Journal of Sound and Vibration 428 (2018) 44e58

Contents lists available at ScienceDirect

Journal of Sound and Vibration

journal homepage: www.elsevier.com/locate/jsvi

Extraction of bridge fundamental frequency from estimated

vehicle excitation through a particle filter approach
Haoqi Wang, Tomonori Nagayama*, Junki Nakasuka, Boyu Zhao, Di Su
The University of Tokyo, Hongo 7-3-1, Bunkyo-ku, Tokyo, 113-8656, Japan

a r t i c l e i n f o a b s t r a c t

Article history: A bridge's natural frequencies are important dynamic properties reflecting the structural
Received 12 October 2017 condition of the bridge. Numerous studies have been conducted in the field to extract a
Received in revised form 22 March 2018 bridge's natural frequencies from responses of passing vehicles. The bridge frequency
Accepted 23 April 2018
peaks are, however, not easily observed, because pavement roughness often influences the
Handling Editor: Z. Su
spectra of vehicle responses. In this research, a method that extracts the fundamental
frequency of a bridge from the responses of an ordinary vehicle with its parameters
calibrated in advance is proposed. The method is based on the idea that the vehicle passing
Keywords:
Bridge frequency extraction
across a bridge is excited by two sources, i.e., pavement roughness and bridge vibration.
Indirect methods The excitation inputs to the vehicle, i.e., displacement inputs at the front and rear tire
Vehicleebridge interaction locations, are estimated from vehicle responses using a particle filter method. The esti-
Pavement roughness mated displacement inputs at the front and rear tires are then subtracted from each other
Particle filter after shifting by a wheelebase distance to eliminate the roughness influence, which
commonly appears in both signals. The signal after the subtraction contains only the
bridge vibration influence and is used to extract the fundamental frequency of the bridge.
This indirect method of bridge frequency extraction is investigated through numerical
simulations. A field measurement was also conducted, and it showed that the bridge's
fundamental frequency was successfully extracted with a good accuracy for several
driving-speed cases.
© 2018 Elsevier Ltd. All rights reserved.

1. Introduction

The demand for monitoring a bridge's structural condition has increased worldwide in recent years [1]. Among various
bridge dynamic properties, the fundamental frequency is an important indicator to reflect the actual conditions of a bridge
[2,3]. From the structural-health-monitoring point of view, it is necessary to keep monitoring the bridge's fundamental
frequency. The conventional methods, also known as direct methods, install sensors on the bridge to measure bridge re-
sponses under various loads, including ambient vibration, traffic loads, and seismic motion [4e9]. The measured bridge
responses are then used to obtain the bridge's dynamic properties of interest. However, it is usually not practical to install
sensors on a large number of bridges, because the installation procedure is always costly and time consuming with site-
specific technical difficulties [10,11].
However, the idea of indirect methods, which uses a moving vehicle as both the exciter and the receiver of the bridge
vibration, was first proposed by Yang et al. [12]. It has been investigated by many researchers in recent years for different

* Corresponding author.
E-mail address: nagayama@bridge.t.u-tokyo.ac.jp (T. Nagayama).

https://doi.org/10.1016/j.jsv.2018.04.030
0022-460X/© 2018 Elsevier Ltd. All rights reserved.
 H. Wang et al. / Journal of Sound and Vibration 428 (2018) 44e58 45

purposes [13], including bridge frequency extractions [14e18] and bridge damage detection [19e24]. However, the effec-
tiveness of these indirect methods is often influenced by pavement roughness, whose frequency components are distributed
over a wide frequency range, making the bridge vibration components unnoticeable in the vehicle responses, which may lead
to the failure of indirect methods [25]. To eliminate the influence of the bridge pavement roughness, Kong et al. proposed the
idea of “shifting and subtracting” by using sensors on two trailers after a tractor [26,27]. The recorded responses of the trailers
were subtracted from each other considering the time shift between the two signals; the bridge frequencies were then
obtained by frequency analysis of the subtracted signal. Although the proposal of this shift-and-subtraction approach has
been shown to be effective through simulation, the method has not been experimentally verified through field measure-
ments. The complexity of the tractor-and-trailer system, which limits the applicability to real cases, is preferably eliminated.
In this paper, a method to extract the bridge fundamental frequency using the responses of a single ordinary vehicle with
calibrated vehicle parameters passing over the bridge is proposed. The idea of shifting and subtracting is adopted. The method
is based on the idea that the excitation to the vehicle consists of two parts: pavement roughness and bridge vibration. The
excitation inputs to the vehicle, i.e., displacement inputs at the front and rear tire locations, are identified using a particle filter
method. One of the excitation inputs is shifted to account for the wheelebase distance and then subtracted from the other
input. Because the influence from the pavement roughness commonly appears in both input signals, the subtracted signal
contains only the bridge vibration component and is used to extract the fundamental frequency. Although the estimation of
vehicle excitation from vehicle responses has been proposed with different algorithms and investigated numerically [28e30],
these approaches, to the authors' knowledge, have not been experimentally validated; the uncertainties in the vehicle models
and drive speeds, as well as sensor setup, are among the typical difficulties in implementing vehicleebridge-interaction
(VBI)-related experiments. The feasibility of the proposed method is validated through a numerical analysis and field
measurement.

2. Vehicle and bridge models

2.1. Vehicle model

A half-car vehicle model, which has four degrees of freedom, namely, vehicle body vertical movement ub, two axle
movements uf and ur, and vehicle pitch motion q, is employed herein and shown in Fig. 1 [31].
The equation of motion describing the half-car model is expressed as

€ þ Cv UðtÞ
Mv UðtÞ _ þ Kv UðtÞ ¼ PðtÞ (1)

where Mv, Cv, and Kv are mass, damping, and stiffness matrices of the half-car model, whose detailed definitions are in
Ref. [31]. P(t) is the excitation force of the vehicle system and U(t) is the vehicle response as

T
T
U ¼ ub q uf ur ; P ¼ 0 0 hf ktf hr ktr (2)

The definition of each variable in Eq. (2) is in Fig. 1. mb, mf, mr are vehicle body mass, front tire mass, and rear tire mass,
respectively. kt and kr are the front and rear stiffness of the suspension system, respectively. ktf and ktr are front and rear tire
stiffness, respectively. cf and cr are the front and rear damping coefficients, respectively, of the suspension system. hf and hr are
the input displacements at the front and rear tires, respectively, including bridge roughness and bridge deflection. Iy is the
moment of inertia of the vehicle body.

Fig. 1. Half-car model.

46 H. Wang et al. / Journal of Sound and Vibration 428 (2018) 44e58

2.2. Bridge model

The bridge is modeled by a EulereBernoulli simply supported beam. The equation of motion of the bridge system is as
follows [32].

v2 yðx; tÞ vyðx; tÞ v4 yðx; tÞ

m þ cb þ EI ¼ Lðx; tÞ (3)
vt 2 vt vx4

where x and t are spatial and time coordinates. m is mass per length, cb is the viscous damping parameter, EI is the flexural
stiffness, y is the time- and space-variant bridge displacement response, and L (x,t) is the moving dynamic load on the bridge.
To solve this equation numerically, the bridge response is decomposed as Eq. (4).

Mb q€ þ Cb q_ þ Kb q ¼ Fb (4)

where Mb, Cb, and Kb are the diagonal modal mass, damping, and stiffness matrices, respectively, q contains the bridge
displacement responses of each mode, and Fb is the column vector containing input forces of each mode with the form of:
 
Fb;i ¼ Ff fi xf þ Fr fi ðxr Þ (5)

where fi is the ith mode of the bridge and xf and xr are the places of front and rear tires, respectively, while Ff and Fr are the
time- and space-dependent forces at the front and rear tire, respectively, and are calculated by Eq. (6).
h .  i  
Ff ¼ mb Lr Lf þ Lr þ mf g þ ktf uf  hf
h .  i (6)
Fr ¼ mb Lf Lf þ Lr þ mr g þ ktr ður  hr Þ

where g is the gravitational acceleration.

3. Bridge frequency estimation technique

3.1. Extraction of bridge vibration components

When a vehicle passes over a bridge, the pavement roughness and bridge vibration are the main sources of vertical
excitation. Through the VBI, the vehicle responses are affected by the pavement roughness as well as the bridge vibration. The
bridge dynamic properties are thus reflected in the vehicle vibration.
A two-axle vehicle passing over a bridge is considered, as shown in Fig. 2. The pavement roughness at the front and rear
tires is defined as rf (t) and rr (t), respectively, while the bridge deflection under each tire is yf (t) and yr (t). The total excitation
inputs to the vehicle are expressed as the summation of pavement roughness and bridge deflection, as shown in Eq. (7).

hf ðtÞ ¼ rf ðtÞ þ yf ðtÞ

(7)
hr ðtÞ ¼ rr ðtÞ þ yr ðtÞ

Assuming the front tire and rear tire pass on the same path, the excitation input displacement components due to the
pavement roughness at the front and rear tires are common if one signal is shifted by the wheelebase distance [26,27]. The
excitation input displacement at the rear tire is subtracted from the input displacement at the front tire, and the pavement
roughness components are canceled out as

DðtÞ ¼ hf ðtÞ  hr ðt þ tÞ
¼ rf ðtÞ þ yf ðtÞ  rr ðt þ tÞ  yr ðt þ tÞ (8)
¼ yf ðtÞ  yr ðt þ tÞ

Fig. 2. Vehicle passing over a bridge with bridge vibration.

 H. Wang et al. / Journal of Sound and Vibration 428 (2018) 44e58 47

where t is the time needed for the vehicle to travel a wheelebase distance and D(t) is the remaining bridge vibration
components, from which the dominant frequency can be extracted through Fourier analysis.
Note that this extracted frequency is the dominant frequency of the vehicleebridge coupled system. According to [33],
when a vehicle is passing across the bridge, the vehicleebridge coupled frequency is affected. This effect is more severe if the
mass of the passing vehicle becomes larger. However, because the vehicle used in this study is an ordinary vehicle, the
difference between the dominant frequency and the bridge fundamental frequency is smaller than the difference typically
observed in heavy tractor-and-trailer system cases.
Note that the half-car model parameters of the vehicle are necessary to conduct the proposed method. The vehicle pa-
rameters are obtained through a process called hump calibration, in which a hump with a known size is used as the input to
the vehicle driving over it. With the measurement of the hump-induced vehicle responses, the half-car model parameters are
calibrated through a genetic algorithm, so that the model's simulated responses are close to the measurement. Details of the
hump calibration can be found in Ref. [34]. Fig. 3 shows the flow chart of the proposed method to estimate the bridge's
fundamental frequency.

3.2. Inadequate speeds

The theoretical foundation of the proposed method is that the subtracted signal D(t) from Eq. (8) contains the bridge
frequency components. However, this foundation is not always guaranteed. There are two conditions in which the bridge
frequency component may disappear. The first condition is that, when t is a multiple of the vibration period corresponding to
the bridge fundamental frequency, the bridge vibration component at this frequency is canceled out through the subtraction
in Eq. (8). This situation is described as

dw 1
t¼ ¼ k$ ; ðk ¼ 1; 2; 3/Þ (9)
v fn

where dw is wheelebase distance, v is vehicle speed, and fn is the bridge fundamental frequency.
This equation is rewritten as Eq. (10).

dw fn
v¼ (10)
k

Fig. 3. Flow chart of the proposed method.

48 H. Wang et al. / Journal of Sound and Vibration 428 (2018) 44e58

For the given values of fn and dw, if the vehicle is driven at the speed in Eq. (10), the bridge vibration component at the
fundamental frequency is canceled out through the subtraction in Eq. (8). Thus, the method proposed in Section 3.1 cannot
give bridge fundamental frequency estimation. The speed shown in Eq. (10) is called an “inadequate speed.”
Another type of inadequate speed is explained below. Because the proposed method is based on the spectrum of bridge
vibration, the bridge's fundamental frequency can be estimated only when the bridge vibration, i.e., y (x,t) in Eq. (8), is fully
excited at the fundamental frequency. However, because the bridge is excited by the front-tire force and the rear-tire force
simultaneously, if the frequency component of Ff and Fr in Eq. (6) at the bridge's fundamental frequency are in opposite
phases, this component is minimized in the total input Fb on the bridge. The condition is given below:

dw 2k  1 1
t¼ ¼ $ ; ðk ¼ 1; 2; 3/Þ (11)
v 2 fn

which leads to

2dw fn
v¼ (12)
2k  1

This is the inadequate speed of excitation. Note that this type of inadequate speed also requires that the forces at the front
and rear tires, i.e., Ff and Fr, are close to each other. However, from Eq. (6), this requires the parameters of the vehicle front part
be close to those of the rear part, which is not always the case for normal vehicles [31].
Because of the effect of inadequate speeds, different driving speeds need to be tested when using the proposed method.
The influence of inadequate speeds on the numerical example and the field measurement is shown in Sections 5 and 6,
respectively.

4. Vehicle input estimation by particle filter

The main process of the proposed method is to estimate the front and rear excitation input displacements hf(t) and hr(t). A
data assimilation method known as a particle filter is utilized for estimating this excitation source. Although another method,
the Kalman filter, is also suitable for this problem, the particle filter method is adopted for its feasibility to consider VBI.

4.1. Particle-filtering technique

Data assimilation techniques, including the Kalman filter and particle filter, use observation data and numerical simulation
together based on a model to estimate, step by step, the system state [35,36]. These techniques require the dynamic equation
of the system to be written in state-space form.

xkþ1 ¼ fk ðxk Þ þ wðkÞ (13)

where xk is the state vector at time step k, fk is the state transition function, and w(k) is the system error with a known
distribution. The observation equation is written as

yk ¼ g k ðxk Þ þ vðkÞ (14)

where yk is the measurement vector at time step k and gk is the measurement function representing the relation between
state vector xk and measurement vector yk. The corresponding observation error is expressed as v(k) following a known
distribution, which is independent of system error w(k) and the state vector xk.
In the particle filter method, the probability density function (PDF) of the system state vector is represented by a large
number of particles in a Monte Carlo way. The objective of the particle filter is to evolve particles through the system equation
and use measurement data step by step to estimate the PDF of the state vector. This method is used here to estimate the
vehicle excitation at the front and rear tires for estimating the bridge frequency.

4.2. Implementation of particle filter

The particle filter is applied to the vehicle responses to estimate the excitation input displacement. As in Eq. (15), the
vehicle excitation terms are included in the state vector, together with the vehicle responses.
h iT
X ¼ U U_ U
€ h
f hr (15)

where X is the system vector. The two displacements hf and hr are assumed to be independent of each other. Because the
pavement roughness is considered a random process, a random-walk model is adopted [37], in which the mean value of the
 H. Wang et al. / Journal of Sound and Vibration 428 (2018) 44e58 49

prior PDF of excitation at time step k þ 1 is equal to the estimated value at time step k, assuming there is not a large difference
between two consecutive input values. This relation is expressed as

hf ;k ¼ hf ;k1 þ xf ;k
(16)
hr;k ¼ hr;k1 þ xr;k

in which x is the difference between two consecutive excitation values following normal distribution. The standard deviation
of x for the pavement roughness considered in this simulation is empirically set here to be 0.005 m. The theoretical relation
between this value and pavement roughness is not investigated in this research.
The vehicle response terms in Eq. (15) evolve through Eq. (17) while hf,k, hr,k are calculated from Eq. (16). Jk-1 is the relation
between system responses in two consecutive steps shown in Eq. (A.3) in the Appendix. w(k) is the system error aiming at
making up for the incompleteness coming from the dynamic model and the discretization for the time evolution in the
Newmark method [38].
h iT h iT 
Uk ; U_ k ; U
€
k ¼ Jk1 Uk1 ; U_ k1 ; U
€ ;h ;h
k1 f ;k r;k þ wðkÞ (17)

Eqs. (16) and (17) formulate the system equation in which particles are evolved to obtain the prior PDF of the state. The
state vector of each particle is transferred to the observation vector through the observation equation shown in Eq. (18).

Yk ¼ BXk þ vðkÞ (18)

where B is the observation matrix defined to extract the corresponding observed values from the state vector, and Yk is the
observation vector. In this case, the vehicle body displacement, vehicle angle, vehicle body acceleration, and angular velocity
are included in the observation vector, as shown in Eq. (19). The vehicle body displacement and vehicle angle are calculated by
double integration of vehicle body acceleration and angular velocity as well as by a high-pass filter. The inclusion of vehicle
body displacement and angle is based on the observability analysis [39]. Note that the method may fail if the vehicle body
displacement and angle are not included, because the system will be unobservable.
h iT
Y k ¼ ub q u€b q_ (19)

Each particle predicts a set of vehicle response through the observation equation in Eq. (18) to be compared with the
measured response. Those particles giving better prediction have a higher likelihood of being selected in the next iteration
step. The system state, including the excitation input at the front and rear tires, is then estimated.

5. Numerical example

In this section, the proposed method is numerically examined using MATLAB. A typical pavement roughness is first
generated from ISO 8608 [40]. The method to account for VBI is also explained. Vehicle responses are calculated under the
given pavement roughness input using the vehicle and bridge model described in Section 3 with the consideration of VBI from
a well-known procedure proposed by Green and Cebon [41]. The calculated response is then used as sensor-recorded values
for the estimation of roughness input. The bridge fundamental frequency is extracted following the procedures described in
Section 2.

5.1. Simulation of vehicle responses

The nonflat-pavement roughness acts as the main excitation source in the vertical direction and thus needs to be real-
istically assumed. A 60-m bridge pavement roughness of Class A is generated, as shown in Fig. 4, with a distance interval of
0.05 m, following the power spectral density (PSD) provided in ISO 8608, which expresses the roughness as the summation of

Fig. 4. A 60-m bridge profile of Class A.

50 H. Wang et al. / Journal of Sound and Vibration 428 (2018) 44e58

Table 1
Vehicle parameters [34].

Name mb (kg) Iy (m2) mf (kg) mr (kg) kf (kN/m) kr (kN/m)

Value 2293 3003 289 418 54600 70980
Name ktf (kN/m) ktr (kN/m) cf (N$s/m) cr (N$s/m) Lf (m) Lr (m)
Value 1092000 1092000 6825 6825 1.4 1.45

a series of harmonics [42]. Note that this bridge pavement roughness is taken as deterministic values in this numerical
example.
VBI occurs when the vehicle passes the bridge [25,43e45]. The vehicle is excited by pavement roughness and dynamic
force appears between the vehicle and bridge, leading to a coupled vehicleebridge system. An iteration method proposed by
Green and Cebon [41] is adopted to account for the interaction effect.
The vehicle first passes over the whole bridge, assuming the bridge pavement roughness as the excitation. Corresponding
vehicle responses are calculated. The dynamic forces on the bridge are calculated from vehicle parameters, pavement
roughness, and bridge vibration at the current iteration step. The bridge displacement response excited by the dynamic forces
at each time step is then added to the pavement roughness to formulate a new set of input displacements, and the vehicle is
simulated to pass over the bridge with the new input for another time. This iteration process stops when the difference of
bridge displacement of two consecutive iterations is smaller than a threshold value determined in advance.
A vehicle represented by the half-car model shown in Fig. 1 is simulated to pass over the bridge. The vehicle parameters are
chosen from a hump test for a real vehicle in Ref. [34] and are assumed to be known, as shown in Table 1. The total mass of the
vehicle is 3000 kg and the wheelebase is 2.45 m.
The parameters of the simply supported bridge are m ¼ 1116.9 kg/m and EI ¼ 2.08  1010 N m2. The damping ratios of each
mode are assumed to be identical and equal to 0.005. The length of bridge l is 60 m. The bridge's fundamental frequency is
1.89 Hz.
The first three modes are considered in the modal decomposition method to calculate bridge responses. For the bridge
used in this numerical example, the frequency of the fourth mode is larger than 30 Hz, and the responses are negligibly low
compared with the first three modes due to the low amplitudes of the spectrum of input, i.e., pavement roughness, for high-
frequency components [40].
The vehicle is simulated to pass over the bridge at a speed of 15 km/h, under the roughness input shown in Fig. 4. Ac-
celerometers are assumed to be installed on the vehicle body above the front and rear tires to measure vertical acceleration,
i.e., d ¼ 0 and d ¼ Lf þ Lr in Fig. 1, respectively. The angular velocity is also obtained. The vehicle responses are calculated with
the consideration of VBI and are taken as measured values from sensors. The vehicle responses are shown in Fig. 5.

5.2. Vehicle inputs and bridge frequency extraction

The estimation procedure described above is processed here twice, for the front tire and the rear tire, respectively. Note
that the acceleration data at the vehicle body above the front tire is only used for the front-tire input estimation, whereas the
rear sensor is used only for rear-tire input estimation, because the sensor near the tire is expected to give a more accurate
result [34].
Artificial noise is included in the calculated vehicle responses, shown in Fig. 5, to examine the robustness of the proposed
method against noises. The sensor noise is represented by a Gaussian white noise, whose standard deviation is set as 5% of the
root-mean-square value of the true response, considering noise and signal levels of ordinary sensors. The vehicle body
displacement and angle are calculated through integration from the noise-polluted sensor responses after a high-pass filter
with a cutoff frequency at 0.15 Hz, and they are included in the observation vector in Eq. (19). Although the displacement and

Fig. 5. Calculated vehicle responses at sensor location: (a) acceleration; (b) angular velocity.
 H. Wang et al. / Journal of Sound and Vibration 428 (2018) 44e58 51

Table 2
Artificial error of vehicle parameters [34].

Name mb mf mr Iy kf
Calibrated -0.4% þ7.2% N/Aa þ4.3% þ11.1%
Adopted -0.8% þ14.4% N/A þ8.6% þ22.2%
Name kr ktf ktr cf cr
Calibrated -9.7% -12.2% þ5.1% þ9.1% -3.5%
Adopted -19.4% -24.4% þ10.2% þ18.2% -7.0%
a
mr is obtained by mr ¼ mtotal e mb e mf, where mtotal is considered to be accurate.

angle of the vehicle body are not directly measured but are obtained from integration, the accuracy of the estimation is still
higher than the case with only acceleration and angular velocity, because the vehicle system becomes observable due to the
inclusion of vehicle body displacement and angle [34,39].
Moreover, in real measurement, the vehicle parameters usually cannot be accurately calibrated, leading to vehicle
modeling error. To obtain the vehicle parameters, a simulation of hump test is conducted, which is described in Ref. [34]. The
calibration error of the vehicle model is listed in Table 2 and is artificially added. To account for the possibility of large
calibration error in a real case, the adopted calibration errors in this numerical example are taken as twice those in Ref. [34],
which leads to a 10% decrease of vehicle fundamental frequency, from 1.13 Hz to 1.02 Hz.
The estimation result of vehicle excitation input at the front tire is shown in Fig. 6. The result without sensor noise and
vehicle calibration error is also shown for comparison. A large difference is found around midspan, where the bridge vibration
is expected to be large, as the estimated excitation inputs also contain bridge vibration components. Moreover, the estimation
results of the front and the rear tires are compared in Fig. 7, where the subtraction between the two results is also shown.
The PSD of the subtracted signal in Fig. 7 is calculated and shown in Fig. 8. A clear peak appears at 1.95 Hz, representing the
first bridge fundamental frequency with an error of 3.17%, whereas the true value is 1.89 Hz. The PSD of the estimated
excitation input at the front tire without the subtraction procedure is also shown for comparison. This PSD is significantly
influenced by the components of the pavement roughness, especially in the low-frequency range, making it hard to
distinguish the bridge's fundamental frequency. The PSD before subtraction has higher amplitudes at all frequency compo-
nents due to the roughness input. After subtraction, the main components are canceled out except for those of bridge vi-
bration, giving much smaller signal amplitudes. Therefore, this PSD is factored by 1/2 to make it comparable in Fig. 8. Large
amplitudes are observed at the low-frequency range below 1 Hz, which can be explained by the limited accuracy of the profile
estimation in this frequency range, potentially due to small signal-to-noise ratio and vehicle modeling error.

5.3. Dependency on driving speed

The above process is repeated to check the speed dependency of the proposed method. Numerical cases for driving speeds
from 10 km/h to 60 km/h with an interval of 1 km/h are conducted. The results are shown in Fig. 9 based on peak-picking in
the PSD of bridge vibration components.

Fig. 6. Estimated vehicle excitation input at the front tire.

Fig. 7. The comparison of the estimation results of the front and rear tires.
52 H. Wang et al. / Journal of Sound and Vibration 428 (2018) 44e58

Fig. 8. Spectrum of the subtracted signal.

At low vehicle speeds, satisfactory results are obtained except for the cases around 17 km/h, which is the first type of
inadequate speed of this numerical example calculated through Eq. (10). The second type of inadequate speed from Eq. (12) is
around 33 km/h. However, the second type of inadequate speed is not observed in Fig. 9, because the half-car model used in
this numerical example is not symmetric for parameters of the front and rear vehicle parts, as explained in Section 3.2. As the
vehicle speed becomes higher, a slight shift of the extracted bridge frequency is observed, as in the range between 23 km/h
and 29 km/h, and it is repeated from 30 km/h to 37 km/h. This trend coincides well with results proposed by Yang et al. [12].
Moreover, errors around 10% are observed from approximately 38 km/h. These errors are due to low resolution in the
spectrum, because the time period for vehicle passage over the bridge becomes shorter. The requirement to ensure sufficient
frequency resolution is briefly discussed here.
The frequency resolution df is the reciprocal of passing time T, if the averaging process in PSD calculation is not considered,
as shown in Eq. (20).

1
df ¼ (20)
T

Note that T ¼ l/v, where l is the bridge length, and v is the vehicle speed.
Assuming that p is the acceptable relative error of the extracted bridge frequency, the following requirement about the
frequency resolution df needs to be satisfied to guarantee that the peak in PSD falls in the acceptable range.

df < pfn (21)

Combining Eqs. (20) and (21), the speed limitation due to frequency resolution is derived as:

v < plfn (22)

From the frequency resolution point of view, it is suggested that the vehicle speed should not exceed the limit shown
above. The speed calculated through Eq. (22) of the numerical example is 40 km/h for an acceptable error of 10%. This is the
limit speed from the theoretical frequency resolution perspective. The algorithm itself, as well as the sensor noise and the
vehicle calibration error, may further decrease the acceptable driving speed.

5.4. Influence of speed fluctuation

In real cases, it is usually difficult to maintain the vehicle driving speed as a constant, even when the driver is asked to do
so. The influence of vehicle speed fluctuation is studied in this section. Three cases of linearly changing driving speed around
20 km/h are considered, as shown in Fig. 10.

Fig. 9. Dependency on driving speed of estimated bridge frequency.

 H. Wang et al. / Journal of Sound and Vibration 428 (2018) 44e58 53

The same method from Section 5.1 is processed for these cases. Because the driving speed fluctuates, the subtraction in the
time domain, as expressed in Eq. (8), cannot fully eliminate the influence from the bridge pavement roughness. Therefore,
before shifting and subtracting, the estimated input time histories at the front and rear tires need to be resampled to the
spatial domain by using the recorded instant driving speed. The PSD analysis is conducted in the spatial domain and then
transformed to the time domain by multiplying the average speed at the spatial frequency axis.
The PSD of the subtracted front and rear inputs for each case are shown in Fig. 11. A peak is observed for each speed case.
Due to the fluctuations of speed, the peak does not necessarily appear at the true value, leading to an estimation error.
However, if the subtracted signals for the all the three cases are connected to obtain a much longer time history, the PSD
analysis gives a clear peak at the bridge fundamental frequency, with a relative error of 3.86%.
The above phenomenon is explained as follows. Although fluctuations occur in each driving speed, the speeds of all
passages are within the same range. The common values in the driving speeds of different cases give a peak within a small
range on the frequency axis, while the fluctuations in each driving speed lead to different error components in the frequency
domain. For one single passage, the correct peak may be less noticeable due to the error components, but after connecting the
subtracted signals for all passages, the correct peak is amplified, giving a clear peak in the spectrum. This connection process is
more effective for field measurement analysis, as shown in Section 6, because the real driving speeds for different passages
share more common components.

6. Experimental validation of bridge frequency extraction

6.1. Overview

A field measurement was conducted at the Tsukige bridge located in Chiba prefecture, Japan. This target bridge is a simply
supported box girder bridge with a length of 59 m and a width of 4 m [10,11,46], as shown in Fig. 12. Based on the ambient
vibration measurement, the first bending natural frequency was identified as 2.16 Hz. A Toyota hi-ace van with a weight of
1850 kg and a wheelebase distance of 2.5 m was chosen as the test vehicle for the estimation of the excitation displacements
to the vehicle. iPod Touch devices equipped with accelerometers and gyros were attached to the vehicle body above the front
and rear tires to record vehicle response data, including acceleration and angular velocity. The sampling frequencies of the
sensors were set as 100 Hz. A GPS sensor was installed inside the vehicle to obtain vehicle speed data at every second to detect
possible changes of driving speed. Because the fluctuation of the sampling rate over time of the iPod touch-based sensors was
larger than that of conventional measurement equipment, resampling based on the time stamp was performed [47].
To estimate the excitation to the vehicle at the tires, the half-car model parameters of the test vehicle needed to be known
in advance. A hump test was conducted to calibrate the test vehicle using a portable hump of known geometry as the vehicle
input. The details of the hump test method are described in Ref. [34]. The identified vehicle parameters are listed in Table 3,
from which the vehicle body bouncing frequency is calculated to be 1.67 Hz.

6.2. Experimental results

The sensor-equipped test vehicle was driven to pass over the bridge. Four driving speeds were adopted, including 20, 25,
30, and 35 km/h, to test the speed dependency of the proposed method. The reason for not trying higher-speed cases is that
higher speeds give shorter passing time, which leads to low resolution in frequency analysis. For each speed case, six tests
were conducted. Although the vehicle was supposed to pass at a constant speed, some fluctuations about the driving speed
were observed. The actual speed range is shown in Table 4. Note that the inadequate speed for the chosen vehicle and bridge
was 19.44 km/h, which is close to one of the testing speeds, i.e., 20 km/h.

Fig. 10. Simulated fluctuation of driving speeds.

54 H. Wang et al. / Journal of Sound and Vibration 428 (2018) 44e58

Fig. 11. Effect of connecting multiple signals.

Fig. 12. Experimental setup. (a) target bridge; (b) test vehicle; (c) sensor on vehicle.

The excitation input displacements at the front and rear tires are estimated using the particle filter method. Fig. 13 shows a
typical estimation result at a speed of 20 km/h, after the application of a bandpass filter with the passband of 0.2e10 Hz. As
the vehicle speed fluctuated during the test, GPS speed data were utilized to convert the estimated excitation input dis-
placements in the time domain to those in the spatial domain. Due to the speed fluctuation, the sampling frequency of the
estimated excitation in the spatial domain is time-dependent. Therefore, an interpolation was conducted to obtain the signal
with a constant sampling spacing. The front and rear tire estimations coincided well. Results of other tests and speed cases
were similar to this case. The difference between the front and rear tire estimations was partly explained by the existence of
noise, including sensor noise and vehicle modeling error. More importantly, the excitation input displacements at the front
and rear tires were different from each other because the bridge vibration components were different for the front and rear
tires.
The PSD of the subtracted signal in Fig. 13 is shown in Fig. 14. Although a peak is found around the bridge fundamental
frequency, the frequency resolution is low due to limited length of the signal, leading to an inaccuracy of the result. Therefore,
as explained in Section 5.4, the estimated results of all tests for one speed case are connected together to form a longer time
history. The subtracted input displacements are connected in spatial domain after being converted from time domain using
GPS speed data. The PSD of the connected signal is then calculated. The results for all the four speed cases are shown in Fig. 15
(a)-(d).
 H. Wang et al. / Journal of Sound and Vibration 428 (2018) 44e58 55

Table 3
Identified vehicle parameters from hump calibration.

Name mb (kg) mf (kg) mr (kg) Iy (m2) kf (N/m) kr (N/m)

Value 1028.6 271.2 550.2 1676.5 80 156 109 890
Name cf (N$s/m) cr (N$s/m) ktf (N/m) ktr (N/m) Lf (m) Lf þ Lr (m)
Value 2673.8 1385.3 1725000 1809 900 0.904 2.5

Table 4
Actual speed range of each case.

Target Speed 20 km/h 25 km/h 30 km/h 35 km/h

Actual Range 15.3e20.5 km/h 21.2e23.4 km/h 24.8e29.8 km/h 28.9e34.3 km/h

Fig. 13. Comparison of front and rear tire estimated inputs.

Fig. 14. PSD of one subtracted signal.

As shown in Fig. 15, peaks around bridge fundamental frequency at 2.16 Hz are clearly observed. The estimation results and
errors are listed in Table 5. Small peaks are also observed around the vehicle frequency, and the PSD amplitudes are higher at
the low-frequency range around 1 Hz. This indicates that the influences from the pavement roughness and vehicle responses
are still not fully eliminated due to various errors and noises. However, the peak of bridge fundamental frequency is clear and
larger than other noise peaks at possible frequency range.
The results prove that the proposed method is practical for frequency extraction for real bridges. It is worth noting that,
even in the inadequate speed case of 20 km/h, a bridge fundamental frequency peak also appears clearly in the spectrum,
which is inconsistent with the simulation results in Section 5. This is because, when the vehicle is passing over the bridge, the
driving speed fluctuates around 20 km/h, making the speed not always inadequate. Thus, the bridge's fundamental frequency
component is not completely canceled out, giving a clear peak in the spectrum.

7. Conclusions

A novel indirect method of extracting bridge fundamental frequency was proposed. The excitation at the front and rear
tires of an ordinary vehicle were estimated from the vehicle responses using a particle filter approach and were then shifted
and subtracted from each other to eliminate the influence of pavement roughness. The bridge fundamental frequency peak
was clearly shown in the spectrum of the signal after the subtraction. Numerical analysis was conducted to confirm the
56 H. Wang et al. / Journal of Sound and Vibration 428 (2018) 44e58

Fig. 15. Bridge fundamental frequency estimation result for all speed cases. (a) 20 km/h; (b) 25 km/h; (c) 30 km/h; (d) 35 km/h.

Table 5
Estimation value and estimation error for all speed cases.

20 km/h 25 km/h 30 km/h 35 km/h

Estimation Value 2.13 Hz 2.17 Hz 2.19 Hz 2.23 Hz
Estimation Error 1.4% 0.5% 1.4% 3.2%

feasibility of this method. The experiment result, with the largest estimation error of 3.2%, shows that this method is practical
for a full-scale bridge. The following conclusions are drawn:

(1) The displacement inputs at the front and rear tires to the vehicle can be estimated through a particle filter in com-
bination with a random walk model to describe system state evolution.
(2) Two types of inadequate speeds of the proposed method, i.e., inadequate speeds due to subtraction and excitation,
were theoretically derived. The influence of the first type is shown in the numerical example, while the second type of
inadequate speed occurs only when the half-car model is symmetric, which is not the case for normal vehicles.
(3) In the field test, the inadequate speeds due to speed fluctuations did not influence the results.
(4) The subtraction between the displacement inputs at the front and rear tires clearly showed the peak at the bridge's
fundamental frequency.
(5) The upper limit of driving speed due to low frequency resolution was determined.
(6) If there is no driving speed fluctuation, the proposed method can be directly processed in the time domain with data
from a single passage. However, when the driving speed fluctuation becomes large, the subtraction needs to be con-
ducted in the spatial domain. Vehicle response data for more passages over the bridge are needed to make the peak at
the bridge's fundamental frequency clear. For each driving speed, six passages are recommended to obtain higher
frequency resolution and to eliminate inadequate speed effects. However, this number of passages is also subject to
bridge length, bridge frequency, and sampling rate.
(7) Large vehicle mass may affect the bridge frequency, while small vehicle mass may not fully excite the bridge vibration.
For the ordinary short-to-medium-span bridge, the vehicle with a mass of approximately 2000 kg was shown to excite
the bridge, and the bridge's natural frequency was estimated with an error of 3.2%.
 H. Wang et al. / Journal of Sound and Vibration 428 (2018) 44e58 57

Acknowledgment

The first author gratefully acknowledges the CSC scholarship (No. 201506260191) for supporting this research. This work
was partially supported by the Council for Science, Technology and Innovation, “Cross-ministerial Strategic Innovation
Promotion Program (SIP), Infrastructure Maintenance, Renovation, and Management” (funding agency: JST) and the Kajima
Foundation's Research Grant.

Appendix A. State space expression

To evolve the vehicle response terms in the particle filter, the vehicle response is also expressed in the state space. The
Newmark-beta average acceleration method is adopted through the following equations [48].
!1
€ ¼ Dt ðDtÞ2
Uk Mv þ Cv þ Kv ðPk þ Cv BC þ Kv BK Þ
2 4
€
U þU€ (A.1)
U_ k ¼ U_ k1 þ k1 k
Dt
2
U€ €
þU
Uk ¼ Uk1 þ U_ k1 Dt þ k1 k
ðDtÞ2
4
where

T
Pk ¼ 0 0 hf ;k ktf hr;k ktr
€
U
BC ¼ U_ k1  k1 Dt
2 (A.2)
€
U
BK ¼ Uk1  U_ k1 Dt  k1 ðDtÞ2
4
Dt ¼ tk  tk1

The above equations are summarized as

T
zk ¼ Uk U_ k €
U k ¼ Jk1 ðzk1 ; Pk Þ (A.3)

where Uk and Pk are vehicle response and excitation at time step k, respectively.

Appendix B. Supplementary data

Supplementary data related to this article can be found at https://doi.org/10.1016/j.jsv.2018.04.030.

References

[1] M. Abdollah, P.J. McGetrick, E.J. OBrien, A review of indirect bridge monitoring using passing vehicles, Shock Vib. (2015), https://doi.org/10.1155/2015/
286139.
[2] O.S. Salawu, Detection of structural damage through changes in frequency: a review, Eng. Struct. 19 (1997) 718e723.
[3] J.T. Kim, N. Stubbs, Crack detection in beam-type structures using frequency data, J. Sound Vib. 259 (2003) 145e160.
[4] O.S. Salawu, C. Williams, Review of full-scale dynamic testing of bridge structures, Eng. Struct. 17 (2) (1995) 113e121.
[5] A. Cunha, E. Caetano, R. Delgado, Dynamic tests on large cable-stayed bridge, J. Bridge Eng. 6 (1) (2001) 54e62.
[6] J.P. Conte, X. He, B. Moaveni, S.F. Masri, J.P. Caffrey, M. Wahbeh, F. Tasbihgoo, D.H. Whang, A. Elgamal, Dynamic testing of Alfred Zampa memorial
bridge, J. Struct. Eng. 134 (6) (2008) 1006e1015.
[7] J.M.W. Brownjohn, P. Moyo, P. Omenzetter, Y. Lu, Assessment of Highway bridge upgrading by dynamic testing and finite-element model updating, J.
Bridge Eng. 8 (3) (2003) 162e172.
[8] D.M. Siringoringo, Y. Fujino, Dynamic characteristics of a curved cable-stayed bridge identified from strong motion records, Eng. Struct. 29 (8) (2010)
2001e2017.
[9] A.W. Smyth, J.S. Pei, S.F. Masri, System identification of the Vincent Thomas suspension bridge using earthquake records, Earthq. Eng. Struct. Dynam.
32 (3) (2003) 339e367.
[10] D.M. Siringoringo, Y. Fujino, Estimating bridge fundamental frequency from vibration response of instrumented passing vehicle: analytical and
experimental study, Adv. Struct. Eng. 15 (2012) 417e434.
[11] T. Nagayama, A.P. Reksowardojo, D. Su, T. Mizutani, Bridge natural frequency estimation by extracting the common vibration component from the
responses of two vehicles, Eng. Struct. 150 (2017) 821e829.
[12] Y.B. Yang, C.W. Lin, J.D. Yau, Extracting bridge frequencies from the dynamic response of a passing vehicle, J. Sound Vib. 272 (2004) 471e493, https://
doi.org/10.1016/S0022-460X(03)00378-X.
[13] Y.B. Yang, K.C. Chang, Y.C. Li, Filtering techniques for extracting bridge frequencies from a test vehicle moving over the bridge, Eng. Struct. 48 (2013)
353e362.
[14] S.Y. Chen, H. Xia, An identification method for fundamental frequency of bridge from dynamic responses due to passing vehicle, Eng. Mech. 26 (8)
(2009) 88e94 (in Chinese).
58 H. Wang et al. / Journal of Sound and Vibration 428 (2018) 44e58

[15] C.W. Lin, Y.B. Yang, Use of a passing vehicle to scan the fundamental bridge frequencies: an experimental verification, Eng. Struct. 27 (2005)
1865e1878.
[16] H.C. Gomez, P.J. anning, M.Q. Feng, S. Lee, Testing and long-term monitoring of a curved concrete box girder bridge, Eng. Struct. 33 (2011) 2861e2869.
[17] Y.B. Yang, K.C. Chang, Extracting the bridge frequencies indirectly from a passing vehicle: parametric study, Eng. Struct. 31 (2009) 2448e2459.
[18] Y.B. Yang, K.C. Chang, Extraction of bridge frequencies from the dynamic response of a passing vehicle enhanced by the EMD technique, J. Sound Vib.
322 (2009) 718e739.
[19] J.Q. Bu, S.S. Law, X.Q. Zhu, Innovative bridge condition assessment from dynamic response of a passing vehicle, J. Eng. Mech. ASCE 132 (12) (2006)
1372e1379.
[20] P.J. McGetrick, A. Gonzalez, E.J. OBrien, Theoretical investigation of the use of a moving vehicle to identify bridge dynamic parameters, Insight 51 (8)
(2009) 433e438.
[21] K.V. Nguyen, H.T. Tran, Multi-cracks detection of a beam-like structure based on the on-vehicle vibration signal and wavelet analysis, J. Sound Vib. 329
(2010) 4455e4465.
[22] S.D. Wang, J.Q. Bu, G.C. Lou, Bridge damage identification by dynamic response of passing vehicle, J. Chang'an Univ. (Nat. Sci. Ed.) 28 (3) (2008) 63e67
(in Chinese).
[23] Z. Xiang, X. Dai, Y. Zhang, Q. Lu, The tap-scan method for damage detection of bridge structures, Interact. Multiscale Mech. 3 (2) (2010) 173e191.
[24] S.H. Yin, C.Y. Tang, Identifying cable tension loss and deck damage in a cable stayed bridge using a moving vehicle, J. Vib. Acoust. 133 (2011), 021007.
[25] X.Q. Zhu, S.S. Law, Recent developments in inverse problems of vehicle bridge interaction dynamics, J. Civ. Struct. Heal Monit. 6 (2016) 107e128.
[26] X. Kong, C.S. Cai, L. Deng, W. Zhang, Using dynamic responses of moving vehicles to extract bridge modal properties of a field bridge, J. Bridge Eng. 22
(6) (2017), 04017018.
[27] X. Kong, C.S. Cai, B. Kong, Numerically extracting bridge modal properties from dynamic responses of moving vehicles, J. Eng. Mech. 142 (6) (2016),
04016025.
[28] E.J. OBrien, C. Bowe, P. Quirke, D. Cantero, Determination of longitudinal profile of railway track using vehicle-based inertial readings, Proc. Inst. Mech.
Eng. - Part F J. Rail Rapid Transit 231 (2017) 518e534, https://doi.org/10.1177/0954409716664936.
[29] E.J. Obrien, J. Keenahan, Drive-by damage detection in bridges using the apparent profile, Struct. Contr. Health Monit. 22 (2015) 813e825, https://doi.
org/10.1002/stc.1721.
[30] E.J. OBrien, P.J. Mcgetrick, A. Gonza lez, A drive-by inspection system via vehicle moving force identification, Smart Struct. Syst. 13 (2014) 821e848,
https://doi.org/10.12989/sss.2014.13.5.821.
[31] R.N. Jazar, Vehicle Dynamics: Theory and Applications, third ed., Springer International Publishing AG, Switzerland, 2008.
[32] A.K. Chopra, Dynamic of Structures: Theory and Applications to Earthquake Engineering, second ed., Prentice-Hall, 2007.
[33] D. Cantero, D. Hester, J. Brownjohn, Evolution of bridge frequencies and modes of vibration during truck passage, Eng. Struct. 152 (2017) 452e464,
https://doi.org/10.1016/j.engstruct.2017.09.039.
[34] B. Zhao, T. Nagayama, IRI Estimation by the frequency domain analysis of vehicle dynamic responses, Procedia Eng. 188 (2017) 9e16, https://doi.org/
10.1016/j.proeng.2017.04.451.
[35] M.S. Arulampalam, S. Maskell, N. Gordon, T. Clapp, A tutorial on particle filters for online nonlinear/non-Gaussian Bayesian tracking, IEEE Trans. Signal
Process. 50 (2002) 174e188.
[36] J. Carpenter, P. Clifford, P. Fearnhead, Improved particle filter for nonlinear problems, IEE Proc. - Radar, Sonar Navig. 146 (2) (1999).
[37] M. Doumiati, A. Charara, A. Victorino, D. Lechner, B. Dubuisson, Vehicle Dynamics Estimation Using Kalman Filtering: Experimental Validation, 2012,
https://doi.org/10.1002/9781118578988.
[38] K. Matsuoka, K. Kaito, M. Tokunaga, T. Watanabe, M. Sogabe, Estimation of bridge deflection response under passing train loads based on acceleration,
Struct. Eng. Earthq. Eng. 69 (2013) 527e542 (in Japanese).
[39] M.N. Chatzis, E.N. Chatzi, A.W. Smyth, On the observability and identifiability of nonlinear structural and mechanical systems, Struct. Contr. HLMT 22
(2015) 574e593.
[40] ISO-8608 1995, Mechanical Vibration-Road Surface Profiles-Reporting of Measured Data, 1995.
[41] M.F. Green, D. Cebon, Dynamic interaction between heavy vehicles and highway bridges, Comput. Struct. 62 (1997) 253e264.
[42] M. Agostinacchio, D. Ciampa, S. Olita, The vibrations induced by surface irregularities in road pavements - a Matlab® approach, Eur. Trans. Res. Rev. 6
(2014) 267e275.
[43] Y.B. Yang, J.D. Yau, Vehicle-bridge interaction element for dynamic analysis, J. Struct. Eng. ASCE 123 (1997) 1512e1518, https://doi.org/10.1061/(ASCE)
0733-9445(1997)123:11(1512).
[44] J. Kim, J.P. Lynch, Experimental analysis of vehicleebridge interaction using a wireless monitoring system and a two-stage system identification
technique, Mech. Syst. Signal Process. 28 (2012) 3e19.
[45] F. Cerda, J. Garrett, J. Bielak, R. Bhagavatula, J. Kovacevic, Exploring indirect vehicle-bridge interaction for bridge SHM, in: 5th International IABMAS
Conference, 2010.
[46] H. Wang, T. Nagayama, B. Zhao, D. Su, Identification of moving vehicle parameters using bridge responses and estimated bridge pavement roughness,
Eng. Struct. 153 (2017) 57e70, https://doi.org/10.1016/j.engstruct.2017.10.006.
[47] T. Nagayama, B. Spencer, K. Mechitov, G. Agha, Middleware services for structural health monitoring using smart sensors, Smart Struct. Syst. 5 (2009)
119e137.
[48] Y.W. Kwon, H. Bang, The Finite Element Method Using MATLAB, second ed., CRC Press, 2000.