2000
0.5
1.5
2.5
3.5
4.5
Time (ns)
4000
6000
8000
(c)
Figure 7: Frequency spectrum at 2.1 ns (a), 2.6 GHz STFT spectrogram (b), and GPR A-scan (c), at day 7.
5. Findings
5.1. Time Domain A-Scans. Figure 9 shows the A-scans
collected at dierent days after fresh mixing. The collection
of data was started once every day from fresh state to day
7, and then once at days 14, 28, 56, and 90 Positions of the
bar reflections were identified by comparing the reflections
when the antenna was captured on top of the bar and on
top of other locations with plain concrete only. The bar
reflections were ill defined in time domain from the fresh
state to day 4 because high water content caused more
attenuation suered by the high-frequency GPR antennae.
This is improved after day 5, where the reflections started to
be well defined, and therefore only data from days 5 to 7,
14, 28, 56, and 90 are presented in Figures 9 and 10. Along
with age, the reflections from the bar traveled much faster
(shorter time) and exhibited a larger intensity. It is because
the majority of free water was progressively consumed and
bound to formulate the structure of calcium silicate hydrate,
4G
3G
2.5 G
2G
1.5 G
Bar
Time zero
Bar
3.5 G
Frequency (Hz)
3G
Time zero
Frequency (Hz)
3.5 G
150 k
4G
2.5 G
75 k
2G
1.5 G
1G
1G
500 M
500 M
500 p
2n
4n
500 p
5n
2n
Time (s)
4n
5n
Time (s)
(a)
(b)
Magnifation factor
35
30
25
20
15
10
5
0
0.5
0.5
1.5
2.5
3.5
4.5
Time (s)
(c)
Figure 8: Original STFT spectrogram (a) and the enhanced STFT spectrogram (b) after applying Gaussian gain in JTFA domain at every
frequency bin (c).
Direct wave
6000
700
4000
300
0
0.5
1.5
2.5
Time (ns)
4000
Amplitude
Amplitude
2000
2000 0
Bar reflection
500
Bar reflection
100
100
1.8
2.2
2.4
2.6
2.8
3
Time (ns)
300
6000
8000
500
10000
700
Day 5
Day 6
Day 7
Day 14
Day 28
Day 56
Day 90
Day 5
Day 6
Day 7
Day 14
Day 28
Day 56
Day 90
Figure 9: 2.6 GHz A-scan in time domain (reflections captured when the antenna was on top of the bar).
1.8 GHz for 2.6 GHz FT (shaded region in Figure 11), and
from 800 MHz to 1.2 GHz for 1.5 GHz FT (shaded region
in Figure 12). The localities beyond the upper threshold of
middle locality are termed as the highest frequency locality.
The FT middle frequency locality is corresponding to the bar
1500
Bar reflection
500
2000 0
4000
0.5
1.5
2.5
Time (ns)
6000
Amplitude
Amplitude
4000
2000
0
Bar reflection
1000
0
1.6
500
1.8
2.2
2.4
2.6
2.8
8000
Time (ns)
1000
10000
1500
Day 28
Day 56
Day 90
Day 5
Day 6
Day 7
Day 14
Day 5
Day 6
Day 7
Day 14
Day 28
Day 56
Day 90
Figure 10: 1.5 GHz A-scans in time domain (reflections captured when the antenna was on top of the bar).
Lowest locality
Middle locality
Lowest locality
Highest locality
Steel bar
Direct
wave
0.8
Normalized amplitude
Normalized amplitude
Clutter and
signal ringing
1
Middle locality
0.6
0.4
0.2
0
1000
Initial
Day 1
Day 2
Day 3
Day 4
Day 5
2000
Frequency (MHz)
3000
4000
Day 6
Day 7
Day 14
Day 28
Day 56
Day 90
Highest locality
3000
3500
4000
Day 7
Day 14
Day 28
Day 56
Day 90
4G
Initial
4G
Day 1
4G
Day 2
4G
3G
3G
3G
3G
2G
2G
2G
2G
1G
1G
1G
1G
750p
4G
2n
Day 4
4n
750p
4G
2n
Day 5
4n
750p
4G
4n
750p
4G
3G
3G
3G
2G
2G
2G
2G
1G
1G
1G
1G
4G
2n
Day 14
4n
750p
4G
2n
Day 28
4n
750p
4G
4n
750p
4G
3G
3G
3G
2G
2G
2G
2G
1G
1G
1G
1G
2n
4n
750p
2n
4n
750p
50k
2n
Day 7
4n
2n
Day 90
4n
2n
4n
2n
Day 56
3G
750p
100k
2n
Day 6
3G
750p
Day 3
2n
4n
750p
Figure 13: 2.6 GHz STFT spectrogram. x-axis is time (s), y-axis is frequency (Hz), and z-axis is the signal intensity with the same scale
for all STFT spectrograms. The broken vertical line is the location of time zero. Gaussian gains were applied at the peak position of the bar
reflections, which is approximately at 2.0 ns.
10
4G
Initial
4G
Day 1
Day 2
4G
3G
3G
3G
3G
2G
2G
2G
2G
1G
1G
1G
1G
0
750p
4G
2n
Day 4
4n
750p
4G
2n
Day 5
4n
Day 3
4G
150k
750p
2n
Day 6
4G
750p
4n
4G
3G
3G
3G
3G
2G
2G
2G
2G
1G
1G
1G
1G
0
750p
4G
0
2n
Day 14
4n
750p
4G
0
2n
Day 28
4n
2n
Day 56
4G
750p
4n
4G
3G
3G
3G
2G
2G
2G
2G
1G
1G
1G
1G
0
2n
4n
750p
2n
4n
2n
Day 7
4n
2n
Day 90
4n
2n
4n
750p
3G
750p
300k
750p
2n
750p
4n
Figure 14: 1.5 GHz STFT spectrogram. x-axis is time (s), y-axis is frequency (Hz), and z-axis is the signal intensity with the same scale
for all STFT spectrograms. The broken vertical line is the location of time zero. Gaussian gains were applied at the peak position of the bar
reflections, which is approximately at 2.5 ns.
Cement hydration
1
0.9
1
0.9
0.8
0.7
0.6
0.5
0.4
0.3
0.2
0.1
0.8
0.7
0.6
0.5
0.4
0.3
0.2
0.1
Normalized amplitude
Normalized amplitude
Cement hydration
0
0
1000
2000
3000
4000
500
Frequency (MHz)
(a)
1500
2000 2500
3000
3500
4000
Frequency (MHz)
Day 6
Day 7
Day 14
Day 28
Day 56
Day 90
Initial
Day 1
Day 2
Day 3
Day 4
Day 5
1000
Day 6
Day 7
Day 14
Day 28
Day 56
Day 90
Initial
Day 1
Day 2
Day 3
Day 4
Day 5
(b)
Figure 15: (a) 2.6 GHz STFT frequency spectra centered at the peak of direct wave. (b) 2.6 GHz STFT frequency spectra centered at the peak
of steel bar.
11
See remark A
1
0.8
0.7
0.6
0.5
0.4
0.3
0.2
0.1
0
Normalized amplitude
Normalized amplitude
1
0.9
500
1000
1500
Frequency (MHz)
2000
2500
Day 6
Day 7
Day 14
Day 28
Day 56
Day 90
Initial
Day 1
Day 2
Day 3
Day 4
Day 5
0.9
0.8
0.7
0.6
Cement hydration
0.5
0.4
0.3
0.2
0.1
0
500
1000
1500
Frequency (MHz)
2500
Day 6
Day 7
Day 14
Day 28
Day 56
Day 90
Initial
Day 1
Day 2
Day 3
Day 4
Day 5
(a)
2000
(b)
Figure 16: (a) 1.5 GHz STFT frequency spectra centered at the peak of direct wave. (b) 1.5 GHz STFT frequency spectra centered at the peak
of steel bar reflection. In time (Figure 10) and JTFA domains (Figure 14), low frequency component rises since bar reflection approached to
the trails of low-frequency direct wave which is stronger in later than early days.
2000
1800
1600
1400
1200
1000
800
600
0
20
40
60
80
100
Days
2.6 GHz ground wave
2.6GHz bar
Figure 17: Change of peak frequency with the early age curing of
concrete.
12
Case 2
Case 1
Amplitude
Day 56
Amplitude
Day 7
Amplitude
Initial
Case 3
Frequency
Frequency
Frequency
Normalized amplitude
0.9
0.8
0.7
0.6
0.5
0.4
0.3
0.2
0.1
0
0
500
1000
1500
2000
2500
3000
3500
4000
Frequency (MHz)
Initial
Day 7
Day 56
Figure 18: Spectrum change due to change from free to bound water in early-aged concrete. Case 1: All water is in free form. Case 2: Water is
hydrated and bounded; free water still remains and absorbed in capillary. Case 3: More water is hydrated and bounded, but smaller amount
of free water still exists in the capillary.
10
20
30
40
50
60
70
80
13
90
100
20
40
60
80
100
120
140
160
180
200
Time (days)
Cover-cured concrete cured in room
temperature & humidiity
Water cured (100% humidity)
6. Conclusions
The STFT method reported in this paper was used to
analyze the time-varying GPR signals. This was achieved
by measuring the changes of frequency spectra centered at
in each time step over the radar time window, rather than
applying classical fourier transform depicting the frequency
spectra over the whole time window. With appropriate
signal processing method, GPR can be used to study the
hydration properties in OPC concrete. It is because the free
or bound form of water yields a dierent but simultaneous
eect on both dielectric relaxation and cement hydration
mechanisms. The method can be further extended to
study dierent hydration mechanisms of dierent concretes
nondestructively, and on a much larger volume of materials
compared to the small transmission line method. Most
importantly, the method is not limited to OPC hydration,
but may also be applied in various types of cement, water to
cement ratios, and materials replacing part of OPC (such as
fly ash and silica fume) that exhibit pozzolanic reactions.
References
[1] J. H. Bungey, Sub-surface radar testing of concrete: a review,
Construction and Building Materials, vol. 18, no. 1, pp. 18,
2004.
[2] V. M. Malhotra and N. J. Carino, Eds., Handbook on Nondestructive Testing of Concrete, CRC Press, Boca Raton, Fla, USA,
2nd edition, 2006.
[3] H. C. Rhim, Condition monitoring of deteriorating concrete
dams using radar, Cement and Concrete Research, vol. 31, no.
3, pp. 363373, 2001.
[4] C. Kohl and D. Streicher, Results of reconstructed and fused
NDT-data measured in the laboratory and on-site at bridges,
Cement and Concrete Composites, vol. 28, no. 4, pp. 402413,
2006.
[5] T. Kind, Study of the behaviour of a multi-oset antenna
array on railway ballast, in Proceedings of the 12th International Conference on Ground Penetrating Radar (GPR 08),
Birmingham, UK, 2008.
[6] I. Al-Qadi, W. Xie, and R. Roberts, Optimization of antenna
configuration in multiple-frequency ground penetrating radar
system for railroad substructure assessment, NDT and E
International, vol. 43, no. 1, pp. 2028, 2010.
[7] Ch. Maierhofer and J. Wostmann, Investigation of dielectric
properties of brick materials as a function of moisture and
salt content using a microwave impulse technique at very high
frequencies, NDT and E International, vol. 31, no. 4, pp. 259
263, 1998.
[8] D. M. McCann and M. C. Forde, Review of NDT methods in
the assessment of concrete and masonry structures, NDT and
E International, vol. 34, no. 2, pp. 7184, 2001.
[9] A. P. Annan, Electromagnetic principles of ground penetrating radar, in Ground Penetrating Radar: Theory and
Applications, H. M. Jol, Ed., chapter 1, Elsevier, Amsterdam,
The Netherlands, 2009.
[10] W. L. Lai, S. C. Kou, W. F. Tsang, and C. S. Poon, Characterization of concrete properties from dielectric properties using
ground penetrating radar, Cement and Concrete Research, vol.
39, no. 8, pp. 687695, 2009.
[11] K. Gorur, M. K. Smit, and F. H. Wittmann, Microwave study
of hydrating cement paste at early age, Cement and Concrete
Research, vol. 12, no. 4, pp. 447454, 1982.
[12] M. A. Rzepecka, M. A. L. Hamid, and A. H. Soliman, Monitoring of concrete curing process by microwave terminal
measurements, IEEE Transactions on Industrial Electronics,
vol. 19, no. 4, pp. 120125, 1972.
[13] V. A. Beek, Dielectric properties of young concrete, Ph.D. thesis,
Delft University, 2000.
[14] M. N. Soutsos, J. H. Bungey, S. G. Millard, M. R. Shaw,
and A. Patterson, Dielectric properties of concrete and their
influence on radar testing, NDT and E International, vol. 34,
no. 6, pp. 419425, 2001.
[15] G. Klysz and J.-P. Balayssac, Determination of volumetric
water content of concrete using ground-penetrating radar,
Cement and Concrete Research, vol. 37, no. 8, pp. 11641171,
2007.
[16] S. Laurens, J. P. Balayssac, J. Rhazi, G. Klysz, and G. Arliguie,
Non-destructive evaluation of concrete moisture by GPR:
experimental study and direct modeling, Materials and
Structures/Materiaux et Constructions, vol. 38, no. 283, pp.
827832, 2005.
[17] W. L. Lai and W. F. Tsang, Characterization of pore systems
of air/water-cured concrete using ground penetrating radar
(GPR) through continuous water injection, Construction and
Building Materials, vol. 22, no. 3, pp. 250256, 2008.
[18] J. L. Davis and A. P. Annan, Ground-penetrating radar
for high-resolution mapping of soil and rock stratigraphy,
Geophysical Prospecting, vol. 37, no. 5, pp. 531551, 1989.
14
[19] M. D. Knoll, A petrophysical basis for ground-penetrating
radar and very early time electromagnetics, Ph.D. thesis, The
University of British Columbia, 1996.
[20] G. C. Topp, J. L. Davis, and A. P. Annan, Electromagnetic
determination of soil water content: measurements in coaxial
transmission lines, Water Resources Research, vol. 16, no. 3,
pp. 574582, 1980.
[21] P. Debye, Polar Molecules, Chemical Publication Company,
1929.
[22] P. Hoeskstra and A. Delaney, Dielectric properties of soils
at UHF and microwave frequencies, Journal of Geophysical
Research, vol. 79, no. 11, pp. 16991708, 1974.
[23] Concrete Society, Guidance on radar testing of concrete
structures, Tech. Rep. 48, Concrete Society, 1997.
[24] W. L. Lai, W. F. Tsang, H. Fang, and D. Xiao, Experimental
determination of bulk dielectric properties and porosity of
porous asphalt and soils using GPR and a cyclic moisture
variation technique, Geophysics, vol. 71, no. 4, pp. K93K102,
2006.
[25] P. Gu and J. J. Beaudoin, Dielectric behaviour of hardened
cementitious materials, Advances in Cement Research, vol. 9,
no. 33, pp. 18, 1997.
[26] D. J. Daniels, Ed., Ground Penetrating Radar, The Institution
of Electrical Engineers, London, UK, 2nd edition, 2004.
[27] N. R. Peplinski, F. T. Ulaby, and M. C. Dobson, Dielectric
properties of soils in the 0.3-1.3-GHz range, IEEE Transactions on Geoscience and Remote Sensing, vol. 33, no. 3, pp. 803
807, 1995.
[28] A. M. Neville, Properties of Concrete, Longman, Harlow, UK,
1995.
[29] J. PH. Poley, J. J. Nooteboom, and P. J. de Waal, Use of V.H.F.
dielectric measurements for borehole formation analysis, Log
Analyst, vol. 19, no. 3, pp. 830, 1978.
[30] A. Von Hippel, Ed., Dielectric Materials and Applications,
Technology Press of MIT, Cambridge, Mass, USA, 1995.
[31] J. Q. Shang and J. A. Umana, Dielectric constant and
relaxation time of asphalt pavement materials, Journal of
Infrastructure Systems, vol. 5, no. 4, pp. 135142, 1999.
[32] K. S. Cole and R. H. Cole, Dispersion and absorption in
dielectrics I. Alternating current characteristics, The Journal
of Chemical Physics, vol. 9, no. 4, pp. 341351, 1941.
[33] G. P. de Loor, Dielectric properties of wet materials, IEEE
Transactions on Geoscience and Remote Sensing, vol. 21, no. 3,
pp. 364369, 1982.
[34] J. B. Hasted, Aqueous Dielectrics, Chapman and Hall, London,
UK, 1973.
[35] N. J. Cassidy, Electrical and magnetic properties of rocks,
soils and fluids, in Ground Penetrating Radar: Theory and
Applications, H. M. Jol, Ed., chapter 2, Elsevier, Amsterdam,
The Netherlands, 2009.
[36] J. R. Wang and T. J. Schmugge, An empirical model for the
complex dielectric permittivity of soils as a function of water
content, IEEE Transactions on Geoscience and Remote Sensing,
vol. 18, no. 4, pp. 288295, 1980.
[37] J. A. Huisman, S. S. Hubbard, J. D. Redman, and A. P. Annan,
Measuring soil water content with ground penetrating radar:
a review, Vadose Zone Journal, vol. 2, pp. 476491, 2003.
[38] L. J. West, K. Handley, Y. Huang, and M. Pokar, Radar
frequency dielectric dispersion in sandstone: implications for
determination of moisture and clay content, Water Resources
Research, vol. 39, no. 2, pp. 10261037, 2003.
Research Article
Strain and Cracking Surveillance in Engineered Cementitious
Composites by Piezoresistive Properties
Jia Huan Yu1 and Tsung Chan Hou2
1 School
2 Department
1. Introduction
Engineered Cementitious Composite (ECC) is an ultraductile fiber reinforced cement based composite which has
metal-like features when loaded in tension and exhibits
tough, strain hardening behavior in spite of low fiber volume
fraction. The uniaxial stress-strain curve shows a yield point
followed by strain-hardening up to several percent of strain,
resulting in a material ductility of at least two orders of
magnitude higher than normal concrete or standard fiber
reinforced concrete [1]. ECC provides crack width to below
100 m even when deformed to several percent tensile strain
(Figure 1). Fiber breakage is prevented and pullout from the
matrix is enabled instead, leading to tensile strain capacity
in excess of 6% for PVA-ECC containing 2% by volume Poly
Vinyl Alcohol (PVA) fiber which is a unique implementation
by Yu and Dai [2].
Cracking in cementitious composite can result from
a variety of factors including externally applied loads,
shrinkage, and poor construction methods. Identification of
cracks can be used to evaluate the long-term sustainability
of structural elements made of cementitious composite. For
120
100
80
60
40
20
Stress (MPa)
0
0
Strain (%)
Stress-strain
Crack width
(a)
(b)
Figure 1: Typical stress-strain-crack width relationship and saturated crack pattern of PVA-ECC.
1000
35 days
900
800
Resistivity (kohmcm)
28 days
700
21 days
14 days
600
7 days
500
400
300
1 day
200
100
0
0
200
400
600
Time (s)
Cement
1.0
Silica Sand
0.8
Fly Ash
1.2
Water
0.66
Superplasticizer
0.013%
Section A-A
12 cm
7.51.25cm2
Copper
electrode
2 cm 2 cm
30 cm
2 cm 2 cm
Aluminum
grip plates
5
880
4
D
E
Resistivity (kohmcm)
Stress (MPa)
840
F
800
D
760
720
0.4
0.8
Strain (%)
1.2
1.6
0.4
0.8
Strain (%)
1.2
1.6
(a)
(b)
(c)
(d)
(e)
(f)
(g)
Figure 7: (a) Photo of the specimen after crack localization (at point G); (b)(g), Cracking patterns at loading point B through G,
respectively.
The higher initial resistivity encountered as the specimens cure can be easily explained. Since more and more
ions are trapped by the hardening hydration byproducts, it
is harder to mobilize the ions, which is consistent with a
higher resistivity. The electric properties of the cementitious
material are characterized chiefly by their initial resistivity at
early stage.
A-B
6.55
B-C
9.53
C-D
13.39
D-E
11.64
E-F
8.32
F-G
12.58
5. Conclusion
This study exploits the piezoresistive properties of engineered
cementitious composites (ECCs) so that they can be used
as their own sensors to quantify the resistivity-strain relationship. ECC plate specimens were monotonically loaded
in axial tension to induce strain hardening behavior in the
material. As a result of linear changes in electrical resistance
due to tension strain, ECC specimens could potentially selfmeasure their strain in the field. The resistivity of ECC
specimens at dierent times after casting was monitored by
4-point probe resistivity measurement. The initial resistivity
changes with hydration degree and increases with DC
polarization. An interesting feature of the material lies in
the detectable change in resistance-strain sensitivity when
strain hardening initiates. The change in piezoresistivity
correlates the cracking in the ECC matrix and results in a
nonlinear change in the material conductivity. Additional
work is underway exploring the theoretical foundation for
ECC piezoresistive behavior.
Acknowledgment
Financial supports from Laboratory of Novel Building Materials Manufacturing and Inspection in Shenyang Jianzhu
Universiry are gratefully acknowledged. The authors would
like to express their gratitude to Professor V. C. Li and J. P.
Lynch, University of Michigan, for their helpful discussion
on properties of ECC.
References
[1] J. H. Yu and V. C. Li, Research on production, performance
and fibre dispersion of PVA engineering cementitious composites, Materials Science and Technology, vol. 25, no. 5, pp. 651
656, 2009.
[2] J. H. Yu and L. Dai, Strain rate and interfacial property eects
of random fibre cementitious composites, Journal of Strain
Analysis for Engineering Design, vol. 44, no. 6, pp. 417425,
2009.
[3] S. P. Shah and S. Choi, Nondestructive techniques for studying
fracture processes in concrete, International Journal of Fracture,
vol. 98, no. 3-4, pp. 351359, 1999.
[4] S. Mindess, Acoustic emission methods, in Handbook on
Nondestructive Testing of Concrete, V. M. Malhotra and N. J
Carino, Eds., CRC Press, Boca Raton, Fla, USA, 2004.
[5] G. Park, H. H. Cudney, and D. J. Inman, Impedance-based
health monitoring of civil structural components, Journal of
Infrastructure Systems, vol. 6, no. 4, pp. 153160, 2000.
[6] D. Lecompte, J. Vantomme, and H. Sol, Crack detection
in a concrete beam using two dierent camera techniques,
Structural Health Monitoring, vol. 5, no. 1, pp. 5968, 2006.
[7] D. D. L. Chung, Damage in cement-based materials, studied
by electrical resistance measurement, Materials Science and
Engineering R, vol. 42, no. 1, pp. 140, 2003.
[8] H. J. Kong, S. G. Bike, and V. C. Li, Eects of a strong
polyelectrolyte on the rheological properties of concentrated
cementitious suspensions, Cement and Concrete Research, vol.
36, no. 5, pp. 851857, 2006.
[9] C. C. Perry and H. R. Lissner, The Strain gage Primer, McGrawHill, New York, NY, USA, 1962.
Research Article
Heuristic Enhancement of Magneto-Optical Images for NDE
` Salvatore Calcagno,
Matteo Cacciola, Giuseppe Megali, Diego Pellicano,
Mario Versaci, and Francesco Carlo Morabito
DIMET Department, Faculty of Engineering, University Mediterranea of Reggio Calabria, Via Graziella Feo di Vito,
89100 Reggio Calabria, Italy
Correspondence should be addressed to Matteo Cacciola, matteo.cacciola@unirc.it
Received 31 December 2009; Revised 15 May 2010; Accepted 26 August 2010
Academic Editor: Joao Marcos A. Rebello
Copyright 2010 Matteo Cacciola et al. This is an open access article distributed under the Creative Commons Attribution
License, which permits unrestricted use, distribution, and reproduction in any medium, provided the original work is properly
cited.
The quality of measurements in nondestructive testing and evaluation plays a key role in assessing the reliability of dierent
inspection techniques. Each dierent technique, like the magneto-optic imaging here treated, is aected by some special types
of noise which are related to the specific device used for their acquisition. Therefore, the design of even more accurate image
processing is often required by relevant applications, for instance, in implementing integrated solutions for flaw detection and
characterization. The aim of this paper is to propose a preprocessing procedure based on independent component analysis (ICA)
to ease the detection of rivets and/or flaws in the specimens under test. A comparison of the proposed approach with some other
advanced image processing methodologies used for denoising magneto-optic images (MOIs) is carried out, in order to show
advantages and weakness of ICA in improving the accuracy and performance of the rivets/flaw detection.
1. Introduction
For both improving manufacturing quality and guaranteeing
safety, devices, components, and structures are usually
inspected to detect the presence of defects or faults which
may threaten their integrity. Nondestructive testing and
evaluation (NDT&E) are industrial methodologies which
couple the ability of detecting defects and characterizing
them usually by means of noninvasive procedures [1]. In
experimental NDT&E, the measurements can give details
on the structural properties of the sample [2], like cracks,
flaws and phase transformations that may develop in discontinuous deformations. Typically, the main challenge is
to detect and characterize flaws starting from experimental
measurements through the solution of suitably formulated
inverse problems. Because of the limitations of the measurement recording and the presence of noise, the problem to
be solved through inversion is commonly ill-posed as well
as ill-conditioned in its numerical counterpart. The critical
decision for a pattern recognition system is the selection
of appropriate features to be extracted from the image for
classification. These features should be unique, informative,
2
aircraft fuselage skin for detecting subsurface corrosion and
cracks at rivet sites. In aircraft inspection, data (images) from
a large number of rivets need to be analysed. Currently it
is manually performed by scanning the MO imager over
the inspection surface. The main advantage of MOI (see
[7, 8] about MOI designed by physical research instruments
(PRI)) can inspect large areas at high speed compared with
classical eddy current methods. Other advantages are its fast
and easy inspection capabilities in comparison with other
conventional nondestructive inspection instruments. An
automated real-time method for evaluating the MO images
for structural defects can reduce human error and increase
accuracy and speed of the inspections. Such a method
requires the elimination of background (image) noise due
to presence of magnetic domain walls in the MO sensor
and enhancing the image for subsequent interpretation.
This paper presents an image processing and automated
classification algorithm that classifies the MO images for
both surface and subsurface cracks in aircraft skins under
various excitation frequencies. Within this framework, progresses have been made in developing the MOI for detection
of subsurface cracks and corrosion by improvements in
instrument design. The key to the instruments capability in
detecting the relatively weak magnetic fields associated with
subsurface defects is the sensitivity of the MO sensor. The
goal of this work is to introduce an enhanced MOI-based
approach for imaging magnetic fields of NDT&E. Within this
framework, tests have been carried out and the presented
methods have been evaluated by experimental results. The
paper is organized as follows. Section 2 briefly describes
the basic principles of MOI, with a particular overview
of domain noise in MO images. Section 3 introduces and
deepens the denoising of MO images by using two image
processing techniques: the former is based on ICA [911],
the latter is a self-implemented adaptive homomorphic filter
(AHF). Both were compared with the motion-based filtering
(MBF) [12, 13], a well-known technique able to reduce static
noise in images. Finally, in Section 4, the performances of the
proposed approaches are discussed.
K M
f
,
K M
(1)
Light source
Analyzer
Polarizer
Bias coil
Sensor
Lap joint
Induction foil
Frame 125
Frame 126
Frame 127
Figure 2: Three consecutive MOI images; dark disks represent rivets, which are moving to the right, while the sensor is moving to the left.
Dark bands show a substrate seam into the airplane structure.
Crack
Frame 33
Frame 34
Frame 35
Rivet with defect
Normal rivet
Frame 36
Frame 37
Figure 5 depicts the exported sets. The reason of composing each set by 5 dierent frames will be described within
the next subsection. As it is possible to denote, the sets cover
the whole time length of the movie, in order to give as much
generality as possible to the proposed approach. Moreover,
let us remark how, on each set, the images composing the
set itself show the same number of objects, that is, rivets,
and the same objects, except for the 5th set: frames 127,
128, and 129 show, in fact, dierent number of objects
and/or just dierent objects. The 5th set has been voluntarily
added to the experimental set, in order to evaluate the ICAs
performance in similar applicative contexts.
3.2. ICA for Enhancing MOI. The problem of source separation has been deeply analysed in electrical engineering;
many algorithms exist, depending also on the nature of the
mixed signals. The problem faced by the BSS is more dicult
because it is not possible to design appropriate preprocessors
in order to optimally separate mixed signals without any
knowledge about them. Even in NDT&E it is possible to meet
problems involving mixed signals and BSS. For instance, let
us consider the problem of noisy measurements in MOI: in
this framework, noise is an additive eect with respect to
the useful information. In many cases of practical relevance,
often for the presence of nonlinear phenomena, or when a
noise source is not strictly gaussian (e.g., the lift-o eect
in eddy current testing), it is very dicult to separate the
informative signal from the uninformative one. In this case,
Frame 33
Frame 34
Frame 36
Frame 37
Frame 35
Frame 50
Frame 51
Frame 53
Frame 54
Frame 52
Frame 75
Frame 76
Frame 78
Frame 79
Frame 77
Figure 5: Continued.
Frame 115
Frame 116
Frame 118
Frame 119
Frame 117
Frame 125
Frame 126
Frame 128
Frame 129
Frame 127
Frame 166
Frame 167
Frame 169
Frame 170
Frame 168
Rate
367 [Kbps], 10.0 [fps]
Resolution
320240 (4 : 3)
Codec
MS-MPEG standard (MPEG4)
Quality
20%
Figure 6: Independent components showing the informative signals, extracted from the in-study sets of magneto-optic images by the
proposed approach.
a a ... a
2j T
s .
x = 21 22
... ... ... ...
(2)
a j1 a j2 ... a j j
Under some general hypotheses, it is possible to recover the
set of j sources by calculating a suitable mixing matrix A,
H(u, v)
L
D(u, v)
that is, the matrix with elements akh . Once the matrix A is
calculated, it is possible to obtain its inverse A1 and retrieve
the independent components (ICs) having non-Gaussian
distributions [17]. The fixed point algorithm [17] has been
exploited in order to calculate and extract the independent
components from each one of the in-study sets. Figure 6
shows results of our experimentations. Our analysis has
been based on SNR evaluation since it relates the power
of signal useful to the noise of the acquisition system. It
can be considered as a measure of sensitivity performance
and it has a remarkable importance in many applications in
NDE. Within this framework, SNR is a key parameter for
evaluation of proposed image processing and its importance
cannot be underestimated. The SNR directly aects error
probability and ultimately the detection of flaws in NDE.
Within this framework, we propose the increment of SNR
compared with the original SNR of the proposed sets in
Table 2. The averaged SNR is more or less the same for the
dierent sets, with a small decrement for the latter set. SNR
is usually taken to indicate an average signal-to-noise ratio,
as it is possible that (near) instantaneous signal-to-noise
ratios would be considerably dierent. The concept can be
understood as normalizing the noise level to 1 (0 dB) and
measuring how far the signal stands out. Finally let us note
that, according to the method for calculating the SNR [15], it
depends on the range of values given by the filtered images:
thus, it is not important to consider the values of SNR as its
increment due to preprocessing procedure.
The performances shown by the application of the ICA
are generally remarkable in terms of increment of the
SNR, but the evaluation of the reliability of the proposed
method must be carried out by a joint consideration of
the SNR increment itself with the assessment provided by
a visual inspection of the useful independent components.
In fact, the 5th set shows a result which can be assessed
as an irregularity by a visual inspection, in spite of the
SNR increment. The failure was expected, since the rivets
depicted within the selected frames, as above explained, are
dierent or in varying number. It introduces an irregularity
in evaluating the mixing matrix and therefore its inverse
demixing matrix, during the calculation procedure of the
independent components. The result is a sort of mirroring
the depicted rivets within the useful component. Really, it is
3.3. AHF for Enhancing MOI. The aim of image enhancement is to improve the interpretability or the perception
of information in images for human viewers or to provide
the better input for other automated image processing
techniques. But, unfortunately, there is no general rule or
any mathematical criterion for determining what is good
image enhancement when it comes to the human perception.
If the image looks good, it is good. This section considers the
homomorphic filtering for the measurement of the degree
of the enhancement of images in NDT&E. Classical theory
about filtering makes use of linear filters for the improvement
of SNR. Our implementation regards a nonlinear system
based on a generalized principle of linearity. White and
black images can be represented by means of a 2-variable
system. Images are composed by reflection of the light from
physical objects. The process of realization of an image can be
modelled like a product about a lighting ( fi ) and a reflection
( fr ) function [18]:
f x, y = fi x, y fr x, y .
(3)
I f x, y
= I fi x, y I fr x, y ,
(4)
I z x, y
= Z(u, v) = I ln f x, y
+ I ln fr x, y .
= I ln fi x, y
(5)
(6)
s x, y = fi x, y + fr x, y
where fi0 (x, y) = e fi (x,y) and fr0 (x, y) = e fr (x,y) are the
illumination and the reflectance components of the output
In
FFT
H(u, v)
FFT1
Exp
Img f
Visual comparison
Fair
Good
Optimal
Sucient
Failure
Sucient
1 + D0 / u2 + v2
2n ,
(8)
4. Conclusions
In this paper, the application of ICA for enhancing the
quality of magneto-optic images has been discussed and
compared with AHF and MBF. The MOI inspection technique is subjected to a special kind of measurement noise
and also bubbles, mazes and other background static noise,
due to the tape itself, that invariably influence the quality
of the acquired images. These noise sources can be thought
of as disturbing signals, with Gaussian and/or non-Gaussian
probability density distributions, convolved with the useful
signal. Accordingly, the practically relevant need of suitably
denoising MO images can be approached as a problem of
BSS. To deal with it, we decide to separate the underlying
components of the signal by making use of a well-known
algorithmic implementation of ICA. Comparing the performances of dierent algorithms, as reported in Table 3,
we find that the standard performance of ICA is higher
10
ICA result
ICA result
ICA result
ICA result
ICA result
ICA result
Nomenclature
:
f :
K:
M:
References
[1] Y. Yin, G. Y. Tian, G. F. Yin, and A. M. Luo, Defect identification and classification for digital X-ray images, Applied
Mechanics and Materials, vol. 1012, pp. 543547, 2008.
[2] S. Yamada, M. Katou, M. Iwahara, and F. P. Dawson,
Eddy current testing probe composed of planar coils, IEEE
Transactions on Magnetics, vol. 31, no. 6, pp. 31853187,
1995.
11
[3] C. F. Morabito, Independent component analysis and feature
extraction techniques for NDT data, Materials Evaluation,
vol. 58, no. 1, pp. 8592, 2000.
[4] L. Udpa and S. S. Udpa, Application of signal processing
and pattern recognition techniques to inverse problems in
NDE, International Journal of Applied Electromagnetics and
Mechanics, vol. 8, no. 1, pp. 99117, 1997.
[5] M. Cacciola, A. Gasparics, F. C. Morabito, M. Versaci, and V.
Barrile, Advances in signal processing to reduce lift-o noise
in eddy current tests, PIERS Online, vol. 3, no. 4, pp. 517521,
2007.
[6] S. Simms, MOI: magneto-optic/eddy current imaging,
Materials Evaluation, vol. 51, no. 5, pp. 529532, 1993.
[7] G. L. Fitzpatrick, Flaw Imaging in Ferrous and Nonferrous
Materials Using Magneto-Optic Visualization, US Patents no.
4,755,752, 1988.
[8] G. L. Fitzpatrick, D. K. Thome, R. L. Skaugset, E. Y. C. Shih,
and W. C. L. Shih, Magneto-optic/eddy current imaging of
aging aircraft: a new NDI technique, Materials Evaluation,
vol. 51, no. 12, pp. 14021407, 1993.
[9] A. Hyvarinen, J. Karhunen, and E. Oja, Independent Component Analysis, John Wiley & Sons, New York, NY, USA, 2001.
[10] G. Yang, G. Y. Tian, P. W. Que, and T. L. Chen, Independent
component analysis-based feature extraction technique for
defect classification applied for pulsed eddy current NDE,
Research in Nondestructive Evaluation, vol. 20, no. 4, pp. 230
245, 2009.
[11] G. Y. Tian, A. Sophian, D. Taylor, and J. Rudlin, Waveletbased PCA defect classification and quantification for pulsed
eddy current NDT, IEE Proceedings: Science, Measurement and
Technology, vol. 152, no. 4, pp. 141148, 2005.
[12] P. Ramuhalli, F. Yuan, U. Park, J. Slade, and L. Udpa,
Enhancement of magneto-optic images, in Proceedings of
the International Workshop on Electromagnetic Nondestructive
Evaluation, p. 199, 2003.
[13] U. Park, L. Udpa, and G. C. Stockman, Motion-based filtering
of magneto-optic imagers, Image and Vision Computing, vol.
22, no. 3, pp. 243249, 2004.
[14] K. Zvezdin and V. Kotov, Modern magnetooptics and
magnetooptical materials, Journal of the Optical Society of
America B, vol. 22, no. 1, article 187, 1997.
[15] I. T. Young, J. J. Gerbrands, and L. J. van Vliet, Fundamentals
of Image Processing, Delft University of Technology, Delft, The
Netherlands, 1995.
[16] T.-W. Lee, M. Girolami, and T. J. Sejnowski, Independent
component analysis using an extended infomax algorithm
for mixed subgaussian and supergaussian sources, Neural
Computation, vol. 11, no. 2, pp. 417441, 1999.
[17] A. Hyvarinen, Fast and robust fixed-point algorithms for
independent component analysis, IEEE Transactions on Neural Networks, vol. 10, no. 3, pp. 626634, 1999.
[18] V. I. Ponomarev and A. B. Pogrebniak, Image enhancement
by homomorphic filters, in Applications of Digital Image
Processing XVIII, vol. 2564 of Proceedings of SPIE, pp. 153159,
San Diego, Calif, USA, July 1995.
[19] R. C. Gonzales and R. E. Woods, Digital Image Processing,
Addison Wesley, San Francisco, Calif, USA, 1993.
Research Article
A Machine Learning Approach for Locating Acoustic Emission
N. F. Ince,1 Chu-Shu Kao,2 M. Kaveh,1 A. Tewfik (EURASIP Member),1 and J. F. Labuz2
1 Department
2 Department
1. Introduction
Rapidly changing environmental conditions and harsh mechanical loading are sources of damage to structures. Resulting damage can be examined based on local identification
such as the presence of small cracks (microcracks) in a component or global identification such as changes in natural
frequency of the structure. Continuous health monitoring
process may involve both global and local identification.
Generally, local damage, such as cracks in critical components, is inspected visually. This type of inspection is slow
and prone to human error. Therefore, automated, fast, and
accurate techniques are needed to detect the onset of local
damage in critical components to prevent failure.
In this scheme, nondestructive testing and monitoring
should be employed so that the damage can be inferred
through analysis of the signals obtained from inspection.
Acoustic emission (AE) events can serve as a source of
information for locating the damage, particularly as caused
by the initiation and propagation of microcracks [13]. The
spatial distribution of AE locations can provide clues about
the position and extent of the damage [4]. In practice, the
AE
Location
estimation
with TOA
SVM-based
P-wave detection
Hierarchical
clustering
Feature extraction
Averaging
Envelope detection
Figure 1: Schematic diagram of the signal processing and classification system. The AE signals were preprocessed with a median filter. In
the following step they are grouped with a hierarchical clustering procedure. An averaging step was implemented in each cluster to improve
the SNR. This is followed by a feature extraction procedure in time and frequency domains. On the test data, the feature extraction and
classification steps were executed when the signal envelope exceeded a predefined threshold. The TOA is calculated by detecting the P-waves
with an SVM classifier.
X
(a)
300
Amplitude (mV)
(b)
P-wave
200
100
0
100
200
300
100
200
300
400
Samples
500
600
(c)
10
5
0
5
10
900 1000
Normalized amplitude
Normalized amplitude
Samples
Original data
900 1000
Samples
Corrected data
(a)
(b)
Figure 3: Original signal on (a) corrupted with spikes. At (b), the corrected signal with a median filter.
equipment, consisting of four two-channel modular transient recorders (LeCroy model 6840) with 8-bit analog to
digital converter (ADC) resolution and a sampling rate of
20 MHz. The data acquisition system was interfaced with
eight piezoelectric transducers (Physical Acoustics model
S9225), and eight preamplifiers with bandpass filters from
0.1 to 1.2 MHz and 40 dB gain were used for conditioning the
raw AE signals. The frequency response of these transducers
ranged from 0.1 to 1 MHz, with a diameter of approximately
3 mm. All channels were triggered when the signal amplitude
exceeded a certain threshold on the first sensor. This sensor
is referred to as the anchor sensor. AE data were acquired
in a more or less continuous fashion until 128 Kbytes of
a digitizer memory were filled; then the AE data were
transferred to the host computer, with approximately four
seconds of downtime. The entire waveforms were stored
automatically and sequentially with a time stamp. This
experiment was repeated twice using two very similar rock
specimens with dimensions of 62 mm (X) 93 mm (Y )
80 mm (Z) labeled as SR1 and SR2. A sample AE signal
recorded with the system is presented in Figure 2(c). In total,
2176 and 1536 AE events were recorded in the experiments
SR1 and SR2, respectively. This number includes both real
AE and spurious (noise) events.
Several events contained spikes (Figure 3), which probably originated from ADC sign errors. Consequently, a
median filter was employed to remove the spikes from the AE
recordings. The median filter is a nonlinear digital filtering
technique that has found widespread application in image
processing. In this study, each sample was replaced with the
median value of a window covering three pre- and postsamples. A representative corrupted signal and median filter
output is shown in Figure 3. The median filter successfully
corrected the events with consecutive spikes.
3. Clustering of AE Events
In practice, the crack locations are inspected visually by
projecting on a plane the locations of individual AE events,
which are estimated from the TOA information at the
sensors [7]. The TOA is determined by comparing the
signal amplitude to a predefined threshold, where the earliest
arrival is due to the P-wave, as shown in Figures 2 and 4(a).
1
x[n]y[n + k],
(N k)x y n
|k | 256.
(1)
Normalized amplitude
10
5
0
5
10
500
1000
1500
5
0
5
10
2000
500
Samples
1000
1500
2000
Samples
(a)
(b)
Normalized amplitude
10
5
0
5
10
500
1000
1500
2000
Samples
(c)
Figure 4: Sample AE recordings. (a) High SNR with clear baseline. (b) Corrupted baseline. (c) Pseudo-AE (noise).
around the P-wave arrival were examined. The energy of Pwaves was generally found to be located in lower frequency
bands. This wave was followed by large oscillations with
similar spectral characteristic (the 1st row in Figure 6(a)).
Sample waveforms and spectra related to a typical Pwave (center frame in the 1st row, Figure 6(a)) and those
windows preceding and following this wave are presented
in frames 1 and 3 in Figure 6(a). The same analysis related
to a segment that may be recognized as a pseudo-Pwave is also given (Figure 6(b)). It is observed that the
pseudo-P-waves were not followed by large oscillations.
In addition, their frequency spectrum indicates that these
waveforms had a certain amount of energy in mid-frequency
bands. In the following, we describe three approaches
for determining features to be used in a classifier. The
identification of the features was implemented on a training
set by selecting around 20 multichannel super AE events
from each data set. The eectiveness of these features and
their combinations are examined on testing datasets in
Section 5.
4.1. Discrete Fourier Transform-Based Features. Based on the
above observations on the frequency characteristics of Pwaves and noise and within the spirit of [10], so-called
Mel Scale, subband energy features were extracted from
the spectrum of each time window using a fast Fourier
transform. A Blackman-Tukey window was used during the
estimation of spectra of segments. In total, five subbands
0.9
500
0.9
500
Event number
Event number
0.8
1000
0.7
0.8
0.7
1000
1500
0.6
0.6
2000
1500
0.5
0
500
1000
1500
2000
0.5
0
500
Event number
1000
Event number
(a)
1500
(b)
Ch-1
Ch-8
100
200
300
400
500
600
Samples
(c)
Figure 5: Correlation matrices of (a) SR1 and (b) SR2. (c) Overlap plot of AE events related to a particular cluster with four members.
Normalized amplitude
6
4
log power
20
40
Samples
60
20
40
Samples
60
10
10
10
10
10
10
20
20
20
20
60 80 100 120
Samples
Frequency (MHz)
Frequency (MHz)
40
Frequency (MHz)
Normalized amplitude
(a)
4
log power
20
40
Samples
60
20
40
Samples
60
10
10
10
10
10
10
20
2
4
Frequency (MHz)
Left
20
2
4
Frequency (MHz)
Center
20
20
40
60 80 100 120
Samples
2
4
Frequency (MHz)
Right
(b)
Figure 6: (a) Waveforms and log power spectra of 64-sample long time window preceding the P-wave, centered around P-wave, and a 128sample long window after the P-wave; (b) Raw data and spectra of noise segments that may be recognized as a pseudo-P-wave.
Left
Right
1 2 3
Level
SR2
Center
Left
0
(a)
1 2 3
Level
Right
4
(b)
Figure 7: The WP subband tiling for datasets SR1 (a) and SR2 (b). Each selected subband is weighted with the corresponding log scaled
Euclidean distance between classes. The darker nodes have higher discrimination power.
AICc = AIC +
2p p + 1
,
N p1
(2)
0.44
0.42
AICc
4
5
6
7
0.4
0.38
0.36
0.34
10
12
10
12
Model order
Model order
(a)
(b)
Figure 8: (a) The corrected Akaike Information criterion is computed for both datasets SR1 and SR2 and then averaged. The AICc criterion
indicated a model order between 6 and 8, where the minimum was at p = 8. (b) ROC curve related to prediction error of the AR model on
the training data was computed in the center and right windows and averaged over both datasets SR1 and SR2.
SR1
Center
1
0.8
0.8
0.8
0.6
AUC = 0.21
0.4
TP rate
0.2
AUC = 0.47
0.6
0.4
0.2
0
0.2
0.4
0.6
0.8
0.4
0
0
0.2
FP rate
0.4
0.6
0.8
(b)
0.8
0.8
0.4
0.2
TP rate
0.8
TP rate
AUC = 0.44
0.6
0.4
0.2
0
0.6
FP rate
(d)
0.8
0.8
AUC = 0.39
0.6
0.4
0.2
0
0.4
0.8
Right
AUC = 0.27
0.6
(c)
SR2
Center
Left
0.2
0.4
FP rate
0.2
FP rate
(a)
0.6
AUC = 0.43
0.6
0.2
0
0
TP rate
Right
TP rate
TP rate
Left
0
0
0.2
0.4
0.6
FP rate
(e)
0.8
0.2
0.4
0.6
FP rate
(f)
Figure 9: The ROC curves related to the model order p = 6 computed on the training data in the left, center, and right windows. Note that
the discrimination in the center and right windows is better than the left window.
5. Results
As a first step, on each training set, the decision characteristics of the SVM classifiers were examined by visualizing
the ROC curves related to their outputs. We individually
investigated the ROC curves of each feature extraction
method described above and computed the area between
the diagonal line. In addition, we also considered the
classification performance of SVM when the raw AE data
in these consecutive windows are applied. The ROC curves
related to the training data for SR1 and SR2 are depicted in
Figure 10. We note that the maximum area in both datasets
were obtained with the WP method (0.496 for dataset SR1
and 0.481 for SR2). The second most discriminative features
were Mel scale subband energies obtained with FFT (AUC =
0.489 and 0.477 for datasets, SR1 and SR2, resp.). On both
datasets, adaptive selection of frequency subbands provided
better performance. We note that the SVMs trained with 256dimensional raw AE data had quite poor performance, where
the AUC was 0.39 and 0.31 for datasets SR1 and SR2.
9
We also examined the performance of a combination
of feature sets. Interestingly, the features computed with
WP method did not provide any better discrimination
performance when they are combined with other features.
For dataset SR1, the best performance was obtained with
those features computed with WP method only. We note
that the best separation performance was obtained with the
combination of Mel Scale, AR model error, and spectrum
variance features on the dataset SR2 (AUC = 0.483). Based on
these observations, we trained the SVM classifiers either with
only WP features or with the combination of Mel Scale, AR
model error, and spectrum variance features. These classifiers
were applied on the test samples we describe below.
In this study, it is desirable to have a system with low false
positive rates since there exist several peaks in the baseline
preceding the P-waves that can be potentially recognized as a
P-wave. For this particular purpose, we used the probability
output of the SVM classifier. We only accepted those points
as P-Wave arrivals when the posterior probability exceeds a
threshold of 0.9. The threshold can also be moved to more
stringent levels. However, this may result in the classifier
missing the P-waves which will yield low TP rates. One
can also select that time as P-wave arrival point, where the
posterior probability of the SVM classifier is maximum on
the whole AE signal. However, this caused the system to miss
the P-waves and identify those regions in the post-P-wave as
they share similar characteristics. Therefore, we selected the
first point as P-wave when the posterior probability exceeded
the 0.9 threshold.
As indicated in earlier sections, the SVM classifier was
trained on the features using the data set of one of the
experiments and applied on the other dataset. Using this
strategy, we evaluated the generalization capacity of the
system on similar specimens. At this point, it is dicult
to numerically quantify the classification accuracies of both
datasets due to the lack of true labels of the test data. The
true labels can be obtained by manually marking the P-waves.
However, several clusters with low number of members
had poor SNR. It was dicult to visually identify the Pwaves in these records. Consequently, we elected to study
the classification accuracy on those clusters with four or
more members. The algorithm identified 13 and 9 clusters
with four or more members in the datasets SR1 and SR2,
respectively. The super AEs obtained from these clusters had
much higher SNR, and the P-waves were mostly visually
observable. We manually marked the locations of P-waves
and when the classifier identified a region in 10 samples
around the marked location. We provided such a tolerance
region because the P-wave location was not clearly visible
on small number of records due to low SNR, and the expert
manually marked these positions as possible P-wave location.
The success of the system in recognizing the P-waves with
WP features was 97.1% when SR2 was used as training and
SR1 as testing set. While using SR1 as training and SR2
as testing set, the success on recognizing the P-waves was
94.5%. The combination of features yielded classification
accuracies of 93.3% and 94.5% using the same training
and testing procedure for these datasets, respectively. We
note that similar recognition accuracies were obtained with
0.8
0.8
0.6
0.6
TP rate
TP rate
10
0.4
0.4
0.2
0.2
0
0
0.2
0.4
0.6
0.8
0.2
0.4
FP rate
Wavelet packets
Raw data
Mel subbands
AR
Spectrum variance
0.6
0.8
FP rate
Wavelet packets
Raw data
Mel subbands
AR
Spectrum variance
(a)
(b)
Figure 10: The training classification performance of dierent feature sets on the dataset SR1 (a) and SR2 (b). The best performance was
obtained with WP approach. The performance of the raw AE data was quite poor compared to other methods.
Ch-1
Ch-8
100
200
300
400
500
600
Samples
Figure 11: Sample cluster average and detected arrivals from eight
sensors of SR1. TOA is marked with a vertical line on each channel.
Note that the SVM classifier was trained on SR2.
11
90
80
80
70
70
60
60
50
80
50
Z-axis
Y -axis
Y -axis
SR1
90
60
40
40
30
30
20
20
20
0
0
10
10
0
20
40
60
80
60
20
0
20
40
is
ax
X-
40
60
40
40
X-axis
SR2
X-axis
90
90
80
80
70
70
60
60
20
60
is
ax
Y-
40
50
Z-axis
Y -axis
Y -axis
80
50
40
60
40
30
30
20
20
20
0
0
10
10
0
20
40
X-axis
(a)
60
20
40
X-axis
(b)
60
is
ax
X-
80
20
60
40
40
20
60
s
axi
Y-
0
(c)
Figure 12: Estimated locations of the AE events for SR1 (first row) and SR2 (bottom row). Each blue circle represents the location of a
particular cluster. The diameter of the circle is proportional to the number of AE in the cluster: (a) The locations of all clusters; (b) the
locations of those clusters with at least four members. Note that the locations are very close to the free surface; (c) the 3D view of the
locations given in the second column.
12
Y -axis
60
50
Y = 50
40
Y
30
20
10
X
0
0
20
40
60
X-axis
SR2
90
80
70
Y -axis
60
50
40
30
Y = 50
20
10
Z
0
20
40
60
X-axis
(a)
(b)
(c)
Figure 13: (a) Photos of the observed cracks at the upper part of the free surface (X = 62 mm). (b) The observed cracks mapped onto X-Y
plane, where the free surface is on the right-hand side (X = 62 mm). Note that the cracks on SR2 sample are hairline thin. (c) Photos of the
observed cracks on the Z = 80 mm surface.
13
SR2
90
90
80
80
70
70
60
60
Y (mm)
Y (mm)
SR1
50
40
50
40
30
30
20
20
10
10
20
40
60
X (mm)
20
40
60
X (mm)
(a)
(b)
Figure 14: AE locations calculated with classic algorithm on SR1 and SR2 without clustering analysis and SVM technique.
Acknowledgments
6. Conclusions
References
Novel approaches based on hierarchical clustering and support vector machines (SVM) are introduced for clustering
AE signals and detecting P-waves for microcrack location
in the presence of noise. Prior to feature extraction and
classification process, spikes from the AE data are removed
by employing a median filter. Clusters of AE events are
identified by inspecting their pairwise correlation. After
identifying clusters, an averaging step was implemented
to obtain super AE with improved SNR. Characteristic
features were extracted from the data in time and frequency
domains to identify P-waves for time of arrival (TOA). SVM
classifiers with probabilistic outputs were trained with these
features to recognize P-waves for TOA determination. The
location of each AE cluster was estimated accordingly.
The proposed machine learning technique with clustering analysis and SVM showed that the estimated clusters
can successfully indicate the location of failure observed in
surface instability tests, in which the cracks were promoted
to occur close to the front free surface of the specimen. This
approach, compared to the classic AE algorithm that gave a
very disperse pattern and was not indicative of the region of
failure, also presents the capability of filtering noisy signals
and enhance the SNR to obtain more reliable AE cluster
locations. The preliminary results show that the method
has the potential to be a component of a structural health
monitoring system.
Partial support was provided by the National Science Foundation, Grant no.CMMI-0825454. The authors express their
appreciation for the constructive comments provided by the
referees, which served to considerably improve the paper.
14
[9] M. Y. Park, T. Hastie, and R. Tibshirani, Averaged gene
expressions for regression, Biostatistics, vol. 8, no. 2, pp. 212
227, 2007.
[10] A. E. Cetin, T. C. Pearson, and A. H. Tewfik, Classification
of closed and open shell pistachio nuts using principal
component analysis of impact acoustics, in Proceedings of the
IEEE International Conference on Acoustics, Speech, and Signal
Processing (ICASSP 04), pp. 677680, May 2004.
[11] N. Saito, Local feature extraction and its applications using a
library of bases, Ph.D. thesis, Department of Mathematics, Yale
University, New Haven, Connm USA, December 1994.
[12] N. F. Ince, F. Goksu, A. H Tewfik, I. Onaran, A. E. Cetin, and
T. Pearson, Discrimination between closed and open-shell
(Turkish) pistachio nuts using undecimated wavelet packet
transform, Biological Engineering Journal, American Society of
Agricultural and Biological Engineers (ASABE), vol. 1, no. 2, pp.
159172, 2008.
[13] K. P. Burnham and D. R. Anderson, Model Selection and Multimodel Inference: A Practical Information-Theoretic Approach,
Springer, New York, NY, USA, 2nd edition, 2002.
[14] J. C. Platt, Probabilistic outputs for support vector machines
and comparisons to regularized likelihood, in Advances in
Large Margin Classifiers, A. Smola, P. Bartlett, B. Scholkopf,
and D. Schuurmans, Eds., MIT Press, Cambridge, Mass, USA,
1999.
[15] J.H. Kruz, S. Koppel, L. Linzer, B. Schechinger, and C. U.
Grosse, Source localization, in Acoustic Emission Testing:
Basics for Research-Applications in Civil Engineering, C. U.
Grosse and M. Ohtsu, Eds., chapter 6, Springer, Berlin,
Germany, 2008.
[16] K. R. Shah and J. F. Labuz, Damage mechanisms in stressed
rock from acoustic emission, Journal of Geophysical Research,
vol. 100, no. 8, pp. 1552715539, 1995.