You are on page 1of 10

Spectral Modeling of Longitudinal Road Profiles

Semiha Türkay and Hüseyin Akçay


Department of Electrical and Electronics Engineering, Anadolu University
Eskisehir 26555, Turkey

Abstract
In this paper, for a quarter-car model, we derive covariance expressions for the MIRA spectrum and calculate the root-mean
square outputs for the ISO 8608 classified roads. The derived results are compared with existing works in the literature. We
carry out similar calculations for a line spectrum model. The linear and the piecewise-linear spectrum models are also contrasted
with the spectrum models delivered by the subspace identification and the compressive sensing algorithms.
Index Terms
road profile; quarter-car; power spectrum; identification; motor industry research association; ISO 8608

E-mails: {huakcay,semihaturkay}@anadolu.edu.tr. Tel: +90 222 335 0580. Fax: +90 222 323 9501. This paper can be accessed at www.ccece2015.org
subject to copywright protections.
I. I NTRODUCTION
The terrain profile is an indexed set of heights that can be considered a signal and represented in vector form. This signal
can be used directly as excitation in vehicle simulations when simulating specific events. It is computationally impractical
to simulate a vehicle travelling over long stretches of measured terrain, requiring choices to be made about the physical
terrain features that are most appropriate for a given application. Therefore, it is advantageous to consider the terrain profile
as a particular realization of an underlying stochastic process.
Since 1960s, a lot of studies and researches have been done on this subject and new measuring equipments and techniques
are still being developed. The terrain profile is a consistent excitation to the chassis, even if the chassis design changes.
Knowledge of this excitation provides the chassis designer with a powerful tool to make informed design decisions early in
the design process while changes are relatively inexpensive to implement. This will, in turn, shorten vehicle development
time and reduce overall development costs.
The use of ISO 8608 standard [1] for road modeling is based on the assumption that a given road has identical statistical
properties along a section to be classified. Vehicle engineers are interested in representative road data to help them design
cars that ride well with minimum vibration and road noise. Global road unevenness can be described statistically by a power
spectral density (PSD) function [2]. While the spectral analysis of road profiles are relatively new to road engineering, it has
been employed by automotive engineers for several decades. In the ISO 8608 report, eight classes of roads are identified.
By comparing the PSDs associated with the classes, roads that have a minor degree of roughness are defined of best quality
while roads that have a high degree of roughness are regarded as very poor. For practical purposes, an artificial road profile
with prespecified PSD from the ISO 8608 classification may be generated.
The simplicity of the ISO spectrum makes it attractive to use in vehicle development. However, often the spectrum model
picked from the ISO 8608 classification does not provide an accurate description of real road spectrum [3].
A two slope spectrum, known also as the MIRA (Motor Industry Research Association) [4] spectrum, fits much better
to the observed spectrum than the simpler single slope spectrum. The PSDs of different classes of road roughness have
been approximated by using two slope spectrum, which expresses a mathematical fit to the empirical data, based on the
extensive measurements collected on the European and the American roads. This simple parametric PSD may not accurately
approximate the road roughness spectrum for the whole range of frequencies, however it will correctly estimate the energy
for the frequencies in the range which may excite the vehicle response. A modified version called the British Standards
Institution (BSI) spectrum [5] is regarded as another industry standard. The MIRA and the ISO spectra are dominant classes
among those used in the industry.
The vertical road input is the most important load for durability assessments of vehicles. The load which a vehicle axle
transfers on the pavement surface is not constant in time and space, but is variable in function of several factors such as vehicle
mass, speed, suspension type, road surface irregularities. For practical purposes, the quarter-car model can effectively be
used to study the dynamic interaction between vehicle and road roughness, and therefore in the study of vibrations generated
by road traffic [6], [7].
The contents of this paper are as follows. For the quarter-car model and the MIRA spectrum, covariance expressions
are derived. The root-mean square (RMS) outputs are calculated for the ISO 8608 classified roads and compared with
existing results in the literature. Similar calculations are made for a linear spectrum model. The linear and the piecewise-
linear spectrum models are contrasted with the spectrum models delivered by the subspace [8] and the compressive sensing
subspace [9], [10] identification algorithms.
II. T HE QUARTER - CAR MODEL
The two-degrees–of–freedom quarter-car model used to study vibration characteristics of a typical lightly damped passenger
car is shown in Figure 1. The parameter values chosen for this study are shown in Table I. In this model, the car body and
the wheel are represented by the sprung and the unsprung masses ms and mu , respectively. The suspension is modeled by
the passive suspension elements ks and cs connected in parallel and the tire kT .
TABLE I
T HE VEHICLE SYSTEM PARAMETERS FOR THE QUARTER - CAR MODEL .

Sprung mass ms 240 kg


Unsprung mass mu 36 kg
Damping coefficient cs 980 Ns/m
Secondary suspension stiffnesses ks 16,000 N/m
Primary suspension stiffness kT 160,000 N/m

The measured longitudinal road profile is considered as a realization h(l) of the random function H(l) along the road track
of distance l. Provided that this function is homogeneous, centered, and Gaussian, its PSD will provide its full statistical
description. A profile consists of different wavelengths, varying from a few centimeters to hundreds of meters. Often, it is
z
Sprung
mass mS

cS kS

y
Unsprung
mass mu

kT h

Fig. 1. The quarter-car model

assumed that frequencies Ω < 0.01 cycle/m for wavelengths above 100 m do not affect the vehicle dynamic response and
hence can be removed from the spectrum. Similarly, high frequencies Ω > 10 cycle/m for wavelengths below 10 cm are
filtered out by tire and also are not included in the spectrum.
While vehicle travels at a constant speed v, on an uneven road profile it is exposed to the road displacement input
h(l) = h(vt) characterized by its elevation PSD GH (Ω). Then the function defined in the spatial domain changes into
equivalent function in time domain by the relation
GH (ω) = v−1 GH (Ω) (1)
where ω is the frequency in rad/s and Ω = ω/v.
The natural frequencies of the heave and the wheel hop modes of the quarter-car model are approximately 1.2 and 11
Hz. Humans are particularly sensitive to vertical oscillatory motion in the frequency range 4 to 8 Hz (vibratory discomfort)
and 0.1 to 0.5 Hz (motion sickness). As seats typically attenuate vibrations at frequencies above 6 Hz, we can expect
car suspensions to be designed to minimize body motion in the frequency ranges 0.1–0.5 Hz and 4–8 Hz. The vehicle
response characteristics serve to amplify profile frequencies around the heave and the wheel-hop modes, and attenuate
profile frequencies well removed from those of the response modes.
The vehicle response variables are as follows:
(a) z1 = z̈, the vertical acceleration of the sprung-mass;
(b) z2 = z − y, the suspension travel;
(c) z3 = kt (y − h), the dynamic tire force Fdyn .
Let Hzh ( jω) be the frequency response function from the road displacement input h to the output vector z. The PSD Gz ( jω)
of the output z obeys the following relation
Gz (ω) = |Hzh ( jω)|2 GH (ω). (2)
As can be seen from Eq. 2 it is necessary to compute from the time histories of the excitations their second-order moments,
or in effect their spectral densities, to obtain a sufficient statistical description of the excitation process. Thus a spectral
description of the road, together with a knowledge of traversal velocity and of the dynamic properties of the vehicle, will
provide a vibration response analysis. This analysis will be discussed in the next section.
III. C LASSIFICATION OF R ANDOM ROAD P ROFILES
The international directive ISO 8608 standard is used to facilitate the compilation and comparison of measured vertical
road profile data from various roads, highways etc. The standard specifies how measurements shall be reported, but not
how the measurements shall be made. In ISO 8608, a two parameter spectrum GH (Ω) has been used to describe the road
unevenness h(l) by a simple expression,
 −δ

GH (Ω) = C . (3)
Ω0
where C = GH (Ω0 ) is the unevenness index, δ is the road waviness, and Ω is the spatial frequency. The parameters C and δ
are independently specified and C is proportional to the unevenness variance. The interpretation of δ is as follows: for δ < 2
the shortwaves prevalent; for δ = 2 the individual wavelengths in the road elevation PSD present in similar proportions; for
δ > 2 the long waves prevalent. Although the analytical model of road unevenness in Eq. (3) is simple, it covers a variety
of examples appearing in the traffic practice.
In ISO 8608, curve-fitting is suggested to estimate C and δ by statistical procedures. Although both C and δ are necessary
for full description of the unevenness PSD, a road classification based on 106 ×C for δ = 2 and Ω0 = 1 rad/m is given in
Table II.
TABLE II
ISO 8608 ROAD CLASSIFICATION ACCORDING TO 106 × GH (Ω0 ).

Road Class Geometric Lower Upper


mean limit limit
A 1 ... 2
B 4 2 8
C 16 8 32
D 64 32 128
E 256 128 512
F 1,024 512 2,048
G 4,096 2,048 8,192
H 16,384 8,192 ...

Typical roads are grouped into Classes A-H. By comparing the PSDs associated with the classes we see that Class A
with δ = 2 and C = 10−6 m2 /rad/m characterizes very smooth highways that have a minor roughness, and therefore for the
purpose of producing vibrations, can be defined of best quality whereas in Class H are included all roads that have a high
degree of roughness for C = 16, 384 × 10−6 m2 /rad/m and regarded as terribly poor.
However, in extensive measurements of road networks in the United States, Germany, and Sweden it was revealed that
except the practically and analytically preferred value of waviness δ = 2, many roads exhibit waviness value in a rather wide
range from 1.5 to 3.5. Road profile spectra generated by changing δ are plotted in Figure 2. The low frequency contents
of GH (Ω) don’t diverse much from the high frequency contents. However, as the waviness increases the effect of high
frequencies decreases in the road spectrum while that of the low frequencies become more pronounced. The conversion
Ω = 2π/l translates the spatial frequency range 0.069–17.77 rad/m to the waveband 0.3534–90.9 m.
The questions of how important is the broad range of road waviness in the vehicle vibration context or if it ever plays a
role immediately arise from Figure 2. In the sequel, a parametric study for the vehicle response variables will be presented
for changing waviness values.
For a vehicle travelling with a constant speed on the random road profile, for example 20 m/s, the long waves prevailing in
the road spectrum when δ > 2, excite predominantly lower eigenfrequency of the vehicle. On the contrary, the short waves
prevailing in the road spectrum when δ < 2 excite predominantly higher eigenfrequency of the vehicle. Small waviness will
activate the high frequency modes of the vehicle which in turn will affect ride comfort, the suspension deflection, and the
dynamic tire forces of the output responses negatively especially in active suspension design studies.

−2
10
δ= 1.5
10
−3 δ= 2
δ= 2.5
−4
δ= 3
10 δ=3.5

−5
10
GH (Ω)

−6
10

−7
10

−8
10

−9
10

−10
10
−1 0 1
10 10 10
Ω (rad/m)

Fig. 2. Simulated road profiles with constant unevenness index C = 10−6 rad/m and variable waviness.

A. Linear approximation to power spectrum


In many simulation studies, when studying the dynamics of vehicle-road interaction, for simplicity the road roughness is
typically specified as an integrated white-noise model. However, the effect of the waviness is more complex than simply
assuming δ = 2 and its impact on the vehicle vibration needs to be examined. For the road model in Eq. 3, the components
of the PSD in Eq. 2 can be expressed in terms of the waviness and the roughness index as
Gzi (ω) = |Hzi h ( jω)|2 ω −δ Cvδ −1 , i = 1, 2, 3. (4)
Thus, for a pre-specified δ , each PSD is proportional to Cvδ −1 .
By omitting the constant terms and keeping only the
frequency dependent terms in Eq (4), the influence of the waviness on the vertical acceleration, the suspension travel, and
the dynamic tire force can be understood by studying
|Hzi h ( jω)|2 ω −δ , i = 1, 2, 3. (5)
The vertical acceleration frequency response for several values of δ , five different road types, and C = 10−6 are plotted
in Figure 3 for the frequency range [0, 200] rad/s. From the figure, observe that the waviness directly effects the magnitude
of the frequency response function. The larger the waviness, the larger the wavelength; hence, the low frequency behaviors
prevalent. See, Figure 2.

4
10
δ= 1.5
δ= 2
δ= 2.5
2
10 δ= 3
δ=3.5
−δ
| Hz (jω)| (ω)
2

0
10
1

−2
10

−4
10

−6
10
0 50 100 150 200
Frequency (rad/s)

Fig. 3. Vertical acceleration response of the quarter-car for the road model in Eq. (3) with 1.5 ≤ δ ≤ 3.

When a simple description of the system response in terms of the output variance is required, this relation can be obtained
from
Cvδ −1 ∞
Z
Rzi (0) = |Hzi h ( jω)|2 ω −δ dω, i = 1, 2, 3 (6)
2π −∞

which is the RMS of the ith output. The RMS values are generally used in vehicle procurement specifications and performance
analysis. They are easy to compute and provide a good ride quality estimation on profiles whose probability distributions
are stationary. The RMS vehicle responses are displayed in Table III.
TABLE III
T HE RMS VALUES OF zk FOR k = 1, 2, 3 AND v = 20 M / S .

RMS δ = 1.5 δ =2 δ = 2.5 δ =3 δ = 3.5


z1 0.1652 0.1408 0.1343 0.1442 0.1401
z2 0.0011 0.0012 0.0014 0.0017 0.0022
z3 93.520 68.534 53.463 46.702 47.707

From Table III, observe that the vertical acceleration is minimized at δ = 2.5. This is not predictable from Figure 3,
verifying the fact the RMS responses are complicated functions of the waviness and the vehicle parameters. The RMS
suspension travel monotonically increases with the waviness. The RMS dynamic tire forces, on the other hand, monotonically
decreases with the waviness except δ = 3.5.
Numerical experiments were carried out for a range of vehicle speeds. Our findings can be summarized as follows:
(i) At low travel speeds, the RMS acceleration increases as δ decreases, more specifically for all δ < 2;
(ii) As far as the RMS acceleration is concerned, high values of the waviness are tolerable at low travel speeds.
Recall that when δ < 2 short waves dominate the road spectrum and they will excite the high resonance frequencies of the
vehicle with a negative impact on ride handling which may result in potential loss of road contact and marked increase in
dynamic tire forces. On the other hand, for long wave components, i.e., δ > 2 the body resonance frequencies will be more
pronounced in the output spectrum (see Figure 3) which in turn will influence ride comfort of the vehicle negatively. In
summary, it is advised to avoid the extreme values of the waviness in vehicle modeling, both the lowest and the highest
values.
In Figure 4, the PSD of a typical road profile and its approximation defined by Eq. 3 for C = 0.76 × 10−5 , δ = 2 and
Ω0 = 0.9870 rad/m is shown. It is clear from the figure that the fit with a single-slope modeling is rather poor; especially
at the low frequency range. Linear approximation to spectrum, generally with waviness δ = 2, is widely used in vehicle
development. In spite of its simplicity, it is not suitable for the majority of roads and the analysis in [12], [13] showed that
it is too coarse to predict the behavior of the vehicle subject to random excitations.

0
10

−1
10

−2
Spectral density (m2/rad/m)

10

−3
10

−4
10

−5
10

−6
10

−7
10

−8
10
−3 −2 −1 0 1 2
10 10 10 10 10 10
Ω (rad/m)

Fig. 4. The spectral data [14] and its approximate modeling with a single-slope approximation for δ = 2.

B. Piecewise-linear approximation to power spectrum


A two-slope spectrum approximation proposed by Dodds and Robson [14] and standardized in ISO 8606 with the general
form
  −δ1
 σ2
 Ω
, Ω ≤ Ω0 ,
u,0 Ω0
GH (Ω) =  −δ2 (7)
 σ2
 Ω
, Ω ≥ Ω 0 ,
u,0 Ω0

is another approach that also yields good fit to typical measured road data with a longitudinal elevation variance σu,0 2 and

cut-off frequency Ω0 .
For the measured spectral data in Figure 4, a more accurate description of the road roughness can be obtained by using
a two-slope spectrum approximation as shown in Figure 5 with waviness values δ1 = 3.1 for the long and δ2 = 2.2 for the
short wave bands of the road profile.
2
The ISO classification identifies five road types displayed in Table IV with possible values of the roughness parameters σu,0
and Ω0 . This classification is very important in practice. A fairly large class of power spectral densities can be generated
by using Eq. (7) and Table IV as depicted in Figure 6. By suitably selecting the roughness parameters, one obtains a
mathematical fit to the measured empirical data.
TABLE IV
ROUGHNESS PARAMETERS FOR THE ROAD TYPES CLASSIFIED BY ISO 8608 [1].

Road Class 2 × 10−6


σu,0 Ω0 (rad/m) δ1 δ2
Very good 1–4 1 2 1.5
Good 4 – 16 1 2 1.5
Average 16 – 64 1 2 1.5
Poor 64 – 256 1 2 1.5
Very poor 256 – 1,024 1 2 1.5
0
10

−1
10

−2

Spectral density (m2/rad/m)


10

−3
10

−4
10

−5
10

−6
10

−7
10

−8
10
−3 −2 −1 0 1 2
10 10 10 10 10 10
Ω (rad/m)

Fig. 5. The spectral data [14] and its approximate modeling with a two-slope approximation.

60
Very good
Good
50
Average
Poor
40 Very Poor
GH(Ω)(dB)

30

20

10

−10

−20
−1 0 1
10 10 10
Ω (rad/m)

Fig. 6. 2 .
Power spectral densities for the road types in Table IV. The lower bounds are selected in Table IV for σu,0

The effect of varying road conditions from very good to very poor is studied next by using the quarter-car model in
Figure 1. The output spectral density of the vehicle for the two-slope spectrum approximation defined by Eq. (7) can be
expressed for i = 1, 2, 3 as
(
2 vδ1 −1 |H ( jω)|2 ω −δ1 , ω ≤ Ω v;
σu,0 zi h 0
Gzi (ω) = 2 vδ2 −1 |H ( jω)|2 ω −δ2 , ω ≥ Ω v. (8)
σu,0 zi h 0

2 , δ , and δ . The output covariance functions are


in terms of the vehicle transfer function and the roughness parameters σu,0 1 2
then given by for i = 1, 2, 3,
Z
2 δ1 −1
Rzi (τ) = σu,0 v (2π)−1 |Hzi h ( jω)|2 ω −δ1
|ω|≤Ω0 v
Z (9)
2 δ1 −1
+σu,0 v (2π)−1 |Hzi h ( jω)|2 ω −δ2 .
|ω|>Ω0 v

The dependence of the covariances on the longitudinal elevation variance, the waviness, the vehicle parameters and the
speed are quite complicated and closed-form solutions are not tractable. Instead, a numerical study is carried out and the
results are plotted in Figure 7. It illustrates influence of road roughness on the vertical acceleration.
Table V shows the RMS values of the vertical acceleration, the suspension travel, and the dynamic tire force of the
quarter-car model computed for the roads in Table IV with the lower bounds for σu,0 2 . The purpose of these calculations is

to show how ride comfort, handling, and packing requirements quantified in terms of the RMS vehicle responses change
from one road type to another.
1
10
Very good
0 Good
10 Average
Poor
−1
10 Very Poor

−2
10

Gz (ω)
1
−3
10

−4
10

−5
10

−6
10

−7
10
0 20 40 60 80 100 120 140 160 180
Frequency (rad/s)

Fig. 7. The vertical acceleration power spectral density of the quarter-car for the road types in Table IV.

TABLE V
T HE RMS VALUES OF zk , k = 1, 2, 3 FOR THE QUARTER - CAR MODEL SUBJECT TO THE ROAD INPUTS IN TABLE IV AT V =20 M / S .

RMS(z1 ) RMS(z2 ) 103 ×RMS(z3 )


Very good 0.171 0.0012 0.0932
Good 0.342 0.0024 0.1865
Average 0.685 0.0050 0.3729
Poor 1.370 0.0099 0.7456
Very poor 2.741 0.0198 1.492

Speed also affects riding quality since it changes frequency content of excitations at the very base of the vehicle. The
higher the vehicle speed, the larger is the excitation; resulting in a worsened vibration response even on smooth roads. As
the vehicle speed changes, due to the relation l = 2πv/ω, the body heave and the wheel-hop modes are primarily excited
in the wavelength bands 1.25 to 9 m and 0.15 to 1.15 m, respectively, as if perceiving a rough profile. Generally speaking,
the wheels are quite damped and they contribute little to the total response. However, the wheel-hop mode of the dynamic
tire force response is significantly amplified as the vehicle speed increases. To sum up, under harsh conditions, i.e., very
rough roads and at high vehicle speeds, performance deterioration is significant.
The two-slope methodology gives an opportunity for approximating and making an initial judgment of the road surface
roughness. However, the model can further be developed by dividing the spectrum in Eq. (7) into three wavebands:
  −δ1
 σ 2 Ω
, 0 ≤ Ω ≤ Ω1 ;
u,0
 Ω
 0−δ2



GH (Ω) = 2
σu,0 Ω
, Ω1 ≤ Ω ≤ Ω2 ; (10)
 Ω0
−δ3

  
 σ2
 Ω
, Ω2 ≤ Ω ≤ Ω3 .
u,0 Ω0

Even if the three-line approximation shows a better fitting to the empirical road spectrum, the medium wavelength band
with the waviness δ2 is too coarse and differs slightly from the first wavelength band with the waviness δ1 . Simulation
studies were carried out for the three-slope spectrum approximation; but the results were similar to those obtained with the
two-slope approximation, hence omitted.
Generally real roads, especially of lower surface quality, often exhibit considerably irregular courses and linear or piece-
wise linear roughness representations which form the basis part of the ISO 8608 standard, often remain overly simplistic.
As the characterized profiles are employed in suspension design and testing, usually more sophisticated approximations are
required.
IV. R ATIONAL APPROXIMATION TO ROAD SPECTRUM
In this section, we will briefly study approximation of the spectral data [14] plotted in Figures 4 and 5 by the subspace-
based identification algorithm in [8] and the hybrid identification algorithms in [9], [10]. The latter are a mixture of the
subspace and the more recent compressive sensing algorithms. The reader is referred to [8], [9], [10] and the references
therein for details.
Estimation problems on [0, ∞) are first carried to ones on the interval [0, π/2) via the bilinear map Ω = µ tan(θ /2) where
we chose µ = 0.2. The number of the spectrum samples N equals to 63. Thus, we set p = 30 for the row dimension
of a Hankel matrix appearing in these algorithms to extract the so-called observability range space. Spectral factors were
extracted from spectral summand estimates by using the re-weighted and regularized nuclear norm minimization scheme
(RRNH) proposed in [11]. The convex surrogate and the regularization parameters δ and λ were fixed to 10−1 and 104 ,
respectively. Only in two iterations, convergence took place in all cases. To estimate the spectral summands, the regularized
nuclear norm heuristic (RNH) in [9] and the RRNH were started respectively with δ = 10−1 , λ = 10 and δ = 10−1 , λ = 102 .
The RRNH converged in three steps.
The spectral factor orders returned by the algorithms were 2, 9, and 3, respectively, for the subspace algorithm, the RNH,
and the RRNH. The misfit errors in the quadratic norm are respectively 0.0079, 0.0043, and 0.0045. Thus, the best rational
models are the ones returned by the subspace algorithm and the RRNH due to their small orders. In Figure 8, the estimation
results are plotted. The fits to the data achieved by the three algorithms are much better than the linear or piece-wise linear
fits in the waveband [0, 0.05]. This waveband can be pushed further down by increasing model orders though we will not
pursue this issue here.

−2
10
PSD (m /cycle/m)

−4
10
2

−6
10

−2 −1 0
10 10 10
Wave number (cycle/m)

Fig. 8. The spectral data and its modeling by rational spectra: “line and circle” the road data; “line and plus” the subspace estimate; “line and square”
the RNH; “line and triangle” the RRNH.

ACKNOWLEDGMENTS
This work was supported by the Scientific & Technological Research Council of Turkey under Grant 112E264.
V. C ONCLUSIONS
Terrain profiles, whether measured or modeled, are a useful tool for many groups, for example, civil engineers use terrain
profiles of highways for health monitoring to determine when a highway needs to be resurfaced and vehicle engineers use
terrain profiles for physical simulations on shaker rigs to test the durability of a vehicle. Since they are the consistent input
into vehicle suspensions, in order to properly set the vehicle response targets, it is necessary to understand the underlying
road.
In this paper simple analytical approximations to the road profiles by using linear and piece-wise linear spectrum modeling
were discussed. Usually, these simple parametric spectral densities will not accurately approximate the road roughness
spectrum for the frequency range of interest. Then, we introduced more sophisticated and realistic terrain models based on
system identification and compressive techniques.
R EFERENCES
[1] ISO 8608: Mechanical vibration-road surface profiles-Reporting of measured data, International Organization for Standardization, 1995.
[2] http://www.umtri.umich.edu/erd/roughness/lttp-erd.html, Long-term pavement performance road profile data, University of Michigan Transportation
Research Institute (UMTRI), accessed March, 2014.
[3] P. Andrén, “Power spectral density approximations of longitudinal road profiles,” Int. J. Vehicle Design, Vol. 40, pp. 2–14, 2006.
[4] R. P. La Barre, R. T. Forbes, and S. Andrews, “The measurement and analysis of road surface roughness,” Report 1970/5, Motor Industry Research
Association, Detroit, MI, 1970.
[5] BSI MEE/158/3/1, Proposals for Generalized Road Inputs to Vehicles, British Standards Institution (BSI72/34562), London, England, 1972.
[6] O. Kropáč and P. Múčka, “Non-standard longitudinal profiles of roads and indicators for their characterization, Int. J. Veh. Design, vol. 36 , pp.
149–172, 2004.
[7] W. Wen, Road roughness detection by analyzing IMU data, M. Sc. Thesis in Geodesy, KTH, Sweden, pp. 1–101, 2008.
[8] H. Akçay and S. Türkay, “Frequency domain subspace-based identification of discrete-time power spectra from non-uniformly spaced measurements,”
Automatica, vol. 40, pp. 1333–1347, 2004.
[9] H. Akçay, “Subspace-based spectrum estimation in frequency-domain by regularized nuclear norm minimization,” Signal Processing, vol. 99, pp.
69–85, 2014.
[10] H. Akçay and S. Türkay, “Identification of power spectra by re-weighted and regularized nuclear norm minimization,” submitted to the 10th Asian
Contr. Conf., Kota Kinabalu, Malaysia, May-June 2015.
[11] H. Akçay and S. Türkay, “Positive Realness in Stochastic Subspace Identification: A Regularized and Re-weighted Nuclear Norm Minimization
Approach,” submitted to ECC’15, Linz, Austria, July 2015.
[12] O. Kropáč and P. Múčka, “Effect of longitudinal road waviness on vehicle vibration response,” VSD, vol. 47, pp. 135–153, 2009.
[13] S. Türkay and H. Akçay, “A study on random vibration characteristics of the quarter-car model,” JSV, vol. 282, pp. 111–124, 2005.
[14] C. J. Dodds and J. D. Robson, “The description of Road Surface Roughness,” JSV, vol. 31, no 2. pp. 175–183, 1973.

You might also like