Series of Full-Scale Field Vibration Tests and Bayesian Modal

Identification of a Pedestrian Bridge

Y. C. Ni1; F. L. Zhang, A.M.ASCE2; and H. F. Lam, M.ASCE3

Abstract: Many spectacular pedestrian bridges were designed and constructed recently. Owing to their special shapes, it is expected that vari-
Downloaded from ascelibrary.org by University of Central Florida on 11/23/20. Copyright ASCE. For personal use only; all rights reserved.

ous types and a wide range of vibration frequency components will be induced by pedestrians. To avoid accidents and reduce risk, the vibration
characteristics of pedestrian bridges during their service life must be carefully assessed. The most direct and reliable way to study the vibration
characteristics of a structural system is through field vibration tests. In this paper, a series of full-scale field vibration tests (including ambient,
forced, and free vibration tests) were carried out on a pedestrian bridge at City University of Hong Kong (CityU). The difficulties encountered
in the field tests are reported. The recently developed Bayesian methods were utilized to determine the modal parameters of the bridge based
on measurements from all three kinds of tests. In addition to the most probable values (MPVs) of modal parameters, the associated posterior
uncertainties were also analytically calculated. Four modes were identified, including three vertical bending modes and one torsional mode.
The accuracy of the identified modal parameters was assessed through the posterior uncertainty. Because the amplitudes of the vibration in the
three kinds of tests were different, the modal parameters determined from these kinds of tests were compared and discussed. Suggestions
related to the proper use and potential vibration problems during the lifecycle of pedestrian bridges were provided based on the analysis results.
DOI: 10.1061/(ASCE)BE.1943-5592.0000857. © 2016 American Society of Civil Engineers.
Author keywords: Pedestrian bridge; Field vibration tests; Modal identification; Bayesian; Posterior uncertainty.

Introduction Brownjohn 2003; Choi et al. 2010; Au and Zhang 2012) are the
most convenient. In this kind of test, dynamic data are acquired
In the past, the construction of pedestrian bridges could not attract when a structure is under unknown but statistically random loading.
as much attention as that of long-span bridges. With some recent Its success hinges on the achievable signal-to-noise (s/n) ratio and
incidents concerning pedestrian bridges, researchers and engineers whether interference from unknown colored excitations in the fre-
have started to pay more attention to the design and construction of quency band of interest is significant (Au 2011; Au and Zhang
this kind of bridge, especially their dynamic behaviors. One exam- 2012). Another popular method for research is the forced vibration
ple of an incident is the large swaying motion of the Millennium test, where some known excitation forces is artificially applied to
Bridge, which is a famous footbridge situated in London, U.K. the structure (De Sortisa et al. 2005; Brownjohn and Pavic 2007;
When the bridge was first opened, an unexpected and uncomfort- Racic et al. 2010). One advantage of this kind of method is that the
able swaying motion was felt by the pedestrians. It took approxi- vibration level can be increased, and this significantly improves the
mately 2 years for the bridge owners to solve this vibration problem. s/n ratio of the collected data. However, this kind of test is usually
Another example is the collapse of a pedestrian bridge in expensive, and special equipment is required. Furthermore, if the
Chongqing, China, which linked the old and new development excitation is not properly controlled, unexpected damage to the
zones of Qijiang City. It collapsed suddenly on January 4, 1998, structure may result. The last one is the free vibration test, which
when 22 armed police soldiers were training on this structure, caus- fits the structural dynamics theory best, but it is not easy to imple-
ing 24 deaths; 16 people were seriously injured, and more than 10 ment in civil engineering structures (Magalhães et al. 2010; Au
people went missing. et al. 2005; Ülker-Kaustell and Karoumi 2011). Based on the col-
A possible way to assess the vibration characteristics of a struc- lected field vibration data, modal identification can be carried out to
tural system is through different kinds of vibration field tests, such obtain the modal parameters, which represent the current dynamic
as ambient, forced, and free vibration tests. Among them, the full- characteristics of the target full-scale structure (Sohn et al. 2003).
scale ambient vibration tests (e.g., Peeters and De Roeck 2001a; The parameters that characterize the responses of a structure sub-
jected to dynamic loads involve natural frequencies, damping ratios,
1
Research Institute of Structural Engineering and Disaster Reduction, and mode shapes. The values of these parameters and the trend of
College of Civil Engineering, Tongji Univ., Shanghai 200092, China. how modal properties change with environmental conditions yield
2
Research Institute of Structural Engineering and Disaster Reduction, important physical insights about structural dynamics (Farrar et al.
College of Civil Engineering, Tongji Univ., Shanghai 200092, China (cor- 1997; Alampalli 2000; Catbas, et al. 2008; Spiridonakos and Chatzi
responding author). E-mail: [email protected]; fengliangzhang@ 2014a, b). Specifically, knowing the natural frequencies of a struc-
tongji.edu.cn ture is useful to avoid the phenomenon of resonance, which amplifies
3
Dept. of Architecture and Civil Engineering, City Univ. of Hong the vibration response of a structure significantly, causing serious
Kong, Hong Kong.
damage and risk (Lutes and Sarkani 1997). It can also help with
Note. This manuscript was submitted on December 2, 2014; approved
on September 15, 2015; published online on February 1, 2016. Discussion assessing potential vibration problems of the structure and providing
period open until July 1, 2016; separate discussions must be submitted for design solutions to alleviate them. Damping is another important
individual papers. This paper is part of the Journal of Bridge quantity affecting vibration level and energy dissipation of structures
Engineering, © ASCE, ISSN 1084-0702. (Hart 1996; Tamura and Suganuma 1996; Satake et al. 2003; Liang

© ASCE C4016002-1 J. Bridge Eng.

J. Bridge Eng., 2016, 21(8): C4016002

 and Feeny 2006; Au et al. 2012). The traditional manner in structural most probable value (MPV) of modal parameters, the associated
design is to assume the system is classically damped and set the posterior uncertainty can also be calculated analytically, which can
damping ratio to a nominal value, such as 1%. However, the damping be characterized by a posterior covariance matrix. The results iden-
ratio is subjected to modeling error, and its identified values are dif- tified from three kinds of field tests were obtained separately to
ferent for different structures, for different modes, and for different assess the structural performance. Detailed comparison was imple-
tests in reality. Field vibration tests and modal identification provide mented among these modal parameters obtained under different
an effective and reliable way to obtain damping ratios. Furthermore, field tests.
the identified damping ratios of the structure provide baseline data
that may be useful for future design and construction of similar
types of structures. Mode shape is another dynamic property that Description of the City University of Hong Kong
reflects the distribution of stiffness and mass as well as boundary Pedestrian Bridge
conditions (Au and Zhang 2011, 2012). The change in modal prop-
erties across major loading events has been explored for model The target structure is a concrete pedestrian bridge situated in the
Downloaded from ascelibrary.org by University of Central Florida on 11/23/20. Copyright ASCE. For personal use only; all rights reserved.

updating (Lam et al. 2014a; Yang et al. 2015), damage detection campus of the City University of Hong Kong (CityU), as shown in
(Lam et al. 2014b), or more generally, structure health monitoring Fig. 1, which serves as the main road between the mass transit rail-
(Chang et al. 2003; Sohn et al. 2003; Van der Auweraer and Peeters way (MTR) station and the main campus. Fig. 2(a) shows the sche-
2003; Brownjohn et al. 2005; Ko and Ni 2005; Koh and Dyke 2007). matic diagram of the bridge. The dimension information of this pe-
In recent years, much research has been performed on bridge destrian bridge can be found in this figure. The bridge has three
modal identification with different types of vibration tests to esti- spans, divided according to the six supporting columns. The first
mate the dynamic performance of structures. An ambient vibration and second spans are shown in Fig. 3(a). The third span cannot be
test was performed on a stress-ribbon footbridge to develop and directly observed, as it is blocked by a laboratory and a greenhouse.
The deck levels for different spans of the pedestrian bridge are not
implement a structural health monitoring system, and the modal
the same. From the first span to the second span, the level increases
properties were estimated by the stochastic subspace identification
gradually from 10.20 to 12.45 m, whereas the third span is flat with
method (Hu et al. 2013). An all-fiber reinforced polymer composite
a height of 12.45 m.
pedestrian bridge was studied, and dynamic tests were conducted.
This bridge has two special features that make it worth studying.
The identification techniques used to extract structural dynamic
The first is that the bridge is not an independent structure. On both
characteristics were the peak-picking method and stochastic sub-
sides of the third span near the third and fourth columns, the bridge
space identification method (Bai and Keller 2008). Ambient vibra-
is connected to two 6-story buildings [i.e., Mong Man Wai (MMW)
tion tests were performed on a well-known iron arch bridge,
and Fung Yun Wa (FYW) buildings], as shown in Fig. 3(b). These
Paderno Bridge, which was erected in 1889 and is protected by the
connections will very likely influence the mode shape and the vibra-
Italian Ministry of Cultural Heritage, to investigate the invariance
tion of the bridge. Another interesting point is that the construction
of the dynamic characteristics. The characteristics of the bridge
of a lift connecting the ground to the bridge deck levels (near the
were identified by frequency domain decomposition and stochastic
second span of the bridge) was started after the completion of vibra-
subspace identification approaches (Gentile and Saisi 2013). A free
tion tests. The completed lift system is shown in Fig. 4. It is clear
vibration test was performed on a large cable-stayed bridge, Vasco from the figure that a steel beam was constructed on the right-hand
da Gama Bridge, to evaluate the dynamic behavior of the bridge, side to strengthen the original bridge. This may change the modal
especially the damping factors (Cunha et al. 2001). Some highway properties of the bridge. If another set of measured vibration data is
bridges were investigated through ambient vibration tests to iden- available for the bridge after the construction of the lift system, one
tify the dynamic characteristics. Enhanced frequency domain can study the effects of the newly installed lift system on the pedes-
decomposition and stochastic subspace identification were used to trian bridge. This paper focuses on the vibration tests and the corre-
analyze the data (Altunis ik et al. 2011). An ambient vibration test sponding modal identification of the bridge before the construction
was performed on a Guadiana River cable-stayed bridge linking of the lift system.
Portugal and Spain. Modal parameters were identified by peak pick-
ing, frequency domain decomposition, covariance-driven stochastic
subspace identification, and data-driven stochastic subspace identi-
fication for the comparison of these modal identification methods
(Magalhães et al. 2007).
In this study, for the purpose of investigating the performance of
the target pedestrian bridge, a series of full-scale vibration tests,
including the ambient, forced, and free vibration tests, were per-
formed. In the forced vibration test, two different kinds of excitation
cases were performed involving traditional frequency resonant ex-
citation (frequency sweep with sine wave excitation) at a group of
specified frequencies and pseudorandom excitation covering all the
modes of interest in one setup. The difficulties encountered in the
field tests are discussed. Recently developed Bayesian methods,
including the fast Bayesian method for ambient vibration data, fast
Bayesian method for forced vibration data, and fast Bayesian
method for free vibration data, were utilized to analyze the meas-
ured acceleration. Bayesian approaches allow the information con-
tained in the measured data to be processed rigorously, consistent
with modeling assumptions and probability logic, to yield inference Fig. 1. Overall view of pedestrian bridge (image by F. L. Zhang)
information on the modal parameters of interest. In addition to the

© ASCE C4016002-2 J. Bridge Eng.

J. Bridge Eng., 2016, 21(8): C4016002

 CityU
CityU
13m

1m
First span 10 9

4.13m
Second span 8 7
(Instrumented

20m
segment)

55m
6
Downloaded from ascelibrary.org by University of Central Florida on 11/23/20. Copyright ASCE. For personal use only; all rights reserved.

20m
MMW FYW Guralp Sensor
4 3
Kistler Sensor
Third span
2 1 Shaker

Column 0.5m
13m MTR
(a) MTR (b)

Fig. 2. Plan of pedestrian bridge: (a) schematic diagram; (b) setup plan of tested segment

Field Tests of sensors in the field to serve as the reference of the subsequent
measurement. In the ambient vibration test, to reduce the noise
Three kinds of field tests were conducted: the ambient vibration level, the whole measurement was performed at midnight with a
test, forced vibration test, and free vibration test. All tests focused relatively quiet environment to guarantee the data quality.
on the second span of the bridge with dimensions of approximately The equipment used in this measurement included 10 acceler-
20 m long and 13 m wide, as indicated in Fig. 2(b). The vibration ometers, National Instruments (NI) data acquisition system with 12
tests are reported in detail one by one in the following subsections. channels, and a laptop. Two kinds of accelerometers were used,
Guralp (Guralp Systems Limited, Reading, U.K.) and Kistler
Ambient Vibration Test K8330 (Kistler Group, Winterthur, Switzerland) accelerometers, as
shown in Figs. 5(a and b), respectively. The Guralp sensor is usually
The ambient vibration test measured 10 degrees of freedom (DOFs) used in p earthquake engineering, with a lower noise level of
ffiffiffiffiffiffi
located on both sides of the span symmetrically. Fig. 2(b) shows the 0:1 m g= Hz, whereas the noise level of Kistler accelerometers is
detailed setup plan. In the determination of sensor locations, the approximately 10 times higher. Thus, the overall channel noise of
pffiffiffiffiffiffi
characteristics of the modes of interest should be taken into account. accelerometers is approximately 1 m g= Hz. Considering the fact
In this case, sensors on both ends and at the middle of the span were that Guralp sensors provide data with a higher s/n ratio, they were
needed to obtain information for identifying the first bending mode used at locations where the structure response is expected to be rela-
of the bridge, whereas the sensors located at the quarter span and tively small, or the noise level is believed to be relatively high.
both ends were indispensable for identifying the second bending Finally, Guralp triaxial accelerometers [see Fig. 5(a), only vertical
mode. The sensors were symmetrically distributed on both sides of channel used], indicated by circles in Fig. 2(b), were used at
the bridge along the longitudinal direction for capturing the tor- Locations 1, 2, 3, 7, 9, and 10. The remaining locations were
sional mode. Finally, 10 locations were measured, with five on each equipped with Kistler K8330 uniaxial accelerometers [see Fig. 5
side. The setting out of sensor locations was performed with tapes (b)], indicated by squares. In this test, 20 min of digital data was
and a laser level meter, which was used to check if the surrounding acquired with a sampling rate of 2,048 Hz (controlled by hardware).
concrete walls were straight enough to work as references. In gen- To reduce the file sizes, the measured data were decimated by 16 to
eral, it is a common practice to use surrounding walls or other con- a sampling rate of 128 Hz for analysis.
struction partitions as references in setting out sensor locations.
After checking, all the sensors were kept 0.48 m away from the con-
Forced Vibration Test
crete walls on the two sides. The locations at both ends in the longi-
tudinal direction were first determined, and then sensors on one side Known Pseudorandom Excitation
of the bridge could be determined by a measuring tape connecting In the forced vibration test with known pseudorandom excitation, a
these two end points with a spacing of 4.13 m. The same procedure setup was designed to focus on the modes identified in the ambient
was repeated for setting out the other five sensor locations on the vibration test. Because the input force could be calculated by the
other side of the bridge. Before the measurement, all sensors were measured shaker mass acceleration, the term pseudorandom was
put together as Setup 0 to measure 10-min data that could provide used here. The sensor locations were the same as those in the ambi-
information on environment noise level and the working performance ent vibration test. A long-stroke electromagnetic shaker (APS 113,

© ASCE C4016002-3 J. Bridge Eng.

J. Bridge Eng., 2016, 21(8): C4016002

Downloaded from ascelibrary.org by University of Central Florida on 11/23/20. Copyright ASCE. For personal use only; all rights reserved.

Fig. 5. Sensors used (images by Y. C. Ni): (a) triaxial Guralp sensor

Fig. 3. Details of pedestrian bridge (images by Y. C. Ni): (a) support- (only vertical used); (b) Kistler sensor
ing columns; (b) part connecting with buildings

targeted modes. In this case, the shaker was put in the quarter span
of the bridge [see the triangle in Fig. 2(b)], where all the modes of
interest have obvious deflection, and the shaker location was kept
unchanged during the entire test. All of the targeted modes were
excited with only a pseudorandom signal, which covered the fre-
quency bands of all targeted modes. The artificial vertical loading
used was generated by a payload of 113 N. The output force and the
excitation frequency were controlled by altering the amplitude and
frequency of input voltage by the LabVIEW software on the laptop.
The acceleration of the shaker mass was measured by a Kistler 8776
uniaxial accelerometer for calculating the input force to the system.
The acceleration responses of the 11 DOFs (10 on the pedestrian
bridge and one on the shaker mass) were collected for modal
identification.
Fig. 4. Lift and steel beam installed after the completion of the mea- According to the results from the ambient vibration test, the ex-
surement (image by F. L. Zhang) citation frequency range was set as 3.5–14 Hz. A broadband pseu-
dorandom signal was generated by filtering a Gaussian white noise
through a band-pass filter to focus on the frequency range of inter-
APS Dynamics, San Juan Capistrano, CA) with a moving mass of est. The pseudorandom excitation was designed to have a flat-root
13.2 kg was used to generate the excitation, as shown in Fig. 6. power spectral density (PSD) in the frequency range. It must be
Fig. 7 shows the set of other equipment used in this forced vibration pointed out that the input voltage to the shaker cannot be too high,
test. Shaker location was designed based on the mode shapes of the especially at the beginning of the test. Otherwise, a large, sudden

© ASCE C4016002-4 J. Bridge Eng.

J. Bridge Eng., 2016, 21(8): C4016002

 three setups were designed to study the three targeted modes. For
each mode, more than 10 excitation frequencies were considered.
At each excitation frequency, a sinusoidal excitation force was
generated by the shaker (see Fig. 6). The measurement for each
excitation frequency lasted for approximately 1 min. It must be
pointed out that additional excitation frequencies would be added
onsite based on the measured amplitude of vibration responses. If
the change in the vibration amplitudes in two succeeding excita-
tion frequencies was large, additional excitation frequency would
be added between the two succeeding frequencies to increase the
resolution of the data points near resonance. Unlike the pseudor-
andom excitation tests, the shaker location in this test was differ-
ent in different setups. In Setup 1, the shaker was put in the middle
Downloaded from ascelibrary.org by University of Central Florida on 11/23/20. Copyright ASCE. For personal use only; all rights reserved.

of the bridge (middle in both length and width directions) to

excite the first mode at approximately 4.67 Hz. In Setup 2, the
shaker was moved to the quarter position in width and middle
position in length for the torsional mode at approximately 6.62
Hz. In Setup 3, the shaker was moved to the quarter position in
length and middle position in width to excite the bending mode at
approximately 10.97 Hz.

Free Vibration Test

During the frequency resonant excitation forced vibration test, at
the moment when the shaker was turned off, the bridge started to go
Fig. 6. Shaker used in forced vibration test (image by Y. C. Ni) into free vibration. Therefore, the free vibration data could be con-
sidered as the byproduct of the forced vibration tests. There were
more than 10 excitation frequencies in each setup of the frequency
resonant excitation test. For each excitation frequency, the vibration
responses of the bridge were continuously measured for 1 min after
the shaker was switched off. This portion of data were treated as
free vibration test data.

Modal Identification Methods

Three sets of vibration data for the CityU pedestrian bridge were
obtained after the series of field tests. Each of them required a differ-
ent modal identification method in the analysis. In this paper, all three
methods are based on the Bayesian approach, which aims to identify
not only the most probable values of modal parameters but also the
corresponding posterior uncertainties. These three Bayesian meth-
Fig. 7. Measurement equipment, forced vibration (image by Y. C.
ods are introduced in the following subsections.
Ni)

Fast Bayesian Modal Identification Method for

impact force may be generated, and it may damage the target bridge. Ambient Vibration
To ensure the safety of the experiment and enable a smooth transi-
tion of the shaker from rest to motion, 5 s were reserved at the be- The fast Bayesian fast Fourier transform (FFT) method for ambient
ginning and at the end of the test to allow the input force to increase modal identification (Au et al. 2013) is first outlined here. In this
to the target amplitude and decrease to zero in a linear form gradu- method, the input forces and the responses are modeled as stationary
ally. The excition frequency first increased linearly from 3.5 to 14 stochastic processes. The antialiasing filters were used to remove fre-
Hz in 20 s and then decreased linearly back to 3.5 Hz in another 20 s quency components higher than the cutoff frequency, which is usually
in an antisymmetric manner. Therefore, the total excitation duration set as the Nyquist frequency or lower. The MPV of modal parameters
was 5 þ 20 þ 20 þ 5 = 50 s. was obtained by maximizing the posterior probability density function
(PDF) of modal parameters conditional on the selected frequency
Frequency Resonant Excitation band of FFT data, which was calculated from the measured responses.
Apart from the pseudorandom excitation forced vibration test, Therefore, if sufficient data are provided, the number of FFT points
traditional frequency resonant excitation (frequency sweep) does not directly influence the MPV but the posterior uncertainty of
forced vibration tests were also performed. For each mode of in- the modal parameters, which can be quantified and assessed analyti-
terest, the target frequency range near the natural frequency was cally. This method has been used in practical implementations, and
selected from the PSD spectra, which were obtained from ambi- interested readers are redirected to Au (2011, 2012a, b) and Zhang
ent vibration test data. The target frequency range was then di- and Au (2013) for the detail formulation, and Au and Zhang (2012)
vided into a group of specified excitation frequencies. A total of and Au et al. (2013) for its applications in various existing structural

© ASCE C4016002-5 J. Bridge Eng.

J. Bridge Eng., 2016, 21(8): C4016002

 systems. To be self-contained, the method is briefly introduced in the where UðiÞ 2 Rn (i = 1, …, m) is the ith mode shape confined to the
following (Au and Zhang 2012). measured DOFs; and h€ i ðtÞ = ith modal response. Its FFT is given by
Let h denote the modal parameters to be identified, which include
natural frequencies (f ), damping ratios (z ), mode shapes (U), PSD of X
m
F k ðhÞ ¼ Fðfk Þ hik Mi1 UðI; iÞUðiÞ (5)
modal forces (S), and PSD of prediction error (Se ). Zk ¼ i¼1
½ReF k ; ImF k  2 R2n denotes an augmented vector of the real and
imaginary part of F k , where F k is the FFT of measured data; n denotes where Fðfk Þ = FFT of the measured input load; hik ¼
the number of measured DOFs. The collection within the selected fre- ½ð b 2ik  1Þ þ ið2 z i b ik Þ1 = (complex) transfer function of the ith
quency band is denoted by fZk g. Based on Bayes’ theorem, the poste- mode evaluated at frequency fk with b ik ¼ fi =fk ; Mi = modal mass
rior PDF of h given the data can be expressed as (Au 2011) of the ith mode; and UðI; iÞ = mode shape at DOF I (excitation loca-
tion) of the ith mode. The vibration data used for modal identifica-
pðhjfZk gÞ / pðhÞpðfZk gjhÞ (1)
tion should include the portion of response data before and after the
Downloaded from ascelibrary.org by University of Central Florida on 11/23/20. Copyright ASCE. For personal use only; all rights reserved.

forced vibration phase to well satisfy the assumption of Eq. (5).

where pðhÞ = prior PDF that reflects the plausibility of h in the absence Given the measured acceleration data of the structure (fx^ €j g) and
of data. Assuming uniform prior information, the posterior PDF the input force (fFj g), the modal identification problem infers infor-
[pðhjfZk gÞ] is proportional to the likelihood function [pðfZk gjhÞ]. mation about the set of modal parameters (h). Accordingly, the neg-
The MPV of the modal parameters (h) can be determined by maximiz- ative log-likelihood function (NLLF) is
ing pðhjfZk gÞ and hence pðfZk gjhÞ for uniform prior.
For large amounts of data and a small time interval, the FFT at LðhÞ ¼ nNf lnp þ nNf lnSe þ S1
e JðhÞ (6)
different frequencies can be shown to be asymptotically independ-
ent and follow a Gaussian distribution. The likelihood function where Nf = number of FFT ordinates in the selected frequency
[pðfZk gjhÞ] can be expressed as band, and
  X
Y 1
pðfZk gjhÞ ¼ ð2p Þn ðdetCk Þ1=2 exp  ZTk C1 k Z k (2) JðhÞ ¼ ½F^ k  F k ðhÞ ½F^ k  F k ðhÞ (7)
k
2 k

where detð:Þ = determinant, and where F^ k is the FFT of x^

€j . The derivative of LðhÞ with respect to h
    is equal to that of JðhÞ multiplied by Se1 . This means that the MPV
1 U Re Hk Im Hk UT Se of h can be obtained as the minimizer of JðhÞ, and it will not depend
Ck ¼ þ I2n (3)
2 U Im Hk Re Hk U T
2 on Se . In contrast, the difficulties caused by the high power of
UðI; iÞ (Au and Ni 2014) can also be solved by expressing it in
denotes the covariance matrix of Zk , where U ¼ ½/1 ; /2 ; :::; /m  2 terms of other modal parameters. Finally, the number of parameters
Rnm denotes the mode shape matrix; m = number of modes in a to be numerically optimized is only 2m, which does not depend on n
selected frequency band; I2n 2 R2n = identity matrix; and Hk 2 and is often not a large number. This is because m is the number of
Rmm = transfer matrix containing the modal parameters to be iden- closely spaced modes in a given frequency band, and its value rarely
tified (i.e., natural frequency, damping ratio, and PSD of modal exceeds 3. In addition to the MPV of modal parameters, the
force). Theoretically, the MPV of modal parameters can be obtained Bayesian formulation also allows their posterior uncertainty to be
by maximizing pðfZk gjhÞ in Eq. (2). assessed quantitatively. The analytical expressions for the deriva-
To address the ill-condition problem and overcome the compu- tives of the NLLF allow the Hessian and hence the posterior covari-
tational difficulties caused by the large number of measured DOFs ance matrix to be determined accurately and quickly without resort-
in the optimization, fast algorithms were developed, which allow ing to a finite-difference method. Interested readers are redirected to
Au and Ni (2014) for detailed formulations and related discussions.
the MPV to be obtained almost instantaneously in the case of well-
separated modes (Au 2011; Zhang and Au 2013) and closely
spaced modes (Au 2012a, b). The posterior uncertainty can also be Fast Bayesian Modal Identification Method for
obtained analytically, which makes it possible to assess the accu- Free Vibration
racy of the MPV and decide whether additional data are needed Based on similar ideas, a fast Bayesian method for free vibration data
for improving the accuracy of the identified modal parameters. was recently developed (Zhang et al. 2016; Ni et al. 2015). The modal
parameters and the associated posterior uncertainty can be deter-
Fast Bayesian Modal Identification Method for mined efficiently and accurately. Because of space limitation, the
Forced Vibration detail formulations are not repeated in this paper. Interested readers
are redirected to Zhang et al. (2016) and Ni et al. (2015) for details.
A fast Bayesian frequency domain method for modal identification
in the context of a forced vibration test with a single shaker input
and multiple output acceleration response measurements is briefly Data Analysis
introduced as follows (Au and Ni 2014). This method is applicable
for cases of well-separated and closely spaced modes even with a By following the aforementioned modal identification methods, the
large number of measured DOFs. identified results are discussed and compared in this section.
Assuming linear classically damped dynamics with m contribut-
ing modes, the theoretical model acceleration response is given by Ambient Vibration Data
X
m A typical time history of the collected data is shown in Fig. 8(a),
x€ðhÞ ¼ UðiÞ h
€ i ðtÞ (4) where the structural response is found to be stationary in general.
i¼1 Figs. 8(b and c) show the root PSD and root singular value (SV)

© ASCE C4016002-6 J. Bridge Eng.

J. Bridge Eng., 2016, 21(8): C4016002

 -4
x 10
5

Acceleration (g)
0

-5
0 200 400 600 800 1000 1200
(a) Time (s)
Downloaded from ascelibrary.org by University of Central Florida on 11/23/20. Copyright ASCE. For personal use only; all rights reserved.

-4
10

-5
10
[g/sqrt(Hz)]

-6
10

-7
10
0 5 10 15
(b) Frequency (Hz)

-3
10
-4 1 2
10 4
3
[g/sqrt(Hz)]

-5
10
-6
10
-7
10 | | | | | || |
-8
10
0 5 10 15
(c) Frequency (Hz)

Fig. 8. Ambient vibration: (a) typical time history; (b) root PSD spectra of measured DOFs; (c) root SV spectra of measured DOFs

spectra of the data. The SV spectra are utilized for locating the ini- Table 1. Modal Identification Results of the CityU Pedestrian Bridge
tial guesses of the natural frequencies and the frequency bands for (Ambient Excitation)
modal identification. In Fig. 8(c), the number on the peaks, the dot,
Modal
and the horizontal line under the peaks denote the mode numbers,
parameter MPV and COV Mode 1 Mode 2 Mode 3 Mode 4
the initial guesses of natural frequencies, and the selected frequency
bands, respectively. Modal identification was performed for each f MPV (Hz) 4.688 6.690 10.346 10.966
frequency band separately. COV (%) 0.063 0.038 0.12 0.064
Table 1 summarizes the identified modal parameters, including z MPV (%) 1.20 0.64 1.71 1.37
natural frequency, damping ratio, PSD of modal force, and PSD of COV (%) 5.87 6.19 12.7 7.46
prediction error. In this study, four modes are presented: the three S MPV (1012 g2 =Hz) 7.43 3.36 0.23 0.50
bending modes (i.e., Modes 1, 3, and 4) and the torsional mode COV (%) 3.56 2.88 20.0 9.44
(i.e., Mode 2), whose frequencies range from 4 to 11 Hz. The natu- Se MPV (1012 g2 =Hz) 2.60 2.96 4.32 3.59
ral frequencies of Modes 1, 2, and 4 were consistent with the reso- COV (%) 0.98 0.85 1.51 1.27
nance peaks in PSD and SV spectra, whereas the resonance peak a 1-EMAC (%) 0.0002 0.0002 0.0001 0.0031
for Mode 3 was not obvious. The damping ratios were all less than
2%, with the smallest damping at Mode 2. The PSD of modal
force was of an order of magnitude of 1012 g2 =Hz, and there was The order of magnitude was similar to that of the PSD of modal
a descending trend with an increase in mode number. This might force. These two quantities are not the structural properties of the
be attributed to the first mode dominating the vibration of the sys- bridge. They only reflect the effect from the environment. In Table
tem. The PSD of prediction error in the four modes was similar. 1, the posterior coefficients of variation (COVs) for each MPV

© ASCE C4016002-7 J. Bridge Eng.

J. Bridge Eng., 2016, 21(8): C4016002

 (posterior coefficient of variation = posterior standard derivation/ spectra. The last row of Table 1 shows the 1 – EMAC value, which
MPV) of the modal parameters are also presented. The COV of is utilized to assess the posterior uncertainty of mode shape.
natural frequency was small (the highest one equals 0.12%), imply- EMAC represents the expected value of the modal assurance crite-
ing that the identified natural frequencies were accurate. The poste- rion (Au and Zhang 2011).The closer the EMAC value is to unity,
rior COVs for damping ratio and PSD of modal force were similar the more accurate the identified mode shape will be. It is clear that
with an order of magnitude of a few percent, which were larger a is close to zero for all four modes, showing that the uncertainty
than those of PSD of prediction error. It is worth mentioning that, associated with the identified mode shapes was very low.
in general, the posterior COVs of the modal parameters in the third Fig. 9 summarizes the modal properties (MPV) of Modes 1–4.
mode are larger than those in other modes. This might be attributed The dots show the measured locations. In the process of modal
to the s/n ratio of this mode being smaller, as reflected in the PSD identification in this work, all mode shapes were normalized to

Mode 1 f=4.688 Hz, z=1.2%

Downloaded from ascelibrary.org by University of Central Florida on 11/23/20. Copyright ASCE. For personal use only; all rights reserved.

Z
Z

20
0 10
10
20 0 Y (m) 0 5 10 15 20
X (m)
X (m)

Mode 2 f=6.69 Hz, z=0.6%

Z
Z

20
0 10
10
20 0 Y (m) 0 5 10 15 20
X (m)
X (m)

Mode 3 f=10.346 Hz, z=1.7%

Z
Z

20
0 10
10
20 0 Y (m) 0 5 10 15 20
X (m)
X (m)

Mode 4 f=10.966 Hz, z=1.4%

Z
Z

20
0 10
10
20 0 Y (m) 0 5 10 15 20
X (m)
X (m)

Fig. 9. Mode shapes of Modes 1–4, ambient vibration

© ASCE C4016002-8 J. Bridge Eng.

J. Bridge Eng., 2016, 21(8): C4016002

 unity. For the first mode, the natural frequency is 4.69 Hz with a symmetrical. The maximum deformation is approximately at the
damping ratio of 1.2%. The mode shape is shown in two different quarter positions along the length, whereas the minimum is near
views. The mode shape on the left and right sides of the bridge is the center of the bridge. The mode shape on the two sides of the
represented by blue solid and red dashed lines, respectively. It is bridge is consistent. Based on the identified mode shapes in Fig. 9,
clear from Mode 1 of the figure that the mode shapes on the two the modal identification results are reasonable.
sides of the bridge are symmetric and almost coincide with each Another two modal identification methods, frequency domain
other. The discrepancy between the left and right parts of the decomposition (FDD) method (Brincker et al. 2001) and stochas-
mode shape is small, and it may be caused by the construction tic subspace identification (SSI) method (Peeters and De Roeck,
handiwork (e.g., the bending stiffness on the two sides of the 2001b), were also used to identify the modal parameters of this
bridge is not identical). In this mode, the deformation of the bridge bridge. Table 2 shows the identified modal parameters of the
reaches the maximum at the center and minimum at both ends. tested structure with three different methods. It can be concluded
This is a typical deform shape of the fundamental bending mode that the natural frequencies identified by different methods are
of a simple beam. The natural frequency of the second mode is very similar for Modes 1, 2, and 4. For Mode 3, the natural fre-
Downloaded from ascelibrary.org by University of Central Florida on 11/23/20. Copyright ASCE. For personal use only; all rights reserved.

6.69 Hz with a damping ratio of 0.83%. It is clear from the figure quencies calculated by the Bayesian and FDD methods are close,
that this is a torsional mode, with the mode shapes on the two sides whereas the one obtained by the SSI method is smaller. This
in opposite signs (or directions). The torsional center is along the might be attributed to the small s/n ratio of this mode as men-
central line in the longitudinal direction. The mode shape on each tioned. For the damping ratios, larger discrepancies can be
side looks like an arc with the maximum deformation at the center observed, and they are obvious for the third and fourth modes,
and minimum deformation at the two ends. It is shown in Fig. 9 as where the damping ratios calculated using the Bayesian FFT
an antisymmetric mode. The third mode is also a bending mode method and SSI method have a similar order of magnitude, but
with a natural frequency of 10.35 Hz and a damping ratio of the values identified using the FDD method are much smaller,
1.71%. With the 10 measured DOFs, the mode shape of this mode especially for the fourth mode. From the identified results, differ-
looks very similar to that of the first mode, with a modal assurance ent methods give different results. This confirms the existence of
criterion (MAC) of 0.97. Based on the measured information, it is uncertainty in the modal identification results. Therefore, it is
difficult to identify a reason for the similarity. For the fourth necessary to assess these uncertainties for future model updating
mode, the natural frequency is 10.97 Hz with a damping ratio of or damage detection, especially in the case with minor damage,
1.37%. This mode is another bending mode, as shown in Fig. 9. where it is difficult to judge whether the variation of modal pa-
The mode shape on the two sides of the pedestrian bridge is rameters is attributed to structure change or uncertainty. The fast
Bayesian FFT method provides an efficient way to quantify these
uncertainties, so it has the potential to improve the reliability of
Table 2. Modal Identification Results Using Different Methods structural model updating and damage detection (Au and Zhang
Mode 1 Mode 2 Mode 3 Mode 4 2015; Zhang and Au 2016).
Method used f (Hz) z (%) f (Hz) z (%) f (Hz) z (%) f (Hz) z (%)
Forced Vibration Data
Bayesian 4.688 1.20 6.690 0.64 10.346 1.71 10.966 1.37
SSI 4.698 0.89 6.704 0.66 10.070 2.79 10.970 1.21 Known Pseudorandom Excitation
FDD 4.720 1.10 6.696 0.88 10.410 0.71 10.930 0.06 Fig. 10 shows the change in excitation frequency as a function of
time and a typical time history at DOF 3. The root PSD spectrum

14
Excitation frequency (Hz)

12
Shaker Shaker Shaker
on
9 off off

3.5
-4
x 10
5
Acceleration (g)

-5
0 10 20 30 40 50 60
Time (s)

Fig. 10. Excitation frequency and a typical time history, forced vibration, pseudorandom excitation

© ASCE C4016002-9 J. Bridge Eng.

J. Bridge Eng., 2016, 21(8): C4016002

 0
10

-1
10

[g/sqrt(Hz)]
-2
10

-3
10
0 5 10 15
(a) Frequency (Hz)
Downloaded from ascelibrary.org by University of Central Florida on 11/23/20. Copyright ASCE. For personal use only; all rights reserved.

-3
10

-4
10
[g/sqrt(Hz)]

-5
10

-6
10

-7
10
0 5 10 15
(b) Frequency (Hz)

-3
10
1 2
-4
10 3 4
[g/sqrt(Hz)]

-5
10
-6
10
-7
10 | | | | | |
-8
10
0 5 10 15
(c) Frequency (Hz)

Fig. 11. Spectra of data, forced vibration, pseudorandom excitation: (a) root PSD spectra of shaker mass; (b) root PSD spectra of measured DOFs; (c)
root SV spectra of measured DOFs

of the input excitation is shown in Fig. 11(a). The spectrum was Table 3. Modal Identification Results of the CityU Pedestrian Bridge
flat between 3.5 and 14 Hz as expected. The amplitudespffiffiffiffiffiffi in the
(Pseudorandom Excitation)
frequency band of interest are of an order of 101 g Hz. In Figs. Modal
11(b and c), the root PSD and SV spectra of the measured struc- parameter MPV and COV Mode 1 Mode 2 Mode 3 Mode 4
tural acceleration are shown. Three frequency bands were
selected in Fig. 11(c), and their FFT data were used for identify- f MPV (Hz) 4.693 6.673 10.264 10.940
ing the modes within that frequency band. Each of the first two COV (%) 0.012 0.005 0.051 0.089
bands contains one mode, whereas the third band contains two z MPV (%) 1.13 0.88 2.44 2.98
closely spaced modes. Because there were common characteris- COV (%) 1.61 0.93 4.25 6.53
r MPV (105 ) 2.37 3.26 1.42 0.32
tics for these two modes, it would cause biased results if they
COV (%) 2.34 1.51 5.11 21.7
were identified separately. Based on the fast Bayesian FFT
Se MPV (1012 g2 =Hz) 48 59 36 36
method using forced vibration data, these two modes were identi-
COV (%) 2.76 2.64 3.31 3.31
fied in one band with m = 2.
a 1 – EMAC (%) 0.027 0.009 0.126 0.251
The modal identification results are listed in Table 3. The first
mode is at 4.693 Hz with a damping ratio of 1.1%. From the
identified mode shapes in Fig. 12, it is obvious that this mode is deformation at the two ends, which is consistent with the results
the first bending mode in the vertical direction with the maximum of the ambient vibration test. Mode 2 is the torsional mode with a
deflection at approximately the midspan and the minimum natural frequency at 6.673 Hz and a damping ratio of 0.9%.

© ASCE C4016002-10 J. Bridge Eng.

J. Bridge Eng., 2016, 21(8): C4016002

 Mode 1 f=4.693 Hz, z=1.1%

Z
Z
20
0 10
10
20 0 Y (m) 0 5 10 15 20
X (m)
X (m)
Downloaded from ascelibrary.org by University of Central Florida on 11/23/20. Copyright ASCE. For personal use only; all rights reserved.

Mode 2 f=6.673 Hz, z=0.9%

Z
Z

20
0 10
10
20 0 Y (m) 0 5 10 15 20
X (m)
X (m)

Mode 3 f=10.264 Hz, z=2.4%

Z
Z

20
0 10
10
20 0 Y (m) 0 5 10 15 20
X (m)
X (m)

Mode 4 f=10.94 Hz, z=3%

Z
Z

20
0 10
10
20 0 Y (m) 0 5 10 15 20
X (m)
X (m)

Fig. 12. Mode shapes of Modes 1–4, forced vibration, pseudorandom excitation

Mode 3 has a natural frequency of 10.264 Hz and damping ratio mode in the ambient vibration test. The identified natural frequen-
of 2.4%. Similar to the ambient vibration test, it was found that cies of Modes 3 and 4 were close, as expected from the PSD
the mode shape of this mode was similar to Mode 1 at the meas- spectra. In general, four modes were identified from the pseudor-
ured DOFs although the natural frequencies are different. These andom excitation data, and Mode 3 was the most special. Further
two modes could not be separated by looking at the identified investigation needs to be conducted.
mode shape alone; however, from the PSD spectra, it could Generally speaking, from the identification results in Table 3
clearly be seen that these two modes were two totally different involving both MPV and posterior uncertainty of the modal pa-
modes corresponding to two different peaks. Mode 4 appeared as rameters, for natural frequency (f ) of different modes, the posterior
an antisymmetric mode along the length of the bridge at 10.940 COVs are very small of an order of magnitude of 104 . Damping
Hz with a damping ratio of 3.0%. It corresponds to the fourth ratios ( z ) for the first two modes are approximately 1%, which

© ASCE C4016002-11 J. Bridge Eng.

J. Bridge Eng., 2016, 21(8): C4016002

 were smaller than those in the third and fourth modes (approxi- 13(b)], the structural response in Mode 1 is dominant and clearly
mately 2 and 3%). Their posterior COVs are of an order of mag- shown in Figs. 13(c and d).
nitude of 102 . For mass ratio (r), the identified values are of an In the frequency resonant excitation forced vibration test, a num-
order of magnitude of 105 , and the posterior COVs are smaller ber of predefined excitation frequencies (near the resonance fre-
(of an order of magnitude of 102 ) for the first three modes and quency) were tested one by one. In this study, the list of excitation fre-
are larger (of an order of magnitude of 101 ) for the fourth quencies was selected from the power spectral density spectra of the
mode. Because the mass ratio of the fourth mode is relatively ambient vibration data. Because the peak for the mode at approxi-
smaller, it is more sensitive, resulting in a higher uncertainty. In mately 10.3 Hz (i.e., Mode 3) was not obvious, this mode was not
general, the MPV of the modal mass ratio is very small, but the considered in the frequency resonant excitation test. Table 5 shows
shaker force is sufficient to generate a response significantly the identified modal parameters and the associated posterior uncer-
above the ambient level by means of resonance. This can be tainty from the frequency resonant excitation forced vibration test.
observed from the large peaks in the root PSD spectra. The PSD The natural frequencies for these three modes were well matched
of prediction error (Se ) is small at approximately 1010 g2 =Hz. with the corresponding peaks in the PSD spectra. The damping ratio
Downloaded from ascelibrary.org by University of Central Florida on 11/23/20. Copyright ASCE. For personal use only; all rights reserved.

The posterior uncertainty of the mode shape is quite small, as for the second mode was smaller than those of the other two modes,
reflected by the value of a, indicating a high accuracy in the which is consistent with the identified results in other field tests. The
identified mode shape for each mode based on the data collected. posterior COVs in Mode 1 and Mode 2 were approximately 0.1%,
The last mode has a higher a value, which means that the mode indicating that the identification of damping ratio in these two modes
shape for the fourth mode is less accurate when compared to the was accurate. Similar observation about posterior uncertainty was
other three modes. found for the modal mass ratio and PSD of prediction error. There
In the forced vibration tests of the pedestrian bridge, a previous was an ascending trend for the posterior uncertainty of identified
modal parameters from Mode 1 to Mode 4, implying that the identifi-
study was performed by the same team with a chirp signal excita-
cation results tend to be less accurate in higher modes. This is reason-
tion. It was found that the MPVs of the natural frequencies, damp-
able because it is believed that higher energy is needed in exciting
ing ratios, and mass ratios identified in Table 3 are well matched
higher modes; however, the energy inputs to the shaker for exciting
with the results in Au and Ni (2014). The main difference in the
different modes are very similar. Fig. 14 shows the identified mode
results from the two studies lies in the posterior uncertainties.
shape for all the three modes. Similar to the results from the other two
Table 4 shows the ratios of posterior uncertainties for pseudoran-
field tests, the first mode is the fundamental bending mode, the sec-
dom excitation over chirp excitation. The posterior COVs for pseu- ond one is the torsional mode, and the fourth one is the second bend-
dorandom excitation were generally larger than those for chirp exci- ing mode.
tation. An exception appears in the modal mass ratio of Mode 4. By comparing the COVs in Table 3 to those in Table 5, it is
The difference between the results under the two types of excita- clear that the posterior uncertainties of results from the pseudor-
tions may be caused by the different loading manners. For the pseu- andom excitation test are higher than those of results from the fre-
dorandom excitation, the movement in the next instant cannot be quency resonant excitation test. In the latter test, all data in differ-
foreseen because it is random; for the chirp excitation, the shaker ent exciting frequencies were used in modal identification, and
mass moves gradually, allowing a smooth transition from rest to the total measurement duration is more than 10 times that in the
motion. The latter may produce larger excitation. In general, the pseudorandom excitation test. It is reasonable that more measured
identified modal properties from the pseudorandom and chirp exci- data points could result in more accurate identification results.
tations are consistent with each other and sufficiently accurate to This also shows the ability of the Bayesian modal identification
represent the dynamic properties of the structure at the time of methods to analytically calculate the posterior uncertainties of the
instrumentation. identified results.

Frequency Resonant Excitation Free Vibration Data

Based on the data collected in the frequency resonant excitation
forced vibration test, the modal parameters were identified using This subsection focuses on the modal parameters identified using the
free vibration data. Fig. 15(a) shows a typical time history of Setup 1
the same fast Bayesian FFT method for forced vibration data as
for Mode 1, where a free vibration response can be observed. Figs.
the pseudorandom excitation case. Fig. 13(a) shows a typical time
15(b and c) show the root PSD and SV spectra, in which the selected
history in a frequency resonant excitation test, where the struc-
frequency band for Mode 1 is indicated. The modal identification
tural response was significantly increased by an excitation fre-
was individually performed for each frequency band. Because only
quency near resonant. Fig. 13 also shows the root PSD and SV
Modes 1, 2, and 4 were considered in the frequency resonant excita-
spectra of measured data for Mode 1. Because the excitation is
tion forced vibration test, the free vibration test also considered these
mainly around the natural frequency of this mode [shown in Fig. three modes. The identified modal parameters corresponding to a
typical setup (Setup 1) are shown in Table 6. The identified results
show that the identified natural frequencies are quite close to the val-
Table 4. Ratios of Posterior COVs for Pseudorandom Excitation over ues identified from ambient and forced vibration tests. The damping
Chirp Excitation ratios vary from 0.8 to 1.5%, with COVs of 1–11%.
Modal Similar to the results in the forced vibration tests, the posterior
parameter Mode 1 Mode 2 Mode 3 Mode 4 uncertainties of the natural frequency, damping ratio, and mode
shape of the fourth mode were much larger than those in the first
f 1.572 1.689 1.607 1.536
and second modes. The mode shapes of the three modes of Setup
z 1.723 1.801 1.580 1.511
1 in terms of their MPV are shown in Fig. 16. The first and fourth
r 1.672 1.791 1.609 0.949
modes are bending modes in the vertical direction, whereas the
Se 1.081 1.081 1.084 1.084
second mode is a torsional mode. More than 10 sets of free vibra-
a 2.781 3.084 2.655 2.094
tion data with different excitation frequency were recorded in each

© ASCE C4016002-12 J. Bridge Eng.

J. Bridge Eng., 2016, 21(8): C4016002

 -3
x 10
1.5
1

Acceleration (g)
0.5

0
-0.5

-1

-1.5
0 10 20 30 40 50 60 70
Downloaded from ascelibrary.org by University of Central Florida on 11/23/20. Copyright ASCE. For personal use only; all rights reserved.

(a) Time (s)

2
10

0
10
[g/sqrt(Hz)]

-2
10

-4
10
0 1 2 3 4 5 6 7 8 9 10
(b) Frequency (Hz)

-2
10

-3
10
[g/sqrt(Hz)]

-4
10

-5
10

-6
10
0 1 2 3 4 5 6 7 8 9 10
(c) Frequency (Hz)

-2
10
-3 1
10
[g/sqrt(Hz)]

-4
10
-5
10
-6
10
| |
-7
10
0 1 2 3 4 5 6 7 8 9 10
(d) Frequency (Hz)

Fig. 13. Forced vibration, sine wave frequency resonant excitation: (a) typical time history; (b) root PSD pectra of shaker mass; (c) root PSD spectra
of measured DOFs; (d) root SV spectra of measured DOFs

mode. Through the investigation of the effect of excitation fre- natural frequency of the system. This observation is reasonable
quency performed by Ni et al. (2015), it was found that the poste- because the closer the excitation frequency is to the natural fre-
rior uncertainties of natural frequency and damping ratio tend to quency, the higher the s/n ratio is that can be obtained in the meas-
increase when the excitation frequency is far away from the ured vibration data.

© ASCE C4016002-13 J. Bridge Eng.

J. Bridge Eng., 2016, 21(8): C4016002

 Comparison of Modal Identification Results from higher than 1%, with the maximum difference approximately 10%.
Different Kinds of Tests The identified natural frequencies for the second mode from all tests
were very close to 6.7 Hz, with a relative difference of 1%. The
The modal identification results of the three common modes (i.e.,
Modes 1, 2, and 4) in all three kinds of tests are summarized in damping ratios identified were consistent in all tests, with a value
Table 7 for easy comparison. For the first mode, the identified natu- lower than 1%. For the fourth mode, the identified natural frequen-
ral frequencies from all three field tests were all close to 4.7 Hz. cies from all tests were consistent. The difference was in the damp-
Regarding the damping ratio, all identified results were a little ing ratios, which ranged from 1.4 to 3%. It is found by other
researchers that the damping of steel and concrete structures
increases with amplitude of vibration (Magalhães et al. 2010; Fang
Table 5. Modal Identification Results of the CityU Pedestrian Bridge et al. 1999). This was reflected in the identification results. The iden-
(Frequency Resonant Excitation) tified damping ratios for Modes 2 and 3 were smaller in the ambient
Modal
vibration test when compared to those in the two forced vibration
tests.
Downloaded from ascelibrary.org by University of Central Florida on 11/23/20. Copyright ASCE. For personal use only; all rights reserved.

parameter MPV and COV Mode 1 Mode 2 Mode 4

In general, the posterior uncertainties of natural frequency in
f MPV (Hz) 4.664 6.620 10.966 all tests were very small (less than 0.15%). This indicates that
COV (%) 0.001 0.001 0.022 the identification of natural frequency is quite accurate based
z MPV (%) 1.25 0.92 1.88 on the collected data. For the damping ratio, the posterior
COV (%) 0.15 0.13 2.69 COVs were approximately a few percent for the ambient vibra-
r MPV (105 ) 2.45 3.42 0.46 tion test, and their values were smaller in the forced and free
COV (%) 0.22 0.22 4.25 vibration tests. The data length used in modal identification in
Se MPV (1012 g2 =Hz) 442 1,631 5,650
the ambient test was much longer than that used in modal
COV (%) 0.69 0.70 1.64
identification in other tests. Although the data length was rela-
a 1 – EMAC (%) 0.0002 0.0002 0.0427
tively short, modal identification in forced and free vibration

Mode 1 f=4.664 Hz, z=1.3%

Z
Z

20
0 10
10
20 0 Y (m) 0 5 10 15 20
X (m)
X (m)

Mode 2 f=6.62 Hz, z=0.9%

Z
Z

20
0 10
10
20 0 Y (m) 0 5 10 15 20
X (m)
X (m)

Mode 4 f=10.966 Hz, z=1.9%

Z
Z

20
0 10
10
20 0 Y (m) 0 5 10 15 20
X (m)
X (m)

Fig. 14. Mode shapes of Modes 1, 2, and 4, forced vibration, sine wave frequency resonant excitation

© ASCE C4016002-14 J. Bridge Eng.

J. Bridge Eng., 2016, 21(8): C4016002

 -3
x 10
2

Acceleration (g)
0

-1

-2
0 1 2 3 4 5 6 7 8 9 10 11 12 13
(a) Time (s)
Downloaded from ascelibrary.org by University of Central Florida on 11/23/20. Copyright ASCE. For personal use only; all rights reserved.

-3
10

-4
10
[g/sqrt(Hz)]

-5
10

-6
10

-7
10
0 1 2 3 4 5 6 7 8 9 10
(b) Frequency (Hz)

-2
10

-3
10
[g/sqrt(Hz)]

-4
10

-5
10
| |
-6
10
0 1 2 3 4 5 6 7 8 9 10
(c) Frequency (Hz)

Fig. 15. Mode 1, Setup 1, free vibration: (a) typical time history; (b) root PSD spectra of measured DOFs; (c) SV spectra of measured DOFs

Table 6. Modal Identification Results of the CityU Pedestrian Bridge Table 8 shows the MAC values of the mode shapes identified
(Free Vibration) from different kinds of field tests. It is clear that the MAC values for
Modal Modes 1 and 2 were all larger than 0.99. For Mode 4, the MAC val-
parameter MPV and COV Mode 1 Mode 2 Mode 4 ues were a little smaller, but they were all larger than 0.94.

f MPV (Hz) 4.683 6.642 10.917

COV (%) 0.006 0.004 0.148 Conclusions
z MPV (%) 1.15 0.85 1.49
COV (%) 0.5 0.5 11.2 The procedures of ambient, forced, and free vibration field tests
Se MPV (1012 g2 =Hz) 65.71 144.82 168.17 of the CityU pedestrian bridge were reported in detail in this work.
COV (%) 5.3 5.1 6.0 The modal properties of the bridge were identified utilizing meas-
a 1 – EMAC (%) 0.0029 0.0025 0.5887 ured data from three kinds of field tests by the corresponding
recently developed frequency domain Bayesian modal identifica-
tion methods. The main findings are summarized as follows.
tests could also achieve similar accuracy. This is attributed to In the resonant excitation forced vibration test with sine wave
the fact that the input force information was known in the frequency sweep and the free vibration test, three modes (i.e.,
forced vibration tests, and the s/n ratios were large as a result Modes 1, 2, and 4) were identified, including two bending modes
of the artificial excitation in the forced and free vibration tests. and one torsional mode. In the ambient vibration test and the forced

© ASCE C4016002-15 J. Bridge Eng.

J. Bridge Eng., 2016, 21(8): C4016002

 Mode 1 f=4.683 Hz, z=1.2%

Z
Z
20
0 10
10
20 0 Y (m) 0 5 10 15 20
X (m)
X (m)
Downloaded from ascelibrary.org by University of Central Florida on 11/23/20. Copyright ASCE. For personal use only; all rights reserved.

Mode 2 f=6.642 Hz, z=0.9%

Z
Z

20
0 10
10
20 0 Y (m) 0 5 10 15 20
X (m)
X (m)

Mode 4 f=10.917 Hz, z=1.5%

Z
Z

20
0 10
10
20 0 Y (m) 0 5 10 15 20
X (m)
X (m)

Fig. 16. Mode shapes of Modes 1, 2, and 4, Setup 1, free vibration

Table 7. Comparison of the Modal Identification Results from Three Kinds of Field Tests

MPV Uncertainty, posterior COV

12 2
Mode Field test f (Hz) z (%) Se (10 g =Hz) f (%) z (%) Se (%) a (%)
1 Ambient vibration 4.688 1.20 2.60 0.063 5.87 0.98 0.0002
Force pseudorandom vibration 4.693 1.13 48 0.012 1.61 2.76 0.027
Force frequency resonant excitation 4.664 1.25 442 0.001 0.15 0.69 0.0002
Free vibration 4.683 1.15 65.71 0.006 0.5 5.3 0.0029
2 Ambient vibration 6.690 0.64 2.96 0.038 6.19 0.85 0.0002
Force pseudorandom vibration 6.673 0.88 59 0.005 0.93 2.64 0.009
Force frequency resonant excitation 6.620 0.92 1631 0.001 0.13 0.70 0.0002
Free vibration 6.642 0.85 144.82 0.004 0.5 5.1 0.0025
4 Ambient vibration 10.966 1.37 3.59 0.064 7.46 1.27 0.0031
Force pseudorandom vibration 10.940 2.98 36 0.089 6.53 3.31 0.251
Force frequency resonant excitation 10.966 1.88 5,650 0.022 2.69 1.64 0.0427
Free vibration 10.917 1.49 168.17 0.148 11.2 6.0 0.5887

vibration test with known pseudorandom excitation, an additional the second span of the bridge was instrumented), it is difficult to
mode (i.e., Mode 3) was identified. The mode shape of Mode 3 was explain the similarity between Modes 1 and 3. It is believed that the
similar to the fundamental bending mode (Mode 1) but with a mode shapes for Modes 1 and 3 are different in the first and third
totally different natural frequency. With the measured DOFs (only spans of the bridge.

© ASCE C4016002-16 J. Bridge Eng.

J. Bridge Eng., 2016, 21(8): C4016002

 Table 8. Modal Assurance Criteria of the Mode Shapes from Three Kinds of Field Tests

Frequency resonant
Mode Case Ambient Pseudorandom excitation Free
1 Ambient 1.0000 1.0000 1.0000 0.9974
Pseudorandom 1.0000 1.0000 1.0000 0.9974
Frequency resonant excitation 1.0000 1.0000 1.0000 0.9974
Free 0.9974 0.9974 0.9974 1.0000
2 Ambient 1.0000 1.0000 0.9999 0.9922
Pseudorandom 1.0000 1.0000 0.9999 0.9922
Frequency resonant excitation 0.9999 0.9999 1.0000 0.9933
Free 0.9922 0.9922 0.9933 1.0000
4 Ambient 1.0000 0.9470 0.9928 0.9904
Downloaded from ascelibrary.org by University of Central Florida on 11/23/20. Copyright ASCE. For personal use only; all rights reserved.

Pseudorandom 0.9470 1.0000 0.9523 0.9649

Frequency resonant excitation 0.9928 0.9523 1.0000 0.9778
Free 0.9904 0.9649 0.9778 1.0000

By comparing the results of the three modal identification meth- the installation and operation of the lift system will alter the
ods (i.e., the Bayesian methods, SSI method, and FDD method), dynamic characteristics of the bridge. The research team is planning
one can conclude that the identified modal parameters from all for another field test for measuring the vibration of the same bridge.
methods were similar, with the Bayesian and SSI methods provid- The effects of the lift system on the bridge can be studied by com-
ing more reliable results in this case. It is worth mentioning that the paring the identified modal parameters of the bridge before and after
Bayesian method provides an efficient way to calculate the posterior the installation of the lift system.
uncertainty analytically without resorting to a finite-difference
method, making it possible to assess the reliability of the identified
modal parameters directly.
The damping ratios of Mode 2 identified in all tests were smaller Acknowledgments
than those identified in Modes 1 and 4, showing good consistency
across different kinds of tests. Through comparison of the identifica- The paper is sponsored by the National Natural Science
tion results of the three kinds of field tests, the natural frequencies Foundation of China (Grant No. 51508407 and 51508413),
identified from different tests were very consistent. Regarding damp- Fundamental Research Funds for the Central Universities, China
ing, Modes 2 and 4 have relatively larger damping ratios in forced (Grant No. 2014KJ040), Shanghai Pujiang Program (Grant No.
vibration tests than in ambient vibration test, reflecting the amplitude 15PJ1408600), and Research Grants Council of the Hong Kong
dependence properties of damping. The natural frequency could be Special Administrative Region, China [Project No. 9041758
identified accurately, whereas the identified damping ratio had a rela- (CityU 110012)]. The authors thank Siu-Kui Au, professor at the
tively large posterior uncertainty. Because of the known input force University of Liverpool, for giving valuable suggestions during
information and high s/n ratio, short measurement durations in forced the field tests. The authors thank the anonymous reviewers for
and free vibration tests could achieve similar accuracy when com- their constructive comments.
pared to the ambient vibration test with long measurement duration.
In summary, for the estimation of natural frequency, all three kinds
of tests can be used to identify the most probable values reasonably References
with relatively small posterior uncertainty. With economy and effi-
ciency considerations, the ambient vibration test is recommended. Alampalli, S. (2000). “Effects of testing, analysis, damage, and environment
The analysis results show that the discrepancies in identified damp- on modal parameters.” Mech. Syst. Sig. Process., 14(1) 63–74.
ik, A., Bayraktar, A., and Sevim, B. (2011). “Output-only system
Altunis
ing ratio from different tests are relatively large when compared to
identification of posttensioned segmental concrete highway bridges.” J.
the discrepancies in identified natural frequency. The forced vibra-
Bridge Eng., 10.1061/(ASCE)BE.1943-5592.0000150, 259–266.
tion tests are recommended for damping ratio identification, as they Au, S. K. (2011). “Fast Bayesian FFT method for ambient modal identifica-
could provide more accurate results with smaller uncertainty. For the tion with separated modes.” J. Eng. Mech., 10.1061/(ASCE)EM.1943
identification of mode shape, similar to the natural frequency, the am- -7889.0000213, 214–226.
bient vibration test is recommended. Au, S. K. (2012a). “Fast Bayesian ambient modal identification in the fre-
The investigation of the modal properties of the CityU pedestrian quency domain, Part I: Posterior most probable value.” Mech. Syst. Sig.
bridge provides a reference for the Bayesian modal identification of Process., 26, 60–75.
this kind of structure based on different kinds of field tests. The Au, S. K. (2012b). “Fast Bayesian ambient modal identification in the fre-
implementation of these vibration tests provides valuable experience quency domain, Part II: Posterior uncertainty.” Mech. Syst. Sig. Process.,
of the design of an effective and safe field test for this kind of pedes- 26, 76–90.
trian bridge in the future. It is worth mentioning that the experimental Au, S. K., Ng, C. T., Sien, H. W., and Chua, H. Y. (2005). “Modal identifi-
cation of a suspension footbridge using free vibration signatures.” Int. J.
aspects are certainly important and decisive for the success of a diag-
Appl. Math. Mech., 1(4), 55–73.
nostic investigation, but that the most delicate issues are those con- Au, S. K., and Ni, Y. C. (2014). “Fast Bayesian modal identification of
nected to the interpretation of the experimental data in connection structures using known single-input forced vibration data.” Struct.
with some analytical or numerical models of the structural system Control Health Monit., 21(3), 381–402.
under study, which will be investigated in the near future. Au, S. K., and Zhang, F. L. (2011). “On assessing the posterior mode shape
A lift system was constructed on the CityU pedestrian bridge af- uncertainty in ambient modal identification.” Probab. Eng. Mech.,
ter the completion of the field tests in this paper. It is believed that 26(3), 427–434.

© ASCE C4016002-17 J. Bridge Eng.

J. Bridge Eng., 2016, 21(8): C4016002

 Au, S. K., and Zhang, F. L. (2012). “Ambient modal identification of a pri- Lam, H. F., Peng, H. Y., and Au, S. K. (2014b). “Development of a practical
mary-secondary structure by Fast Bayesian FFT method.” Mech. Syst. algorithm for Bayesian model updating of a coupled slab system utiliz-
Sig. Process., 28, 280–296. ing field test data.” Eng. Struct., 79, 182–194.
Au, S. K., and Zhang, F. L. (2016), “Fundamental two-stage formulation for Liang, J. W., and Feeny, B. F. (2006). “Balancing energy to estimate damp-
Bayesian system identification, Part I: General theory.” Mech. Syst. Sig. ing parameters in forced oscillators.” J. Sound Vib., 295(3–5), 988–998.
Process., 66–67, 31–42. Lutes, D. L., and Sarkani, S. (1997). Stochastic analysis of structural and
Au, S. K., Zhang, F. L., and Ni, Y. C. (2013). “Bayesian operational modal mechanical vibrations, Prentice Hall, Uppser Saddle River, NJ.
analysis: Theory, computation, practice.” Comput. Struct., 126, 3–14. Magalhães, F., Caetano, E., and Cunha, Á. (2007). “Challenges in the appli-
Au, S. K., Zhang, F. L. and To, P. (2012). “Field observations on modal cation of stochastic modal identification methods to a cable-stayed
properties of two tall buildings under strong wind.” J. Wind Eng. Ind. bridge.” J. Bridge Eng., 10.1061/(ASCE)1084-0702(2007)12:6(746),
Aerodyn., 101, 12–23. 746–754.
Bai, Y., and Keller, T. (2008). “Modal parameter identification for a GFRP Magalhães, F., Cunha, A., Caetano, E., and Brincker, R. (2010). “Damping
pedestrian bridge.” Compos. Struct., 82(1), 90–100. estimation using free decays and ambient vibration tests.” Mech. Syst.
Brincker, R., Zhang, L., and Anderson, P. (2001). “Modal identification of Sig. Process., 24(5), 1274–1290.
Downloaded from ascelibrary.org by University of Central Florida on 11/23/20. Copyright ASCE. For personal use only; all rights reserved.

output-only systems using frequency domain decomposition.” Smart Ni, Y. C., Zhang, F. L., Lam, H. F., and Au, S. K. (2016). “Fast Bayesian
Mater. Struct., 10(3), 441–455. approach for modal identification using free vibration data, Part II—
Brownjohn, J. M. W. (2003). “Ambient vibration studies for system identifi- Posterior uncertainty and application.” Mech. Syst. Sig. Process., 70–71,
cation of tall building.” Earthquake Eng. Struct. Dyn., 32(1), 71–95. 221–244.
Brownjohn, J. M. W., Moyo, P., Omenzetter, P., and Chakraborty, S. Peeters, B., and De Roeck, G. (2001a). “One-year monitoring of the Z24-
(2005). “Lessons from monitoring the performance of highway bridge.” Bridge: Environmental effects versus damage events.” Earthquake Eng.
Struct. Control Health Monit., 12(3–4), 227–244. Struct. Dyn., 30, 149–171.
Brownjohn, J. M. W., and Pavic, A. (2007). “Experimental methods for esti- Peeters, B., and De Roeck, G. (2001b). “Stochastic system identification for
mating modal mass in footbridges using human-induced dynamic exci- operational modal analysis: A review.” J. Dyn. Syst. Meas. Contr.,
tation.” Eng. Struct., 29(11), 2833–2843. 123(4), 659–667.
Catbas, F. N., Susoy, M., and Frangopol, D. M. (2008). “Structural Racic, V., Brownjohn, J. M. W., and Pavic, A. (2010). “Reproduction and
health monitoring and reliability estimation: Long span truss bridge application of human bouncing and jumping forces from visual marker
data.” J. Sound Vib., 329(16), 3397–3416.
application with environmental monitoring data.” Eng. Struct., 30(9),
Satake, N., Suda, K. I., Arakawa, T., Sasaki, A., and Tamura, Y. (2003).
2347–2359.
“Damping evaluation using full-scale data of buildings in Japan.” J.
Chang, P. C., Flatau, A., and Liu, S. C. (2003). “Health monitoring of civil
Struct. Eng., 10.1061/(ASCE)0733-9445(2003)129:4(470), 470–477.
infrastructure.” Struct. Health Monit., 2(3), 257–267.
Sohn, H., Farrar, C. R., Hemez, F. M., Shunk, D. D., Stinemates, D. W. and
Choi, S., Park, S., Hyun, C. H., Kim, M. S., and Choi, K. R. (2010). “Modal
Nadler, B. R. (2003). “A review of structural health monitoring litera-
parameter identification of a containment using ambient vibration meas-
ture: 1996-2001.” Rep. LA-13976-MS, Los Alamos National
urements.” Nucl. Eng. Des., 240(3), 453–460.
Laboratory, Los Alamos, NM.
Cunha, A., Caetano, E., and Delgado, R. (2001). “Dynamic tests on large
Spiridonakos, M., and Chatzi, E. (2014a). “Polynomial chaos expansion
cable-stayed bridge.” J. Bridge Eng., 10.1061/(ASCE)1084-0702(2001)6:
models for SHM under environmental variability.” Proc., Eurodyn
1(54), 54–62. 2014, 9th Int. Conf. on Structural Dynamics, Faculty of Engineering,
De Sortisa, A., Antonaccib, E., and Vestronic, F. (2005). “Dynamic identifi-
Univ. of Porto, Porto, Portugal.
cation of a masonry building using forced vibration tests.” Eng. Struct., Spiridonakos, M., and Chatzi, E. (2014b). “Stochastic structural identification
27(2), 255–265. from vibrational and environmental data”, Encyclopedia of earthquake en-
Fang, J. Q., Jeary, A. P., Li, Q. S., and Wong, C. K. (1999). “Random damp- gineering: Springer Reference, M. Beer, E. Patelli, I. Kougioumtzoglou,
ing in buildings and its AR model.” J. Wind Eng. Ind. Aerodyn., and I. Au, eds., Springer, Berlin, 1–16.
79(1–2), 159–167. Tamura, Y., and Suganuma, S. Y. (1996). “Evaluation of amplitude-de-
Farrar, C., Doebling, S., Cornwell, P., and Straser, E. (1997). “Variability of pendent damping and natural frequency of buildings during strong
modal parameters measured on the Alamosa Canyon bridge.” 15th Int. winds.” J. Wind Eng. Ind. Aerodyn., 59(2–3), 115–130.
Modal Analysis Conference (IMAC XV), Society for Experimental Ülker-Kaustell, M., and Karoumi, R. (2011). “Application of the continuous
Mechanics, Orlando, FL. wavelet transform on the free vibrations of a steel-concrete composite
Gentile, C., and Saisi, A. (2013). “Operational modal testing of historic railway bridge.” Eng. Struct., 33(3), 911–919.
structures at different levels of excitation.” Constr. Build. Mater., 48, Van der Auweraer, H., and Peeters, B. (2003). “International research proj-
1273–1285. ects on structural health monitoring: An overview.” Struct. Health
Hart, G. C. (1996). “Random damping in buildings.” J. Wind Eng. Ind. Monit., 2(4), 341–358.
Aerodyn., 59(2–3), 233–246. Yang J. H., Lam H. F., and Hu J. (2015). “Ambient vibration test, modal
Hu, W. H., Caetano, E., and Cunha, Á. (2013). “Structural health monitor- identification and structural model updating following Bayesian frame-
ing of a stress-ribbon footbridge.” Eng. Struct., 57, 578–593. work.” Int. J. Struct. Stab. Dyn., 15(7).
Ko, J. M., and Ni, Y. Q. (2005). “Technology developments in structural Zhang, F. L., and Au, S. K. (2013). “Erratum for 'Fast Bayesian FFT
health monitoring of large-scale bridges.” Eng. Struct., 27(12), method for ambient modal identification with separated modes' by Siu-
1715–1725. Kui Au.” J. Eng. Mech., 10.1061/(ASCE)EM.1943-7889.0000501,
Koh, B. H., and Dyke, S. J. (2007). “Structural health monitoring for flexi- 139(4), 545–545.
ble bridge structures using correlation and sensitivity of modal data.” Zhang, F. L., and Au, S. K. (2016). “Fundamental two-stage formulation for
Comput. Struct., 85(3–4), 117–130. Bayesian system identification, Part II: Application to ambient vibration
LabVIEW [Computer software]. National Instruments, Austin, TX. data.” Mech. Syst. Sig. Process., 66–67 (Jan.), 43–61.
Lam, H. F., Hu, Q., and Wong, M. T. (2014a). “The Bayesian methodology Zhang, F. L., Ni, Y. C., Au, S. K., and Lam, H. F. (2016). “Fast Bayesian
for the detection of railway ballast damage under a concrete sleeper.” approach for modal identification using free vibration data, Part I—Most
Eng. Struct., 81, 289–301. probable value.” Mech. Syst. Sig. Process., 70–71 (Mar.), 209–220.

© ASCE C4016002-18 J. Bridge Eng.

J. Bridge Eng., 2016, 21(8): C4016002