Damage Detection Using Optical Measurements

and Wavelets

S. Patsias* and W.J. Staszewskiy

Dynamics Research Group, Department of Mechanical Engineering, University of

Sheffield, Sheffield, S1 3JD, United Kingdom

The paper presents an application of the wavelet transform for damage detection based on optical
measurements. A number of important issues that need to be considered when image sequences are
used for vibration analysis are discussed. These include: correspondence of image features from
image to image, image calibration and spatial resolution. The principles of image edge detection are
discussed and a comparison between the wavelet approach and the classical method is presented.
A novel damage detection method based on optically measured mode-shape data is proposed. The
method is illustrated using a simple cantilever beam experiment. The major advantage of the method
is the significantly increased number of discrete points used to describe mode shapes. This is in
contrast to classical techniques where in practice a small number of measurement points are obtained
from a limited number of sensors.

Keywords dynamic testing  damage detection  non-contact measurements  optical measure-

ments  image edge detection  wavelets

1 Introduction Image Sequences operating at elevated temperatures. Often

and Dynamic Measurements
Vibration tests are widely performed as part of
the design and development procedure of almost
all engineering systems. In practice, modal tests
are employed to identify natural frequencies,
damping ratios and modeshapes of structures
[1,2]. This information can further be used to
validate Finite Element results [3].
The traditional approach to vibration testing
uses transducers, such as accelerometers, force
transducers and strain gauges, which need to be
attached to a structure during experiments.
However, in many cases this is either not
desirable or not possible, for example, when a
structure is tested under hazardous conditions or
*Author to whom correspondence should be addressed.
E-mail: s.patsias@sheffield.ac.uk
y
E-mail: w.j.staszewski@sheffield.ac.uk
no interference with a structure is allowed.
This problem is particularly relevant to small-
scale structures, e.g., Micro-Electro-Mechanical
Systems (MEMS). One solution is to employ
non-contact instruments such as eddy current
probes or optical devices including laser displace-
ment probes and laser Doppler vibrometers.
Unfortunately, like traditional accelerometers, all
these devices measure single point values. The
only exception is when a laser Doppler vibro-
meter is used in scanning mode resulting in a
number of measurements (assuming a non-transi-
ent event).
Different solutions can be proposed when
optical measurements are used [4]. Recently,
high-speed video systems have become affordable
Copyright 2002 Sage Publications,
Vol 1(1): 000522
[1475-9217 (200207) 1:1;522; 027110]

5
6 Structural Health Monitoring 1(1)
and powerful, making possible the use of optical
measurements in a number of applications. Such
systems often employ markers so that the process
is more robust; in other words the so-called
correspondence problem is avoided. This ensures
that analysed image features (or markers) can be
matched, i.e., their trajectories can be tracked
from image to image. However, in practice one
would like to avoid markers on structures. This
can be achieved using a single camera with
motion constrained in a 2D plane and for a
known geometry with clearly identified features.
It is common for the time step between image
sequences to be constant. The sequences are
further used to assemble trajectories of
the analysed features (or markers). This is
preceded by an identification routine that locates
the markers coordinates [5]. The trajectories
can then be assembled using a number of
techniques, for example, the differences between
consecutive frames [68]. The analysis using
image sequences is usually restricted to a
limited number of attached markers. Recent
application examples include markers used to
investigate flutter mode shapes in an aeroelastic
wind-tunnel [9].
This paper brings together recent develop-
ments in image edge detection and proposes a
novel approach by which wavelet coefficients are
used to locate or detect damage in a structure.
The main objective of the paper is to present a
damage detection procedure using image
sequences. Edge contours are obtained from
image sequences using the classical wavelet-based
approach. This provides a number of measure-
ment points that are significantly larger than
in the case of classical accelerometer-based
measurements. The novelty of the proposed
method is that edge contours, which describe the
structures movement, are utilised for further
damage detection analysis. The orthogonal
wavelet transform is also applied to damage
identification. Preliminary results from this
work were initially reported in [10] where
the presence of seeded damage in a beam was
detected. The work presented in this paper extends
the previous results in that more sophisticated
equipment has been used for optical measure-
ments and hence higher quality modeshape data
were obtained. Additionally, the orthogonal
wavelet transform is applied to modeshapes for
damage identification.
The structure of the paper is as follows:
Section 2 briefly describes image-processing
principles and introduces the wavelet approach
used for image edge detection, and terminology
and definitions are provided. Section 3 introduces
the proposed damage detection, method based on
the Orthogonal Wavelet Transform of mode-
shapes. The method is illustrated using a simple
cantilever beam experiment in Section 4. This
analysis includes a comprehensive discussion
related to image sequence processing. Finally, the
paper is concluded in Section 5.
2 Image Edge Detection:
Classical Methods and the
Wavelet Approach
Image edge detection is well known in the image
processing community, although the analysis is
still a marginal subject of research in engineering
applications. Therefore, for the sake of complete-
ness, this section briefly introduces the general
concept and the wavelet-based approach to edge
detection. The reader is referred to [11] for more
detailed analysis.
2.1 Image Edge Detection
Research in image edge detection shows a wide
diversity of proposed techniques. The theory of
image detection can be traced back to Marr and
Hildreth in 1980 [12,13]. In 1986 Canny proposed
the method that today is the de-facto technique
used in image analysis [14]: the use of the wavelet
transform expanded to a wide range of areas and
Mallat suggested a wavelet transformation that
can be implemented in detecting edges both in
one dimensional signals and images [15].
Edges in images often appear where large
variations in the intensity exist. An edge can be
defined as a step discontinuity in a 2D signal
(e.g., an image). The edges can be located by
searching for these discontinuities: either as a
 Patsias & Staszewski
local maximum of the first derivative or as the
zero-crossings of the second derivative of the
image. Therefore, for a continuous image f(x, y),
its derivative assumes a maximum in the direction
of the edge. It can be shown [16] that the
direction of the edge of the continuous image
shown in Figure 1 is given by,


1 fy

g tan
fx
and the magnitude of the edge is given by,
q operating. The Laplacian of a Gaussian (LoG),
M fx2 fy2
where, fx @f =@x, fy @f =@y, @f =@x and @f =@y
are the rates of change of the image function f in
the two perpendicular directions. Clearly with
digital images, the gradient has to be computed
using the difference approximations of either the
orthogonal gradients fx, fy or the directional
gradient @f =@r.
Operators that rely on detecting the local
maxima include those of Prewitt and of Roberts
are discussed by Jain in [16]. It is important to
point out that the first derivative operators rely
on a single scale, that of the image, whereas Marr
and Hildreth [13] proposed a multiscale edge
operator. The second derivative is computed
Figure 1 The gradient of the image function along the
direction of the edge [16].
Optical Measurements and Wavelets 7
through the use of a smoothing filter so that
robustness is ensured. It has been shown that the
Gaussian distribution performs the task ade-
quately [13] and is given below,
2 2
=2 2
gx, y ex y 3
1 where (x, y) are the image coordinates and  is
the standard deviation of the probability distribu-
tion. The standard deviation is the variable that
controls the scale at which the Gaussian is
2 which is also called the MarrHildreth operator
is defined as,
r2 gx, y,  f x, y or r2 gx, y,  f x, y
4
Locating the zero-crossings of the second
derivative is easier than locating the maxima of
the first derivative. The drawback is the fact that
the influence of noise on the second derivative is
more profound and hence false edge location is a
possibility. This naturally leads to the question of
which is the best edge operator? Previous
research that tackled this issue by comparing first
and second derivative approaches concluded that,
No one single edge detector was best overall; for
any given image it is difficult to predict which
detector will be best [17]. There was, however, a
clear distinction between the better edge quality
of the second derivative operators over the first
derivative ones.
The classical Cannys operator is in fact a
second derivative multiscale method. Therefore,
the use of the wavelet transform in image edge
detection is closely related to Cannys operator
and for a particular class of wavelets it is
equivalent to detecting the wavelet transform
maxima [15,18]. The scale parameter in Cannys
method is controlled by the standard deviation of
the Gaussian, whereas in the wavelet method it is
controlled by the scale parameter. The multiscale
edges are extracted by locating the local maxima
of the second derivative of the smoothed image
at various scales [19,20].
8 Structural Health Monitoring 1(1)

2.2 Wavelet Image Edge Detection that they are almost identical and hence they are
both approximated by a unique smoothing func-
When a wavelet function (x) is the derivative of
tion
(x,y) [18]. The smoothing and wavelet
a smoothing function
(x),
functions used in the paper are defined as [11],
d

x
x: 5 _ sin!=2 n
dx
! , n 2, 3, 4, 5, . . . 11
!=2
The wavelet transform of a signal at a scale
2 j is proportional to the derivative of the and
smoothed signal at the same scale [11],
_
 ! 
sin!=2n
! i , n 2, 3, 4, 5, . . .
d
j 2 !=2
W2 j f x 2
2 j f x: 6
dx 12

Let functions 1 and 2 be defined as the respectively. Figure 2 gives an example of graphi-
2D wavelets, such that 1 approximates the cal representation of both functions.
partial derivative of a smoothing function with The local maxima of M2 j f x, y along
respect to x and 2 approximates the partial the gradient direction given by A2 j f x, y are
derivative of a smoothing function with respect detected in the same way as in Cannys edge
to y, operator [14]. The image coordinates of the local
maxima as well as the values of the modulus
1 @
x, y 2 @
x, y M2 j f x, y and of the argument A2 j f x, y are
x, y and x, y : 7
@x @y stored. Hence, the local maxima at scale 2 j are a
set of quadruples:
From the above property these can also be
 
defined in a Multiresolution Analysis (MRA) x, y, M, A : 13
framework as [21,22],
 x y x The image of a bridge, shown in Figure 3
1 1 1 2 1 2 y
2j , and 2j , 8 was downloaded from [24] and is used here to
2j 2 j 2j 2j 2 j 2j demonstrate the edge operation for Canny and
the wavelet approach. The Discrete Wavelet
The Fast Wavelet Transform (FWT) as dis-
Transform was applied to this image using
cussed by Mallat [23] uses the same scaling and
Daubechies (with four vanishing moments [25]) as
wavelet functions to calculate the wavelet coeffi-
the wavelet function. The detail subimage (LH),
cients along the rows and the columns of the
which corresponds to the horizontally smoothed
image. These wavelets can then be defined as,
image and the detail subimage (HL vertically
1 2
smoothed image) as well as the diagonally
x, y x y and x, y x y 9 smoothed (HH) subimage is shown in Figure 4
along with the image approximation (LL) at the
where, (x) is a one dimensional smoothing 2nd decomposition level. This decomposition (in
function. The smoothing functions can then be Figure 4) is perhaps a more conventional way of
expressed as, displaying the DWT decomposition.
For image edge detection following Mallats

1 x, y
xy and
2 x, y x
y: 10 approach [26], the second decomposition level
was computed using the functions shown in
(x) is selected to be close to
(x) but not equal, Figure 2. The horizontally smoothed and vertical
the same applies to
1(x, y) and
2(x, y). However, smoothed images are shown in Figure 5 (a and b
in practice they can be assumed to be so close respectively), whereas the magnitude and argument
 Patsias & Staszewski
Figure 2 The smoothing function and scaling
function used for wavelet edge detection,
(a) Smoothing function: piecewise quadratic spline;
(b) Wavelet function: corresponding to the derivative
of the scaling function shown in (a).
Figure 3 The original image of a bridge, downloaded from [24].
Optical Measurements and Wavelets 9
images are shown in Figure 5 (c and d respectively).
The edge contour is then obtained by a procedure
that selects the pixel locations with the largest
wavelet coefficients in the direction indicated by
the corresponding argument value. The edge
representation at each scale can be selected
by either a simple thresholding operation, i.e.,
keeping those points with a magnitude values
higher than a set value, or by more advanced
thresholding approaches [27].
Figure 6(b) shows the edge representation of
the original image in Figure 3 using the wavelet
edge detection procedure described above. The
second level (or scale) of the wavelet decomposi-
tion was used with the threshold level at 2.8
units. Figure 6(a) also shows the corresponding
edge contour produced by Cannys method using
Matlabs Image Processing Toolbox [27]. The
image of this segment of the bridge is a compli-
cated scene: the clouds in the background for
instance, interfere with the depiction of the
bridge. Both methods produced a clear edge
representation, the wavelet representation, how-
ever is of higher quality in the sense that less of
the bridge information is lost, while reducing the
background information (i.e., the clouds or
the background shade at the right hand side
of the image). Both techniques employ a multi-
scale edge detection operation, but the wavelets
are by their nature localised functions. This
allows better feature localisation across the
10 Structural Health Monitoring 1(1)
Smoothed Image - LL
Vertically Smoothed - HL
Figure 4 Fast wavelet transform decomposition of image in Figure 3 2nd
decomposition level.
decomposition scales (e.g., from small to large
image features). Therefore, the wavelet-based
approach is used in the current studies for
vibration analysis: extraction of modeshape data.
3 Damage Detection Using Modal
Curvatures and Wavelets
A number of modal-based techniques have been
developed over the years for identifying and
is approximately propor-
locating damage in structures as summarised in
[28]. This section introduces the modal curvature
approach and discusses the proposed DWT
analysis of modeshape to identify damage. Both
algorithms are applied to image sequence data.
2 14
3.1 Modal Curvature Method
The principle employed in identifying damage is
based on the fact that damage in a structure will
reduce the flexural rigidity in the vicinity of the
Horizontally Smoothed - LH
Diagonally Smoothed - HH
damage: in other words the damage will create a
weak spot. Under vibration testing the mode-
shapes of the damaged structure will contain
information about the damage: since damage will
also affect the dynamic characteristics, for
instance reduction of the natural frequency or
increase in modal damping, when compared to
the undamaged case.
Pandey et al [29] proposed using curvature
functions of modeshapes to locate damage. The
curvature function
tional to the second derivative of the modeshape
^ with respect to the distance xc along the
cantilever and is given as,
^
1 d 2
R dxc
where R is the radius of the curvature. The
curvature and bending strain for beams are
 Patsias & Staszewski Optical Measurements and Wavelets 11

(a)

(b)

(c)

Figure 6 Edge contours of image in Figure 3 using

(a) Cannys operator and (b) the Wavelet Transform
approach (utilising the argument and magnitude images
in Figures 5(c) and (d) respectively).

directly related as,

y
"
y 15
R
(d) where " is the strain and y is the distance from
the neutral axis of the cantilever.
It has been shown that absolute changes in
modeshape curvatures indicate the presence of
damage in cantilevers [29]. One can obtain the
curvature by simply differentiating the modeshape
twice. In the work carried out here and [4,10] this
is computed from the displacement modeshapes
using the central difference approximation
defined as,

^ ^ ^
Figure 5 Decomposing the image in Figure 3 using the ^ 00 q1  2q q1
 16
i 2
wavelet transform shown in Figure 2. The partial o
derivative in the horizontal and vertical directions are
shown in (a) and (b) respectively. The argument and where o is the distance between the two locations
magnitude images are shown in (c) and (d) respectively. (q  1) and (q 1) [28].
12 Structural Health Monitoring 1(1)
Figure 7 (a) Theoretical modeshape curvature of the
cantilever for the fundamental flexural mode; (b) the
simulated damaged curvature function at 0.6 normal-
ised length from the clamped end (the length was also
normalised).
It is well-known that localised damage will
produce a discontinuity in its curvature [29].
Figure 7(a) shows an example of the intact
curvature of the first flexural modeshape of a
cantilever. Similar curvature for a modeshape
with damage present is shown in Figure 7(b). The
damage here is clearly located at 0.6 normalised
distance from the clamped-end.
3.2 Wavelet Transform Method
The wavelet transform is well suited for detecting
localised events in signals, due to the localised
nature of wavelet basis functions. Discrete wave-
lets are particularly useful for decomposition,
would expect the orthogonal wavelet transform
to be suitable for damage detection and localisa-
tion from vibration modeshapes. The multiscale
nature of the orthogonal wavelet transform
means that a number of signal representations,
corresponding to different decomposition wavelet
levels implemented, are possible. At each level the
so-called approximation and detail versions of
the input signal can be obtained through the use
of Low-Pass (LP) and High-Pass (HP) filters,
respectively.
The modeshape corresponding to the
damaged curvature function shown in Figure 7(b)
is used to demonstrate the application of the
orthogonal wavelet transform for damage detec-
tion. Daubechies wavelets with 20 vanishing
moments [25] were employed to obtain the first
three high-frequency wavelet levels from this
modeshape. The algorithm supplied in Matlabs
wavelet toolbox was used for the analysis [25].
Figure 8 shows the approximation and detail
signals from the orthogonal wavelet decomposi-
tion. Here, the approximation signals given in the
left column clearly exhibit the displacement
modeshape. Additionally, the detail signals given
in the right column show the presence of the
seeded damage. This clearly indicates that
the differentiation procedure associated with the
modal curvature method, presented in Section 3.1,
is no longer required. If one recalls problems
related to signal differentiation [30], this is indeed
a significant result.
4 Experimental Results from a
Damaged Cantilever
The damage detection procedure described in
Section 3 is illustrated here using a simple canti-
lever experiment. This section brings together the
optical measurement technique, the wavelet-based
edge detection and the orthogonal wavelet trans-
form for damage identification.
4.1 Optical Measurements
The experimental set-up used in the paper is shown
in Figure 9. It comprises a steel cantilever beam
compression and feature selection. Therefore, one (0.6  0.025  0.001 m) and an electrodynamic
 Patsias & Staszewski Optical Measurements and Wavelets 13

Figure 8 Discrete Wavelet Transform decomposition of the damaged mode-

shape (obtained from the curvature in Figure 7(b)). Daubechies wavelets with 20
vanishing moments were used.

Clamp

Steel frame

Exciter

Stinger rod

Free-end of cantilever

Figure 9 The cantilever under first bending mode excitation. Image captured by a
commercial camcorder (SONY) [31].

exciter connected to it via a stinger rod to
minimise exciter/structure interaction. The beam
and the electrodynamic exciter are both sus-
pended from a very stiff steel frame. In Figure 9
the cantilever is under excitation corresponding
to the first mode of vibration and the image was
captured by a commercial SONY camcorder [31].
One of the important issues associated with
optical measurements is the quality of the cap-
tured data. The image captured in Figure 9
clearly shows the problem the free-end of the
cantilever is very fuzzy. The reason for this is the
rather low image sampling frequency of 25
frames per second (fps). This camera was used
14 Structural Health Monitoring 1(1)
with some success in the initial study presented
in [10], although a Kodak high-speed camera
system [32] was used to capture image sequences
used for damage detection in order to overcome
this problem. Such systems are these days afford-
able due to the falling cost of computers
Random Access Memory (RAM) devices that
store captured images. The Kodak camera
system, used in the paper, supports a wide range
of image sampling rates from 40 to 600 fps.
Figure 10 shows a typical image captured by
the Kodak system. Note that the orientation of
the camera has now been changed. The brightness
of the image is very low, and high reflectivity
points in the image are avoided by covering the
cantilever with a thin non-reflective tape. The use
of additional lighting was discarded since this
could create high reflectivity areas. The length of
the beam (0.6 m) was such that efforts to diffuse
the light proved to be very difficult.
Also, the image size is variable from 160  60 to Figure 11 summarises the procedure
640  240 pixels (picture elements). The camera
has a finite amount of RAM and hence the
image sampling frequency and image size deter-
mine the number of images that can be captured
at one time.
employed in capturing an image sequence and
performing the subsequent image analysis.
Having captured the image sequence, the wavelet-
based edge detection procedure was carried out
for each image. Then the edge representations
were identified and correspondence between the

Figure 10 A typical captured frame by the KODAK system, depicting the cantilever under
vibration.

Figure 11 The feature-based image sequence analysis employed in

this paper.
 Patsias & Staszewski Optical Measurements and Wavelets 15

images was ensured [33]. This is indeed a very
important part of the analysis: it can be described
from a dynamic testing viewpoint as ensuring
that the transducers are measuring the correct
quantities. The free-end of the cantilever was
displacement device (LD 1605-20) measuring the
cantilevers vibration. This response and its PSD
are shown in Figure 14(a) and Figure 14(b),
respectively. In both cases the signals were
obtained without disturbing the structure, the
used as a marker to ensure that all of the image
coordinates would be correctly estimated. The
final step was to correct for any image distortions
due to the equipment used [34,35] and convert
the measurements from pixels into real world
measurements (mm). The method employed was
to use a grid of known dimensions, obtain the
edges, identify the grid locations and then obtain
only difference is that with the Epsilon laser
displacement device, a longer time signal was
recorded. The identified natural frequencies are
shown in Table 1, where the maximum difference
between the two tests is 1.36%, which is within
the expected experimental error. The natural
frequency of the fundamental flexural mode of
vibration was found to be 2.45 Hz from the
the calibration factor. More information on this
can be found in [4].
The first task before any modeshape measure-
ments can be performed and the seeded damage
identified, is to validate the optical measurements.
This was achieved by carrying out an impulse test
on the cantilever (without the exciter) and identi-
fying the natural frequencies of the beam from the
captured image sequence and comparing against
another non-contact measurement method.
The impulse was initiated at 30 mm from the
free end of the cantilever and an image sequence
of approximately three seconds in length was
recorded by the Kodak system at the maximum
sampling frequency (600 Hz). The image sequence
was analysed with the procedure described above.
A frame from this image sequence is shown in
Figure 12(a). For the wavelet transform the
partial derivatives in the horizontal and vertical
directions (shown in Figures 12(b) and (c) respec-
tively), were used to calculate the argument and
magnitude images. These are shown in Figures 12(d)
and 12(e), respectively. The final edge representa-
tion is shown in Figure 12(f) after a threshold
operation was performed. The same operation
was repeated for all images and the required
vibration signal was obtained from a routine
developed in [4] that locates the coordinates of
the free-end of the cantilever. The outcome is the
displacement of the cantilever as a function of the
distance from the clamped end for each image.
This is shown in Figure 13(a) with the corre-
sponding Power Spectral Density (PSD) in Figure
13(b). The indicated frequencies corresponding
to the first four modeshapes are clearly marked.
image sequence and 2.42 Hz from the Epsilon
laser displacement device test.
4.2 Experimental Modeshapes
Having demonstrated that the optical measure-
ments can provide reliable vibration results, the
next task was to identify the fundamental flexural
modeshape of the cantilever. In this case, the
electrodynamic exciter was attached as shown in
Figure 9 and an image sequence was captured
while the cantilever was under sinusoidal excita-
tion at its first mode of vibration.
The procedure described in Section 2.2 was
carried out and the modeshapes were extracted
using a line tracking algorithm that follows the
vibrating beams outline. The quality of the edge
contours is of such high quality that the origin of
the cantilever is always at the same image
location and used as the starting point of this
algorithm. More information about this algo-
rithm can be found in [4]. The image sequence
recorded the motion of the cantilever at the
maximum deflection (in both directions). To
identify the modeshapes at the extreme deflec-
tions, only the relevant images were analysed.
It is important to highlight the fact that the
modeshape consists of a number of discrete
values (the pixels). This can be compared to a
classical dynamic test, where transducers are
attached to the cantilever. In this case a max-
imum of 68 lightweight accelerometers can be
used (with the penalty of altering the dynamics of
the system), whereas the image contour that
describes the modeshapes has definitely more
The test was repeated with a Epsilon laser data points (in excess of 500 in this case).
16 Structural Health Monitoring 1(1)

(a) The first image from the impulse test image sequence.

(b) Partial derivative in the horizontal direction of image in (a).

(c) Partial derivative in the vertical direction in (a).

(d) Argument image of the horizontal and vertical derivatives in (b) and (c) respectively.

(e) Magnitude of the horizontal and vertical derivatives in (b) and (c) respectively.

(f) Final edge representation after thresholding removed small details from image in (d).

Figure 12 Wavelet edge detection images, (a) first image from the
impulse test; (b) partial derivative in the horizontal direction; (c) partial
derivative in the vertical direction; (d) argument; and (e) the
corresponding magnitude image. The final edge representation using
a threshold operation is shown in (f).

Clearly, one has to remember that these points
are discrete values. The smoothed modeshape is
obtained by curve fitting the discrete modeshape
points, using Matlabs standard polynomial curve
fitting routine (finding the coefficients of a poly-
nomial of a set order and fitting the data in a
least squares manner) [36,37]. The conversion of
modeshape into mm (as described in Section 4.1)
is performed prior to the curve fitting. The
modeshape of the cantilever for the first vibra-
tional mode is shown in Figure 15, where the
fitted curve is also overlayed. The stepwise
manner of the contour (i.e., discrete locations) is
exaggerated by the use of approximately equal
 Patsias & Staszewski Optical Measurements and Wavelets 17

Figure 13 (a) Decay signal from impulse test on the cantilever. The
displacement signal at the free-end was extracted from the image sequence;
(b) The corresponding PSD.

Figure 14 (a) Decay signal (free-end) from the impulse test on the cantilever
recorded by the Epsilon displacement device; (b) The corresponding PSD.

axes, used to display 600 mm and 48 mm for the
length and the displacement of the cantilever
respectively.
The next question is whether the extracted
modeshape really corresponds to the first flexural
mode. The theoretical and the experimental
modeshapes are overlayed in Figure 16, where the
displacements and length have been normalised.
It is clear that the two curves look very similar.
However, there are some differences. The reason
for these differences can be traced back to the
experimental set-up. The electro-dynamic exciter
18 Structural Health Monitoring 1(1)

Table 1 Comparison of natural frequencies identified attached through the stinger rod to the cantilever
from the image sequence and the signal recorded by the has indeed affected its dynamic characteristics:
laser displacement device.
when the cantilever is at its maximum and mini-
Frequencies Frequencies mum displacement positions the exciter affects its
Identified Identified motion the most. Such problems are common in
from from
dynamic testing; more information can be found
Image Laser Device Difference
Mode Sequence Signal (%) in [1]. Figure 16 shows the location of stinger rod
(50 mm in length) and at 40 mm from the
1 2.45 2.42 1.22
clamped end (0.0667 in normalised length).
2 14.74 14.54 1.36
3 40.82 40.79 0.07 To prove this conclusively the cantilever was
4 79.56 79.9 0.43 used to capture an image sequence with the
exciter removed: the cantilever was displaced by
approximately 10 cm and released so that the free
vibration of the fundamental flexural mode was
recorded. A procedure that adjusts modeshapes
biased by the effect of an attached exciter was
developed and described in [4]. This routine
was applied on the modeshape obtained and is
shown in Figure 16. Its effectiveness can be seen
in Figure 17, where the adjusted versus the
analytical modeshapes for the first mode are
overlayed: the MAC [3] was calculated as 1.00,
i.e., perfect correlation was achieved.

4.3 Damage Detection

All of the required test procedures up to this point
have been discussed and proved to be reliable.
Figure 15 Curve fitted on the extracted first flexural The final and most crucial part is to identify
modeshape from an image frame. a seeded damage condition on the cantilever.

Figure 16 A visual comparison between the theoretical Figure 17 The fundamental flexural vibration mode-
and experimental first flexural modeshape. The arrow shape of the cantilever adjusted for the exciter effect. Also
indicates the location of the stinger rod (excitation point). overlayed is the theoretical modeshape.
 Patsias & Staszewski Optical Measurements and Wavelets 19

The damage was created by scoring a groove damage location can be readily obtained. The
across the cantilevers width at 252 mm (or 0.42 potential of the proposed method is twofold:
normalised length) from the clamped end of the
beam in order to simulate the local change in 1. Optical measurements do not affect the struc-
stiffness. This is a typical type of damage tures dynamic behaviour while providing a far
observed in practice and studied in the literature greater number of measurement points when
(e.g., cracked beams or opening and closing compared to classical measurements taken with
cracks [38,39]). The same procedure was followed accelerometer transducers;
as previously: i.e., the cantilever was excited at its
fundamental natural frequency (2.45 Hz); the
image sequence was wavelet edge detection
analysed and the modeshape was extracted. A
spline function was fitted to the modeshape and
the displacement curvature was computed using
Equation (16), and the modeshape was adjusted
for the exciter effect (as described above). The
normalised curvature for the damaged cantilever,
shown in Figure 18, clearly exhibits the damage
at the 0.42 normalised length.
Finally, the same modeshape was analysed
using the orthogonal wavelet transform with the
Daubechies wavelet (20 vanishing moments). The
wavelet decomposition is shown in Figure 19, where
the damage can also be identified at the expected
location in the first three detail signals. What this Figure 18 The curvature function of the damaged
demonstrates is the ability to identify damage from modeshape of the cantilever (fundamental flexural mode
the wavelet transformation on a modeshape: the of vibration).

Figure 19 Discrete wavelet transformation of the damaged cantilever mode-

shape (fundamental vibration mode). The Daubechies wavelet family with 20
vanishing moments were used. The first three decomposition levels are shown
(approximation and detail signals).
20 Structural Health Monitoring 1(1)
2. The wavelet transformation of measured
modeshapes can be used to identify possible
damage while avoiding all problems associated
with curvatures. The former directly uses the
modeshapes for damage detection, whereas the
later relies on curvature differentiation (see
Equation 16); this is not an easy task as
reported in [30].
It has been shown in [4] that the existing test
set-up could successfully identify the first four
flexural modeshapes of an undamaged cantilever
(with the same geometry as the one used in this
paper). Tests on the damaged cantilever were
limited to the fundamental vibrational mode. The
reason for this is due to an important feature of
optical measurements: that temporal resolution
alone does not suffice to establish the useful
bandwidth of an optical system. The Kodak
system has a maximum image sampling frequency
of 600 Hz, but this does not actually mean that
vibration up to 300 Hz can be recorded (based on
the Nyquist criterion).
The image sequences form spatio-temporal
data, so the spatial resolution must also be
adequate. As discussed in [4], for the fourth
flexural vibration mode, the amplitude of vibra-
tion will be much lower than the fundamental
mode, but the spatial resolution used would be
the same in both cases. This means that for the
amplitude measurements of the modeshapes, one
would have the same number of pixels per mm.
At the limit of the spatial resolution this means
that one pixel value would exist for amplitudes
corresponding to smaller quantities. Therefore,
modeshape measurements will be prone to
significant errors and damage detection from
such data would not be possible. A possible
method to increase the spatial resolution (apart
from different equipment) would be to use
sub-pixel interpolation methods described in [40].
5 Conclusions
This paper introduced two new aspects into
damage detection: the use of optical measure-
ments based on the wavelet edge detection
technique and the use of the orthogonal wavelet
transform to identify the presence and location of
damage. It was shown that the damage in a
cantilever was successfully located using the
above methods.
One of the key advantages of the optical
measurements is the number of discrete points
used to describe the modeshape. The number of
measurement points is limited only by equipment
used, but is by far higher than the number of
transducers that could practically be attached to
a structure. Assuming one resolves the correspon-
dence and the spatial resolution issues, data
collected from optical measurements would
almost certainly offer more measurement
points in vibration modeshapes. Clearly the
higher the number of the measurement points
(or sensors) the better the approximation of the
modeshape. Also, due to the fact that numerical
differentiation tends to amplify errors in the data,
a spatially under-sampled modeshape would be
prone to such errors and might even obscure the
damage. Therefore, optical measurements offer
an advantage over the traditional transducer for
damage detection.
With the existing test arrangement (i.e., the
cantilever), the work should be extended to cover
higher modes than the fundamental flexural mode
of vibration with the appropriate equipment. The
preliminary work presented in this paper is based
on a simple cantilever beam structure. Further
work is required with more complex structures to
confirm the results. High-speed camera systems
with better specifications than the one used
here (or even multiple camera systems) could
be employed for real engineering structures.
However, the authors realise that significant
deflections are required for the proposed metho-
dology. This in fact is the main limitation of
the proposed technique. Also, damage detection
procedures require significant changes in stiffness
(and/or in modal characteristics) due to damage;
clearly, local cracks will not be detected using
the proposed method. One possible area of
application includes Micro-Electro-Mechanical-
Systems (MEMS), where vibration, predomi-
nantly constrained in the 2D plane, needs to
be analysed using non-contact devices. The pro-
posed methodology would be ideal for such a test
framework.
 Patsias & Staszewski
Furthermore, the wavelet transform analysis
should involve an appropriate wavelet basis
function that can be applied in a 3D domain. In
this approach one can take image cross sections
to identify modeshapes. The modeshape data
11.
could then be used for damage detection.
12.
References
1. Dynamic testing agency modal testing primer (1993). 13.
Primer on best practice in dynamic testing. London:
Chameleon Press Ltd. 14.
2. Maia, N.M.M., Silva e, J.M.M., He, J., Lieven, N.A.J.,
Lin, R.M., Skingle, G.W., To, W.-M. and Urgueira,
A.P.V. (1997). In N.M.M. Mai and J.M.M. Silva e 15.
(Eds.), Theoretical and experimental modal analysis.
England: Research Studies Press Ltd, John-Wiley & 16.
Sons Inv.
3. Ewins, D.J. (2000). Modal testing: theory, practice and 17.
application. (2nd ed.) England: Research Studies Press
Ltd. September.
4. Patsias, S. (November 2000). Extraction of dynamic 18.
characteristics from vibrating structures using image
sequences. PhD Thesis, University of Sheffield, United
Kingdom.
5. Walton, J.S., Nalodka, E.C., Adam, W.R., Darling,
R.S., Fung, S. and Gleason, J.T. (1989). Over-the-road 19.
video based measurements of suspension kinematics.
SAE Technical Paper Series, 801143. Proceedings of the
1989 Noise and Vibration Conference (pp.175183).
Traverse City, Michigan, May 1618. 20.
6. Cafforio, C. and Rocca, F. (1976). Methods for
measuring small displacements of television images.
IEEE Trans. On Information Theory, IT- 22(5), 21.
573579.
7. Houkes, Z. (1983). Motion parameter estimation in
TV-pictures, In: T.S. Huang, (Ed.), Nato advanced
study institute on image sequence and dynamic scene 22.
analysis. Image sequence processing and dynamic scene
analysis (pp. 249263). Berlin: Springer.
8. Lenz, R. and Gerhard, A. (1983). Image sequence
coding using scene analysis and spatio-temporal inter- 23.
polation. In T.S. Huang, (Ed.), Nato advanced study
institute on image sequence and dynamic scene analysis.
Image sequence processing and dynamic scene analysis
(pp. 264274). Springer: Berlin. 24.
9. Kuttenkeuler, J. (2000). Optical measurements of flutter
mode shapes. J. Aircraft, 37(5), 846849.
fireballx5/buildings/structures/pics/
10. Patsias, S., Staszewski, W.J. and Tomlinson, G.R.
(1997). Wavelet based edge detection for structural 25.
dynamics. CD-ROM Proceedings of the Symposium on
Optical Measurements and Wavelets 21
Time-Frequency and Wavelet Analysis at the 16th
Biennial Conference on Mechanical Vibration and
Noise. ASME Design Engineering Conferences.
Sacramento, California, Paper No. DETC97/VIN-
4138.
Zhong, S. (September 1990). Edge representation from
wavelet transform maxima. Ph.D. Thesis, New York
University.
Marr, D. (1982). Vision, New York: W.H. Freeman and
Company; Pattern Analysis and Machine Intelligence,
(July 1992), 14(7), 710732.
Marr, D. and Hildreth, E. (1980). Theory of edge
detection. Proc. R. Soc. London, B 207 (pp. 187217).
Canny, J. (November 1986). A computational approach
to edge detection. IEEE Trans. on Pattern Analysis and
Machine Intelligence, PAMI-8(6), 679698.
Mallat, S. and Zhong S. Characterization of signals
from multiscale edges. IEEE Transactions on.
Jain, A.K. (1989). Fundamentals of digital image
processing. USA: Prentice-Hall, Inc.
Nair, D., Mitiche, A. and Aggarwal, J.K. (1995). On
comparing the performance of object recognition
systems. Intern. Conference on Image Processing.
Mallat, S. and Zhong, S. (1992). Wavelet transform
maxima and multiscale edges. In M.B. Ruskai, G.M.
Beylkin, R. Coifman, I. Daubechies, S. Mallat, Y.
Meyer and L. Raphael (Eds.), Wavelets and their
applications (pp. 67104). Jones and Bartlett.
Hummel, R. and Moniot, R. (December 1989).
Reconstruction from zero-crossings in scale-space.
IEEE Trans. on Acoustic Speech and Signal
Processing, 37(12), 21112130.
Mallat, S. (July 1991). Zero-crossings of a wavelet
transform. IEEE Trans. on Informations Theory, 37(4),
10191033.
Mallat, S.G. (July 1989). A theory for multiresolution
signal decomposition: the wavelet representation. IEEE
Trans. on Pattern Analysis and Machine Intelligence,
11(7), 674693.
Mallat, S.G. (July 1989). Multifrequency channel
decompositions of images and wavelet models. IEEE
Trans. on Acoustics, Speech, and Signal Processing,
37(12), 20912110.
Mallat, S. and Zhong, S. (1991). Compact image coding
from edges with wavelets. Proceedings of the
International Conference on Acoust. Speech and Signal
Processing (pp. 27452748).
Image of a bridge. (2002) Downloaded from internet
site: www.freeimages.co.uk, Hypertext link: www.btin-
ternet.com/
bridge04090075.jpg, 1st, February.
Mathworks Inc. (1997). Wavelet Toolbox for used with
Matlab V1., January.
22 Structural Health Monitoring 1(1)
26. Wave2 Downloaded from Internet ftp site, ftp://
cs.nyu.edu/oub/wave/software.
27. Mathworks Inc. (1998). Image Processing Toolbox for
use with Matlab V2.1, January.
28. Doebling, S.W., Farrar, C.R., Prime, M.B. and Shevitz,
D.W. (1996). Damage identification and health mon-
itoring of structural and mechanical systems from
changes in their vibration characteristics: A literature
review. Report LA-13070-MS, UC-900, Los Alamos
National Laboratory, New Mexico 87545, USA. May.
29. Pandey, A.K., Biswas, M. and Samman, M.M. (1991).
Damage detection from changes in curvature mode
shapes. Journal of Sound and Vibration, SIAM, 145(2),
321332.
30. Worden, K. (1990). Data processing and experiment
design for the restoring force surface method. part 1:
integration and differentiation of measured time data.
Mechanical Systems and Signal Processing, 4(4),
295319.
31. Sony Corporation. (1996). Operating Instructions
Video Camera Recorder video Hi8 CCD-TR760E/
TR810E Hi8, France.
32. Eastman Kodak Company Kodak MotionCorder
Analyzer. (1996). Users Manual Model 1000, San
Diego, California.
33. Lee, C.-H. and Joshi, A. (1993). Correspondence
problem in image sequence analysis. Pattern
Recognition, 26(1), 4761.
34. Tsai, R.Y. (August 1987). A versatile camera calibra-
tion technique for high-accuracy 3D machine vision
metrology using off-the-shelf TV cameras and lenses.
IEEE Journal of Robotics and Automation, RA-3(4),
323344.
35. Weng, J., Cohen, P. and Herniou, M. (October 1992).
Camera calibration with distortion models and accu-
racy evaluation. IEEE Trans. on Pattern Analysis and
Machine Intelligence, 14(10), 965980.
36. Chapra, S.C. and Canale, R.P. (1998). Numerical
methods for engineers with programming and software
applications. Singapore: WCB/McGraw-Hill.
37. Mathworks Inc. (1999). Matlab Function Reference
(Volume1: Reference), version 5, June.
38. Chance, J.E. (1996). Structural fault detection employ-
ing linear and nonlinear dynamic characteristics. PhD
Thesis, School of Engineering, University of
Manchester.
39. Krawczuk, M. and Ostachowicz, W.M. (1992).
Parametric vibrations of a beam with crack. Archives
of Applied Mechanics, 62, 463473.
40. Astrom, K. and Heyden, A. (2001). Stochastic model-
ling and analysis of image acquisition and sub-pixel
edge detection. Downloaded from internet site http://
citeseer.nj.nec.com/2563.html, September.