You are on page 1of 11

Engineering Structures 26 (2004) 39–49

www.elsevier.com/locate/engstruct

Applications of wavelet transforms to damage detection in frame


structures
A.V. Ovanesova, L.E. Suárez ∗
Civil Engineering Department, University of Puerto Rico, Mayagüez, PR 00681, USA

Received 24 October 2002; received in revised form 26 July 2003; accepted 13 August 2003

Abstract

This paper presents applications of the wavelet transform to detect cracks in frame structures, such as beams and plane frames.
A simplified review of the theory behind the wavelet transform is given, and the ability of wavelets to detect crack-like damage
in structures is demonstrated by means of several numerical examples. The method requires the knowledge of only the response
of the damaged structure, i.e. no information about the original undamaged structure is required. In addition, it is shown that the
procedure can detect the localization of the crack by using a response signal from static or dynamic loads. Moreover, the response
needs to be obtained only at the regions where it is suspected that the damage may be present. The results of the simulation show
that if a suitable wavelet is selected, the method is capable to extract damage information from the response signal in a simple,
robust and reliable way.
 2003 Elsevier Ltd. All rights reserved.

Keywords: Wavelet analysis; Continuous wavelet transform; Discrete wavelet transform; Damage identification; Crack detection; Nondestructive
methods

1. Introduction to detect damage are based directly on the examination


of the response signals. Most of these methods are based
The detection of crack-like defects in mechanical sys- on the use of the Fourier transform, which breaks down
tems and civil engineering structures is a problem that a signal into constituent sinusoids of different fre-
received considerable attention from engineering quencies. The Fourier analysis transforms the signal
researchers in the last decades [4]. The research effort from a time-based or space-based domain to a fre-
devoted to this topic is due to its practical importance. quency-based one. Unfortunately, in transforming to the
For example, one of the most important requirements for frequency domain, the time or space information is lost
the proper maintenance of machinery or civil infrastruc- and it is impossible to determine when or where a parti-
ture systems is the detection of crack-like damage at the cular event took place. To correct this deficiency, the
early stages of growth. This leads to the necessity of method that became to be known as the short-time Four-
developing nondestructive techniques for crack detection ier transform (STFT) was proposed by Dennis Gabor
that are both practical and accurate. [1,5,6]. This windowing technique analyzes only a small
There are a large number of nondestructive methods section of the signal at a time. The STFT maps a signal
for crack detection that are based on the changes in the into a 2-D function of time or space and frequency. The
dynamic properties of the structure (frequencies, mode transform has the disadvantage that the information
shapes, transfer functions, etc) caused by the damage. about time or space and frequency can be obtained with
They are, for example, reviewed in [4]. Other techniques a limited precision that is determined by the size of the
window. A high resolution in time or space and fre-
quency cannot be obtained simultaneously since once the

Corresponding author. Tel.: +1-787-832-4040, ext. 3669; fax: +1- window size is chosen, it is the same for all frequencies.
787-833-8260. A more flexible method with a variable size window
E-mail address: lsuarez@ce.uprm.edu (L.E. Suárez). is required to determine with greater precision any parti-

0141-0296/$ - see front matter  2003 Elsevier Ltd. All rights reserved.
doi:10.1016/j.engstruct.2003.08.009
40 A.V. Ovanesova, L.E. Suárez / Engineering Structures 26 (2004) 39–49

cular features of a signal in time or space as a function Wavelet analysis starts by selecting from the existing
of frequency. The wavelet transform is precisely a new wavelet families (see Section 4) a basic wavelet function
way to analyze signals that overcomes the problems that that can be a function of space x or time t. In this paper
other signal processing techniques exhibit. Wavelets are we will consider that the independent variable is x. This
composed of a family of basis functions that are capable basic wavelet function, called the ‘mother wavelet’
of describing a signal in a localized time (or space) and y (x), is then dilated (stretched or compressed) by a and
frequency (or scale) domain [1,3,6,10,12,18]. translated in space by b to generate a set of basis func-
The wavelet transform, like the STFT, is a two-para- tions ya,b(x) as follows:

冉 冊
meter transform. For time signals, the two domains of
the wavelet transform are time t and scale a. The scale 1 x⫺b
ya,b(x) ⫽ y
冑a
. (1)
a can be approximately related to the frequency w [12]. a
For the applications we are interested in, the signals to
be analyzed are function of space and the two parameters The function is centered at b with a spread proportional
of the wavelet transform are space (or position x) and to a. The wavelet transform (in its continuous or discrete
scale a. version) correlates the function f(x) with yab(x). The
The main advantage gained by using wavelets is the continuous wavelet transform (CWT) is the sum over all
ability to perform local analysis of a signal, i.e. to zoom- time of the signal multiplied by a scaled and shifted ver-
in on any interval of time or space. Wavelet analysis is sion of a mother wavelet
thus capable of revealing some hidden aspects of the data
that other signal analysis techniques fail to detect. This
property is particularly important for damage detection
applications. Moreover, due to the availability of a fast
C(a,b) ⫽
冑冕
1
a

⫺⬁
f(x) y 冉 冊
x⫺b
a
dx
(2)
transform version, the computational effort to perform
the signal transformation is reduced. Because of these
features, the wavelet transform is rising as a promising
⫽ 冕 ⬁

⫺⬁
f(x) yab(x)dx,

new method for damage identification in structures. A where the scale a and the position b real numbers and
report on the new wavelet-based methods for damage a ⫽ 0. The results of the transform are wavelet coef-
identification is presented in [15]. ficients that show how well a wavelet function correlates
The objective of this paper is to show the ability of with the signal analyzed. Hence, sharp transitions in f(x)
the wavelet transform to detect cracks by applying the create wavelet coefficients with large amplitudes and this
technique to a number of structural members and simple precisely is the basis of the proposed identification
structures under different loading conditions. The dam- method.
age identification is implemented by means of numerical The CWT has an inverse: the inverse CWT permits
simulations. Beam finite elements with a single non- to recover the signal from its coefficients C(a,b) and is
breathing crack that runs across the width of the member defined as
[8,13] are used in the neighborhood of the crack location
whereas the rest of the structure is modeled with conven-
tional beam elements. Although there are a number of
monographs devoted to the subject of wavelets
f(x) ⫽
1
Ky 冕 冕

a ⫽ ⫺⬁

b ⫽ ⫺⬁
da
C(a,b) ya,b(x) db 2 ,
a
(3)

[1,3,10,11,14,16,18], most of them are highly mathemat- where the constant Ky depends on the wavelet type. One
ical in nature and focused on theoretical aspects. There- of the drawback of the CWT is that a very large number
fore, a brief introduction to the wavelet transform that of wavelet coefficients C(a,b) are generated during the
should be accessible to structural engineers is first analysis. Moreover, few wavelets have an explicit
presented. expression, and most are defined with recursive equa-
tions. It can be shown [3] that the CWT is highly redun-
dant, in the sense that it is not necessary to use the full
2. The wavelet transforms domain of C(a,b) to reconstruct f(x). Therefore, instead
of using a continuum of dilations and translations, dis-
In a nutshell, wavelets are localized waves, in other crete values of the parameters are used. The dilation is
words signals with a zero average value that drop to zero defined as a = 2j and the translation parameter takes the
after a few oscillations. The origin of the wavelet analy- values b = k 2j, where (j, k)苸Z, and Z is a set of integers.
sis as it is known today can be traced to the work of This sampling of the coordinates (a,b) is referred to as
Alex Grossmann and Jean Morlet [9,10]. Wavelets can dyadic sampling because consecutive values of the dis-
be real or complex functions. Since real wavelets are crete scales differ by a factor of 2. Using the discrete
useful to detect sharp signal transitions [10], this paper scales one can define the discrete wavelet transform
deals exclusively with them. (DWT)
A.V. Ovanesova, L.E. Suárez / Engineering Structures 26 (2004) 39–49 41

Cj,k ⫽ 2⫺j/2 冕 ⬁

⫺⬁
f(x) y(2⫺jx⫺k)dx (4)
The new coefficients cAJ(k) are known as the level-J
approximation coefficients. The discrete version of the
reconstruction formula, eq. (7), becomes

⫽ 冕 ⬁

⫺⬁
f(x) yj,k(x)dx.
f(x) ⫽ 冘冉冘
J ⬁

cDj(k) yj,k(x) 冊 (10)


j ⫽ ⫺⬁ k ⫽ ⫺⬁
The signal resolution is defined as the inverse of the
scale 1 / a = 2⫺j, and the integer j is referred to as the

level. As the level and the scale decrease, the resolution ⫹ cAJ(k) fj,k(x).
k ⫽ ⫺⬁
increases and the smaller and finer components of the
signal can be accessed. The function in the parentheses is known as the detail
The signal can be reconstructed from the wavelet function at level j:
coefficients Cj,k and the reconstruction algorithm is


called the inverse discrete wavelet transform (IDWT):
Dj(x) ⫽ cDj(k) yj,k(x). (11)

冘 冘
⬁ ⬁
k ⫽ ⫺⬁
f(x) ⫽ Cj,k 2⫺j/2y(2⫺jx⫺k). (5)
j ⫽ ⫺⬁ k ⫽ ⫺⬁
The function defined by the second summation in eq.
(10) is called the approximation function at level J
Suppose that the wavelet transform Cab is only available
for small scales, say a ⬍ a 0 and we want to recover

the function f(x). In this case we need the complement AJ(x) ⫽ cAJ(k) fj,k(x). (12)
of the information corresponding to C(a,b) for a ⬎ k ⫽ ⫺⬁

a 0. To obtain this information another function f(x), Eq. (10) thus tells that the original function can be
referred to as the scaling function, needs to be introduced expressed as the sum of its approximation at level J plus
[10]. Substituting y(x) by f(x) in eq. (2) one obtain a all its details up to the same level. i.e.


function D(a0,b)

冕 冉 冊
⬁ f(x) ⫽ AJ(x) ⫹ Dj(x). (13)
x⫺b 1
D(a0,b) ⫽ f(x) f jⱕ J
冑a 0 a 0
dx
⫺⬁
(6) For this study, we are interested in the detail signals. As

⫽ 冕⬁

⫺⬁
f(x) fa0,b(x) dx.
it will be shown with the numerical examples, if f(x) is
a response signal, typically the deflection curve, the sig-
nals Dj(x) contain the information necessary to detect the
It can then be shown that the inverse CWT in eq. (3), i.e. cracks in the structure.
the reconstruction formula, can be divided as follows: Eqs. (8) and (9) describe the DWT whereas eqs. (11),

冕 冉冕 冊
a0 ⬁
(12) and (13) constitute the IDWT. Note that the signal
1 da f(x) is still continuous. In practice, only a sample of the
f(x) ⫽ C(a,b)ya,b(x)db (7)
Ky 0 ⫺⬁ a2 signal is available, and the numerical implementation of

冕 ⬁ the DWT and its inverse is done by means of the fast


1
⫹ A(a0,b)fa0,b(x) db. wavelet transform (FWT). The FWT is a set of algor-
Kya0 ⫺⬁ ithms discovered by Stéphane Mallat [9] which make
The scaling function does not exist for every wavelet. use of the concept of low-pass and high-pass digital fil-
The existence of the function f(x) is important for the ters. These filters, referred to as filter banks, implement
numerical implementation of the fast wavelet transform a fast orthogonal transform that requires only O(N) oper-
discussed later. ations for a signal discretized as a sequence f(n) of length
Suppose now that the dyadic scale is used for a and N. For the applications of the wavelet transform
b, and consider a reference level J. Applying eq. (4) for described in this paper, it is not necessary to know the
this case one obtains a set of coefficients details the digital filter interpretation of the FWT. The


interested readers are referred to the book by Strang and

Nguyen [16] and the ‘User’s Guide of the Wavelet Tool-
cDJ(k) ⫽ f(x) yJ,k(x) dx. (8)
⫺⬁ box’ [12].
The coefficients cDJ(k) are known as the level-J detail
coefficients. Using the dyadic scale and level J, eq. (6)
yields another set of coefficients 3. Properties of wavelets

cAJ(k) ⫽ 冕 ⬁

⫺⬁
f(x) fJ,k(x) dx. (9) Many types of wavelets were developed whose fea-
tures and performance vary based on their associated
42 A.V. Ovanesova, L.E. Suárez / Engineering Structures 26 (2004) 39–49

properties [2,3]. Some of the most relevant properties to carry out the DWT using the FWT. The disadvan-
are: (1) regularity; (2) support; (3) number of vanishing tages of this type of wavelets are their poor regularity
moments; and (4) symmetry. A brief definition of these as well as that symmetry and exact reconstruction are
properties, including references for a more detailed incompatible (except for the Haar wavelet) if the
explanation, is given next. same finite impulse response (FIR) filters are used for
reconstruction and decomposition [7,16]. Symmetry
1. Regularity is defined as follows: if r is an integer and in an FIR filter guarantees linearity in the frequency
a function is r-time continuously differentiable at x0, response phase, a very desirable property in many
then the regularity is r. If r is not an integer, let n be applications [14].
the integer such that n ⬍ r ⬍ n + 1, then function 4. To avoid the limitations of the orthogonal wavelets,
has a regularity of r in x0 if its derivative of order n the biorthogonal wavelets were introduced [2,3,12].
resembles (x⫺x 0)r⫺n locally around x0. This property The biorthogonal wavelets are compactly supported
is useful for getting nice features, such as smoothness, biorthogonal spline wavelets for which symmetry and
of the reconstructed signals [11,18]. exact reconstruction are possible with FIR filters. Two
2. The support of a function is the smallest space-set wavelets and scaling functions, instead of one, are
(or time-set) outside of which function is identically used: ȳj,k is used for the analysis to calculate the coef-
zero [1,10,17,18]. ficients C̄j,k of a signal f:


3. The number of vanishing moments of wavelets deter-
mines the order of the polynomial that can be approxi- C̄j,k ⫽ f(x)ȳj,k(x)dx
mated and is useful for compression purposes
[1,3,11].
and yj,k is used for the synthesis to reconstruct signal
4. The wavelet symmetry relates to the symmetry of the
f(x) from the coefficients C̄j,k
filters and helps to avoid dephasing in image pro-
cessing. Among the orthogonal families, the Haar
wavelet is the only symmetric wavelet. For biorthog-
f(x) ⫽ 冘 C̄j,kyj,k(x).
j,k
onal wavelets it is possible to synthesize wavelet
functions and scaling functions that are symmetric or Examples are the biorthogonal wavelets of order
antisymmetric [3,7,16]. Nr.Nd, where Nr and Nd are the wavelet orders for
reconstruction and decomposition, respectively.

4. Classification of existing wavelets


5. Procedure for crack detection
This section provides a classification of some of the
most well known wavelets. The classification is restric- 5.1. Criteria for optimal wavelet selection
ted to those available in the Matlab Toolbox [12]
because this will be the software used later to implement To apply the wavelet transform in any application it
the identification methodology. is important to select the most appropriate (or optimal)
wavelet for the analysis. The selection is usually done
1. The Gaussian, Mexican Hat, Morlet and Shannon by trial and error, but by examining the properties one
wavelets represent examples of wavelets for which can discard many candidates and expedite the process.
the wavelet function y has an explicit expression To analyze the response signals in this study the optimal
[3,12]. The scaling functions j do not exist, and thus wavelet was chosen according to the following criteria:
the DWT, FWT, and the discrete reconstruction are
unavailable. The analysis is limited to the CWT. 1. From the available wavelets, eliminate those that do
2. The Meyer wavelet is an example of infinitely regular not allow carrying out a FWT. They are the Gaussian,
wavelet [2,3,11,12]. This wavelet does not have an Mexican Hat, Morlet, Shannon and Meyer wavelets.
explicit expression for the wavelet function y. The 2. The orthogonal and biorthogonal wavelets remain
scaling function j exists and the DWT is possible, after the first elimination. Both permit to apply the
although a fast algorithm (FWT) is unavailable. FWT using filter banks.
3. The Haar, the Daubechies wavelets of order N, the 3. The requirement to satisfy symmetry and exact recon-
Symlet wavelets of order N, the Coiflet wavelets of struction of the analyzed signal limit the choice to the
order N are examples of orthogonal wavelets Haar wavelet and biorthogonal wavelets.
[2,3,12,17]. These wavelets do not have an explicit 4. The two candidate wavelets have a significant differ-
representation for the wavelet function y, except for ence in regularity. Thus, regularity is the last property
the Daubechies wavelet of order 1, which is the Haar that singles out the wavelet to perform the analyses.
wavelet. The existence of scaling functions j permits The irregularity of the Haar wavelet leaves the bior-
A.V. Ovanesova, L.E. Suárez / Engineering Structures 26 (2004) 39–49 43

thogonal wavelets as the choice for the analyses (see will be detected by the distribution of coefficients on
Case 1 (dynamic analysis) in Section 6). the wavelet coefficients plot. In order to observe the
5. The regularity of the different biorthogonal wavelets signal discontinuities, it is recommended to perform
increase with the wavelet order N (see point (1) in a low-scale analysis for the CWT or a low-level
Section 3). The wavelet available with the highest analysis for the DWT.
Nr.Nd (bior6.8) is selected as the most suitable (see 4. In the case of the DWT, the level at which the wavelet
Fig. 1). analysis must be performed cannot be determined
beforehand. It depends on the nature of the signal, the
5.2. Wavelet methodology for damage detection characteristics of the structure, the location, type and
severity of the damage, etc. It is suggested to try with
The main idea behind the use of wavelets for struc- different levels. As it will be shown in Section 6, in
tural identification purposes is based on the fact that the order to identify discontinuities due to cracks in sim-
presence of cracks introduces small discontinuities in the ple structures, it is sufficient to perform only one level
structural response at the damaged sites [8,13,19]. Often of analysis or consider only the first level detail sig-
these discontinuities cannot be observed from the exam- nal.
ination of the structural response, but they are detectable
from the distribution of the wavelet coefficients obtained
by the CWT or the detail signals form the DWT. The 6. Numerical studies
following procedure is proposed to detect crack-like
damage in a structure by the CWT or DWT as follows. The ability of wavelets to detect cracks in a structural
member and in a simple structure is demonstrated with
1. Measure (or for numerical simulations, calculate) the several numerical examples. In particular, special atten-
signal associated with the static or dynamic structural tion is given to the performance of the method when the
response, usually the transverse displacements. For crack is close to a support or to a point of geometric
the numerical simulations of frame structures, one can discontinuity such as a corner in a frame. The structures
use at the crack location short length elements with chosen to present numerical results include a beam with
reduced area and moment of inertia, or the specially two fixed ends and a one-story one-bay plane frame.
developed cracked beam elements available in the Results for beams with different support conditions can
technical literature. Alternatively, if 3D solid be found elsewhere [13]. It was discussed in Section 5.1
elements are used, it is possible to simulate the crack that the bior6.8 wavelet is the most suitable wavelet for
simply by removing elements or decreasing their elas- analysis. However, the Haar wavelet was also used in
tic modulus. some cases for comparison purposes.
2. Using the measured signal, compute its continuous
wavelet coefficients from eqs. (2) or (4), or the detail 6.1. Example 1: fixed-end beam
signals associated with the discrete transform from
eqs. (8) and (11). The fixed-end uniform beam shown in Fig. 2 is con-
3. Plot and examine the wavelet coefficients for the sidered for the first example. The beam is subjected to
CWT or the detail signals in the case of the DWT. either a dynamic load consisting of a half-sine impulse
Provided that a suitable wavelet is selected for the with magnitude 4.5 kN or a concentrated static load with
analysis (see Section 5.1), any signal discontinuity magnitude 3.5 kN both applied at the distance L f = L /

Fig. 1. Biorthogonal wavelet: bior6.8.


44 A.V. Ovanesova, L.E. Suárez / Engineering Structures 26 (2004) 39–49

Fig. 2. Fixed-end beam under a dynamic or static load.

2 from the left support. The geometric properties of the


beam are the following: the length is L = 3 m, the height
is h = 0.15 m, and the depth is w = 0.15 m. The crack
depth d and the crack location Lcr are varied during the
analysis. The material properties are as follows: modulus
of elasticity E = 31 GPa, mass density r = 2.3 kg / m3, Fig. 4. Fixed-end beam deflection at time t = 0.048 s after impact.
and Poisson ratio n = 0.2.

6.1.1. Frequency analysis 6.1.2. Case 1 (dynamic analysis)


To estimate the effect of the crack on the natural fre- The fixed-end beam in Fig. 2 is subdivided into 100
quencies, a free vibration analysis of the fixed-end beam finite elements and the crack with depth d = 1.5 ×
without and with crack was undertaken. The finite 10⫺3m is introduced at distances L cr = L / 12 (node 9)
element model used to represent the structure with a and L/25 (node 4) from the left support. The dynamic
crack is described in [13]. The eigenfrequencies of the response of the uncracked and cracked beam at time t
damaged beam were obtained for various crack depths = 0.048 s after an initial impact is applied is given in
and locations. Only the fundamental natural frequencies Fig. 4. The deflection of the beam is smooth except near
are plotted in Fig. 3. It is observed that the crack depth the position of the crack (nodes 4 and 9) where a very
has a small effect on the frequencies, and although the small perturbation can be seen. Most likely, these pertur-
location of the crack has a more notable effect, the vari- bations will go undetected if the response is measured
ations are within the range of the measurement errors. experimentally and a simple visual inspection is conduc-
This confirms that the natural frequencies alone cannot ted. When the deflection signals of the cracked beam are
be used to detect cracks, unless the magnitude of the analyzed with the DWT, these small disturbances are
cracks is such that its presence is obvious. Although not detected by the bior6.8 wavelet (Figs. 5 and 7). Note
presented here, the situation is the same with the that in this and in the following examples, the exact
vibration modes [13]. crack location is identified by the peak with the
maximum absolute value in the plots of the detail signal
calculated with the DWT. To show the effect of the type

Fig. 3. First natural frequency of the fixed-end beam versus crack


depth. Fig. 5. Case 1: DWT by the bior6.8 wavelet (crack at node 9).
A.V. Ovanesova, L.E. Suárez / Engineering Structures 26 (2004) 39–49 45

Fig. 6. Case 1: DWT by the Haar wavelet (crack at node 9). Fig. 8. Case 1: DWT by the Haar wavelet (crack at node 4).

of wavelet on the detection capabilities, Figs. 6 and 8


show the results obtained applying the first wavelet dis-
covered, the well-known Haar wavelet [3,10]. In this
case the method is not quite reliable: when the crack is
at node 9, a peak is clearly evident in Fig. 6; however,
if the crack is at node 4, the corresponding peak cannot
be detected from Fig. 8. The poor regularity of the Haar
wavelet explains the inconsistent results during the
analysis compared to the bior6.8 wavelet (see point (1)
in Section 3 and Section 5.1).

6.1.3. Case 2 (static analysis)


The beam in Fig. 2 is now subdivided into 500 finite
elements and the crack is introduced at a distance L cr
= L / 12 from the left support (at node 42). The deflection
shape of the beam under a static force without and with Fig. 9. Fixed-end beam deflection due to a static load.
crack is shown in Fig. 9. After the DWT is applied to
the response signal of the uncracked beam, wavelet
the load, applied at node 250, and due to the supports
analysis was able to pick up signal disturbances due to
at the ends (Fig. 10). Next, a crack with depth d = 1.5
× 10⫺3m is introduced to the beam at L cr = L / 12 (at
node 42) and both its presence and exact location are
detected from the cracked beam response signal by the
bior6.8 wavelet (Fig. 11).
In order to provide an alternative to the DWT as well
as to show the importance of using low-scale analysis,
the CWT is applied to the static deflection signal.
Although, as discussed previously, the CWT is redun-
dant, it can help to interpret the results. Figs. 12 and 13
show plots of the wavelet coefficients C(a,b) for the case
of the undamaged and damaged beam, respectively. In
these figures, the color at each point of the a–b domain
is associated to the magnitude of the wavelet coefficients
with the lighter color corresponding to the larger coef-
ficients and darker color to the smaller ones. In the case
of the uncracked beam, the bior6.8 wavelet detected dis-
continuities at the supports and at the point of the applied
Fig. 7. Case 1: DWT by the bior6.8 wavelet (crack at node 4). load (Fig. 12). In the case of the cracked beam, the dis-
46 A.V. Ovanesova, L.E. Suárez / Engineering Structures 26 (2004) 39–49

Fig. 10. Case 2: DWT of the uncracked beam response by the Fig. 13. Case 2: CWT by the bior6.8 wavelet (crack at node 42).
bior6.8 wavelet.

continuity due to the crack is clearly detected (Fig. 13),


especially in a low-scale analysis (Fig. 14).

6.2. Example 2: simple plane frame

In this example the one-story, one-bay plane frame in


Fig. 15 subjected to a single static load at different pos-
itions is chosen as a case of study. The frame dimensions
are: columns’ height L column = 1.5 m, beam’s length
L beam = 2.0 m, and the crack depth is d = 1.5 × 10⫺3m.
The cross-sectional area and the moment of inertia of all
the members are A = 0.0225 m2 and I = 4.2 × 10⫺5
m4. The crack is introduced at various locations into the
column and beam. In this example, the effect of changes
in the geometry on the wavelet performance is exam-
ined. To obtain the structural response numerically, each
element of the frame is subdivided using 50 finite
Fig. 11. Case 2: DWT by the bior6.8 wavelet (crack at node 42). elements.

Fig. 12. Case 2: CWT of the uncracked beam response by the Fig. 14. Case 2: low-scale analysis by the bior6.8 wavelet (crack at
bior6.8 wavelet. node 42).
A.V. Ovanesova, L.E. Suárez / Engineering Structures 26 (2004) 39–49 47

Fig. 15. One-bay plane frame under a static load.

Fig. 17. Case 1: zoom-in on details in Fig. 16.

6.2.1. Case 1
The frame with a cracked column (member 1) and
subjected to a horizontal static load applied at joint B is 6.2.2. Case 2
first considered. The crack is introduced in the column We consider one more time the frame with a cracked
(member 1) at a distance L cr = L column / 4 (node 13) from column (member 1) subjected to a horizontal static load
the fixed support A (Fig. 15). The wavelet transform of applied at joint B. The purpose here is to analyze the
the response signal is computed using the bior6.8 wave- effect of the distance between the crack and the point
let and the results (the level-1 detail signal) are shown of geometry change on the performance of the wavelet
on Fig. 16. The disturbances indicate the location in analysis. For this we locate the crack at a distance L cr
space of the discontinuities due to the crack (node 13) = L column / 5 (node 40) close to the corner joint B (Fig.
and due to the changes in the frame geometry (nodes 15). The signal coming from the static response is ana-
51, 101 and 151). A local analysis of the signal is perfor- lyzed using the bior6.8 wavelet. Small peaks due to the
med in order to estimate the exact crack location by zoo- crack discontinuity can be seen on Fig. 18 and after zoo-
ming-in at the suspected area of damage. The plot in ming-in on the level-1 detail signal, the drastic changes
Fig. 17 now shows more clearly the exact location in indicate the exact crack location (Fig. 19). Of course,
space of the discontinuity caused by the crack at node the much larger oscillations are known to be caused by
13. the change in geometry and thus are ignored.

Fig. 16. Case 1: DWT of frame response by the bior6.8 wavelet Fig. 18. Case 2: DWT of frame response by the bior6.8 wavelet
(crack at node 13). (crack at node 40).
48 A.V. Ovanesova, L.E. Suárez / Engineering Structures 26 (2004) 39–49

Fig. 19. Case 2: zoom-in on details in Fig. 18. Fig. 21. Case 3: zoom-in on details in Fig. 20.

6.3. Discussion of results


6.2.3. Case 3
In the last example we consider the plane frame with A fixed-end beam and a three-member plane frame
a crack in the beam (member 2) at a distance L cr = were subjected to dynamic and static concentrated loads
L beam / 4 (node 64) from joint B and subjected to a verti- and their responses were numerically calculated.
cal static load applied at the middle of the beam Although the detection procedure does not require
(member 2, node 75), as depicted in Fig. 15. The first knowledge of the response of the undamaged structures,
level detail signal obtained by applying the DWT to the the wavelet analysis was also applied to their response
frame deflected shape is shown in Fig. 20. Once again, signals. The analysis of the undamaged structures was
the bior6.8 wavelet was used for this purpose. Here the done to show that the discontinuities in the response sig-
supports A and D (nodes 1 and 151), as well as the cor- nal introduced by the supports, the corners, and the
ner joints B and C (nodes 51 and 101) introduce large points of application of the external loads are all detected
spikes in the signal. However, a small disturbance at the by the wavelet transform. The static and dynamic
crack location (node 64) can also be observed. In order responses were calculated again after introducing a
to evaluate the exact crack location it is necessary to crack. The responses (i.e. deflected shapes) were pro-
zoom-in the detail signal, as shown in Fig. 21. cessed with the wavelet transform to detect any feature
in the signals other than those associated with the normal
discontinuities. It was found that a sudden peak, which
occurs due to unknown sources, pinpoints the presence
of a crack.
For the case of a plane frame, and in order to test the
method, the responses were obtained with a load applied
at one of the corners and in the middle of the beam. It
was noticed that the position of the applied load does
not affect the performance of the proposed procedure
significantly. However, the discontinuities in the
response due to the crack are more evident if the load
is applied at the middle of the beam or column. Never-
theless, all the numerical simulations presented in the
paper and additional examples [12] not reported here
showed that the wavelet analysis is capable of revealing
the discontinuities in the response signal.
The results of the study indicated that the effective-
ness of wavelet analysis and the interpretation of the
results are sensitive to the members’ boundary con-
Fig. 20. Case 3: DWT of frame response by the bior6.8 wavelet. ditions (discontinuity due to the supports or the geometry
(crack at node 64). changes at the beam-column joints) and distance from
A.V. Ovanesova, L.E. Suárez / Engineering Structures 26 (2004) 39–49 49

structural supports and joints. In some cases, a local Acknowledgements


analysis of the signal in the vicinity of the area where the
damage is suspected was helpful to identify the crack. It This research was supported by the Civil Engineering
is concluded that, among those available for the present Department of the University of Puerto Rico at Mayag-
study, the biorthogonal wavelet of order 6.8 (bior6.8) is uez. This support is gratefully acknowledged.
the most appropriate wavelet for crack detection in
beams and frames. However, more research is needed
to verify the performance of this wavelet for damage References
identification in more complex structural systems.
[1] Chui CK. Wavelets: a mathematical tool for signal analysis.
SIAM Monographs on Mathematical Modeling and Computation.
Philadelphia, PA: SIAM, 1997.
[2] Cohen A, Daubechies I, Feauveau JC. Biorthogonal basis of com-
7. Concluding remarks pactly supported wavelets. Comms on Pure and Appl Mathemat-
ics 1992;45:485–560.
[3] Daubechies I. Ten lectures on wavelets. CBMS–NSF Conference
Unlike other damage identification techniques, wave- Series, 61. Philadelphia, PA: SIAM, 1992.
let-based methods can be applied not only to structural [4] Doebling SW, Farrar CR, Prime MP, Shevitz DW. Damage
members but also to full structures. In addition, the dam- identification and health monitoring of structural and mechanical
systems: a literature review. Technical Report No. LA-13070.
age can be detected using the static response of the struc- Los Alamos, New Mexico: Los Alamos National Laboratories;
tures. This is a very useful feature of the method since 1996.
it is much easier and more inexpensive to measure the [5] Gabor D. Theory of communication. IEEE Journal
static response compared to the dynamic one. Another 1946;93:429–57.
fact that makes the wavelet-based methodology easy to [6] Gurley K, Kareem A. Application of wavelet transforms in earth-
quake, wind and ocean engineering. Engng Struct
implement in practice is that it can be used for structural 1999;21:149–67.
monitoring at the expected areas of damage only. All [7] Herley C, Vetterli M. Wavelets and recursive filter banks. IEEE
these properties make the method a potentially reliable Transa Sig Process 1993;41(8):2536–56.
and cost-effective assessment technique that can be [8] Liew KM, Wang Q. Application of wavelet theory for crack
applied to the maintenance of the built infrastructure identification in structures. J Engng Mech 1998;124(2):152–7.
[9] Mallat SGA. Theory for multiresolution signal decomposition:
(bridges, water and gas distribution pipelines, electrical the wavelet representation. IEEE Trans Pattern Anal and Mach
and transmission towers, dams and other hydraulic Intell 1989;11(7):674–93.
structures). However, further analytical research is [10] Mallat SG. A wavelet tour of signal processing, 2nd ed. London:
needed to advance the method: for example, to establish Academic Press, 1999.
a rigorous methodology for the selection of the optimal [11] Meyer Y. Wavelets and operators. Cambridge Studies in
Advanced Mathematics, 37. Cambridge: Cambridge University
wavelet and to derive a procedure to eliminate the effects Press, 1992.
of the boundary conditions on the wavelet-based crack [12] Misiti M, Misiti Y, Oppenheim G, Poggi J-M. Wavelet tool-
identification. We have shown that the method can be box — user’s guide. Natick, MA: The Math Works Inc, 2000.
applied when the crack is near a support or a corner [13] Ovanesova AV. Applications of wavelets to crack detection in
joint. However, the case where the damage is actually frame structures. Ph.D. Thesis, Civil Engineering Department.
College of Engineering, University of Puerto Rico, Mayagüez,
in the joint itself was not considered. This can be Puerto Rico, 2000.
important, for example, to detect damage in the beam- [14] Rao RM, Bopardikar AS. Wavelet transforms — introduction to
column connections of steel buildings. Other future theory and applications. Reading, MA: Addison Wesley Long-
extensions of the method are its applications to bi- man, 1998.
dimensional structures such as plates and shells. More [15] Staszewski WJ. Structural and mechanical damage detection
using wavelets. The Shock and Vibration Digest
importantly, experimental studies are urgently needed to 1998;30(6):457–72.
validate the results obtained from the numerical simula- [16] Strang G, Nguyen T. Wavelets and filter banks. Wellesley, MA:
tions. Real signals obtained by measuring the structural Wellesley–Cambridge Press, 1996.
response should be analyzed by the wavelet transform [17] Walker JS. A primer on wavelets and their scientific applications.
in order to demonstrate the reliability and robustness of Boca Raton, FL: Chapman and Hall/CRC, 1999.
[18] Walter GGW. Wavelets and other orthogonal systems with appli-
the technique in the presence of noise and the unavoid- cations. Boca Raton, FL: CRC Press, 1994.
able uncertainties associated with experimental pro- [19] Wang Q, Deng X. Damage detection with spatial wavelets. Int J
cedures. of Solids and Struct 1999;36:3443–68.

You might also like