See discussions, stats, and author profiles for this publication at: https://www.researchgate.

net/publication/264146634

Signal processing based damage detection in structures subjected to random

excitations

Article in STRUCTURAL ENGINEERING AND MECHANICS · December 2011

DOI: 10.12989/sem.2011.40.6.745

CITATIONS READS
18 118

1 author:

Luis Montejo
University of Puerto Rico at Mayagüez
59 PUBLICATIONS 898 CITATIONS

SEE PROFILE

All content following this page was uploaded by Luis Montejo on 15 November 2019.

The user has requested enhancement of the downloaded file.

[Montejo, Luis A. "Signal processing based damage detection in structures subjected to random excitations." Structural Engineering and Mechanics 40.6
(2011): 745-762.]
Author preprint
Final published version: http://www.koreascience.or.kr/article/ArticleFullRecord.jsp?cn=KJKHB9_2011_v40n6_745
Related computer codes: https://sites.google.com/a/upr.edu/montejo/computer-cod

Signal Processing Based Damage Detection in Structures Subjected

to Random Excitations

Luis A. Montejo*

Department of Engineering Science and Materials, University of Puerto Rico at Mayagüez, Mayagüez, Puerto Rico

Abstract. Damage detection methodologies based on the direct examination of the nonlinear-nonstationary

characteristics of the structure dynamic response may play an important role in online structural health

monitoring applications. Different signal processing based damage detection methodologies have been proposed

based on the uncovering of spikes in the high frequency component of the structural response obtained via

Discrete Wavelet transforms, Hilbert-Huang transforms or high pass filtering. The performance of these

approaches in systems subjected to different types of excitation is evaluated in this paper. It is found that in the

case of random excitations, like earthquake accelerations, the effectiveness of such methodologies is limited.

An alternative damage detection approach using the Continuous Wavelet transform is proposed to overcome

this limitation.

Keywords: Damage detection; Wavelet transform; Empirical mode decomposition; Butterworth filters; Hilbert

Transform.

1. Introduction

Although visual or localized experimental damage-detection methods (e.g. acoustic, ultrasonic,

magnetic or thermal methods) are capable of identifying and determining the extent of damage, effective

application of all of these techniques require previous knowledge of the damage vicinity and the portion of the

structure under inspection to be easily reached (Doebling et al. 1998). Moreover, a comprehensive application

of localized damage detection methods to a complex structure is a lengthy and expensive assignment. An

inspection of the Golden Gate Bridge main cables, anchorages and tie-downs was performed at a cost of

$226,900 in 1998 (Sohn et al. 2003). Visual inspections after the M6.3 Abruzzo earthquake (Italy, April 6, 2009)

lasted more than 4 months and deployed about 1500 inspectors daily (EERI 2009).

*E-mail: [email protected]
 The aforementioned localized damage detection methods limitations have been one of the main

motivations for the continuous research and development of quantitative global damage-detection methods.

Damage identification based upon changes in vibration characteristics has been the preferred approach to

monitor changes in the structure on a global basis. Vibration based damage detection (VBDD) fundamental

principle is that changes in the physical properties of the structure (mass, energy dissipation mechanisms and

stiffness) will cause changes on the modal parameters (frequencies, mode shapes, and modal damping). Detailed

literature reviews concerning these techniques and response parameters used for damage identification can be

found elsewhere (Doebling et al. 1996, Farrar and Doebling 2001, Sohn et al. 2003)

Despite the apparently intuitive nature of VBDD, different researchers and practitioners have found

significant technical challenges: (1) In order to avoid premature collapse modern civil structures are designed

to be highly redundant, therefore the occurrence of damage (which is typically a local phenomenon) may not

significantly influence global response of a structure (Farrar and Doebling 2001, Farrar and Cone 1995). (2)

Since structural damage is usually reflected in the high frequency response of the structure, methods that monitor

changes in the shape modes require a large spatial resolution in the sensor network in order to capture high order

modes. Such a dense sensor network array is not usually available in complex civil structures. (3) Some of the

available methodologies largely depend on the updating of a detailed finite element (FE) model. Due to large

computational demands and difficulties involved in obtaining the FE model, this approach may become

unfeasible for complex structures (Rodgers and Çelebi 2006). (4) Many algorithms presume that a data set from

the undamaged structure is available and the damage detection is performed by comparison of the dynamic

response characteristics of the pristine and damaged structures. This type of approaches has the inconvenient

that the events that occurred between dynamic measures are lost and significant information can be missed.

Todorovska and Trifunac (2007) investigated the frequency variations during the 1979 Imperial Valley

earthquake of a six-story reinforced concrete structure that was severely damaged by the earthquake. They

detected a decrease in the system frequency of about 44% that can be attributed to structural damage followed

by a 35% increase at the end of the recorded shaking. Notice that in this case the severity of damage would be

underestimated if the damage assessment was to be performed based only on the vibration characteristics of the

structure before and after the earthquake. (5) Civil structures are in continuous interaction with the soil they are

Montejo, 2010 Signal-Processing Based Damage Detection in Structures Subjected to Random Excitations
founded on and with the environment. Experimental results have shown that the changes in vibration properties

as a result of such interactions can be of larger proportions than the changes induced in the early stages of

damage (Huth et al. 2005, Clinton et al. 2006).

In order to overcome the issues of conventional VBDD methodologies some authors have explored the

possibility of identifying damage from the direct examination of the nonlinear-nonstationary characteristics of

the dynamic response, eliminating in this way the dependency on large and detailed FE models or on the prior

knowledge of the undamaged structure vibration characteristics. Analysis of the registered structural response

is performed using high pass filtering or signal processing tools that allow for simultaneous time frequency

examination, e.g. Wavelets (Mallat 1989) and Hilbert-Huang transforms (Huang et al. 1998). The occurrence

of damage is associated with any changes in the vibration parameters (instantaneous frequency or damping) or

with the occurrence of singularities in the high frequency response. This paper focuses on the second alternative,

i.e. the detection of singularities (novelties or spikes) in the high frequency component of the structural response.

Investigations on the parameter identification capabilities of these transforms are available elsewhere (Kijewski

and Kareem 2003, Yang et al. 2004, Yan and Miyamoto 2006, Kijewski-Correa and Kareem 2006).

2. Simultaneous time-frequency analysis

A time-frequency analysis of a signal can provide information about how the frequency content of the

signal evolves with time, thus providing a tool to dissect and interpret strongly non-stationary signals. Methods

like the Short Time Fourier transform allow for the construction of a time-frequency map of the signal, however,

its constant resolution time-frequency window is not suitable for detecting certain phenomena out of the

resolution of the window. More recently developed methodologies like Wavelet and Hilbert Huang transforms

were devised to overcome these limitations. These methods are able to provide information about the slow

varying phenomena of a signal and at the same time, permit the detection of sudden discontinuities.

2.1 Wavelet Transforms

Montejo, 2010 Signal-Processing Based Damage Detection in Structures Subjected to Random Excitations
 As with the Fourier transform, there are two types of wavelet transform: the continuous and discrete

version. The Continuous Wavelet Transform (CWT) takes a function x(t) of a real variable and transforms it to

a function C(s, p) of two real variables s and p:

∞
C ( s , p ) = ∫ x (t )ψ s , p (t )dt (1)
−∞

The wavelet coefficients C(s, p) contain information about the function x(t) at the scale s around the time

position p. The functions ψs,p(t) are defined by translating along the time axis and stretching or compressing a

“mother wavelet” Ψ(t):

1 t − p
ψ s , p (t ) = Ψ  (2)
s  s 

In order to recover the original signal f(t) it is necessary to use the “reconstruction formula”:

1
∞  ∞ dp  1
∞
x(t ) =
Kψ ∫s=0  p=∫−∞ C (s, p)ψ s , p (t ) s 2 ds = Kψ ∫ D(s, t )ds (3)
  0

where Kψ is a constant that depends on the mother wavelet. The functions within parenthesis in Eq. (3) are

referred to as the “detail functions” D(s,t) and have a dominant frequency that depends on the type of wavelet.

In the CWT the wavelet coefficients are calculated at scale values that vary continuously, the result is a highly

redundant and time demanding representation of the signal. In the Discrete Wavelet Transform (DWT) the

values of scale and position are choose based on a dyadic scale, for some special mother wavelets ψ(t) the

corresponding discretized wavelets ψs,p(t) constitute an orthonormal basis. Mallat (1989) developed a fast

wavelet decomposition and reconstruction algorithm for the DWT using a two-channel subband coder. In the

DWT, a signal can be represented (Eq. 4) by its approximations (A) and details (D) at different levels of

decomposition (j). The approximations are the high-scale, low-frequency components of the signal. The details

are the low-scale, high-frequency components.

x(t ) = A j + ∑ Di (4)
i≤ j

Montejo, 2010 Signal-Processing Based Damage Detection in Structures Subjected to Random Excitations
 To demonstrate the application of the wavelet transforms, the signal displayed in Fig. 1a was artificially

generated as the sum of a sinusoidal of frequency 8Hz and a chirp signal starting with a frequency of 0.5Hz and

increasing linearly up to a frequency of 4Hz at time 8 seconds. At time 4 seconds a small discontinuity is

introduced by multiplying the signal amplitude at that point by 0.99. The Fourier spectrum of the signal is

presented in Fig.1b, it is seen that the Fourier spectrum captured the correct range of frequencies of the signal,

but the evolution in time of the frequency content is missed. An analysis in the time-frequency domain using

the CWT, shown in Figs. 2, allows us to identify the time where each event took place. Fig. 2a shows the real

values of the coefficients C(s,p) that result from the application of the CWT (Eq.1) using the Complex Morlet

Wavelet (Eq. 5). The Morlet Wavelet is essentially a sinusoid modulated by a Gaussian envelope offering

optimal resolution in both the time and frequency domain. Fig. 2b is similar to Fig. 2a, but now the wavelet

coefficients C(s,p) are shown in a two-dimensional graph with their absolute values plotted. This is the usual

form to graph the coefficients and it is called a Wavelet Map. The darker colors indicate higher values of the

wavelet coefficients. By observing the darker colors in the graph, one can perceive the frequency content of the

signal at any time instant. A more precise estimation of the instantaneous dominant frequencies can be obtained

by identifying each component by a distinct ridge in the time-frequency plane. There are several techniques to

identify these ridges (Carmona et al. 1997), the instantaneous dominant frequencies in Fig. 2c were obtained by

locating the local maximas at each time instant. It is seen that the dominant frequencies and its evolution in time

are successfully identified for most part of the signal. However, a deviation in the identified instant frequencies

(IF) is noticed at the beginning and end of the signal due to the end effects. Different methodologies haven

proposed to meliorate the end effects in the CWT, e.g. padding the beginning and end of the signal with surrogate

values (Kijewski and Kareem 2003). The results displayed in Figs. 2 were obtained from the original signal

without padding.

ψ (t ) = e i 2πf e −t / 2σ 2
2
o
(5)

If well the frequency content evolution of the signal was properly captured by the CWT analysis, the

discontinuity at 4 seconds was not detected. This is because small discontinuities are usually reflected in the

high frequency range and the analysis performed was limited to the range 0-10 Hz. One may extend the

Montejo, 2010 Signal-Processing Based Damage Detection in Structures Subjected to Random Excitations
frequency range over which the CWT is performed until traces of the discontinuity are detected, however this

is computationally expensive and the preferred approach is the decomposition of the signal in details and

approximations via DWT (Eq. 4). Note that the Morlet wavelet does not comply with the orthogonality

requirements to perform a DWT analysis; therefore the DWT was implemented using the Biorthogonal (Bior)

6.8 basis. The Biorthogonal wavelets are compactly supported biorthogonal spline wavelets for which symmetry

and exact reconstruction are possible. Moreover, the Bior 6.8 has been successfully used in the past to uncover

discontinuities (Ovanesova and Suarez 2004, Todorovska and Trifunac 2009). Contrary to the Morlet wavelet,

Bior wavelets do not have an explicit expression for the wavelet function. The results for the first level

decomposition are displayed in Figs. 3. The figure on top shows the low frequency component of the signal, i.e.

the approximation function (A), and the figure on the bottom shows the high frequency component, i.e. the detail

function (D). It is seen that the discontinuity in the signal can be associated with the spike at 4 seconds in the

detail function.

2.2 Hilbert Huang Transform

The instantaneous frequency (IF) of a monocomponent signal (i.e. signals in which only one frequency

is present at a given time) can be calculated using the Hilbert transform, through which the complex conjugate

y(t) of the real valued function x(t) can be determined by:

∞
P x( s )
H [ x(t )] =
π ∫
−∞
t−s
ds (6)

where P is the Cauchy principle value. With the Hilbert Transform, the analytic signal is defined as:

z (t ) = x(t ) + iH [ x(t )] = x(t ) + iy (t ) = a (t )e iθ (t ) (7)

where:

a(t ) = x 2 + y 2 and θ (t ) = arctan( y / x) (8)

Here a(t) is the instantaneous amplitude and θ is the phase function. From the definition in Eq.7, the concept of

IF is defined as the time-varying derivative of the phase (θ):

Montejo, 2010 Signal-Processing Based Damage Detection in Structures Subjected to Random Excitations
 1 d
f (t ) = θ (t ) (9)
2π dt

It must be noticed that the HT cannot be directly applied to multi-component signals (i.e. signals in

which more than a frequency is present at a given time), as the result will be only one (average) instant

frequency. To successfully apply the HT to multicomponent signals, the signal needs to be preprocessed into

their monocomponent elements (e.g. by bandpass filtering) before implementation of the transform. Huang et

al. (1998) introduced the concept of Empirical Mode Decomposition (EMD) as an alternative way to separate

multicomponent signals into their monocomponent constituents through a progressive sifting process to yield

empirical bases termed intrinsic mode functions (IMF). These IMFs are defined so as to ensure that they have

well-behaved HTs and conform to a narrowband condition. An IMF represents a simple oscillatory mode as a

counterpart to the simple harmonic function, but it is much more general: instead of having a constant amplitude

and frequency, as in a simple harmonic component, the IMF can have a variable amplitude and frequency as

function of time (Huang 2005). The combination of the EMD and the Hilbert spectral analysis is known as the

Hilbert-Huang transform (HHT). Fig. 4 (left) show the first two IMF’s obtained after the EMD is applied to the

signal previously presented in Fig. 1a. The instant frequencies in Fig. 4 (right) are obtained by applying the HT

to each IMF component. It is seen that for this case the resolution capabilities of the HHT are comparable to the

CWT and the signal discontinuity can be identified by the spike at 4 seconds in the instant frequency of the first

IMF. The algorithm used in this work to calculate the HHTs is the one provided in Rilling et al. (2003).

3. Damage detection using signal processing methods

A three storey shear building model (Fig. 5) is used to demonstrate how signal processing has been used

to detect structural damage. Natural frequencies of the undamaged building are 2.50 Hz, 6.90 Hz and 9.74 Hz;

damping ratio is specified as 5% for all three modes. The structural response is obtained by direct integration

using the average acceleration method. Damage is induced at three different times (1/4, 1/2 and 3/4 of the

excitation duration) as sudden relative reductions of 5%, 10% and 20% of the 1st floor stiffness (Fig 5). The

resulting changes in the structure natural frequencies are shown in Table 1. It is seen that the reduction in the

structure first mode vibration frequency caused by the induced damages is small (between 1% and 10%), the

first floor stiffness reduction values used can then be associated with damage related with a local

Montejo, 2010 Signal-Processing Based Damage Detection in Structures Subjected to Random Excitations
mechanism/member not affecting much the global behavior. Fig. 6 shows the acceleration time histories at each

level of the structure when a 3Hz constant amplitude sinusoidal acceleration is applied at the base, sampling

frequency is 200 Hz. Notice that the induced damages are not evident from a simple inspection of the response.

Several researchers (Sone et al. 1995, Al-Khalidy et al. 1997, Hou and Noori 1995, Todorovska and Trifunac

2009) have noticed that the occurrence of spikes in the DWT high frequency components of the dynamic

response coincides with the instant at which damage occurred. Fig. 7 shows the absolute values of the detail

functions (high frequency components, Eq. 4) obtained via DWT with the Bior6.8 wavelet basis for the

acceleration response at each level of the structure in Fig. 5. The spikes in the detail functions correspond to the

time instants where the damage was induced. Although damage was induced in the 1st floor only, it can be

detected from the acceleration histories in any of the floors. Furthermore, the spikes amplitude tends to be

related to the relative proximity to the damage location and the amount of damage induced, i.e. the larger spikes

are detected from the acceleration on the first floor and for the third damage event at 6 seconds.

The high frequency component (first IMF) obtained after an EMD may also depict localized

discontinuities in the structural response (Vincent et al. 1999). Fig. 8 shows the first IMF and its IF via HT for

each of the levels of the structure in Fig. 5. The two largest induced damages can be detected by the spikes

appearing in the IF of the first IMF at the first floor; however the detection of damage instants using the

accelerations in the 2nd and 3rd floors is not as clear as with the DWT approach (Fig. 7). Other authors (Yang

et al. 2004, Xu and Chen 2004) have proposed a different approach to use the first IMF resulting from the EMD

to detect the occurrence of damage, instead of applying the HT to detect the discontinuities, they post process

the first IMF with an intermittency check (Huang et al. 1999). In the intermittency check, data from the first

IMF having frequencies lower than a specified intermittency frequency is removed by a straightforward

counting process. The intermittency frequency should be smaller than the frequency of the discontinuity but

larger than the highest structural frequency. The results obtained after applying an intermittency check at 10 Hz

to the IMFs in Fig. 8 are displayed in Fig. 9, it is seen that damage instants can be identified from the spikes

appearing in the first floor for the two largest damages induced only – no spikes appear in the other floors.

Finally, Bisht (2005) observed that a direct high pass filtering of the measured response without any IMF

separation can be as effective as any other approach for detecting sudden changes in the system. Fig. 10 shows

Montejo, 2010 Signal-Processing Based Damage Detection in Structures Subjected to Random Excitations
the acceleration in each of the floors after a 10th order Butterworth high-pass filter to suppress frequencies

below 75Hz is applied, the results obtained are similar to the obtained using the DWT (Fig. 7).

4. Damage detection in structures subjected to random excitations

In order to evaluate the damage detection capabilities of the previously discussed methodologies when

the structure is subjected to random excitations, the structure is now subjected to the four acceleration histories

shown in Fig. 11 along with their Fourier spectra. The input signals were selected to study the influence of

different frequency contents and non-stationary characteristics in the results of the detection methodologies.

The first two signals (Figs. 11a and 11b) are artificially generated white noises that have been passed through

low pass filters at 25 Hz and 40 Hz; we will refer to these excitations as wnoise25 and wnoise40, respectively.

The other two signals are earthquake acceleration records. The first accelerogram (Fig 11c) was recorded at the

Treasure Island station during the M6.93 1980 Loma Prieta earthquake. The second accelerogram (Fig 11d) was

recorded at the Mcgee Creek station during the M5.82 1984 Round Valley earthquake. Notice from the Fourier

spectra in Fig. 11 that the Round Valley record has a quite spread frequency content compared to the very

narrow Fourier spectrum of the Loma Prieta record. Both records were obtained from the NGA database (PEER

2010). All four acceleration histories have a sampling frequency of 200 Hz and were normalized to a maximum

acceleration of 0.3g.

In the case of the white noise excitations (Figs. 12 and 13), it is seen that the most effective methodology

is the high pass filtering at 90 Hz (Figs. 12d and 13d) since it was able to detect all three damage instants (at

2.5, 5 and 7.5 seconds) from the accelerations at all three levels of the structure. The DWT approach was only

successful detecting damage episodes from the first floor acceleration when the building is excited by wnoise25

(Fig.12a). A large amount of spikes appear in the 2nd and 3rd floor (Fig. 12a) for wnoise25 and at all levels for

wnoise40 (Fig.13a) that are not related with the induced damages. The methodology based on the IF of the first

IMF was not able to detect any damage (Fig. 12b and 13b). The EMD methodology with an intermittency check

at 100 Hz was able to detect the first induced damage from the first floor accelerations when wnoise40 is applied

(Fig. 13c), but did not detected any damage from the accelerations at the other two levels of the structure. In the

case of wnoise25 no spikes appear after the intermittency check (Fig. 12c).

Montejo, 2010 Signal-Processing Based Damage Detection in Structures Subjected to Random Excitations
 Figs. 14 show the results obtained when the structure is excited with the Loma Prieta record. None of

the methodologies was capable of identifying all three damage instants at the three levels of the structure. The

best results were obtained using the high pass filter at 90 Hz (Fig 14d) and the DWT approach (Fig 14a), where

the three damage instants (at 5, 10 and 15 seconds) can be identified from the accelerations at the first level.

The EMD methodology with an intermittency check at 100 Hz was able to detect only the larger induced damage

(at 15 seconds) from the accelerations at the first level. No damage can be identified based on the IF of the first

IMF (Fig 14b).

The results obtained when the structure is excited by the Round Valley record are displayed in Figs. 15.

The total duration of this record is 6.87 seconds, so spikes related to damage are expected around 1.72, 3.44 and

5.16 seconds. None of the 4 methodologies evaluated was able to clearly detect any damage instants from any

of the floor accelerations. The results obtained using the EMD with intermittency check (Fig 15c) show clear,

isolated and congruent spikes in each of the floors; however the times of occurrence do not coincide with the

times the damage was induced. The large amount of spikes obtained using the other three methodologies (Figs

15a, b and d) makes difficult an unambiguous identification of the damage instants.

5. Damage detection using the Continuous Wavelet Transform

It has been shown in the previous sections that signal processing based damage detection methodologies

are successful when the system is excited by deterministic loads. However, problems arise when the system is

excited by random loads like an earthquake. Since the detection of damage is based on the premise that its

occurrence will be reflected in the high frequency component of the structural response, application of the

available methodologies is limited by the frequency content of the load excitation. If the frequency content of

the excitation spans over the frequency range where damage is reflected, the results obtained after a high

frequency analysis may include not only spikes due to structural damage but also spikes proper of the excitation.

To overcome this limitation an alternative approach using a projection of the structure high frequency response

via the CWT instead of the DWT is proposed. Using the CWT has the advantage that the central frequencies at

which it operates can be defined by the user (e.g. Montejo and Suarez 2007, Montejo and Kowalsky 2008),

while the frequency bands of the detail functions obtained via DWT are predetermined by the sampling period

Montejo, 2010 Signal-Processing Based Damage Detection in Structures Subjected to Random Excitations
of the signal. The procedure is exemplified using the first floor accelerations of the structure excited by the

Round Valley earthquake. First, a redundant decomposition at high frequency levels of the excitation load and

the first floor response is performed using the Complex Morlet Wavelet (Eq. 5); the results are displayed in the

WaveletMaps in Figs. 16a and b, respectively. The decomposition was performed at every 1 Hz between 70 Hz

and the Nyquist frequency (100 Hz). Notice that both maps are quite similar, which means that a large part of

the structural response high frequency components arrive directly from the load excitation and are not related

to damage. Nevertheless, a closer look to Fig. 16b can identify a dark vertical trend (around 3.4 seconds) that is

not present on the excitation map (Fig. 16a) and which occurrence time coincide with one of the induce damage

instants. In order to reduce the interference of the high frequency components proper of the base excitation on

the analysis of the structural response, the ratio in absolute values of the wavelet coefficients from the first floor

and base excitation is calculated. The results are displayed in Fig. 16c, it is seen that all three damage instants

become apparent. A more precise estimation of the damage instant can be obtained by finding the ratios maxima

(points where the derivative with respect to time is zero) at each frequency. To avoid the identification of

spurious peaks a threshold criterion is adopted, first the ratios (r) at each time instant (j) are normalized

according to the rule:

rj − µ
zj = (10)
σ

where µ and σ are the mean and the standard deviation of the ratio values. Any instant (j) where the normalized

value (z) is larger than 2 (i.e. the ratio deviates more than 2 standard deviations from the mean value) is treated

as a damage instant. The results obtained are shown in Fig. 17, it is seen that, with exception of the Round

Valley earthquake case, the three damage instants can be detected for any of the other three excitations and at

any of the structure levels. In the case of the Round Valley excitation, the 5% reduction in stiffness can only be

detected from the first floor response (Fig. 17d).

It should be noticed that the first 4 methodologies (based on DWT, HHT, EMD and high pass filtering)

were evaluated on an output-only fashion, that is, only the output signal (system response) was analyzed.

However, in the case of the proposed CWT methodology, the input excitation is also required. To be fair on the

evaluation of the other methods and explore if their performance is similarly improved when "ratios" to the

Montejo, 2010 Signal-Processing Based Damage Detection in Structures Subjected to Random Excitations
input processed by their respective methods are implemented, the same type of analysis was carried out for the

critical input excitation of the Round Valley earthquake for all four methodologies. The results are shown in

Fig.18. It is seen that in the case of the DWT, HHT and EMD based methodologies (Figs. 18a, b and c,

respectively) the performance is not improved as a large number of spurious peaks not related to damage

emerge. However, the high pass filtering methodology is significantly improved when the ratio of the high

passed output to input signals is taken (Fig 18d). It is seen that the results are comparable to the obtained with

the CWT methodology (Fig. 17d) since all three levels of damage can be detected from the first floor response

and the two largest ones can also be detected from the analysis of the second and third floor response.

Nevertheless, some spurious spikes of significant amplitude appear around the damage instants that may

complicate an unambiguous identification.

6. Conclusions

Four available methodologies for model-free signal-processing-based damage detection were examined

using a three storey shear building subjected to different types of load. Damage is introduced as sudden relative

reductions of 5%, 10% and 20% of the 1st floor stiffness. The methodologies examined are based on the

detection of spikes on the high frequency response of the structure obtained via: (1) Discrete Wavelet Transform,

(2) Hilbert Huang Transform, (3) Empirical Mode Decomposition with intermittency frequency and (4)

Butterworth high pass filtering. It was found that successful unambiguous damage detection using such methods

largely depends on the characteristics of the excitation load. Good results are obtained when the system is

excited by deterministic loads. However, damage detection capabilities diminish as the system is excited by

random loads and the high frequency content of the excitation increase. The best results were obtained using

the high pass filtering technique.

A high frequency/highly redundant time-frequency analysis of the structural response and input

excitation showed that a large part of the system response high frequency components may arrive directly from

the load excitation and are not necessarily related to damage. An alternative approach based on a redundant

decomposition of the structural response at high frequencies via the Continuous Wavelet Transform is then

proposed. To reduce the effect of the input load high frequency content, damage instants are detected by the

Montejo, 2010 Signal-Processing Based Damage Detection in Structures Subjected to Random Excitations
spikes resulting from the ratio between wavelet coefficients of the structural response and the excitation. The

results obtained were quite promising as the three levels of induced damage were identified even when the input

load has significant high frequency content. A similar output/input ratio approach was then implemented for the

other four methodologies. Significant improvement was only noticed for the high pass filtering technique,

though some spurious spikes of considerable amplitude emerged around the damage instant making difficult an

explicit identification.

Identification of damage location and relative damage magnitude can be accomplished when the

structure is excited by a stationary load with limited frequency content, e.g. the 3Hz constant amplitude

sinusoidal signal used in this paper. It is seen, for example, that the spikes amplitude from the DWT details (Fig.

7) are larger in the first floor (where damage was induced) and increase as the level of damage increases. When

the system is excited by a nonlinear-nonstationary load with wide frequency content, estimation of the relative

level of damage is not evident from the results obtained. In Figs. 12a and 12b for example, the amplitude of the

spikes do not correlate with the level of damage induced. This is because the spike amplitude not only depends

on the level of damage but also on the external load and damage time. Nevertheless, identification of damage

location is still viable as the larger spikes will be obtained on the measure points closer to the damage locations.

This can be corroborated from Figs.18d and 19, where the amplitude of the spikes emerging at the first floor is

much larger than the spikes amplitude on the upper floors. Fig. 19 is a 3D version of Fig. 17d to show also the

ratio value.

Finally, it should be noticed that a limitation of the proposed methodology is that input measurements

are required and this is not always available in some monitoring applications (e.g., during operational

monitoring under traffic, wind or other ambient loads when damage occurs under service loading). Further

validation is required using real data that usually have signal noise in the high frequency range where the applied

signal processing techniques mainly work.

7. References

Montejo, 2010 Signal-Processing Based Damage Detection in Structures Subjected to Random Excitations
Al-Khalidy, A., Noori, M., Hou, Z., Yamamoto, S., Masuda, A. and Sone, A. (1997), “Health monitoring

systems of linear structures using wavelet analysis”, Proceeding of the 1st International Workshop on Structural

Health Monitoring, Stanford, CA, 164-174.

Bisht, S.S. (2005), Methods for structural health monitoring and damage detection of civil and mechanical

systems, MS Thesis, Department of Engineering Science and Mechanics, Virginia Polytechnic Institute and

State University.

Carmona, R.A., Hwang, W.L. and Torresani, B. (1997), “Characterization of signals by the ridges of their

wavelet transforms”, IEEE Transactions on Signal Processing 54(10), 2586-2590.

Clinton, J.F., Bradford, S.C., Heaton, T.H. and Favela, J. (2006), “The observed wander of the natural

frequencies in a structure”, Bulletin of the Seismological Society of America 96(1), 237-257.

Doebling, S.W., Farrar, C.R., Prime, M.B. and Shevitz, D.W. (1996), Damage identification and health

monitoring of structural and mechanical systems from changes in their vibration characteristics: a literature

review, Los Alamos National Laboratory Report LA-13070-MS.

Doebling, S.W., Farrar, C.R. and Prime, M.B. (1998), A summary review of vibration-based damage

identification methods, The Shock and Vibration Digest 30(2), 91-105.

EERI Earthquake Engineering Research Institute (2009), “Learning from earthquakes: The M6.3 Abruzzo,

Italy, Earthquake of April 6, 2009”, EERI Special Earthquake Report.

Farrar, C.R. and Cone, K.M. (1995), “Vibration testing of the 1-40 bridge before and after the introduction of

damage”, Proceedings 73th international Modal Analysis Conference, Nashville, TN, 203-209.

Farrar, C.R., Doebling, S.W. and Nix DA. (2001), “Vibration based structural damage identification”,

Philosophical Transactions of the Royal Society A 359,131-149.

Hou, Z. and Noori, M. (1999), “Application of wavelet analysis for structural health monitoring”, Proceedings

2nd International Workshop on Structural Health Monitoring, Stanford, CA, 946-955.

Montejo, 2010 Signal-Processing Based Damage Detection in Structures Subjected to Random Excitations
Huang, N.E., Shen, Z., Long, S.R., Wu, M.C., Shih, H.H., Zheng, Q., Yen, N.C., Tung, C.C. and Liu, H.H.

(1998), “The empirical mode decomposition and the Hilbert spectrum for nonlinear and non-stationary time

series analysis”, Philosophical Transactions of the Royal Society A 454, 903-995.

Huang, N.E., Shen, Z. and Long, S.R. (1999), “A new view of nonlinear water waves: the Hilbert spectrum”,

Annu. Rev. Fluid Mech. 31, 417–57.

Huang, N.E. (2005), “Introduction to the Hilbert Huang transform and its related mathematical problems”,

Hilbert Huang Transform and Its Applications, World Scientific, 1-26.

Huth, O., Feltrin, G., Maeck, J., Kilic, N. and Motavalli, M. (2005), “Damage identification using modal data:

experiences on a prestressed concrete bridge”, ASCE Journal of Structural Engineering 131(12), 1898-1910.

Kijewski, T. and Kareem, A. (2003), “Wavelets transforms for system identification in civil engineering”,

Computer Aided Civil and Infrastructure Engineering 18, 339-355.

Kijewski-Correa, T. and Kareem, A. (2006), “Efficacy of Hilbert and Wavelet transforms for time-frequency

analysis”, Journal of Engineering Mechanics 132(10), 1037-1049.

Mallat, S.G. (1989), “Theory for multiresolution signal decomposition: the wavelet representation”, IEEE

Transactions on Pattern Analysis and Machine Intelligence 11(7), 674–93.

Montejo, L.A. and Kowalsky, M.J. (2008), “Estimation of frequency-dependent strong motion duration via

wavelets and its influence on nonlinear seismic response”, Computer Aided Civil and Infrastructure

Engineering 23, 253-264.

Montejo, L.A. and Suarez, L.E. (2007), “Applications of the Wavelet Transform in structural engineering”,

Mecánica Computacional XXVI, 2742-2753 (in Spanish).

Ovanesova, A.V. and Suarez, L.E. (2004), “Applications of wavelet transforms to damage detection of frame

structures”, Engineering Structures 26, 39-49.

PEER Pacific Earthquake Engineering Research Center (2010), Next Generation Attenuation of Ground

Motions NGA database <http://peer.berkeley.edu/nga/index.html>

Montejo, 2010 Signal-Processing Based Damage Detection in Structures Subjected to Random Excitations
Rilling, G., Flandrin, P. and Gonçalvès, P. (2003), “On empirical mode decomposition and its algorithms”,

IEEE-EURASIP Workshop on Nonlinear Signal and Image Processing NSIP-03.

Rodgers, J.E. and Çelebi, M. (2006), “Seismic response and damage detection analyses of an instrumented

steel moment-framed building”, ASCE Journal of Structural Engineering 132(10), 1543-1552.

Sohn, H., Farrar, C.R., Hemez, F.M., Shunk, D.D., Stinemates, D.W. and Nadler, B.R. (2003), A review of

structural health monitoring literature: 1996-2001, Los Alamos National Laboratory Report LA-13976-MS.

Sone, A., Yamamoto, S., Nakaoka, A., Masuda, A. (1995), “Health monitoring system of structures based on

orthonormal wavelet transform” Proceedings of the ASME Pressure Vessels and Piping Conference, 161–167.

Todorovska, M.I. and Trifunac, M.D. (2007), “Earthquake damage detection in the imperial county services

building I: the data and time frequency analysis”, Soil Dynamics and Earthquake Engineering 27, 564-576.

Todorovska, M.I. and Trifunac, M.D. (2009), “Earthquake damage detection in the Imperial County Services

building II: Analysis of novelties via wavelets”, Structural Control and Health Monitoring.

Vincent, H.T., Hu. S.J. and Hou, Z. (1999), “Damage detection using empirical mode decomposition and a

comparison with wavelet analysis”, Proceedings of the 2nd International Workshop in Structural Health

Monitoring, Stanford, CA, 891-900.

Xu, Y.L. and Chen, J. (2004) “Structural damage detection using empirical mode decomposition: experimental

investigation”, Journal of Engineering Mechanics 130(11), 1279-1288.

Yan, B. and Miyamoto, A. (2006), “A comparative study of modal parameter identification based on Wavelet

and Hilbert–Huang transforms”, Computer-Aided Civil and Infrastructure Engineering 21, 9-23.

Yang, J.N., Lei, Y., Lin, S. and Huang, N. (2004), “Hilbert-Huang based approach for structural damage

detection”, Journal of Engineering Mechanics 130(1), 85-95.

Montejo, 2010 Signal-Processing Based Damage Detection in Structures Subjected to Random Excitations
 (a) 2
f(t)

0
-2
0 2 4 6 8
t [s]
(b) 4
|F|

2
0
0 2 4 6 8 10
f [Hz]

Fig. 1 Signal in the time domain (a) and its Fourier spectrum (b)

Montejo, 2010 Signal-Processing Based Damage Detection in Structures Subjected to Random Excitations
 (a) (b) (c)

Fig. 2 3D plot of the wavelet coefficients (a), Wavelet Map (b) and instantaneous frequencies from wavelet ridges
(c)

2
A [m/s 2]

-2
0.01
D [m/s2]

-0.01
0 2 4 6 8
t [s]

Fig. 3 Approximation (A) and detail (D) functions at the first level of decomposition via DWT

2 10
signal

0
8
-2
discontinuity
6
IMF 1

IF law
f [Hz]

0
HHT
4
2
IMF 2

0 2
-2 0
0 2 4 6 8 0 2 4 6 8
t [s] t [s]

Fig. 4 Intrinsic mode function IMFs (left) and instantaneous frequency via HT (right)

Montejo, 2010 Signal-Processing Based Damage Detection in Structures Subjected to Random Excitations
 norm. stiffness
1
0.8
0.6
0.4
0.2

0 0.25 0.5 0.75 1

norm. time

Fig. 5 Shear building model with base excitation and induced damage in the 1st floor

0.4
0
-0.4 3rd floor
Accel. [g]

0.4
0
-0.4 2nd floor
0.4
0
-0.4 1st floor
0 2 4 6 8
t [s]

Fig. 6 Acceleration response when a 3Hz constant amplitude sinusoidal acceleration is applied at the base

0.2 3r floor
0.1
[m/s2]*102

2nd floor
0.2
0.1
15
1st floor
10
5

0 2 4 6 8
t [s]

Fig. 7 First level decomposition detail functions (in absolute values) via DWT using Bior6.8 basis

5 3
0 1.5
First IMF [m/s 2]

-5 3rd floor 3rd floor

5 3
IF [Hz]

0 1.5
-5 2nd floor 2nd floor
5 30 1st floor
0 20
-5 1st floor
0 2 4 6 8 0 2 4 6 8
t [s] t [s]

Fig. 8 First IMF (left) and its IF via HT (right) for each of the levels of the structure in Fig. 5

Montejo, 2010 Signal-Processing Based Damage Detection in Structures Subjected to Random Excitations
 0.2
0
3rd floor
-0.2
0.2

[m/s2]
0
-0.2 2nd floor
0.2
0
-0.2 1st floor
0 2 4 6 8
t [s]

Fig. 9 First IMFs after an intermittency check at 10 Hz

0.1 3rd floor

0.05
[m/s2]*102

0.1 2nd floor

0.05

4 1st floor
5
0
0 2 4 6 8
t [s]

Fig. 10 Floor accelerations (in absolute values) after a 75Hz high pass filter is applied

(a) (b)

0.2 0.2
Accel. [g]

Accel. [g]

0 0

-0.2 -0.2
0 2 4 6 8 0 2 4 6 8
0 t [s] 0 t [s]
10 10
Norm. Amp.

Norm. Amp.

-1 -1
10 10
-2 -2
10 10
-3 -3
10 10
0 20 40 60 80 100 0 20 40 60 80 100
f [Hz] f [Hz]
(c) (d)

Montejo, 2010 Signal-Processing Based Damage Detection in Structures Subjected to Random Excitations
 0.2
0.2

Accel. [g]
Accel. [g]
0
0
-0.2
-0.2
0 5 10 15 20 0 1 2 3 4 5 6
0 t [s] 0 t [s]
10 10
Norm. Amp.

Norm. Amp.
-1 -1
10 10
-2 -2
10 10
-3 -3
10 10
0 20 40 60 80 100 0 20 40 60 80 100
f [Hz] f [Hz]

Fig. 11 Four random excitations and its frequency content via Fourier spectrum: (a) white noise HP 25 Hz; (b)
White noise HP 40Hz; (c) 1989 Loma Prieta earthquake; (d) 1984 Round Valley earthquake

(a)
2 3rd floor
1
[m/s2]*102

2 2nd floor
1

1st floor
4
2

0 2 4 6 8 10
t [s]
(b)

Montejo, 2010 Signal-Processing Based Damage Detection in Structures Subjected to Random Excitations
 80 3rd floor
40

80 2nd floor

IF [Hz]
40

80 1st floor
40

0 2 4 6 8 10
t [s]
(c) 0.5 3rd floor
0
-0.5
0.5 2nd floor
[m/s2]

0
-0.5
0.5 1st floor
0
-0.5
0 2 4 6 8 10
(d) t [s]
0.04
3rd floor
0.02
[m/s2]*103

0.04 2nd floor

0.02

10
1st floor
5

0 2 4 6 8 10
t [s]
Fig. 12 Damage detection when the structure is exited by white noise HP 25 Hz: (a) First level detail functions (in
absolute values) via DWT, (b) IF of the first IMF via HT, (c) IMF 1 with intermittency check at 100 Hz and (d)
floor accelerations (in absolute values) after a 90Hz high pass filter is applied
(a)
40 3rd floor
20
[m/s2]*102

40 2nd floor
20

40 1st floor
20

0 2 4 6 8 10
t [s]
(b)

Montejo, 2010 Signal-Processing Based Damage Detection in Structures Subjected to Random Excitations
 80 3rd floor
40

80 2nd floor

IF [Hz]
40

80 1st floor
40

0 2 4 6 8 10
t [s]
(c)
0.02 3rd floor
0
-0.02

0.02 2nd floor

[m/s2]

0
-0.02

0.02 1st floor

0
-0.02
0 2 4 6 8 10
t [s]
(d)
0.1
3rd floor
0.05

0.1
[m/s2]*103

2nd floor
0.05

20 1st floor
10

0 2 4 6 8 10
t [s]
Fig. 13 Damage detection when the structure is exited by white noise HP 40 Hz: (a) First level detail functions (in
absolute values) via DWT, (b) IF of the first IMF via HT, (c) IMF 1 with intermittency check at 100 Hz and (d)
floor accelerations (in absolute values) after a 90Hz high pass filter is applied
(a)
0.4 3rd floor
0.2
[m/s2]*102

0.4 2nd floor

0.2

10 1st floor
5
0
0 5 10 15 20
(b) t [s]

Montejo, 2010 Signal-Processing Based Damage Detection in Structures Subjected to Random Excitations
 80 3rd floor
40

80 2nd floor

IF [Hz]
40

80 1st floor
40

0 5 10 15 20
t [s]
(c)
1
0
-1 3rd floor

1
[m/s2]

0
-1 2nd floor
0.1
0
-0.1 1st floor
0 5 10 15 20
t [s]
(d) 0.2 3rd floor
0.1
[m/s2]*103

0.2 2nd floor

0.1

30 1st floor
15

0 5 10 15 20
t [s]
Fig. 14 Damage detection when the structure is exited by an acceleration record from the 1989 Loma Prieta
earthquake: (a) First level detail functions (in absolute values) via DWT, (b) IF of the first IMF via HT, (c) IMF 1
with intermittency check at 100 Hz and (d) floor accelerations (in absolute values) after a 90Hz high pass filter is
applied
(a)
30
20 3rd floor
10
30
[m/s2]*102

20 2nd floor
10
30
1st floor
20
10

0 1 2 3 4 5 6 7
t [s]
(b)

Montejo, 2010 Signal-Processing Based Damage Detection in Structures Subjected to Random Excitations
 80 3rd floor
40

80 2nd floor

IF [Hz]
40

80 1st floor
40

0 1 2 3 4 5 6 7
t [s]
(c)
0.1
3rd floor
0
-0.1
0.1
2nd floor
[m/s2]

0
-0.1
0.5
1st floor
0
-0.5
0 2 4 6
t [s]
(d)
16
8 3rd floor
[m/s2]*103

16
2nd floor
10

16
1st floor
8

0 1 2 3 4 5 6 7
t [s]
Fig. 15 Damage detection when the structure is exited by an acceleration record from the 1984 Round Valley
earthquake: (a) First level detail functions (in absolute values) via DWT, (b) IF of the first IMF via HT, (c) IMF 1
with intermittency check at 100 Hz and (d) floor accelerations (in absolute values) after a 90Hz high pass filter is
applied
(a) (b) (c)

Fig. 16 High frequencies WaveletMaps of the load excitation – Round Valley record (a) and the first floor
dynamic response (b), and ratio of the Wavelet coefficients (c)

Montejo, 2010 Signal-Processing Based Damage Detection in Structures Subjected to Random Excitations
 (a)
90
80
3rd floor
70
90
f [Hz]

80
2nd floor
70
90
80
1st floor
70
0 2 4 6 8
t [s]
(b)
90
80
3rd floor
70
90
f [Hz]

80
2nd floor
70
90
80
1st floor
70
0 2 4 6 8
(c) t [s]

Montejo, 2010 Signal-Processing Based Damage Detection in Structures Subjected to Random Excitations
 90
80
3rd floor
70
90

f [Hz]
80
2nd floor
70
90
80
1st floor
70
0 5 10 15 20
t [s]
(d)
90
3rd floor
80
70
90
f [Hz]

2nd floor
80
70
90
1st floor
80
70
0 1 4 2 5 3 6
t [s]
Fig. 17 Damage detection via CWT ratios for different excitations: (a) wnoise25; (b) wnoise40; (c) Loma Prieta
record; (d) Round valley record
4
(a) 3rd floor
2

4
DWT ratio

2nd floor
2

4
1st floor
2

0 1 2 3 4 5 6 7
t [s]
(b)
3rd floor
4
2

2nd floor
IF ratio

4
2

1st floor
4
2

0 1 2 3 4 5 6 7
t [s]
(c)

Montejo, 2010 Signal-Processing Based Damage Detection in Structures Subjected to Random Excitations
 10
3rd floor
0
-10
10

IMF ratio
2nd floor
0
-10
10
1st floor
0
-10
0 1 2 3 4 5 6 7
t [s]
(d) 3
2 3rd floor
1
3
HP ratio

2 2nd floor
1
9
6 1st floor
3
0 1 2
4 53 6 7
t [s]
Fig. 18 Damage detection when the structure is exited by an acceleration record from the 1984 Round Valley
earthquake using floor response to input motion ratios: (a) DWT ratio, (b) IF ratio, (c) IMF ratio and (d) high
passed signals ratio

Montejo, 2010 Signal-Processing Based Damage Detection in Structures Subjected to Random Excitations
 Fig, 19 3D stem plot showing ratio amplitudes for the CWT damage detection methodology (Round Valley
excitation)

Table 1. Natural frequencies change with induced damages

No damage Damage 1 (5%) Damage 2 (10%) Damage 3 (20%)

2.50 Hz 2.47 Hz 2.40 Hz 2.26 Hz
6.90 Hz 6.84 Hz 6.71 Hz 6.47 Hz
9.74 Hz 9.71 Hz 9.64 Hz 9.54 Hz

Montejo, 2010 Signal-Processing Based Damage Detection in Structures Subjected to Random Excitations

View publication stats