RADAR DETECTION OF REBARS INCLUDING USE OF
NEURAL NETWORKS AND HORN ANTENNAS
J. H. Bungey1, S. G. Millard1, T. C. K. Molyneaux1 and Y. Huang2
Department of Civil Engineering, The University of Liverpool, Liverpool, UK L69
3GQ
2
Depa
Department of Electrical Engineering & Electronics, The University of Liverpool,
Liverpool, UK L69 3GJ
ABSTRACT. Work to establish the dielectric properties of concrete at radar frequencies is
described. This is aimed at improving capabilities to identify and accurately locate embedded
reinforcing bars. The work is supported by studies of the characteristics of commonly used
antennas and the use of Artificial Neural Networks to assist interpretation. Current activity
includes development of a wideband horn antenna system to characterise insitu properties
including the effects of moisture gradients.
INTRODUCTION
Subsurface radar has been widely used by testing companies within the civil
engineering industry to investigate construction details and structural integrity of
reinforced concrete. It is becoming increasingly popular as a method that is totally
non-invasive and non-destructive and can often provide information at depths beyond
the range of other techniques. On-going research at Liverpool University commenced
some thirteen years ago to investigate the capabilities and limitations of commercially
available Ground Penetrating Radar (GPR) systems in a variety of structural situations.
One of the most popular applications is the location of steel reinforcing bars and
metallic post-tensioning ducts, where the method offers advantages of speed and depth
penetration compared with conventional "covermeters".
Work has included
measurements of dielectric properties of different concretes from laboratory
transmission line specimens, and study of antenna characteristics. An emulsion
simulation tank has been used to examine the response of buried features including
voids and reinforcing bars, and automatic analysis of scans using neural networks has
been assessed.
Current work is seeking to resolve a major difficulty in extracting useful
quantitative rather than comparative data from insitu test results. This centres around
the need for accurate estimation of speed of propagation of the radar signal through
concrete, which is likely to contain moisture gradients, to permit reliable depth
predictions.
CP657, Review of Quantitative Nondestructive Evaluation Vol. 22, ed. by D. O. Thompson and D. E. Chimenti
© 2003 American Institute of Physics 0-7354-0117-9/03/S20.00
1148
SIMULATION TANK AND TRANSMISSION LINE
Initial work relating to this was based on comparative study of results from
concrete specimens together with the development and use of an oil-water emulsion
simulation tank [1]. A large diameter co-axial transmission line [2] enabled specially
cast laboratory specimens of a wide range of concrete types to be tested at varying ages
under a range of moisture conditions and temperature. This yielded values of relative
permittivity and electrical conductivity over a range of practical frequencies. Relative
permittivity was shown to vary between about 4 and 12 depending on moisture
condition, which was clearly the dominant factor. Electrical conductivity, and hence
attenuation, is also significantly affected by moisture. These results allowed emulsions
of rape-seed oil and water to be produced with addition of an organic salt to represent
particular practical situations ranging from very dry to very moist concrete. A 1.25 x
1.1 x 0.45m all-timber tank was used to minimise signal interference, with reinforcing
bars located on timber supports within the emulsion. A radar antenna was
mechanically scanned across the emulsion surface to enable a wide range of
configurations of bar size, spacing and depths to be rapidly assessed, without the need
for casting and curing of large numbers of concrete specimens. Some typical responses
are shown in Figure 1.
It was recognised that responses were likely to be cleaner than those to be
expected from concretes with large aggregate particles, nevertheless results provided a
useful indication of capabilities and limitations [3]. In particular it was confirmed that
ability to locate a bar was a function of size, spacing and depth together with antenna
orientation. Concrete properties had only a minor influence on resolution capability
and masking effects. For a commonly used IGHz bow-tie antenna, for example, there
was significant response signal overlap from bars at less than about 100mm spacing
whilst for 12mm bars or larger features below the steel were masked out. For a lower
frequency antenna, these effects become important at even greater bar spacing. Bar
depth estimation relies heavy upon a reliable estimate of the dielectric properties of the
concrete, whilst bar sizing is only possible in general comparative terms at a given
depth without the use of specialised measurement techniques. Results of this work
have been incorporated in American Concrete Institute recommendations [4] and a
Concrete Society Technical Report in the UK [5].
(a) widely spaced, large cover
(b) small cover
FIGURE 1. Radar images from 2 subsurface steel bars in emulsion.
1149
NEURAL NETWORKS
Given the skilled and subjective nature of interpretation of GPR results, the use
of a neural network approach to automate and facilitate post-processing has been
explored. A comprehensive data set of radar scans was established covering the normal
range of reinforcing bar diameters (ie 6mm to 32mm) and bar covers between 20mm
and 200mm in 5mm increments using the emulsion simulation tank.
From the typical B-scan shown in Figure l(a) relating to two 12mm bars with a
cover thickness of 140mm a number of features can be seen. At the top of the image is
the cross-talk signal and surface reflection that remain constant. In the centre can be
seen two hyperbolic patterns resulting from the divergent radar signal being reflected
back from the two bars. Beneath these is 'clutter' that cannot easily be associated with
any particular feature. Before trying to apply any type of neural network categorisation
it is advisable to clarify and simplify the image.
Ideally a background image with no sub-surface features should be subtracted,
but if this is not possible then one obtained by smoothing the entire scan can be used as
shown in Figure 2(a).
Following this enhancement, the image was simplified to a
series of lines tracing out the signals of strongest magnitude as in Figure 2(b) using an
edge detection routine.
It was decided that the most convenient way of automatically identifying the
presence of reinforcing bars was to find sections of line containing the basic hyperbola
shape. Each peak trace line was split into a series of overlapping sub-sections where
each sub-section characterised that portion of line as in Figure 3.
(a) Background subtraction
(b) Reduction to line trace
FIGURE 2. Image clarification and simplification.
1. Flat
2. Up
3. Don
4. Peak
5. Wiggle
(a) Trace line subsections
FIGURE 3. Identification of hyperbolas.
(b) Categorisation of lines
1150
A neural network that would take one of these subsections as input and
transform it into a set of numbers that would correspond to one of the above
classifications was required. A number of neural network paradigms are available for
the analysis of data. In this case the multi-layer perceptron (MLP) approach was
considered to be appropriate due to the nature of the GPR images that could be preprocessed prior to the analysis by the MLP network. A network with a single hidden
layer containing 12 nodes was found to be suitable [6].
This automatic hyperbola identification system resulted in lists of classifications
of sub-sections of the trace with their accompanying centre point coordinates. A
program which searched the lists for peaks was thus developed. The result of
hyperbola identification was therefore a set of hyperbola co-ordinates, where each
hyperbola point had as additional attributes its peak trace line and a quality number
defining the severity of the constraints used when finding it.
The identification of hyperbolic trace lines using the neural network
methodology enabled the location of the lateral position of reinforcing bars to be
automatically determined. The remaining problem then was to quantify the cover of
the reinforcing bar, i.e. the distance between the uppermost surface of the bar and the
surface of the concrete. Extensive analysis of emulsion tank data revealed an empirical
linear relationship between the difference between the start time and the hyperbola peak
time and the cover to the bar. In Figure 3(b) the vertical location of the reinforcing bars
has been determined in this way.
Once a neural network system for the automatic identification, and lateral
location and vertical cover, of a reinforcing bar had been developed using the analogue
emulsion tank to model concrete, the system was tested on reinforcing bars in real
concrete slabs.
Of a total of 26 bars investigated, 8 were not detected at all, as the cover was
deeper than 200mm and the signal returning to the surface was sometimes too
attenuated and indistinct to enable a proper analysis. Two other bars at 25mm & 50mm
cover were not found, as the hyperbolic reflection signal was not clear enough for the
neural network procedure to detect. The results for the remaining bars are shown in
Figure 4 and are very encouraging. In all cases the bars were detected with a very high
level of reliability. The covers evaluated can be seen to be very close to the actual
covers.
Bar size estimation was more problematical but general indications were
possible.
Q)
8
100
200
Actual cover (mm)
FIGURE 4. Bar depth predictions.
1151
300
ANTENNA CHARACTERISTICS
Improvement of the accuracy of depth estimation relies upon a knowledge of
the insitu dielectric properties of the concrete, which vary according to concrete type,
moisture condition and frequency of radar signal. Investigations of characteristics of
commonly-used antennas were thus undertaken to seek to improve interpretation.
Findings included confirmation of the suitability of a monostatic approach to
characteristic measurement, as well as assessment of antenna radiation patterns and
frequency characteristics in air, concrete and water and are described elsewhere [7].
The extent of changes to effective signal centre frequency is particularly relevant since
this value reduces significantly when antennas are coupled to a concrete surface rather
than in air. This will in turn affect the effective values of dielectric properties to be
used in calculations if these are based on the transmission line tests at specific
frequencies. Values for two particular antennas are given in Table 1, and could lead to
errors of up to 7% for depth estimations in wet concrete.
WIDEBAND HORN ANTENNA SYSTEM
To overcome the uncertainties of depth estimation insitu and bearing in mind
the likely internal moisture gradients, the approach of measuring the dielectric
properties insitu thus seems to be the only realistic way forward.
A pilot study using an open-ended co-axial transmission line [8] was successful
at high frequencies (IGHz to 6GHz) but could only penetrate a few millimetres in
concrete. This has led to current efforts to measure insitu properties directly with a
wideband horn antenna using a signal inversion process, together with portable batterypowered instrumentation to replace the conventional laboratory-based network
analyser.
A high fidelity transverse electromagnetic (TEM) horn antenna was developed
to give a number of required characteristics:
•
•
•
Wide bandwidth
Near linear phase characteristics
Constant polarisation
Near constant gain
TABLE 1. Antenna centre frequencies.
Antenna Centre Frequency
Test Medium
900 MHz Nominal
1 GHz Nominal
Air
802 MHz
870 MHz
Concrete
500 MHz
666 MHz
1152
These features were needed in order to minimise the ‘ringing’ effects in the time
domain and to allow the use of time gating techniques to process the measurement data.
These features
neededtwo
in order
to minimise
'ringing'
effects
in the time
The prototype
antenna were
comprises
divergent
1mm the
metal
plates,
connected
to a
domain
and
to
allow
the
use
of
time
gating
techniques
to
process
the
measurement
data.
network analyser via an N-type coaxial adaptor (Figure 5) and a robust case will
be
The prototype
antenna use
comprises
two divergent 1mm metal plates, connected to a
incorporated
for practical
in the field.
network analyser via an N-type coaxial adaptor (Figure 5) and a robust case will be
incorporated for practical use in the field.
Measurements of the Voltage Standing Wave Ratio (VSWR) for the prototype
antenna, carried
out using both the laboratory and the portable network analysers, were
Measurements of the Voltage Standing Wave Ratio (VSWR) for the prototype
promising.
Following
theseboth
preliminary
measurements,
tests were
set up
to determine
antenna, carried out using
the laboratory
and the portable
network
analysers,
were
thepromising.
radiation pattern
for
the
antenna
over
the
entire
500MHz-3.0GHz
frequency
range.
Following these preliminary measurements, tests were set up to determine
The
antenna
wasforlocated
on theover
rooftheofentire
a building
to maximise
the freerange.
space
thehorn
radiation
pattern
the antenna
500MHz-3.0GHz
frequency
conditions.
A
second
horn
antenna,
built
as
part
of
an
earlier
part
of
the
project
was
The horn antenna was located on the roof of a building to maximise the free space
used
to
receive
the
transmitted
signals.
Typical
results
of
the
radiation
pattern
in
the
Hconditions. A second horn antenna, built as part of an earlier part of the project was
plane
are
seen
in
Figure
6.
This
shows
a
highly
directional
radar
signal
and
was
used to receive the transmitted signals. Typical results of the radiation pattern in the Hconsidered
thisshows
frequency.
Fuller details
the and
antenna
plane are quite
seen inacceptable
Figure 6. atThis
a highly directional
radar ofsignal
was
development
its properties
patterns canFuller
be found
elsewhere
considered and
quite
acceptableand
at radiation
this frequency.
details
of the[9].antenna
development and its properties and radiation patterns can be found elsewhere [9],
Measurements are taken in the frequency domain. In order to evaluate the
electrical Measurements
properties of the
reconstruct
areconcrete,
taken in the
the inversion
frequencyprocedure
domain. isInrequired
order totoevaluate
the
theelectrical
multi layer
permittivity
that the
produced
these
results. is required to reconstruct
properties
of theprofile
concrete,
inversion
procedure
The layer
complete
inverse profile
modelling
three distinct sections:the multi
permittivity
that process
producedcomprises
these results.
The complete inverse modelling process comprises three distinct sections:i. Forward model
Forward
model
ii. i. Mean
square
error function (MSE) and
Mean optimisation
square error function
(MSE) and
iii.ii. Global
or minimisation
algorithm
iii. Global optimisation or minimisation algorithm
and is described more fully by the authors elsewhere [10]. To validate the procedure,
and is described
by on
theaauthors
[10].
validateplastic
the procedure,
measurements
weremore
firstfully
made
single elsewhere
40mm layer
of To
“PE500”
with the
measurements
firstfrom
madetheon surface.
a single 40mm
layer of
plastic with the
antenna
located were
500mm
The results
of "PE500"
these measurements
are
antenna with
located
500mm from
theinsurface.
The show
results
of agreement.
these measurements are
compared
the modelled
values
Figure 7 and
good
compared with the modelled values in Figure 7 and show good agreement.
These measurements yielded a predicted relative permittivity of 2.23 compared with a
Theseactual
measurements
relative
of 2.23
compared
withthea
known
value of yielded
2.3, anda predicted
a very low
valuepermittivity
of conductivity
compared
with
known
actual
value
of
2.3,
and
a
very
low
value
of
conductivity
compared
with
the
value of zero assumed. The thickness estimate from this test was 43mm.
value of zero assumed. The thickness estimate from this test was 43mm.
FIGURE
5. 5.
Prototype
horn
antenna.
FIGURE
Prototype
horn
antenna.
1153
90o
1
o
900.8
1
0.6
0.8
0.4
0.6
0.2
0.4
180o
0.2
o
180
180'
0o
0o
270o
FIGURE 6. H-plane radiation pattern at 1.0GHz.
270°
270
o
FIGURE6.6. H-plane
H-plane radiation
radiation pattern
pattern at
at 1.0GHz.
l.OGHz.
FIGURE
Similar
measurements
have
subsequently been carried out on two and three
layer systems incorporating ‘Tufnol’ with a higher permittivity and almost zero
Similar measurements
measurements have
have subsequently
subsequently been
Similar
been carried
carried out
out on
ontwo
twoand
andthree
three
conductivity,
yielding results shown
in Figure 8.
layer
systems
incorporating
TufnoP
with
a
higher
permittivity
and
almost
layer systems incorporating ‘Tufnol’ with a higher permittivity and almost zero
zero
conductivity, yielding
yielding results
results shown
shown in
Figure
8.
conductivity,
in
Figure
8.
The programme continues to determine the frequency range that yields the best
correlation
theoretical
data. This
will be the
followed
by validation
the the
inverse
Thewith
programme
continues
to determine
frequency
range that of
yields
best
The
programme
continues
to determine
the
frequencyand
range
that yieldsprofile,
the best
modelling
system
for
a
model
of
linearly
varying
permittivity
conductivity
correlation with theoretical data. This will be followed by validation of the inverse
correlation
with
theoretical
data.concrete,
This will
be followed
validation
of the inverse
together
withsystem
application
to insitu
which
ispermittivity
the mainbyfocus
of the research.
modelling
for a model
of linearly varying
and conductivity
profile,
modelling
system
for
a
model
of
linearly
varying
permittivity
and
conductivity
profile,
together with application to insitu concrete, which is the main focus of the research.
CONCLUSIONS
together with application to insitu concrete, which is the main focus of the research.
CONCLUSIONS
CONCLUSIONS
Results of the work described have identified capabilities and limitations of subsurface radar
to of
locate
steeldescribed
bars in reinforced
concrete.
A knowledge
of dielectric
Results
the work
have identified
capabilities
and limitations
of subofconcrete
the work
identified
capabilities
andvalues
limitations
of
subproperties
of theto
is described
crucial
forhave
reliable
depth
estimation
and
been
surfaceResults
radar
locate
steel
bars in
reinforced
concrete.
A knowledge
ofhave
dielectric
surface
radar
toa wide
locate
steel
bars
inforreinforced
concrete.
A knowledge
dielectric
established
for
range
of
conditions.
Antenna
performance
on
properties
of the
concrete
is crucial
reliable
depth
estimation
andcharacteristics
valuesofhave
been
properties
offor
the
is crucial
for
depth estimation
values
have been
established
aconcrete
wide
of conditions.
Antenna
performance
characteristics
on
concrete
have
also
been range
established
andreliable
the potential
for
use of and
neural
networks
to
established
for also
a wide
ofdemonstrated.
conditions.
Antenna
performance
characteristics
concrete
have
been
established
and the potential
use to
of neural
networks
toon
assist
interpretation
hasrange
been
Current for
work
measure
dielectric
assist interpretation
has established
been for
demonstrated.
Current
work
toof measure
dielectric
concrete
have
and the
potential
usevariations
neuralusing
networks
properties
in thealso
fieldbeen
accounting
internal
moisture
andfor
other
a hornto
properties
in the field
accounting
forbut
internal
moisture
and other
variations
usingdielectric
a horn
assist
interpretation
has
been
demonstrated.
Current
work
to measure
antenna
is yielding
promising
results
requires
considerable
further
development
and
antenna isinyielding
promising
results
but requires
considerable
further
development
properties
the field
accounting
for internal
moisture
and other
variations
using a and
horn
refinement.
refinement.
antenna
is yielding promising results but requires considerable further development and
0
refinement.
S11 (dB)
S11 (dB)
-100
-10
-20
-20
-30
-30
-40
1 Layer Measured
1 Layer Measured
1 Layer Model Theoretical
1 Layer
Layer Model
Theoretical
Measured
-50
-40
0
9
1 Model Theoretical
2
1 Layer
10
Freq
F r e q ((Hz)
Hz)
xx 10
-50 for single 40mm “PE500” layer.
FIGURE
7. 7.Inversion
results
FIGURE
Inversion
results
for single 40mm
1 "PE500" layer.
2
0
Freq (Hz)
FIGURE 7. Inversion results for single 40mm “PE500” layer.
1154
3
9
x 10
3
0
——
S11 (dB)
*»» «
3 L a ye r M e a s u re d
Measured
33 LLayer
a ye r M
o d e l T h e o re tic a l
3 Layer Model Theoretical
-1 0
\ /
-2 0
-20
1
1
F re q (H z)
Freq ( H z )
2
9
x 10
3
x10
FIGURE
FIGURE8.8. Inversion
Inversionresults
resultsfor
for33layer
layer“PE500”/Tufnol/”PE500”
"PE500"/Tufnol/"PE500" system.
system.
ACKNOWLEDGEMENTS
ACKNOWLEDGEMENTS
The
by the
the
The work
work described
described has
has been
been funded
funded both
both by
by the
the European
European Union
Union and
and by
Engineering
and
Physical
Science
Research
Council,
with
current
work
supported
by
Engineering and Physical Science Research Council, with current work supported by
Grant
Ref.
GR/N
34130/01.
Contributions
by
Dr
M
Shaw,
Dr
A
Shaari
,
Dr
M
Grant Ref. GR/N 34130/01. Contributions by Dr M Shaw, Dr A Shaari , Dr M
Nakhkash,
and
Mr
J
Davis
are
also
gratefully
acknowledged.
Nakhkash, and Mr J Davis are also gratefully acknowledged.
REFERENCES
REFERENCES
1.1.
Bungey,
of Concrete
Concrete Research,
Research,
Bungey,J.J.H.,
H., Millard,
Millard, S.
S. G.
G. and
and Shaw,
Shaw, M.
M. R.,
R., Magazine
Magazine of
45,
187-195
(1993).
45, 187-195 (1993).
2.2. Soutsos,
R. and
and Patterson,
Patterson, A.,
A.,
Soutsos, M.
M. N.,
N., Bungey,
Bungey, J.J. H.,
H,, Millard,
Millard, S.
S. G.,
G., Shaw,
Shaw, M.
M. R.
NDT
&
E
International,
34,
419-425
(2001).
NDT & E International, 34, 419-425 (2001).
3.3. Bungey,
and Building
Building
Bungey, J.J. H.,
H., Millard,
Millard, S.
S. G.
G. and
and Shaw,
Shaw, M.
M. R.,
R., Construction
Construction and
Materials,
8,
119-126
(1994).
Materials,8, 119-126(1994).
4.4. American
for Evaluation
of Concrete
in
American Concrete
Concrete Institute,
Institute, NDT
NDT Methods
Methods for
Evaluation of
Concrete in
Structures,
MI,
Structures, ACI
ACI 228.29-98,
228.29-98, American
American Concrete
Concrete Institute,
Institute, Farmington
Farmington Hills,
Hills, MI,
1998.
1998.
5.5. Concrete
TR48,
Concrete Society,
Society, Guidance
Guidance on
on Radar
Radar Testing
Testing of
of Concrete
Concrete Structures,
Structures, TR48,
Concrete
ConcreteSociety,
Society,Slough,
Slough,1997,
1997,88p.
88p.
6.6. Shaw,
Shaw,M.
M.R.,
R.,Molyneaux,
Molyneaux, T.
T. C.
C. K.,
K., Millard,
Millard, S.
S. G.,
G., Bungey,
Bungey, J.
J. H.
H. and
and Taylor,
Taylor, M.
M.
J.,
J., “Automatic
"Automatic analysis
analysis of
of GPR
GPR scans
scans on
on concrete
concrete structures”,
structures", in
in Seventh
Seventh
International
R. Plumb,
Plumb,
International Conference
Conference on
on Ground
Ground Penetrating
Penetrating Radar,
Radar, edited
edited by
by R.
Lawrence,
Lawrence,Ka,
Ka,1998,
1998,pp
pp449-454.
449-454.
7.7. Millard,
Millard,S.S.G.,
G.,Shaari,
Shaari,A.
A.and
andBungey,
Bungey, J.J. H.,
H., INSIGHT,
INSIGHT, 43,
43, 450-457
450-457 (2001).
(2001).
8.8. Thompson,
Thompson,F.,
F.,“Measurements
"Measurementsofofpermittivity
permittivity in
in concrete
concrete samples
samples using
using an
an openopenended
Brit.
endedco-axial
co-axialline”,
line",ininNDT
NDT ininCivil
Civil Engineering,
Engineering, edited
edited by
by J.
J. H.
H. Bungey,
Bungey, Brit.
Inst.
Inst.NDT,
NDT,Northampton,
Northampton,1997,
1997,pp
pp287-302.
287-302.
9.9. Millard,
for
Millard,S.S.G.,
G.5Davis,
Davis,J.,
J., Huang,
Huang, Y.
Y. and
and Bungey
Bungey J.
J. H.,
H., “A
"A wide
wide band
band system
system for
measuring
measuringdielectric
dielectricproperties
propertiesofofconcrete”
concrete"in
in Ninth
Ninth International
International Conference
Conference on
on
Ground
Ground Penetrating
Penetrating Radar,
Radar, edited
edited by
by S.
S. K.
K. Koppenjan
Koppenjan and
and H.
H. Lee,
Lee,Proc.
Proc.SPIE
SPIE
4758,
4758,Bellingham,
Bellingham,WA,
WA,2002,
2002,pp.
pp.416-421.
416-421.
10. Davis,
Davis,J.J.D.,
D.,Huang,
Huang,Y.,
Y.,Millard,
Millard, S.S. G.
G. and
and Nakhkash,
Nakhkash, M.,
M., “Validation
"Validation of
of an
an
10.
algorithm for
for the
the reconstruction
reconstruction of
of an
an inhomogeneous
inhomogeneous dielectric
dielectric profile”,
profile", Proc
Proc
algorithm
IEE
IEERadar
Radar2002
2002Conf.,
Conf., Edinburgh,
Edinburgh,2002
2002(to
(tobe
bepublished).
published).
1155