Download as pdf or txt
Download as pdf or txt
You are on page 1of 26

Journal of Sound and Vibration 457 (2019) 92e117

Contents lists available at ScienceDirect

Journal of Sound and Vibration


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

Road profile estimation, and its numerical and experimental


validation, by smartphone measurement of the dynamic
responses of an ordinary vehicle
Boyu Zhao, Tomonori Nagayama*, Kai Xue
Dept. of Civil Engineering, The University of Tokyo, Hongo-Bunkyo, Tokyo, Japan

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

Article history: The pavement of a road that is subject to deterioration due to vehicle loads needs quan-
Received 25 December 2018 titative and frequent evaluation. The road profile, which determines the ride quality, is an
Received in revised form 21 April 2019 important property, though its estimation is usually costly and thus, infrequent. In this
Accepted 9 May 2019
study, a road profile estimation method using an ordinary-vehicle's responses measured
Available online 15 May 2019
Handling Editor: J. Cheer
by only a smartphone is developed. The algorithm consists of two steps. At first, an or-
dinary vehicle is modelled as a half car (HC) and a genetic algorithm (GA) identifies its
parameters by using the responses of the vehicle passing over a known-size hump. With
Keywords:
Road profile estimation
the estimated vehicle model, an augmented Kalman filter, in which the road profile is
Vehicle model identification included in the state vector, estimates the road profile; Rauch-Tung-Streiber (RTS)
Vehicle dynamics smoothing is employed to improve the accuracy. The observation variables and locations
Kalman filter are determined based on an observability analysis. A numerical simulation is conducted to
Genetic algorithm investigate the profile estimation performance in terms of drive speeds, model error, and
measurement noise. The experiment is carried out on a 13 km road. The profiles estimated
by three types of ordinary vehicles are compared with a reference profile obtained by a
laser profiler to validate the proposed method. Results from both the simulation and the
experiment show that the combination of the vehicle parameter estimation and the profile
estimation methods accurately estimates the road profile with a high degree of accuracy
and robustness.
© 2019 Elsevier Ltd. All rights reserved.

1. Introduction

The quantitative evaluation of a road surface is important in order to maintain the road effectively. The condition of a
pavement influences not only the safety of drivers, but also the comfort level of both drivers and the surrounding envi-
ronment. Without appropriate maintenance, a damaged pavement can further deteriorate; large cracks and potholes can
even cause accidents as well as poor drive comfort and loud noise emissions.
Practical road condition assessments widely employed by road owners can be categorized into two groups: those
employing visual inspection, and those using precise measurement techniques by employing lasers and other sensors. Visual
inspection can be performed without using expensive instruments though the quality of evaluation relies on the skill of the

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

https://doi.org/10.1016/j.jsv.2019.05.015
0022-460X/© 2019 Elsevier Ltd. All rights reserved.
B. Zhao et al. / Journal of Sound and Vibration 457 (2019) 92e117 93

inspectors; the evaluation is subjective. A precise profile measurement typically employs laser and inertia sensors to estimate
the profile at regular intervals [1,2]. The laser-based method provides profile estimation with a high accuracy and resolution.
However, especially with ordinary road networks, their frequent evaluation through the use of the laser-based method is
impractical because of its high cost.
A response-based road roughness condition evaluation that is low cost and also provides an objective assessment in terms
of some indices has been widely studied. Gonza lez et al. [3] propose a power spectrum density (PSD) method for the esti-
mation of the roughness of a road surface, based on the transfer function from the road profile to the vertical acceleration of
the vehicle body. The performance is validated by several simulations using different road roughness values generated by ISO
classes [4]. DRIMS [5e7] evaluates the International Roughness Index (IRI) [8], a ride comfort index defined as the accu-
mulated relative displacement of the suspension spring of the Golden-Car, using an accelerometer installed above a vehicle
axle. A measurement vehicle is modelled as a Quarter-Car (QC); the ratio of the Golden-Car response power spectrum to the
QC response power spectrum is estimated in advance and multiplied by the measured responses. The responses of the
Golden-Car on the measurement target road are thus calculated. IRI is then estimated based on a correlation analysis. To
account for both pitching and bouncing motions, DRIMS has been further improved by using a half-car (HC) model taking into
account the measurements of both vertical acceleration and angular velocity of the vehicle body using a smartphone [9e11].
The IRI estimation was performed in the frequency domain utilizing the ratio of the Golden-Car response power spectrum to
the HC model response power spectrum. Other smartphone implementations to estimate IRI have also been reported [12].
These response-based methods evaluate the road condition in terms of specific indices or classification. However, accurate
road profile estimation remains challenging.
There are some response-based techniques to estimate road profiles. Road roughness reconstructions based on artificial
neural networks have been developed [13,14]. A parametric adaptive observer based on the YK parameterization (Q-
parameterization) has been used to estimate the road profile [15,16]. Sliding mode observers were also employed for road
profile estimation [17]. Combinatorial optimization approaches have been applied to road profile estimation problems [18].
There have been similar applications to railway profile estimation as well [19,20]. A stochastic method was found to have
improved its computational efficiency by using Kalman filters [21,22]; the performance of this estimation technique was
validated by using sensor-equipped vehicles with known dynamic properties. These methods can estimate road profiles with
various levels of accuracy, complexity, and computational cost.
However, to date, those methods with experimental validations require the dynamic properties of test vehicles to be
known in advance from the vehicle manufacturer, or via laboratory tests and/or suspension motions to be directly measured,
which are impractical with smartphone instrumentation. These limitations make these methods impractical to implement on
a large scale using a variety of vehicles for frequent and quantitative road condition assessment. Road profile estimation
techniques measuring only the vehicle body motion, on the other hand, typically employ simplified models (e.g., single DOF
model and Quarter-Car model) and their calibrations [12,23]; the accurate profile estimation capabilities including the cor-
rections in the drive speed dependency and sensor location dependency are thus limited. As far as the authors are aware,
there have been no accurate profile estimation methods capable of correcting the differences due to vehicle characteristics,
drive speeds, and sensor installation locations purely by using of a smartphone or another single sensor installed on the body
of a vehicle with unknown dynamic characteristics.
To address these issues, a response-based road profile estimation using multiple outputs measured by a smartphone
installed on ordinary vehicles is proposed in this study. The algorithm consists of two steps. Initially, the measurement vehicle
is modelled as a HC; its parameters are identified by a genetic algorithm (GA) when the vehicle drives over a hump of a known
size. While the concept of calibrating vehicles using humps has been utilized in numerical simulations in the past for the HC
roll model [18] and implemented in practice using QC [7], a practical implementation has to address issues such as the
synchronization between the hump input and the vehicle response. In particular for the pitch HC model, the drive speed
uncertainties and a large number of unknown parameters make the calibration difficult. These issues are addressed by using a
GA with the response power spectrum as the objective function, which is robust against the synchronization error. Next, with

Fig. 1. Flow chart of the proposed method.


94 B. Zhao et al. / Journal of Sound and Vibration 457 (2019) 92e117

the calibrated vehicle model, an augmented Kalman filter is designed to estimate the road profile. The road profile is included
in the form of augmented state variables and estimated through Kalman filtering. RTS smoothing, a computationally efficient
smoothing algorithm, is employed to improve the estimation accuracy. Note that none of the QC, HC, or full car models can
exactly reproduce the dynamic behavior of real vehicles. Stochastic methods capable of dealing with modeling errors, such as
the Kalman filter, are thus advantageous. As far as the authors are aware, this is the first study to propose an augmented
Kalman filter satisfying the observability condition of the pitch HC model using no more than a smartphone on the vehicle
body, together with RTS smoothing. While considering a large-scale implementation of the proposed algorithm on a variety of
commercial vehicles equipped with smartphones [11,24], the capability to calibrate a variety of vehicles in a simple way and
the ability to estimate profiles at a computationally inexpensive cost are both important. These two steps are validated both
numerically and experimentally in this research work. The flow chart of the proposed method is shown in Fig. 1.

2. Road profile estimation

2.1. Vehicle modeling

The measurement vehicle is modelled as a HC (Fig. 2), which can reproduce the bouncing, pitching, and axle modes of
vehicle motion and therefore is able to account for the sensor installation location along the longitudinal direction. The
parameters of the HC are defined in Table 1. Parameters of the HC model are identified when a vehicle passes over a known-
size hump at a specified drive speed (Fig. 3). The vertical acceleration and pitching angular velocity of the vehicle body are
measured.
The dynamic equation of the HC model is shown in Eq. (1):

Mx€ðtÞ þ CxðtÞ
_ þ KxðtÞ ¼ PyðtÞ (1)


where M, C and K are the system mass, damping, and stiffness matrices; x(t), _
x(t) and x(t) are the acceleration, velocity, and
displacement vectors; P is the matrix prescribing the force; yðtÞ represents the profile as a function of time. These matrices
and vectors are given by
2 3
mH 0 0 0
6 0 Iz 0 0 7
6
M¼4 7; (2)
0 0 mf 0 5
0 0 0 mr
2 3
cf þ cr Lr cr  Lf cf cf cr
6 Lr cr  L c L2f cf þ L2r cr Lf c f Lr cr 7
C¼6
4
f f 7; (3)
c f Lf cf cf 0 5
cr Lr cr 0 cr

Fig. 2. HC model.

Table 1
HC model parameters.

mH mf mr cf cr kf
Vehicle body mass Front tire Rear tire Front suspension Rear suspension Front suspension stiffness
mass mass damping damping
kr ktf ktr Iz Lf Lr
Rear suspension Front tire Rear tire Pitching moment of Distance from the center of gravity to Distance from the center of gravity to
stiffness stiffness stiffness inertial the front axle the rear axle
B. Zhao et al. / Journal of Sound and Vibration 457 (2019) 92e117 95

Fig. 3. Hump.

2 3
kf þ kr Lr kr  Lf kf kf kr
6 L r kr  L k L2f kf þ L2r kr Lf kf Lr kr 7
K¼6
4
f f 7; (4)
kf Lf kf kf þ ktf 0 5
kr Lr kr 0 kr þ ktr
h iT
xðtÞ ¼ x q xf xr ; (5)

 T
0 0 ktf 0
P¼ ; (6)
0 0 0 ktr
 T
yðtÞ ¼ yf yr : (7)

HC model parameters normalized by the total mass are shown in Eq. (8) together with the drive speed V. The drive speed V
is also an unknown parameter because drivers usually cannot keep a target drive speed while driving over the hump. Each
vehicle parameter is explained in Table 1.
  
Gv ¼ kf kr cf cr ktf ktr Lf mH mf Iz mtotal ; V (8)

Unknown parameters including 10 vehicle mechanical parameters and drive speed are estimated through the minimi-
zation of the difference between the frequency spectra of the measured and simulated vehicle responses. A genetic algorithm
(GA) [25] optimization method is adopted. The target function to minimize is based on the vehicle body acceleration and
angular velocity responses measured by a smartphone as shown in Eq. (9).
vffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi ffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi
vZ
uZ n o u n o
u b u b
u S acc ðf Þ  Sacc ðf Þ df u S angv ðf Þ  Sangv ðf Þ df
u u
C ¼u Z þu Z (9)
t t
Sacc ðf Þdf Sangv ðf Þdf

where bS acc ðf Þ and Sacc ðf Þ are the power spectra of the acceleration responses of the HC vehicle model and those obtained by
actual measurement, respectively. Similarly, b S angv ðf Þ and Sangv ðf Þ are the power spectra of angular velocity responses of the
model and those captured by the measurement. In addition, the reasonable natural frequency ranges of the HC model are
specified as constraint conditions. Individuals in the GA whose natural frequencies are out of reasonable ranges are eliminated
(i.e., the 1st and 2nd natural frequencies need to be smaller than 2 Hz; the 3rd and 4th frequencies need to be between 5 and
15 Hz).
Note that the target function employed in the GA is defined in the frequency domain without the phase information.
Optimization with phase or time information in addition to the spectral amplitude information is considered to provide better
accuracy. For example, target functions of the GA optimization can be defined in the time domain. Further, time domain
methods, such as the unscented Kalman filter [26] and particle filter [27], can be employed for parameter identification
instead of GA. However, the time domain parameter identification is practically difficult to implement; the measured vehicle
responses and the model input, that is the hump profile, cannot be accurately synchronized because of an imperfect estimate
of the vehicle position and speed [11]. This synchronization error leads to unreliable and inaccurate identifications, thus
decreasing the robustness of the system. The GA optimization with the power spectrum optimization function is, therefore,
employed. Furthermore, in order to improve the robustness and prevent local convergence, each vehicle repeats the drive
over the hump five times and the GA is conducted three times for each drive test.
96 B. Zhao et al. / Journal of Sound and Vibration 457 (2019) 92e117

2.2. Profile estimation using Kalman filter

The profile estimation is formulated based on a Kalman filter with an augmented state-space model, where the road
profile is represented as state variables [21,22]. The formulation is based on the pitch HC model with only the vehicle body
motion measurement. The state-space model describing the vehicle response to the profile is formulated as in Eq. (10). The
measurement equation is shown by Eq. (11).

X_ c ðtÞ ¼ Ac Xc ðtÞ þ Bc yðtÞ þ wðtÞ (10)

uðtÞ ¼ Cc Xc ðtÞ þ Dc yðtÞ þ vðtÞ (11)

X c ðtÞ ¼ ½ xðtÞ _ 
xðtÞ (12)

 
044 I44
Ac ¼ ; (13)
M1 K M1 C

 
042
Bc ¼ ; (14)
M1 P

where Xc ðtÞ is the state vector consisting of the vehicle responses, u(t) is the measurement vector, and Ac, Bc, Cc, and Dc are the
discrete system matrices. Cc, and Dc depend on the measurement; their components corresponding to specific measurements
are given later. w(t) and v(t) are system noise and measurement noise, respectively. These noises are zero-mean Gaussian
white-noises.
The state vector is then augmented with the profile. The augmented state vector is described by Xa ðtÞ, where the su-
perscript ‘a’ means “augmented”.
h i
Xa ðtÞ ¼ x q xf xr x_ q_ x_f x_r yf yr (15)

The profile at the front tire yf and the rear tire yr are included as the augmented states. These profiles are related to each
other; when one profile is shifted by a time duration corresponding to the wheelbase distance, the other profile is obtained.
However, in this study, these profiles are assumed to be independent of each other because the time duration used in the shift
depends on the drive speed estimate. The drive speed estimates through GPS are not always accurate and have a much lower
sampling frequency than acceleration and angular velocity. The inaccuracy and low sampling rate may result in large error in
the time shift duration estimate and they thus influence the accuracy of the profile estimation.
The augmented state equation is defined as

X_a ðtÞ ¼ Aca Xa ðtÞ þ zðtÞ (16)

uðtÞ ¼ Cca Xa ðtÞ þ vðtÞ (17)

where Aca, Cca and z (t) are as follows:


 
Ac Bc
Aca ¼ (18)
028 022

Cca ¼ ½ Cc Dc  (19)
 
wðtÞ
zðtÞ ¼ (20)
hðtÞ

The time evolution of the input profile is prescribed by a zero-mean Gaussian white-noise h(t).

_ ¼ hðtÞ
yðtÞ (21)

The formulation in the continuous time domain is then converted to the formulation in the discrete time domain as
B. Zhao et al. / Journal of Sound and Vibration 457 (2019) 92e117 97

Xakþ1 ¼ Aa Xak þ zk (22)

uk ¼ Ca Xak þ vk (23)

where dt is the temporal sampling interval and k (0 < k < N-1) represents the time step. The system noise of this discrete
formulation is composed of wk and hk corresponding to wðtÞ and hðtÞ.
 
wk
zk ¼ (24)
hk

The measurement noise vk corresponds to vðtÞ. The continuous and discrete representations have the following
relationship.

Aa ¼ eAca dt (25)

Ca ¼ Cca (26)

Xak a
¼ X ðkdtÞ (27)

uk ¼ uðkdtÞ (28)

Note that the time evolution of the input profile in the discrete time domain is also represented by a zero-mean Gaussian
white-noise, hk , which can be expressed as hðkdtÞdt.

ykþ1 ¼ yk þ hk (29)

The covariance matrices of wk, vk and hk are defined as


h i
E wk wTl ¼ Q dk;l (30)

h i
E vk vTl ¼ Rdk;l (31)

h i
E hk hTl ¼ Sdk;l (32)

where dk,l is the Kronecker delta.


The Kalman filter then recursively estimates the state vector [28]. The posteriori state estimate X ca , and the posteriori error
k
covariance matrix, Pk , are initialized at the time step k ¼ 0 as in Eq. (33) and Eq. (34). The symbol^represents the estimate of
state.

ca ¼ EX a 
X (33)
0 0
h i
T
P0 ¼ E ðX a 0  E½X a 0 ÞðX a 0  E½X a 0 Þ (34)

ca  , is calculated as
The priori estimate of the state, X kþ1

ca  ¼ A X
X ca : (35)
kþ1 a k

The priori error covariance, P


kþ1 at the time step kþ1 is given by

P T
kþ1 ¼ Aa Pk A a þ Q k ; (36)

where Q k is the system noise covariance matrix at the time step k. The Kalman gain at the time step kþ1, Gkþ1 , is obtained as
98 B. Zhao et al. / Journal of Sound and Vibration 457 (2019) 92e117

h i1
Gkþ1 ¼ P T  T
kþ1 Ca Ca Pkþ1 Ca þ R kþ1 ; (37)

where Rkþ1 is the measurement noise covariance matrix at the time step kþ1. The posteriori state estimate at the time step
kþ1 is updated as


ca ca  ca 
X kþ1 ¼ X kþ1 þ Gkþ1 ukþ1  Ca X kþ1 (38)

The posteriori error covariance at the time step kþ1 is calculated as

Pkþ1 ¼ ðI  Gkþ1 Ca ÞP


kþ1 (39)

Because the inclusion of derivative terms of the profile in the state vector is found to result in higher accuracy estimation
h iT
than estimation without these terms, yf y_f yr y_r is taken as the new profile vector yðtÞ and the following augmented
state vector is employed.
h iT
Xa ðtÞ ¼ x q xf xr x_ q_ x_f x_r yf y_f yr y_r (40)

The corresponding system matrix is as follows.


 
Ac Bc 0
Aca ¼ (41)
048 T44
 
044
Bc 0 ¼ (42)
P0
2 3
0 0 0 0
6 0 0 0 07
P0 ¼ 6
4 ktf =mf
7 (43)
0 0 05
0 0 ktr =mr 0
2 3
0 1 0 0
60 0 0 07
T¼4 6 7 (44)
0 0 0 15
0 0 0 0

The measurement matrix Ca, is determined based on the measured variables. In this study, a smartphone is installed on an
ordinary vehicle body to measure its acceleration and angular velocity responses. In addition to vertical acceleration and
pitching angular velocity, which are directly measured, the vertical displacement and pitching angle that are obtained by
numerical integrations, are considered as possible measurement variables. Integration errors associated with measurement
noise are suppressed by high-pass filters. As explained in a later section, the inclusion of the displacement and angle are
necessary. While the inclusion of integrated quantities in the observation to satisfy the observability has also been utilized in
the past [21,22], this study takes advantage of the inclusion to achieve the observability by only using a single smartphone.
With these four observation variables of vertical acceleration, angular velocity, vertical displacement and pitching angle, the
matrix can be written as
 T
C1
Ca ¼ ; (45)
O44
B. Zhao et al. / Journal of Sound and Vibration 457 (2019) 92e117 99

2

3
kf þ kr Lf  d Lr kr  Lf kf
6 7
6  þ 0 1 07
6 mH IZ 7
6

7
6 7
6 Lf  d L2f kf þ L2r kr 7
6 L r kr  L f kf 7
6 þ 0 d  Lf 17
6 mH IZ 7
6 7
6
7
6 Lf  d Lf kf 7
6 kf 7
6 þ 0 0 07
6 mH I 7
6 7
6
7
6 7
6 kr L f  d L r kr 7
6 7
6  0 0 07
6 mH IZ 7
C1 ¼ 6
6

7
7 (46)
6 Lf  d Lr c r  Lf c f 7
6 cf þ cr 7
6  þ 0 0 07
6 mH IZ 7
6 7
6

7
6 Lf  d L2f cf  L2r cr 7
6 Lr cr  Lf cf 7
6 þ 1 0 07
6 mH IZ 7
6 7
6
7
6 7
6
6 cf Lf  d Lf c f 7
7
6 þ 0 0 07
6 mH IZ 7
6
7
6 7
6 Lf  d Lr cr 7
4 cr 5
 0 0 0
mH IZ

In this Kalman filter process, the profiles yf ðtÞ and yr ðtÞ are both obtained as functions of time as a part of the state vector.
The profiles are then converted to the spatial functions using the relationship between the time and the distance. Both the
process and measurement noise are tuning parameters. The measurement noise is primarily the sensor noise and the noise
from vehicle engine vibration. The noise levels are experimentally determined through the smartphone static measurement
test and the vibration measurement on a stationary vehicle while its engine is idling. As for the process noise, the covariances
are set by reference to the rate of the change of state variables in the same manner as in other Kalman filter applications. An
adaptive update of the process and measurement noise covariance matrices is under investigation.

2.3. RTS smoothing for profile estimation

The augmented Kalman filter addressed the optimum linear filtering problem in a straightforward manner to obtain the
ca  is the estimate at time k based on all the mea-
optimal priori and posteriori states estimates. The priori state estimate X k

surements up to time k-1. The posterior state X ca is the estimate at time k based on all the measurements up to time k. In this
k
section, a fixed-interval smoothing algorithm, Rauch, Tung, and Striebel (RTS) smoothing [29], is introduced to improve the
state estimation. The RTS smoothing filters backward and re-estimates the states at time k using the measurements up to L (k
< L  N). Though the smoothing algorithm does not provide real time estimation, the accuracy of the profile estimation is
more of interest in this study. Further, the computational load of the smoothing is similar to that of the Kalman filter; even
with the smoothing, processing of data collected from many vehicles is still practical.
RTS smoothing proceeds as follows. At the first step, the forward filter executes the standard Kalman filter to the observed
data for k ¼ 0,1,2 … N-1. At the second step, the backward RTS smoother initializes the estimation-error covariance matrix PN b
d
and state vector of RTS smoother X RTS N as

PbN ¼ PN (47)

d ca
X RTS N ¼ X N (48)

where N is the last time step. The smoother then propagates backward from k ¼ N as
h i1
Fk ¼ Pk Aa P
kþ1 ; (49)
100 B. Zhao et al. / Journal of Sound and Vibration 457 (2019) 92e117



Pbk ¼ Pk  Fk P
kþ1  Pkþ1 Fk ;
b T
(50)



xd ca þ F xd
RTS ¼ X ca 
kþ1  x kþ1 ;
RTS (51)
k k k

where Fk is the gain of the RTS smoother.

2.4. Observability analysis of the profile estimation

The capability of the profile estimation through the Kalman filter is first investigated through an observability analysis. A
system is observable if the current state can be determined in finite time using only the outputs [21,22,30]. If a system is not
observable, the current values of some of its states cannot be determined through output measurements. A continuous time-
invariant linear state space model with W-dimension states is observable if and only if the rank of the observability matrix O
is equal to W. The observability matrix is written as
h iT
O ¼ CTa ATa C Ta ATa 2 C Ta / ATa W1 T
Ca : (52)

This observability analysis is performed on the pitch HC model. Taking into consideration that vehicle response mea-
surement is carried out using smartphones that are easily installed, only sensor locations on the body are investigated;
smartphone installation on the tire is impractical and excluded. Several measurement combinations related to vehicle body
motion are chosen for the observability analysis and the results are shown in Table 2. More observability analysis results of
different measurement combinations are reported in Ref. [31].
As shown in Table 2, measurements of both the vertical displacement and the pitching angle are needed to satisfy the
observability condition. The vertical acceleration and angular velocity are also included in the outputs as this inclusion made
the system observable and resulted in more numerically accurate estimation, particularly for high frequency component
estimation. Note that the displacement and angle are obtained through numerical integration with high-pass filters. Though
the observability requirement is satisfied with the measurement combination, the low frequency components of the profile
are not obtained. The cut-off frequency of the high-pass filter is set considering the lowest vehicle natural frequency and
profile wavelengths of interest.

3. Numerical validation

The performance of hump calibration and profile estimation are numerically examined herein. The measurement vehicle
in this simulation is a pitch HC model with the natural frequencies of 1.11 Hz, 1.46 Hz, 8.43 Hz, and 10.02 Hz. These frequencies
correspond to the bouncing, pitching, front axle hop, and rear axle-hop modes. Sensors are located at 1.8 m from the front
axle. The parameters are shown in Table 3. Note that the dynamic characteristics of real vehicles are more complicated than
those of the HC model. However, even for the HC model, the performance of the calibration and profile estimation are not
clear. The identification of 10 unknown parameters as well as the drive speed is challenging. There is no guarantee that the
true parameters are obtained for the HC model. The HC model with inaccurately identified parameters does not necessarily
estimate the profile accurately. These issues are numerically investigated. The performance of the calibration and profile
estimation taking into account more complex vehicles as well as measurement and input noises are discussed later in
experimental validation.

Table 2
Observability analysis.

Measurement combination Vertical acceleration Vertical velocity & Vertical displacement Vertical displacement Angle
& angular velocity Angular velocity & Angle
Observability No No Yes No No

Table 3
HC model parameter values.

mH (kg) mf (kg) mr (kg) cf¼ cr(N s/m) kf (N/m) kr(N/m) ktf ¼ ktr(N/m) Iz(kg m2) Lf (m) L(m)
840 106 152 2500 20000 26000 400000 1100 1.4 2.87
B. Zhao et al. / Journal of Sound and Vibration 457 (2019) 92e117 101

3.1. Hump calibration performance in simulation

Hump calibration performance is numerically examined under an ideal condition. First, the vertical acceleration and
angular velocity of the vehicle passing over a hump at 10 km/h is calculated by using the HC model. The hump is 30 cm wide
and 5 cm high. The road profile consists of the hump and the flat surface, which are sampled at intervals of 1 cm. The
simulated responses are resampled at 100 Hz to produce the measurement signal. The parameter ranges and other config-
urations of the GA optimization are shown in Table 4 and Table 5. The sensor noise is zero and the road surfaces, except for the
hump, are perfectly flat. The frequency range of the optimization function is from 0 to 10 Hz. As shown in Fig. 4, the responses

Table 4
Range of unknown parameters.

Parameters Initial range set in GA


kf,kr,cf,cr,ktf,ktr,Iz 0.3e3 times of true value
mH, mf 0.5e2 times of true value
Lf 0.7e1.3 times of true value
V 6e15 km/h

Table 5
GA configuration.

Generation 150
Population size 200
Mutation 1%
Crossover rate 90%
Length of bit 20

Fig. 4. GA optimization results. (a) acceleration time history, (b) angular velocity time history, (c) acceleration power spectrum, and (d) angular velocity power
spectrum.
102 B. Zhao et al. / Journal of Sound and Vibration 457 (2019) 92e117

Table 6
Relative error of estimated parameters.

mH mf mr cf cr kf kr ktf ktr Iz Lf V
(kg) (kg) (kg) (N s/m) (N s/m) (N/m) (N/m) (N/m) (N/m) (kg m2) (m) (km/h)
Estimation 833 114 151 368 391 21900 21660 346800 416000 1168 1.48 10.03
Relative error (%) 0.8 6.9 0.6 8.1 2.2 9.5 8.3 13.3 4.0 6.2 5.9 0.33

of the hump drive at 10 km/h are predominantly below 10 Hz. This is the case for the experimental data too, as shown later.
Taking into consideration that the noise can be larger than the signal in the high frequency range, and the repeatability of
axle-hop modes is not high, the optimization frequency range is determined.
In order to prevent local convergence, the GA process is repeated three times and the best parameter set with the smallest
value of the GA target function is selected. The relative errors between the estimated values and the true values of vehicle
physical parameters are shown in Table 6. Most parameters show errors smaller than 10%. The comparison of power spectra
and time history from the identified model and those obtained by actual measurement are shown in Fig. 4.
The drive speed of 10 km/h is chosen through numerical simulations with different drive speeds, experiments, and
sensitivity analysis that will be explained later. Higher drive speed was avoided because a fast suspension motion may results
in significant nonlinearity in the actual vehicle responses, which is not included in the model; the nonlinearity is assumed to
not be significant during the 10 km/h hump test and ordinary road drives. Further, a slow drive speed allows the response due
the passage of the front wheel over the hump and the response due to the passage of the rear wheel to be separated in the
temporal domain, which makes the interpretation of the calibration responses simple. On the other hand, slower drive speeds
give rise to the practical difficulty of keeping constant drive speeds during the hump passage.

3.2. Road profile estimation

3.2.1. Profile generation


The road profile is generated based on ISO 8608 [4]. The profile can be expressed in terms of the sum of a series of har-
monics [32] as shown in Eq. (53)

D qffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi
X
hðxÞ ¼ 2Gðni ÞDn cosð2pni x þ fi Þ (53)
i

in which x is the distance along the road, h(x) is the road profile, ni is the i-th spatial frequency component, Dn is the frequency
spacing determined by the total length of the profile, and fi is the phase angle which follows a uniform distribution from 0 to
2p. G (ni) is the one-sided PSD defined in Eq. (54) where G (n0) is chosen according to the different road classes defined in ISO
8608. D is the number of harmonics.
2
n
Gðni Þ ¼ Gðn0 Þ i (54)
n0

3.2.2. Profile estimation


Two cases of simulation are conducted. In Case 1, the profile is estimated without the influence of noise and without model
errors. In Case 2, both noise and model errors (Table 6) are included.
The “measured responses” of vertical acceleration and angular velocity are generated by using the HC model shown in
Table 3. A 500 m long pavement with a roughness generated by Eq. (53) and Eq. (54), is used as the excitation. n0 is 0.1 cycle/m.
The spatial interval of the pavement roughness is 0.05 m, resulting in a sampling frequency of 20 cycle/m, and a D value of
10,000 is used. The sensor location is 1.8 m from the front axle. The profile in the spatial domain is transferred to the temporal
domain using the drive speed. The profile in the temporal domain is used as the input to calculate the vehicle responses,
which is then resampled at 100 Hz to obtain the “measured responses”. Finally, the estimated profile in the temporal domain
at 100 Hz is transferred back to the spatial domain; the spatial sampling frequency is 20 cycle/m. The estimated profile is
compared with the raw profile.
Case 1. without noise and model error
Profile estimations with and without RTS smoothing are compared with the true profile as shown in Fig. 5. The drive speed
is 20 km/h. The vertical acceleration and angular velocity are the “measured responses” while the displacement and angle are
obtained through integration. A high-pass filter is not used because measurement under noise-free conditions results in no
integration error.
B. Zhao et al. / Journal of Sound and Vibration 457 (2019) 92e117 103

Fig. 5. Profile estimation without noise and model error. (a) profile estimation zoom-in 50e55 m, (b) profile PSD, (c) zoomed in profile PSD (with smoothing).

As shown in Fig. 5(a) the estimated road profiles show good agreement with the true profile, especially in the low fre-
quency ranges. The discrepancy at a higher frequency range is observed. This difference is further investigated in terms of the
PSD as shown in Fig. 5(b). The estimation based on standard KF shows agreement with the true profile up to 2 cycle/m while
the estimation with the RTS smoother shows a much better performance. The zoomed in comparison of smoothing results are
shown in Fig. 5(c). The estimations show agreement with the true profile up to 7 cycle/m that is slightly below the Nyquist
frequency of 9 cycle/m or 50 Hz.
Case 2. - with 10% noise & model error
Model error is introduced by using the estimated parameters shown in Table 6. The response is simulated by adding noise
components to the unpolluted vehicle responses as.

Ep
xm ¼ x þ N sðxÞ; (55)
100 noise

where Ep is the percentage noise level, Nnoise is a noise with a standard normal distribution, s(x) is the standard deviation of
the “measured” response.
A high-pass filter with a cut-off frequency of 0.15 Hz is used for the removal of the low frequency integration error. Note
that this high-pass filter is also effective in removing the effect of a sudden change in drive speed. The reasoning behind the
cut-off frequency is explained herein.
104 B. Zhao et al. / Journal of Sound and Vibration 457 (2019) 92e117

In this paper, 50 m is set as the longest wavelength of the interest. The corresponding spatial frequency is 0.02 cycle/m.
Very long wavelength or very low spatial frequency components of the road profile can be ignored in practice due to its
negligible influence on road ride comfort.
The high-pass filter is, however, applied in the time domain in the same manner as the Kalman filter, whose system
equations are written in the time domain. While the cut-off frequency is kept constant in the temporal domain, the frequency
in the spatial domain depends on the drive speed. The relationship is described as

w ¼ f =V; (56)

where w is the frequency in the spatial domain and f is the frequency in the temporal domain. While w is expressed in units of
cycle/m, and f is in units of Hz, the velocity V has units of m/s. A low drive speed results in a high cut-off spatial frequency and
underestimates or ignores low frequency components.
A practical low drive speed considered in this study for the purpose of profile estimation is 30 km/h. The temporal fre-
quency corresponding to the 0.02 cycle/m limit under this drive speed is 0.167 Hz. The temporal cut-off frequency of 0.15 Hz is
therefore low enough to be employed. The profile estimation at a drive speed of 30 km/h is shown in Fig. 6.
As shown in Fig. 6 (aeb), the profile estimation accuracy is improved significantly with the RTS smoothing, in particular,
for the high frequency component. The comparison of front and rear tire profile estimations are shown in Fig. 6 (c). The rear
tire profile estimation performs better than the front tire. This is because the parameter estimations of the rear part of the HC
model (mr, kr, ktr, cr) that are for parameters close to the sensor, are more accurate than the estimations for the front part
parameters (mf, kf, ktf, cf) as shown in Table 6. The reasons for the relatively higher estimation accuracy of rear parameters can
be explained by the sensitivity analysis in section 3.3.

3.2.3. Influence of drive speed


The profile estimation of the rear tire in Case 2 with different drive speeds are shown in Fig. 7. When the drive speed is low,
the estimated PSD shows good agreement at high frequency ranges while a large difference can be observed at a low fre-
quency range. This is due to the 0.15 Hz high-pass filter in the time domain, given by Eq. (56), as discussed already. On the
other hand, when the drive speed is high, the estimated PSD shows a good match with the measured values in the low
frequency range and a worse match in the high frequency range. The upper estimation bound in the spatial domain depends
on both the sampling frequency in the time domain and the driving speed. The spatial interval dx of the estimated profile is
calculated by Eq. (57)

dx ¼ Vdt (57)

where dt is the sampling interval of 0.01 s. When the drive speed is high, the spatial interval becomes large. Therefore, the
upper bound in the spatial domain is reduced. Considering the Nyquist frequency (50 Hz in the time domain) and the high-
pass filter (0.15 Hz), the upper and lower bound in the spatial domain at representative drive speeds are summarized in Table
7.
Above all, a drive speed between 30 km/h and 130 km/h is considered to provide reliable profile estimation within the
frequency range from 0.017 cycle/m to 1.4 cycle/m, if there is no noise and no model error. Even with the influence of noise
and model error, such as Case 2, a good accuracy can be obtained in the frequency range from 0.02 cycle/m to 1 cycle/m, as
shown in Fig. 7. The profile estimation performance in this frequency range is examined in the experimental validation.

3.3. Sensitivity analysis

The sensitivity analysis of the responses with respect to the perturbation of system physical parameters provides an
understanding of the contributions of each parameter to the responses [33]. A similar sensitivity analysis of bridge responses
with respect to HC model parameters is discussed in Ref. [27]. The sensitivity analysis of the vehicle responses is performed
herein. Consider the linear time-invariant system given by Eq. (1).
If a system parameter a, which is assumed to be related only to the stiffness matrix K, is perturbed by Da; the perturbed
equation of motion is obtained by differentiating both sides of Eq. (1) with respect to the system parameter. For instance, the
sensitivity of the response to the change in stiffness is obtained by differentiating the equation with respect to the stiffness
perturbation. The partial derivative affects only the stiffness and responses, as follows:
8 9 ( )
<vX€ðtÞ = _    
vXðtÞ vXðtÞ vK
M þC þK ¼ XðtÞ; (58)
: va ; va va va


where fvXðtÞ=v _
ag, fvXðtÞ=vag, fvXðtÞ=vag are the sensitivities of the responses with respect to the stiffness change. Note that
Eq. (58) is in the same form as Eq. (1). The right side of Eq. (58) is considered as an equivalent force. Therefore the sensitivities
of the responses to the parameter changes, (i.e., vXðtÞ € =va, vXðtÞ_ =va; and vXðtÞ =va) can be easily obtained by numerical
integration of Eq. (58). When the sensitivity with respect to mass or damping parameters is considered, the right hand-side of
B. Zhao et al. / Journal of Sound and Vibration 457 (2019) 92e117 105

Fig. 6. Profile estimation with noise and model error. (a) profile estimation zoom-in 50e55 m, (b) profile PSD, (c) zoomed in profile PSD (with smoothing).

Eq. (58) changes correspondingly. Note that the purpose of this sensitivity analysis is to investigate which of several pa-
rameters can be estimated more easily in the hump calibration process. Therefore, the sensitivity comparison among different
parameters is of more interest than the absolute sensitivity amplitude of each parameter.
The sensitivity analyses are performed assuming three different sensor locations in the HC model and during vehicle
passage over the known-size hump at a speed of 10 km/h, as introduced in section 2.1. The three sensor locations are chosen
as 1) above rear axle, 2) 1.8 m from the front axle (close to the rear passenger seat), and 3) the center of gravity. The sensor
location parameter d in each case is thus 2.87 m, 1.8 m, and 1.4 m, respectively. The sensitivities of both vertical acceleration
and angular velocity with respect to the front (kf, cf, mf) and rear (kr, cr, mr) HC model parameters are calculated and shown in
Figs. 8e11.
The sensitivity of angular velocity to the various parameters is shown in Fig. 8. As the vehicle body is assumed to be a rigid
body, the angular velocity measured on the body does not depend on the sensor location. The sensitivity of the angular
velocity is therefore independent of the sensor location. As shown in Fig. 9, when the sensor is above the rear axle, the
sensitivities of vertical acceleration with respect to the rear tire stiffness are clearly larger than the corresponding sensitivities
with respect to the front tire stiffness, while the sensitivities of the rear suspensions parameters are similar to those of the
front suspension parameters. On the other hand, when the sensor is closer to the middle of the vehicle (Figs. 10 and 11), the
sensitivities of the front and rear parameter pairs all show similar values. This indicates that the contributions from the tire
parameters close to the sensor location are dominant with regards to the vertical acceleration. Tire parameters closer to the
sensor location are thus expected to be identified more accurately as in Table 6. Though the sensitivities of the front and rear
suspensions do not show much difference, the large sensitivity of the tire stiffness is considered to contribute to the better
106 B. Zhao et al. / Journal of Sound and Vibration 457 (2019) 92e117

Fig. 7. Profile estimation of different drive speeds.

Table 7
Upper and lower bound frequency in the spatial domain.

Speed (km/h) 10 30 50 70 100 130


Upper bound (cycle/m) 18 6 3.6 2.5 1.8 1.4
Lower bound (cycle/m) 0.05 0.017 0.01 0.007 0.005 0.004

estimation of the suspension parameters because the tire and suspension motions have interactions. In conclusion, those
parameters close to the sensor location are expected to be better identified as in Table 6; the corresponding profile is esti-
mated more accurately than the profile derived at the other tire farther from a sensor location. When the sensitivity analysis is
performed at higher drive speeds, the suspension parameters are found to have relatively small sensitivity. Thus a slower
drive speed is considered advantageous to identify the suspension parameters while extremely slow drive speeds result in
practical difficulties in keeping a constant drive speed during the hump passage.

4. Profile estimation by field test

4.1. Instrumentation

Apple's iPhone and iPod touch [34] are employed as measurement devices with an iOS application called iDRIMS mea-
surement application. The requirements of accurate sampling timing and easy manipulation by operators who are drivers of
commercial vehicles have driven the development of the application (Fig. 12). The application can record acceleration, angular
velocity, and GPS signals simultaneously. The application can also capture photo and movie data. The vibration measurement
is performed at a sampling frequency of 100 Hz and a GPS signal is obtained with a sampling frequency of 1 Hz. The mea-
surement data can optionally be sent to a server. Because the sampling rate of the devices varies from one to another and
fluctuates over time, offline resampling based on the time stamp was performed [35].
Two kinds of shake table tests are conducted to investigate the accuracy of vibration measurement with the 6-th gen-
eration of iPod touch devices. In this research, the 6th generation iPod touch is employed. Because an iPod touch does not
have internal GPS, an external GPS module [36] is connected to obtain the location and speed information. One test is a shake
table test in the horizontal direction to investigate the accuracy of the acceleration measurement by comparison with a high
accuracy servo-type accelerometer (Tokyo Sokushin CV-375; Fig. 13). The other test is a shake table test where one end of a
bar installed with sensors is attached to the shake table and the other end is attached to a pivot. The bar rotates around the
pivot following the motion of the shake table. The accuracy of the angular velocity measurement is thus examined by
comparing it to the angular velocity obtained by a gyroscope (MEMSIC VG440-CA; Fig. 14).
The PSD of the measured signals are shown in Fig. 15 together with the references. The table is excited by a 0e50 Hz and
0e0.6 Hz band-limited white noise inputs. Both acceleration and angular velocity signals show good agreement with signals
from the reference sensors in the range of 0.15e15 Hz. While the agreement with reference signals outside the range of
0.15e15 Hz is not as good as that within the range, the signals show similarities to the reference signals up to 40 Hz. As the
sensor can capture the DC components of acceleration, frequency components below 0.15 Hz are also expected to be captured
when the low frequency components of vehicle acceleration responses are large.
Because both the acceleration and angular velocity of an iPod touch show good agreement in the frequency range of
0.15e15 Hz, the sensor is confirmed to be suitable for road profile estimation in this frequency range; the dominant and
important frequency of road profile estimation is considered to fall in this frequency range.
B. Zhao et al. / Journal of Sound and Vibration 457 (2019) 92e117 107

Fig. 8. Sensitivity of angular velocity to parameters (common to all sensor locations). (a) sensitivities to kr, ktr, kf, ktf; (b) sensitivities to mf, mr, cf, cr.

Fig. 9. Sensitivity of acceleration to parameters (sensor location: above rear tire). (a) sensitivities to kr, ktr, kf, ktf; (b) sensitivities to mf, mr, cf, cr.

Fig. 10. Sensitivity of acceleration to parameters (sensor location: 1.8 m to front axle). (a) sensitivities to kr, ktr, kf, ktf; (b) sensitivities to mf, mr, cf, cr.

4.2. Vehicle types and sensor locations

A field test is performed on a 13 km ordinary road in Chiba city, Japan. In order to investigate the applicability of the
proposed methods to different types of vehicles, three rental cars are tested. These vehicles are a light vehicle that is
representative of post office vehicles, a small size vehicle representative of taxis, and a middle size vehicle representing a
shipping van as shown in Fig. 16 (aec). Each vehicle is installed with multiple iPod touch devices allowing different sensor
locations along the longitudinal direction including the dashboard, front and rear passenger seat floors, and trunk. In order to
108 B. Zhao et al. / Journal of Sound and Vibration 457 (2019) 92e117

Fig. 11. Sensitivity of acceleration to parameters (sensor location: center of gravity). (a) sensitivities to kr, ktr, kf, ktf; (b) sensitivities to mf, mr, cf, cr.

Fig. 12. DRIMS App and iPod touch 6th.

measure accurately the dynamic response of the vehicle and to eliminate the need for coordinate transformations, the sensors
are installed on horizontal surfaces. The profile estimation accuracy of the proposed method is examined by comparing the
estimates with a reference profile obtained by a laser profiler Fig. 16 (d) [37].

4.3. Vehicle modeling

Hump calibration is performed on a 40 m straight road. A road surface without any manholes or potholes is selected. Note
that while the parameter identification algorithm assumes the road is completely flat except for the hump, actual roads are
not completely flat and the vehicle response is non-zero even before the vehicle reaches the hump or long after the vehicle
passes over the hump. Nevertheless, the responses to the hump are still much larger than the vehicle responses to the un-
evenness of the road. The road condition and hump setting are shown in Fig. 17. The vehicle first accelerates to the target

Fig. 13. Horizontal shake table test.


B. Zhao et al. / Journal of Sound and Vibration 457 (2019) 92e117 109

speed (10 km/h) and coasts over the hump. The measurement continues for about 15 m after passing over the hump and free
vibration responses are obtained. The whole process is summarized in Fig. 18.
Each vehicle repeats the hump drive five times. The best HC model of each measurement vehicle is extracted by choosing
the model with the smallest target function calculated by Eq. (9). The response time histories and power spectra of the light
vehicle are shown in Fig. 19. The target function after optimization is 31.4%; the model's time histories and spectra show good
agreement with the actual measurement.

4.4. Profile estimation

The profile estimation accuracy is examined by comparing the estimates with the reference obtained by the road profiler
at an interval of 0.05 m. The accuracy validation is conducted in terms of the profile in the distance domain and its PSD in the
spatial frequency domain.

Fig. 14. Rotation shake table test.

Fig. 15. Sensor performance of iPod touch. (a) acceleration PSD (Excitation frequency 0e50 Hz), (b) angular velocity PSD (Excitation frequency 0e50 Hz), (c)
acceleration PSD (Excitation frequency 0e0.6 Hz), and (d) angular velocity PSD (Excitation frequency 0e0.6 Hz).
110 B. Zhao et al. / Journal of Sound and Vibration 457 (2019) 92e117

The measured vertical acceleration and angular velocity are initially processed by a 0.15 Hz-cutoff high-pass filter and then
integrated to obtain the vertical displacement and pitching angle. The time domain profile estimated by the Kalman filter (KF)
is initially mapped to the spatial domain using the distance obtained from the GPS drive speed signal. The profile is further
interpolated at an interval of 0.05 m.
One 1.8 km-long representative section is extracted from the 13 km road and the iDRIMS profile estimation is compared
with the profiler measurement. On the 13 km road, a laser profiler measured the profile in December 2014 and December
2016 while the iDRIMS measurement was performed in December 2015. There are sections where the two laser profiler data
sets show clear differences due to an occurrence of road repair between the two measurements, laser profiler inaccuracy at
stop locations, and other reasons. Only three sections (i.e. section A from 1080 m to 2020 m, section B from 2820 m to 4620 m,
and section C from 9980 m to 10680 m) show consistency between the 2014 and 2016 profiler measurements in terms of IRI.
The longest of the three sections, Section B, is thus chosen as the representative section and the 2016 profiler data is set as the
reference. Road profile estimation using the sensors on the front passenger seat floors of three vehicles are compared with the
reference values in Fig. 20.
The profile estimations using the three vehicles show good agreement overall with the reference values in terms of both
profiles and their PSDs even when the non-constant drive speeds of the three vehicles are different from each other (see
Fig. 20(a) to (c)). The zoomed-in comparison in Fig. 20(d) shows the agreement of the iDRIMS and profiler measurements in
detail. Results indicate that the proposed method is able to identify road conditions after compensating for differences in
vehicle types and drive speeds.
The International Roughness Index (IRI) [38] is further calculated and compared with the profiler data. IRI has been
proposed by the World Bank as an indicator of drive comfort and is widely used in the evaluation of road roughness. It is
defined as:
8 9
> L=V >,
<Z c =
IRI ¼ jz_s  z_u jdt L; (59)
>
: >
;
0

where zs and zu are the vertical absolute displacement of the sprung mass and unsprung mass of the Golden-Car model, L is
the length of the evaluation section, and Vc is the drive speed that is set as 80 km/h, and t is time. The parameters of the

Fig. 16. Experimental vehicles. (a) light vehicle, (b) small size, (c) middle size, and (d) profiler.

Fig. 17. Hump and road condition.


B. Zhao et al. / Journal of Sound and Vibration 457 (2019) 92e117 111

Fig. 18. Vehicle modeling process.

Golden-Car model are shown in Fig. 21, where ms and mu are the sprung mass and unsprung mass, respectively, ks and cs are
the suspension stiffness and damping, kt is the stiffness of the tire, and h is the road profile.
The IRI of the target road is calculated through the estimated profile and compared with the reference profiler data, shown
in Fig. 22. The calculation procedure follows the definition as shown in Eq. (59); responses of the Golden-Car model are
numerically calculated and converted to IRI. The evaluation length L is 200 m and the interval is 10 m.
As shown in Fig. 22, iDRIMS can accurately estimate IRI using different vehicles and drive speeds. The relative IRI error
between the three vehicles and the road profiler is calculated as
vffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi
uP  
u i IRIest;i  IRItrue;i 2
Eiri ¼ 100%  t (60)
P 2
i IRItrue;i

Fig. 19. Response time history and power spectra. (a) acceleration time history, (b) angular velocity time history, (c) acceleration power spectrum, and (d) angular
velocity power spectrum.
112 B. Zhao et al. / Journal of Sound and Vibration 457 (2019) 92e117

Fig. 20. Profile estimation performance. (a) profile estimation, (b)drive speeds, (c) profile PSD, and (d) profile estimation (zoom-in). (For interpretation of the
references to colour in this figure legend, the reader is referred to the Web version of this article.)
B. Zhao et al. / Journal of Sound and Vibration 457 (2019) 92e117 113

Fig. 21. Golden-Car model.

Fig. 22. IRI calculated from the estimated profile. (For interpretation of the references to colour in this figure legend, the reader is referred to the Web version of
this article.)

Table 8
IRI relative error (%).

Light Small size Middle size


Relative error 8.4 7.8 6.9

where IRIest,i and IRItrue,i are the IRI calculated through the estimated profile and the IRI measured by the road profiler. The
results are shown in Table 8. All the vehicles show IRI relative error smaller than 10%, which indicates that the proposed
method can estimate the road profile with a high accuracy.
The effect of sensor installation location along the longitudinal direction on the profile estimation can be considered by
utilizing a HC model because the model takes the bouncing and pitching modes into account. To check the flexibility of sensor
installation locations, iPod touch devices are installed at some locations along the longitudinal direction, including the
dashboard, front passenger seat, rear passenger seat, and the trunk of each experimental vehicle. As an example, the sensor
layout of the light vehicle is shown in Fig. 23. The distances from the sensor location to the front axle are shown in Table 9. An
800 m road profile estimation from these sensor locations is compared with the profiler reference. The comparison is per-
formed in terms of the relative RMS and IRI errors of the three vehicles at four sensor locations (see Fig. 24 and Table 10). The
relative RMS error is quantified every 200 m as

jb
y rms  yrms j
E¼ ; (61)
yrms
sffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi sffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi
P N P N
2
where b
y rms ¼ y ðxÞ =N ; yrms ¼
b yðxÞ2 =N ; b y ðxÞ and yðxÞ are the estimated and reference profile respectively, along
i¼1 i¼1
the road calculated at 0.05 m intervals. N is the number of samples. Note that the profile estimation error calculated for every
0.05 m interval point is largely influenced by the distance estimation error. Therefore, the accuracy of profile estimation is
quantified using the relative RMS error.
114 B. Zhao et al. / Journal of Sound and Vibration 457 (2019) 92e117

Fig. 23. Sensor layout. (a)Light vehicle sensor layout, (b) dashboard, (c) front passenger seat floor, (d) rear passenger seat floor, and (e) trunk.

Table 9
Sensor layout of the light vehicle (wheelbase: 2.46 m).

Sensor locations Distance to front axle (m)


Dashboard 0.43
Front passenger seat 0.82
Rear passenger seat 1.60
Trunk 2.47

As shown in Table 10, most of the sections show a relative RMS error smaller than 10% that indicates that the proposed
method can estimate the road profile with a high accuracy by using different vehicles and sensor locations. Only at a few
sections are the errors above 10%.
Finally, Fig. 25(a) shows the estimated IRI for a longer section of the 13 km road with the reference profiler data. The
reference profiler data is obtained at the three sections, A, B, and C. The data from the three vehicles show good agreement.
Fig. 25(b) shows the drive speeds of the three iDRIMS vehicles. While the laser profiler cannot provide accurate IRI at the stop
positions, iDRIMS can estimate IRI with reasonable accuracy even when the vehicles stop; there are many locations where
only one or two vehicles stop while the IRI of the three vehicles are consistent. The IRI estimate at the stop position is,
however, not largely different from IRI estimates obtained from vehicles which did not stop. This capability to estimate IRI
B. Zhao et al. / Journal of Sound and Vibration 457 (2019) 92e117 115

Fig. 24. Profile estimation performance and sensor locations. (a) profile estimation, (b) RMS error. (For interpretation of the references to colour in this figure
legend, the reader is referred to the Web version of this article.)

Table 10
RMS error and IRI relative errors (%).

Vehicle type Sensor locations Sections (RMS error) Section (IRI relative error)

0e200 m 200e400 m 400e600 m 600e800 m 0e800 m


Light Dashboard 7.77 3.06 1.23 5.21 6.3
Front seat floor 9.01 5.57 5.99 8.44 6.8
Rear seat floor 15.82 2.46 10.30 3.87 7.4
Trunk 16.19 9.71 13.13 8.58 5.9
Small size Dashboard 8.21 3.11 2.90 9.13 6.6
Front seat floor 12.19 2.84 4.38 7.26 6.7
Rear seat floor 15.02 6.55 6.91 5.33 7.2
Trunk 13.37 5.57 9.98 11.71 6.5
Middle size Dashboard 9.33 3.37 1.71 8.48 6.2
Front seat floor 6.33 7.14 2.68 7.57 6.3
Rear seat floor 5.73 5.58 7.34 13.10 6.9
Trunk 8.01 3.71 8.25 11.55 5.9

even at stop positions is taken into account partly due to the fact that the IRI estimate is obtained with an evaluation length of
200 m; the effect of one stop location in a 200 m stretch does not largely contribute to the 200 m IRI value.

Possible reasons behind some differences between the laser profiler and iDRIMS estimates are summarized herein:
1. Real vehicle suspension systems usually have nonlinearity. In particular, the large displacement and velocity of the sus-
pension motion at high drive speeds and at large road inputs, such as when potholes and manholes are encountered, result
in nonlinear behaviors. The nonlinearity cannot be reproduced by the linear HC model and limits the profile estimation
accuracy. While the Kalman filter can deal with the modeling errors, including the nonlinearity, by increasing the process
noise covariance, the profile estimation accuracy may be sacrificed owing to large modeling errors.
2. Even though the three vehicles and the profiler drove on the same road, their drive paths are slightly different; slight
difference in drive paths can result in discrepancies in road profile estimation; in particular, along the left-wheel path,
there are many manholes, whose influence on the profile estimation largely depends on the drive path.
116 B. Zhao et al. / Journal of Sound and Vibration 457 (2019) 92e117

Fig. 25. IRI calculated from the estimated profile for a longer section. (a) IRI, (b) drive speed. (For interpretation of the references to colour in this figure legend,
the reader is referred to the Web version of this article.)

3. The HC model cannot represent the rolling motion of a real vehicle, which often happens if the road conditions under the
left and right wheels are different. From this point of view, a full car model, which accounts for rolling motion, should
perform better. This improvement is being addressed in the on-going work.
4. Long-wave components of profiles cannot be reconstructed when the drive speed is extremely slow or when the vehicles
are stopped. IRI with the evaluation length of 200 m was, however, not influenced much by such drives and was estimated
at a reasonable accuracy.

5. Conclusion

A response based road profile estimation using the body acceleration and angular velocity of an ordinary vehicle measured
by a single smartphone has been proposed [39]. The proposed road profile estimation method consists of two steps; vehicle
modeling and profile estimation. In vehicle modeling, a pitch HC model, which can reproduce both vehicle bouncing and
pitching motions, is optimized through a GA using the responses of the vehicle passing over a known-size hump. The problem
of the drive speed and the vehicle location inaccuracies during the hump test are addressed by including the drive speed in
the unknown parameters and employing the response power spectra as the optimization function. The road profile is then
estimated by using an augmented Kalman filter and the estimated HC model. The profile estimation accuracy is improved by
employing RTS smoothing. Numerical simulation is conducted to investigate the profile estimation performance with model
error and measurement noise. The frequency range where the IRI estimation accuracy is high is numerically shown to have
relationships with the driving speed, sampling frequency, and the high-pass filter cutoff frequency. Before the imple-
mentation of this proposal on real vehicles, the accuracy of the acceleration and angular velocity measurement using
smartphone sensors is validated by horizontal and rotational shake table tests, respectively. An experiment on a 13 km road is
conducted by comparing the measurements from three rental cars and a laser profiler reference. The proposed method is
shown to accurately evaluate a road profile by compensating for the differences in the vehicles, different sensor locations, and
varying drive speeds.

Acknowledgement

This work was supported by Council for Science, Technology and Innovation, “Cross-ministerial Strategic Innovation
Promotion Program (SIP), Infrastructure Maintenance, Renovation, and Management” (funding agency: JST).

References

[1] M.W. Sayers, S.M. Karamihas, The Little Book of Profiling, the Regent of the University of Michigan, vol. 2, 1998.
[2] M.W. Sayers, S.M. Karamihas, Interpretation of Road Roughness Profile Data, No. FHWA-RD, vols. 96e101, 1996.
B. Zhao et al. / Journal of Sound and Vibration 457 (2019) 92e117 117

[3] A. Gonz alez, E.J. O'brien, Y.Y. Li, K. Cashell, The use of vehicle acceleration measurements to estimate road roughness, Veh. Syst. Dyn. 46 (6) (2008)
483e499.
[4] International Organization for Standardization, Technical Committee ISO/TC, Mechanical Vibration, Shock. Subcommittee SC2 Measurement, Evalu-
ation of Mechanical Vibration & Shock as Applied to Machines, Vehicles, and Structures, Mechanical VibrationdRoad Surface ProfilesdReporting of
Measured Data, 1995.
[5] H. Asakawa, Y. Fujino, T. Nagayama, M. Ohsumi, Development and application of road monitoring system using dynamic response of vehicles, in: 11th
East Asia-Pacific Conference on Structural Engineering and Construction, EASEC, vol. 11, 2008.
[6] H. Asakawa, T. Nagayama, Y. Fujino, T. Nishikawa, T. Akimoto, K. Izumi, Development of A Simple pavement diagnostic system using dynamic re-
sponses of an ordinary vehicle, Journal of Japan Society of Civil Engineers, Ser. E1 (Pavement Engineering) 68 (2012) 20e31.
[7] T. Nagayama, A. Miyajima, S. Kimura, Y. Shimada, Y. Fujino, Road condition evaluation using the vibration response of ordinary vehicles and syn-
chronously recorded movies, in: Proceedings of the SPIE Smart Structures and Materialsþ Nondestructive Evaluation and Health Monitoring, 2013, p.
86923A. https://doi.org/10.1117/12.2010074.
[8] M.W. Sayers, T.D. Gillespie, W.D.O. Paterson, The International Road Roughness Experiment, 1985. World Bank Technical Paper 45.
[9] B. Zhao, T. Nagayama, N. Makihata, M. Toyoda, M. Takahashi, M. Ieiri, IRI Estimation by the Frequency Domain Analysis of Vehicle Dynamic Responses
and its Large-Scale Application, Adjunct Proceedings of the 13th International Conference on Mobile and Ubiquitous Systems: Computing Networking
and Services, ACM, New York, NY, USA, 2016, pp. 41e46. MOBIQUITOUS 2016, https://doi.org/10.1145/3004010.3004013.
[10] B. Zhao, T. Nagayama, IRI estimation by the frequency domain analysis of vehicle dynamic responses, Procedia Engineering 188 (2017). https://doi.org/
10.1016/j.proeng.2017.04.451.
[11] B. Zhao, T. Nagayama, M. Toyoda, N. Makihata, M. Takahashi, M. Ieiri, Vehicle model calibration in the frequency domain and its application to large-
scale IRI estimation, J. Disaster Res. 12 (3) (2017) 446e455.
[12] L. Forslo€fL, H. Jones, Roadroid continuous road condition monitoring with smart phones, Journal of Civil Engineering and Architecture 9 (4) (2015).
https://doi.org/10.17265/1934-7359/2015.04.012.
[13] H.M. Ngwangwa, P.S. Heuns, F.J.J. Labuschagne, G.K. Kululanga, Reconstruction of road defects and road roughness classification using vehicle re-
sponses with artificial neural networks simulation, J. Terramechanics 47 (2010) 97e111.
[14] M. Yousefzadeh, S. Azadi, A. Soltani, Road profile estimation using neural network algorithm, J. Mech. Sci. Technol. 24 (3) (2010) 743e754.
[15] M. Doumiati, S. Erhart, J.-J. Martinez, O. Sename, L. Dugard, Adaptive control scheme for road profile estimation: application to vehicle dynamics,
World Congress 19 (2014) 8445e8450.
[16] J.C. Tudon-Martinez, S. Fergani, O. Sename, J.J. Martinez, R. Morales-Menendez, L. Dugard, Adaptive road profile estimation in semiactive car sus-
pensions, IEEE Trans Control Syst (2015) 2293e2305.
[17] H. Imine, Y. Delanne, N.K. M'sirdi, Road profile input estimation in vehicle dynamics simulation, Veh. Syst. Dyn. 44 (4) (2006) 285e303.
[18] N.K. Harris, A. Gonzalez, E.J. OBrien, P. McGetrick, Characterisation of pavement profile heights using accelerometer readings and a combinatorial
optimisation technique, J. Sound Vib. 329 (5) (2010) 497e508. https://doi.org/10.1016/j.jsv.2009.09.035.
[19] E.J. OBrien, P. Quirke, C. Bowe, D. Cantero, Determination of Railway Track Longitudinal Profile Using Measured Inertial Response of an In-Service
Railway Vehicle, Structural Health Monitoring, 2017, https://doi.org/10.1177/1475921717744479.
[20] E.J. OBrien, C. Bowe, P. Quirke, D. Cantero, Determination of longitudinal profile of railway track using vehicle-based inertial readings, Journal of Rail
and Rapid Transit (2016), https://doi.org/10.1177/0954409716664936.
[21] M. Doumiati, A. Victorino, A. Charara, D. Lechner, Estimation of road profile for vehicle dynamics motion: experimental validation, in: IEEE American
Control Conference (ACC), 2011, pp. 5237e5242. San Francisco, CA.
[22] W. Fauriat, C. Mattrand, N. Gayton, A. Beakou, T. Cembrzynski, Estimation of road profile variability from measured vehicle responses, Veh. Syst. Dyn.
54 (5) (2016) 585e605.
[23] K. Yagi, A measuring method of road surface longitudinal profile from sprung acceleration and verification with road profiler, Journal of Japan Society
of Civil Engineers (pavement engineering) 69 (3) (2013).
[24] Q. Mei, M. Gul, A Crowdsourcing-Based Methodology Using Smartphones for Bridge Health Monitoring, Structural Health Monitoring, 2018, https://
doi.org/10.1177/1475921718815457.
[25] J.H. Holland, Genetic algorithms, Sci. Am. 267 (1) (1992) 66e72.
[26] Y. Ding, B.Y. Zhao, B. Wu, X.C. Zhang, L.N. Guo, Simultaneous identification of structural parameter and external excitation with an improved unscented
Kalman filter, Adv. Struct. Eng. 18 (11) (2015) 1981e1998.
[27] 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.
[28] R.E. Kalman, A new approach to linear filtering and prediction problems, Journal of basic Engineering 82 (1) (1960) 35e45.
[29] H.E. Rauch, F. Tung, C.T. Striebel, Maximum likelihood estimates of linear dynamic systems, AIAA J. 3 (8) (1965) 1445e1450.
[30] H. Robert, A. Krener, Nonlinear controllability and observability, IEEE Trans. Autom. Control 22 (5) (1977) 728e740.
[31] T.J. Saravanan, B. Zhao, D. Su, T. Nagayama, An Observability Analysis for Profile Estimation through Vehicle Response Measurement, ICE Virtual Library
Essential Engineering Knowledge, 2016, https://doi.org/10.1680/tfitsi.61279.357.
[32] M. Agostinacchio, D. Ciampa, S. Olita, The vibrations induced by surface irregularities in road pavementsea Matlab® approach, European Transport
Research Review 6 (3) (2014) 267e275.
[33] Z.R. Lu, S.S. Law, Features of dynamic response sensitivity and its application in damage detection, J. Sound Vib. 303 (1) (2007) 305e329.
[34] Apple Inc, 2017. https://www.apple.com.
[35] T. Nagayama, B.F. Spencer Jr., K.A. Mechitov, G.A. Agha, Middleware services for structural health monitoring using smart sensors, Smart Struct. Syst. 5
(2) (2009) 119e137.
[36] Bad Elf, GPS for Lighting Connector, 2016. http://bad-elf.com/pages/be-gps-1008-detail.
[37] Nichireki corporation. https://www.nichireki.co.jp/product/consult/consult_list05/consult05_01.html.
[38] M.W. Sayers, The International Road Roughness Experiment: Establishing Correlation and a Calibration Standard for Measurements, 1986.
[39] Patent application PCT/JP2018/017182, Road surface profile estimating device, road surface profile estimating system, road surface profile estimating
method, and road surface profile estimating program, 2018.4.27.

You might also like