You are on page 1of 14

Original Article

Proc IMechE Part D:


J Automobile Engineering
226(9) 1171–1184
Image-based approach to optimize the Ó IMechE 2012
Reprints and permissions:
tyre pitch sequence for a reduction in sagepub.co.uk/journalsPermissions.nav
DOI: 10.1177/0954407012439950

the air-pumping noise based on a pid.sagepub.com

genetic algorithm

Eui-Youl Kim, Sung-Wook Hwang and Sang-Kwon Lee

Abstract
The paper presents a novel approach to solving problems involved in the application of a genetic algorithm to determine
the optimal tyre pitch sequence to reduce the tyre air-pumping noise which is generated by the repeated compression
and expansion of the air cavity between the tyre pitch and the road surface. The genetic algorithm was used to deter-
mine the optimal tyre pitch sequence with a low level of tyre air-pumping noise using the image-based air-pumping
model. In the genetic algorithm used in previous studies, there are a number of problems related to the encoding struc-
ture and the selection of an objective function. This paper proposes a single encoding element with five integers, a diver-
gent objective function based on an evolutionary process, and the optimal evolutionary rate based on the Shannon
entropy in an attempt to solve the problems. The results of the proposed genetic algorithm with an evolutionary process
are compared with those of a randomized algorithm. The randomized algorithm is a traditional method used to obtain
the tyre pitch sequence. It was confirmed that the genetic algorithm more effectively reduces the peak value of the pre-
dicted tyre air-pumping noise. The consistency and cohesion of the obtained simulation results are also improved.

Keywords
Tyre air-pumping noise, image-based noise prediction, genetic algorithm, Shannon entropy

Date received: 5 August 2011; accepted: 27 January 2012

Introduction is changed, the finite element model for a CFD analysis


must be modified repeatedly, which naturally increases
Tyre air-pumping noise is one of the major noise the cost and the time needed during the development
sources of noise inside a vehicle. In terms of psycho- process. A square-wave-based method13,14 was pre-
acoustics, the annoyance caused by noise sources inside sented to predict the tyre air-pumping noise roughly by
a passenger car highly depends on the relative relation- Fourier modelling the tyre pitch sequence with a square
ships between the noise sources.1–5 Therefore, in an wave. The computational load is relatively low in this
electric-powered vehicle,6–8 the tyre air-pumping noise method and the effects of minor pitch elements on the
becomes important because other noises such as engine tyre pitch length can be ignored owing to the limitations
noise and transmission noise are not present in the of the Fourier transform method; however, this is not
cabin of a vehicle. In this sense, this research aims to true for the groove elements, which are one of the major
reduce the tyre air-pumping noise which is generated by
pattern elements. An image-based analysis method15
the repeated compression and expansion of the air cav-
was introduced to predict the tyre air-pumping noise
ity between the tyre pitch and the road surface.9–11 It
focuses on how to predict the tyre air-pumping noise
and to optimize the tyre pitch sequence without using Acoustics and Vibration Signal Processing Laboratory, Department of
complicated numerical methods. For a prediction of the Mechanical Engineering, Inha University, Incheon, Republic of Korea
tyre air-pumping noise, a hybrid prediction method12 is
suggested that combines the computational fluid Corresponding author:
Sang-Kwon Lee, Acoustics and Vibration Signal Processing Laboratory,
dynamics (CFD) method and the Kirchhoff integral Department of Mechanical Engineering, Inha University, 253 Yonghyun-
method. However, its computational load is the highest. Dong, Nam-Gu, Incheon 402-751, Republic of Korea.
Moreover, whenever the shape of the tyre pitch length Email: sangkwon@inha.ac.kr
1172 Proc IMechE Part D: J Automobile Engineering 226(9)

Figure 1. The basic structure of the tyre pitch sequence optimization process using a genetic algorithm.
RMS: r.m.s.

directly by modelling the air pumping in a pixel-based computation and the low possibility of local conver-
image format. This method makes it possible to con- gence error. In the genetic algorithm used in previous
sider the effects of the depth and the minor pitch ele- research, there were some problems in the encoding
ments in detail with a relatively small amount of structure and in the selection of an objective function.
computation. The functional requirement to predict the This paper proposes a single encoding element with
tyre air-pumping noise is to ensure sufficient accuracy five integers, a divergent objective function based on an
within a short time in spite of the frequent modifica- evolutionary process, and the determination of the opti-
tions, as the tyre pitch length is frequently modified by mal evolutionary rate based on the Shannon entropy to
the close collaboration between the designer and the solve the aforementioned problems. The results of the
engineer at the beginning of the development process of proposed genetic algorithm with an evolutionary process
the tyre pattern. For this reason, the image-based are compared with those of a randomized algorithm.
method is popular. In the image-based method, the The randomized algorithm is a traditional method used
optimization algorithm for the optimal tyre pitch to obtain the tyre pitch sequence. It was confirmed that
sequence is important. Although the tyre air-pumping the genetic algorithm as an optimal algorithm more
noise is affected by the various design parameters and effectively reduces the peak value of the predicted tyre
the environment conditions, it can be modelled simply air-pumping noise. The consistency and cohesion of the
as a function of the tyre pitch pattern shape, the leading obtained simulation results are also improved.
edge shape and the pitch sequence, as shown in Figure
1. Among these three design parameters, the tyre pitch
Image-based tyre air-pumping noise
sequence is commonly recognized in many research
papers as an independent design parameter having a
analysis
high degree of freedom in reducing the air-pumping In the paper, the image-based air-pumping model is
noise. The performance levels of heuristic algorithms used to increase the efficiency of the noise prediction
were compared in the process of finding the optimal process. It is employed to predict the tyre noise gener-
tyre pitch sequence,13,14 and it was concluded that the ated by the repeated compression and expansion of the
genetic algorithm is more useful than other heuristic air cavity between the tyre pitch and the road surface,
algorithms owing to the relatively small amount of as shown in Figure 2. Its main advantage is the

Figure 2. The image-based tyre air-pumping noise prediction method using a simplified air cavity model.
Kim et al. 1173

possibility of directly predicting the air-pumping noise where A is the space term related to the area and in this
with sufficient accuracy over a short time from a pixel- case A = 1 since unit area is used. Equation (6) is sub-
based tyre pitch image without a complicated numeri- stituted back into equation (5) to obtain
cal process. The generation mechanism of the air-
pumping noise is derived by applying the Euler equa- 1 ∂2 h i, j  r  a
p9 i, j ðr, tÞ = r0 t  ð7Þ
tion for a point monopole source which is modelled as 4pr ∂t2 c
a spherical emitter of a radius element with oscillations
where the height of the cube refers to the depth of the
much less than the radius. If the spherical emitter is
tyre pitch image at the pixel location (i, j), and where
centred at the origin, all the wave propagation para-
the portable grey-map format has 8 bits per pixel to
meters are functions of only the radial distance r and
express the information on the contrast of the pixel-
the time t. The gradient of the pressure is in the radial
based image. Thus, the depth of the tyre pitch length
direction; the fluid velocity vector can also be in the
can be expressed with 256 steps. If the radial distance
radial direction. The wave equation in the radial direc-
and the temperature are constant, as shown in Figure
tion is written as
2, the sound pressure wave emitted from the pixel loca-
  tion (i, j) of the tyre pitch length can be predicted easily
1 ∂2 p9 1 ∂ 2 ∂p9
= 2 r ð1Þ by considering only the second derivative of the height
c2 ∂t2 r ∂r ∂r
change in equation (7). The non-calibrated tyre air-
Here, the gradient of the pressure is in the radial direc- pumping noise, as the sum of the unit area, is calculated
tion, c is the sound of speed, p# is the sound pressure by scanning the tyre pitch image together with the lead-
and the general solution of equation (10) for the sound ing edge and time. This is influenced by the tyre pitch
pressure is given by image, the leading edge shape and the pitch sequence,
as shown in Figure 1. Among the three design para-
fðr  ctÞ
p9ðr, tÞ = ð2Þ meters, the first two parameters are the dependent
r design parameters, as the tyre pitch image is not the
where f can be any function. Assuming that the size of the modification target but the validation target in the pres-
monopole source with radius a is much smaller than the ent step and because the leading edge shape is mainly
wavelength of sound, the surface vibration of the spheri- affected by the design of the tyre body and road condi-
cal emitter related to the sound pressure is written as tions. On the other hand, the tyre pitch sequence,
because of its high level of freedom in minimizing the
a2 ∂ua  r  a tyre air-pumping noise, is considered as the indepen-
p9ðr, tÞ = r0 t ð3Þ
r ∂t c dent parameter. In this study, for tyre pitch optimiza-
where r0 is the density of air and ua is the radial velocity tion, the basic tyre pitch length given by the Nexen
at radius a. The change in the volume V is related to the Tyre Corporation of Korea is used to find the optimal
radial velocity on the surface of the spherical emitter, tyre pitch sequence for the reduction in the tyre air-
according to pumping noise, as shown in Figure 3. The basic tyre
pitch length is modified at intervals of 10% in the direc-
∂V ∂a tion of travel, and a total of five types of tyre pitch
= 4pa2 = 4pa2 ua ð4Þ
∂t ∂t model from 80% to 120% is used for the simulation
and equation (4) is substituted back into equation (3) to based on the genetic algorithm. These pitch sequences
obtain are arranged by using the random function. The effect
of the leading edge shape is not considered here owing
1 ∂2 V  r  a to the use of the leading edge in the form of a straight
p9ðr, tÞ = r0 2 t  ð5Þ
4pr ∂t c line. The harmonic spectra of these pitch sequences are
analysed and their peak value and r.m.s. value were
which means that the acoustic sound pressure emitted
used as the cost function in the genetic algorithm. The
by a point monopole source is proportional to the sec-
r.m.s. value and the peak value were chosen as the vari-
ond derivative of the volume change. Here, as shown in
ables of interest in the harmonic spectrum of the pre-
Figure 2, the volume of a cube is a function of the
dicted sound pressure for the tyre air-pumping noise.
length l, the width w and the height h. To connect the
theoretical concept of the air-pumping noise model to
the pixel-based image format at the pixel location (i, j), Genetic algorithm
the cross-section of the volume should be the unit area.
In addition, the length and the width of the cube are This section presents the novel approach for the appli-
constant values of the pixel. Thus, the second derivative cation of the genetic algorithm to the determination of
of the volume change is simply expressed via the height the optimal sequence of a tyre pitch for low air-
change of the cube. It is given as pumping noise. Figure 4 shows a flow chart of the
genetic algorithm used throughout this research. The
∂2 ∂2 hi, j processes of crossover, mutation and pitch optimiza-
2
Vi, j ðl, w, hÞ = A 2 ð6Þ
∂t ∂t tion are required. In each process, novel approaches to
1174 Proc IMechE Part D: J Automobile Engineering 226(9)

Figure 3. The effects of tyre pitch sequences on the harmonic spectrum of the tyre air-pumping noise and the selected evaluation
criteria.
FFT: fast Fourier transform; RMS: r.m.s.

Figure 5. Structural difference between (a) the binary


encoding model and (b) the proposed encoding model.

encoding elements, in keeping with the eight empty


spaces.
Since only five spaces can be used to correspond to a
tyre pitch if we use the binary structure, three spaces
are always left behind. Therefore, in order to remove
these three empty spaces, in this paper the encoding
scheme based on the decimal number format is pro-
posed. The method uses an integer value between 1 and
5, and only five spaces are used to correspond to a tyre
pitch. In this case there is no empty space like binary
encoding. The five spaces consist of one unit as shown
Figure 4. A flow chart of the genetic algorithm. in Figure 5(b), which corresponds to one tyre pitch. In
order to express information in the five types of pitch as
shown in Figure 3, five units are necessary. One tyre
improve the convergence speed and stability are consists of 50 pitches (i.e. 50 units). In this paper, a
presented. chromosome consists of 50 units and includes informa-
tion on 50 pitches. If we use a binary structure, 150
units would be necessary for the expression of the pitch
Encoding design for tyre pitch sequences sequence of a tyre. A group consisting of 10 chromo-
Binary encoding using 0 and 1 is used widely to express somes is used for the genetic algorithm for optimization
the genetic information of chromosomes, as shown in of the pitch sequence. Ten chromosomes are repetitively
Figure 5(a). In the application of a genetic algorithm used in the evolution process of the genetic algorithm.
using binary encoding, two types of problem can arise Therefore, the proposed encoding structure can
because of the uncertainty of the correspondence express a tyre pitch without a waste of space using only
between the binary encoding structure and tyre pitches. a single encoding element. Thus, it is possible to reduce
The first problem is related to the waste of space inside the required matrix space size during the computational
the encoding structure. To express a tyre pitch, the process of the genetic algorithm, which naturally leads
encoding structure requires at least the three binary to a reduction in the computational load and a
Kim et al. 1175

Table 1. Comparison of the Shannon entropy and the entropy concluded that the proposed encoding structure is more
waste between the binary encoding model and the proposed appropriate for expressing a tyre pitch than the binary
encoding model. encoding structure is.
Encoding type Shannon Entropy
entropy waste
Design of a divergent objective function
Binary encoding (0,1) 3.0000 bits 0.9544 bits A chromosome includes the information on the pitch
Proposed encoding (1, 2, 3, 4, 5) 2.3219 bits 0.0000 bits
sequence in the circumference direction of a tyre. The
characteristic of the air-pumping noise is expressed in
reduction in the time needed for the repeated evolution- terms of the peak and r.m.s. values of the harmonic
ary cycles. The second problem is related to the absence spectrum of each chromosome. In the evaluation pro-
of immediacy in the application of the constraints. cess, the noise quality of a group consisting of 10 chro-
During each evolutionary process, the two types of mosomes is relatively evaluated. The group with good
constraint should be verified between the adjacent noise quality is dominant and the opposite group is
pitches owing to the problems related to the durability recessive. In order to determine whether or not a group
and the tyre air-pumping noise. is dominant throughout the genetic algorithm, the fit-
Because the proposed encoding scheme is based on ness of chromosomes is evaluated from the peak and
the decimal number format, without unnecessary ambi- r.m.s. values of each chromosome. The fitness of each
guity and complexity, the proposed encoding structure chromosome is used by comparing it with those of other
makes it easy to consider the constraints between the chromosomes. This has usually been calculated accord-
adjacent pitches in the application of the genetic algo- ing to a predefined convergent objective function in pre-
rithm. The usefulness of the proposed encoding struc- vious research.13–18 Therefore, the potential of the
ture can also be described by comparing the uncertainty optimization algorithm can be limited because of its
of the correspondence between the two different encod- convergence characteristic as regards the predefined tar-
ing structures quantitatively. In general, the uncertainty get. For example, even when the optimization algorithm
is expressed via the theoretical concept of the Shannon has a sufficiently high potential to optimize the tyre
entropy, which is written as pitch sequence, if the objective function has already
achieved the predefined target, the optimization process
X
N must stop. Moreover, if there is not sufficient informa-
E=  pi log2 (pi ) ð8Þ tion about the optimal absolute criteria, it is difficult to
i=1
define the target and to evaluate the fitness of the chro-
Here, pi is the probability of each case. Table 1 shows mosomes properly. For these reasons, in this paper, a
the Shannon entropy and the waste of space inside each divergent objective function is proposed. In the pro-
encoding structure. The uncertainty of the encoding posed method, the fitness of each chromosome obtained
structure was reduced from 3 bits to 2.322 bits at a level through the genetic algorithm is evaluated by compar-
of 77.4% with the proposed encoding structure. Thus, ing the relative positions in the two-dimensional normal
the waste of the Shannon entropy was reduced overall distribution of the reference chromosomes, as shown in
from 0.9544 bits to 0 bits. From these results, it can be Figure 6. This is calculated through the following

Figure 6. Two-dimensional normal distribution using two types of evaluation criterion.


RMS: r.m.s.
1176 Proc IMechE Part D: J Automobile Engineering 226(9)

procedures. The r.m.s. value x1(k) and the peak value


x2(k) are chosen as the variables of interest in the har-
monic spectrum of the predicted tyre air-pumping noise.
If x1(k) and x2(k) are uncorrelated, the two-dimensional
normal distribution can be decomposed into two parts,
after which the probability density function of each
variable can be written as
 
pffiffiffiffiffiffi (xi  m i )2
p(xi ) = (si 2p)1 exp , i = 1, 2 ð9Þ
2s2i
where mi and si are the mean value and the standard
deviation respectively of the reference chromosomes.
These are calculated using the equations

1 XN Figure 7. Evaluation of each chromosome based on the


mi = xi (k) ð10Þ distance from the reference line in the two-dimensional normal
N k=1
distribution.
1 XN RMS: r.m.s.
s2i = ½xi (k)  mi  2 ð11Þ
N k=1

Here, x1(k) and x2(k) are the r.m.s. value and the peak Thus, the fitness of each chromosome can be inter-
value respectively extracted from the number N of refer- preted as its opportunity to be selected on the roulette
ence chromosomes. Each variable of the chromosomes wheel, as shown in Figure 8. Figure 9 shows the two-
obtained through the genetic algorithm is normalized dimensional normal distribution of the reference chro-
according to mosomes as generated using the randomized algorithm,
and Table 2 shows the statistical information. The ran-
xi (k)  mi domized algorithm is a traditional method used to
zi (k) = ð12Þ
si obtain the tyre pitch sequence. An inverse proportional
where zi refers to the normalized position of each chro- relationship exists between the r.m.s. value and the
mosome in the two-dimensional normal distribution, peak value, as shown in Figure 9. This relationship
and equation (12) is substituted back into equation (9) reflects an acoustic–physical property of the spectrum
to obtain of the tyre air-pumping noise. It is difficult to improve
 2 both of these simultaneously. Here, the main feature of
pffiffiffiffiffiffi zi the proposed divergent objective function is that the
p(zi ) = ( 2p)1 exp ð13Þ
2 potential of the optimization algorithm can be maxi-
mized without any limitation, unlike the existing con-
Because chromosomes having a high value in the direc- vergent objective function, because the optimization
tion of the negative z1 axis and negative z2 axis are con- process started from the initial position in the form of
sidered dominant objects, as shown in Figure 7, the
distance d(k) between a chromosome and the evalua-
tion reference z2(k) = 2z1(k) is defined as the sum of
the squared variables according to

d2 (k) = ½z1 (k)  a(k)2 + ½z1 (k)  b(k)2 ð14Þ


where k is the integer number corresponding to one
chromosome among 10 chromosomes in the one group
and d(k) means one of the noise characteristics of one
chromosome The fitness of each chromosome is calcu-
lated on the basis of the roulette wheel principle19,20
and is a function of the normalized r.m.s. value and the
normalized peak value. According to the roulette wheel
principle, the total area corresponds to the summation
d2(k) of the variation in each chromosome, and the
probability P of each chromosome is expressed by
" #1
X
10
2
P= d (k) ½ d2 (1) d2 (2)   d2 (k)  d2 (10) 
k=1
Figure 8. The roulette wheel method for selecting a pair of
ð15Þ dominant chromosomes in the crossover process.
Kim et al. 1177

Figure 9. The two-dimensional normal distribution of the


reference chromosomes generated by the randomized
algorithm.
RMS: r.m.s.

Table 2. The characteristics of the reference chromosomes


generated by the randomized algorithm in the two-dimensional
normal distribution.

Parameter type Mean Standard deviation Figure 10. A flow chart of the crossover process and the
mutation process with constraints between adjacent pitches.
Parameter I (r.m.s. value) 0.3163 1.9276 3 104
Parameter II (peak value) 2.9102 0.4212

does not come from the parent chromosomes. The


mutation process arises only at a randomly selected
divergence without a predefined target. The dominant position inside each chromosome as an unexpected
chromosomes having a high level of the fitness have a change. Therefore, in the process of the evolutionary
greater opportunity to be selected as the parent chro- convergence, the diversity of the genetic information is
mosomes in an evolutionary process such as crossover more sensitive to the mutation process than to the
and mutation through the proposed genetic algorithm. crossover process. In the application of the genetic algo-
This evolutionary process is explained in the next two rithm, the selection of the evolutionary rate is an impor-
sections. tant process. Here, the evolutionary rates need to be
selected carefully because an inverse proportional rela-
tionship exists between the reduction in the tyre air-
Evolutionary process pumping noise and the computational load. Thus, the
Crossover and mutation are the core parts of the optimal evolutionary rates were selected after identify-
genetic algorithm, as shown in Figure 10. First, the ing the overall tendency obtained through the suitabil-
exchange of the genetic information between a pair of ity analysis for various combinations of parameters.
parent chromosomes is conducted through the cross- Additionally, according to the pitch sequence condition
over process. In general, the dominant chromosomes between the adjacent tyre pitches, the changes in the
having a good fitness have greater opportunities to be genetic information can cause problems in terms of the
selected as parent chromosomes. Therefore, the genetic durability and the noise. Therefore, as shown in Figure
information of the recessive chromosomes cannot be 10, two constraints were considered between the adja-
passed on to the offspring chromosomes and will be cent tyre pitches, as mentioned in the section on the
eventually disappear in the process of evolutionary con- design of a divergent objective function. The first con-
vergence. Whether or not the genetic information is straint suppresses the rapid change between the adja-
exchanged between a pair of chromosomes is wholly cent tyre pitches due to the durability, which was
determined according to the crossover rate. In this limited within a maximum of two steps. The second
paper, five pairs of the parent chromosomes are selected constraint prevents the continuous repetition of the
from among a total of 10 chromosomes, including repe- same tyre pitch for a reduction in the tonal noise com-
tition. Unlike the crossover process, the change in the ponents in the tyre air-pumping noise, which is limited
genetic information caused by the mutation process within a maximum of two repetitions. After exchanging
1178 Proc IMechE Part D: J Automobile Engineering 226(9)

the genetic information between a pair of chromosomes diversity of the genetic information that the offspring
in the crossover process, if these two constraints are sat- chromosomes can have also becomes maximized. In the
isfied, the evolutionary process continues to the muta- repeated evolutionary cycles, the offspring chromo-
tion process. However, if the offspring chromosomes somes, through the crossover process and the mutation
cannot satisfy the constraints, the change in the genetic process, become similar to the specific chromosome
information will be restored to the previous state, after including the dominant genetic information. The diver-
which the crossover process will be repeated until its sity and the Shannon entropy also gradually decrease
constraints are satisfied. The same procedure is also from their initial states in the direction towards zero. In
applied to the mutation process, as shown in Figure 10. the converged state, the diversity of the genetic informa-
tion among the chromosomes no longer exists, and the
Shannon entropy also reaches a value of zero, which
Exit conditions means that the additional evolutionary cycle has no
Excessive repetitions of evolutionary cycles can cause meaning at this point. Based on the evolutionary con-
an abnormal convergence and increase the computa- vergence process and its relationship with the Shannon
tional load. Therefore, the exit conditions of the evolu- entropy, to prevent the excessive repetitions of evolu-
tionary process should be configured carefully, in tionary cycles, two exit conditions were applied as
common with other optimization algorithms. In this shown below. The first condition is as follows: if the
paper, as mentioned in the section on the design of a Shannon entropy value is less than a predefined zero
divergent objective function, the diversity of the genetic threshold, the evolutionary process converges to the
information among the chromosomes is used as a quan- dominant chromosome because 10 chromosomes have
titative basis to determine when the evolutionary cycle the same information about pitch sequences as each
other. The second condition is that, if the evolutionary
should be stopped. The Shannon entropy is used as the
rates have an excessive value, the convergence process
objective indicator to express the proportional relation-
becomes unstable and diverges, after which it will be
ship between the evolutionary convergence process and
repeated continuously. Therefore, upon exceeding 100
the diversity of the genetic information and to express
times, the evolutionary cycle was stopped.
the uncertainty of a system in various fields. Figure 11
shows the changes in the genetic information inside the
chromosomes between the initial state and the con- Selection of the optimal evolutionary
verged state of the evolutionary cycles. In the initial
rates
state, the Shannon entropy value is close to the maxi-
mum entropy value of 2.323 bits, as the five types of The diversity of the genetic information inside the
tyre pitch have equal probabilities. This means that the chromosomes and the convergence characteristics are

Figure 11. The relationship between the Shannon entropy and the diversity of genetic information in the process of evolutionary
convergence.
Kim et al. 1179

affected by the crossover rate and the mutation rate, the mutation rate rather than on the change in the
which are also closely related to the characteristics of crossover rate, and the number of cycles was increased
the tyre air-pumping noise. In general, although the suddenly around a mutation rate of 14%. Figure 12(b)
increase in the evolutionary rates provides more oppor- shows the relationship between the two types of evolu-
tunities for reducing the tyre air-pumping noise, it can tionary rate and the number of reversal phenomena in
also cause a greater computational load. For this rea- the Shannon entropy, which is similar to the results
son, to find the optimal evolutionary rates at which the shown in Figure 12(a). The number of reversal phenom-
diversity of the genetic information and the computa- ena was suddenly increased around a mutation rate of
tional load can be balanced, a suitability analysis was 15%. Figure 12(c) and (d) shows the relationship
conducted. Here, various combinations of parameters between the two types of evolutionary parameter and
were considered as described below. The crossover rate the characteristics of the tyre air-pumping noise. The
was divided into 11 cases from 0% to 100% at intervals r.m.s. value is influenced by the change in the crossover
of 10%, after which the mutation rate was divided into rate, which is between 22.2s and 21.4s. On the other
21 cases from 0% to 20% at intervals of 1%. Therefore, hand, the peak value is influenced by the change in the
231 cases were considered in the process of the suitabil- mutation rate rather than by the change in the cross-
ity analysis. In each case, the genetic-algorithm-based over rate, which is between 21.8s and 21.4s. From
simulation was repeated 100 times to derive the general- these simulation results, it was found that the r.m.s.
ity from the predicted tyre air-pumping noise in terms value and the peak value extracted from the predicted
of the probability. Figure 12(a) shows the relationship tyre air-pumping noise are commonly minimized with
between the two types of evolutionary rate and the the increases in the two evolutionary rates. In particu-
number N of cycles required for evolutionary conver- lar, the overall tendency of the simulation results sensi-
gence. The overall tendency depends on the change in tively depends on the changes in the mutation rate

Figure 12. The effects of the crossover rate and the mutation rate on the evolutionary convergence process: (a) the number
of required evolutionary cycles; (b) the number of reversal phenomenon in the Shannon entropy; (c) parameter I (r.m.s. value);
(d) parameter II (peak value).
1180 Proc IMechE Part D: J Automobile Engineering 226(9)

Figure 13. Suitability assessment for finding the optimal Figure 14. Comparison of the two-dimensional normal
evolutionary rates. distribution generated by the different algorithms.
RMS: r.m.s.

Table 3. The selected crossover rate and mutation rate for the
optimal evolutionary process. Figure 14 shows the normal distributions of 1000
objects obtained using the randomized algorithm and
Crossover Mutation the genetic algorithm. The x axis is the normalized
r.m.s. value and the y axis denotes the normalized peak
pc = 0.70 pm = 0.14 value in the harmonic spectrum of the predicted tyre
air-pumping noise. Through the application of the
genetic algorithm, the normal distribution of 1000
rather than on the changes in the crossover rate, as a objects was moved in the direction of the negative x
change in the genetic information in the mutation pro- axis and the negative y axis and the cohesion increased.
cess, unlike the crossover process, does not come from From these results, it can be concluded that the genetic
the parent chromosomes as unexpected changes. Figure algorithm is more effective at reducing the tyre air-
13 shows a colour map of the global suitability index pumping noise and increasing the consistency of the
for 231 cases. Here, the global suitability S is defined as obtained simulation results than the randomized algo-
the mean value of the local suitability factors sˆi. It can rithm is. However, the inverse proportional relation-
be written as ship between two extracted features in the harmonic
spectrum still exists. Figure 15 shows the change in the
1X N
probability density in the normal distribution. As listed
S= si
^ ð16Þ in Tables 4 and 5, the central position of the normal
N i=1
distribution was commonly moved from 0 to a value
where N = 4 and the local suitability factors are around 2s on the normalized x axis. Figure 16 shows
obtained by normalizing the four types of simulation the changes in the Shannon entropy during the process
result, as shown in Figure 12. This is expressed as of evolutionary convergence; this value decreased from
si 1.963 bits to 0 bits with increasing number of genera-
s^i = for i = 1, 2, 3, 4 ð17Þ tions. When the five types of tyre pitch have equal
ks i k
probabilities of being selected, as listed in Table 6, the
The optimal evolutionary rates were positioned around diversity of the genetic information inside the chromo-
the highest value of the global suitability. As listed in somes can have a maximum Shannon entropy value of
Table 3, a crossover rate of 70% and a mutation rate 2.329 bits. This implies that the initial chromosomes
of 14% were chosen as the optimal evolutionary rates. that were generated using the randomized algorithm
have a high level of diversity close to the maximum
level of the Shannon entropy. During the repeated evo-
Simulation results lutionary cycles, owing to the selected mutation rate,
To compare with the results of the randomized algo- the reversal phenomenon of the convergence curve
rithm in terms of the probability, the genetic algorithm- occurred at around generation 4. The diversity of the
based simulation was repeated 1000 times with the genetic information was increased, which provides
optimal evolutionary rates obtained through the suit- more opportunities for reduction in the tyre air-
ability analysis in the previous section, after which the pumping noise. At generation 13, all the chromosomes
tyre pitch sequences of 1000 objects were obtained. have the same genetic information, and the Shannon
Kim et al. 1181

Figure 15. Comparison of the probability density in the normal distribution: (a) parameter I (r.m.s. value); (b) parameter II (peak
value); (c) parameters I and II.
RMS: r.m.s.

Table 4. Comparison of the characteristics of the chromosomes generated using different algorithms in the two-dimensional
normal distribution.

Randomized algorithm Genetic algorithm


Mean Standard deviation Mean Standard deviation

Parameter I (r.m.s. value) 0.3163 1.9276 3 1024 0.3161 3.1476 3 1026


Parameter II (peak value) 2.9102 0.4212 2.5022 1.6168 3 1022

Table 5. Comparison of the probability densities in the two-dimensional normal distribution generated using different algorithms.

Parameter I Parameter II Parameters I and II


Randomized Genetic Randomized Genetic Randomized Genetic
algorithm algorithm algorithm algorithm algorithm algorithm

3s 0.0020 0.0000 0.0150 0.0000 0.0150 0.0000


2s 0.0510 0.0000 0.0630 0.0013 0.0630 0.0000
s 0.2760 0.0126 0.2020 0.0239 0.2020 0.0000
0 0.3760 0.2290 0.3710 0.1893 0.3710 0.0074
2s 0.2200 0.4580 0.3150 0.5729 0.3150 0.5443
22s 0.0670 0.2509 0.0320 0.2125 0.0320 0.4384
23s 0.0080 0.0466 0.0000 0.0001 0.0000 0.0099
1182 Proc IMechE Part D: J Automobile Engineering 226(9)

Table 7. Comparison of the characteristics in the predicted


tyre air-pumping noise generated using different algorithms.

Algorithm type Parameter I Parameter II


(r.m.s. value) (peak value)

Randomized algorithm 0.3166 3.4588


Genetic algorithm 0.3159 (99.78%) 1.7239 (49.84%)

type of pitch, the peak value of the harmonic spectrum


should appear at around the fiftieth order. However, in
this paper, in order to express the pitch sequence of a
tyre with 50 pitches, five types of pitch model are used.
In the optimal condition optimized by the genetic algo-
rithm, ten pitches consist of one type of pitch model.
Figure 16. The change in the Shannon entropy in the process
Therefore the peak value of the harmonic spectrum
of evolutionary convergence.
appears at the tenth order. This means that the acoustic
energy with a single tone was scattered to avoid the
annoyance due to the tonal air-pumping noise. From
Table 6. The change in the Shannon entropy in the process of
these results, it can be concluded that the genetic algo-
the evolutionary convergence.
rithm is more effective at reducing the peak value than
Generation 1 5 9 13 the r.m.s. value in the harmonic spectrum of the pre-
dicted tyre air-pumping noise. Figure 18 shows the con-
Shannon entropy 1.9634 0.8868 0.2228 0.0000 vergence process in the two-dimensional normal
distribution during the process of the repeated evolu-
tionary cycles. Figure 19 shows the changes in the tyre
entropy inside the chromosomes also reaches a value pitch sequences between the initial chromosomes and
close to zero. Therefore, according to the predefined the converged chromosomes. These results show the
exit conditions, the genetic-algorithm-based simulation relationship between the diversity of the genetic infor-
was stopped at generation 13. Figure 17 and Table 7 mation and the convergence process well.
show the difference in the averaged tyre pitch spectra
between the initial chromosomes before the evolution- Summary and conclusions
ary process based on the random algorithm and the
converged chromosomes after the evolutionary process In this paper, a tyre pitch sequence was optimized using
based on the genetic algorithm. The r.m.s. value was an image-based air-pumping model and a genetic algo-
reduced from 0.3166 to 0.3159 at a level of 99.78%, rithm in an effort to reduce the tyre air-pumping noise,
and the peak value was reduced from 3.4588 to 1.7239 as the tyre air-pumping noise is strongly affected by
at a level of 49.84%. In Figure 17, if the number of changes in the tyre pitch sequence compared with other
pitches used for a tyre is 50 and they all have the same design parameters. First, the image-based air-pumping
model was used to predict the air-pumping noise
directly from a pixel-based tyre pitch length without a
complicated numerical process. This method is more
efficient than the method based on CFD or an experi-
mental method when the tyre pitch changes length fre-
quently. In the application of the genetic algorithm, a
binary encoding model using a value between 0 and 1
incurs two problems related to the complexity of the
existing encoding structure and the ambiguity in the
process of corresponding to the tyre pitch sequence.
Thus, a new encoding model using an integer between 1
and 5 was proposed. The proposed encoding model
makes it easier to consider the constraints between adja-
cent pitches without unnecessary ambiguity in terms of
the probability. Unlike the existing objective function
having convergence characteristics, the divergent objec-
tive function was suggested to evaluate the fitness of the
tyre pitch sequence generated by the randomized algo-
Figure 17. Comparison of the tyre air-pumping noise spectra. rithm or the genetic algorithm. The fitness of each
Kim et al. 1183

Figure 18. The evolutionary convergence process of 10 chromosomes using the genetic algorithm before the evolutionary process
( ) and after the evolutionary process (O): (a) generation 1; (b) generation 5; (c) generation 9; (d) generation 13.
RMS: r.m.s.

the basis of the relative position in the two-dimensional


normal distribution. Here, the r.m.s. value and the peak
value were used as the variables of interest. The opti-
mized evolutionary rates in the genetic algorithm were
selected through a suitability analysis based on the rela-
tionship between the characteristics of the predicted
tyre air-pumping noise and the computational load.
The results of the suitability analysis showed that the
diversity of the genetic information among the chromo-
somes is more sensitive to the mutation process than to
the crossover process in the process of evolutionary
convergence. The diversity of the genetic information in
the chromosomes was used as a quantitative basis in
order to prevent the excessive repetition of the evolu-
tionary cycles, as calculated using the Shannon entropy.
The results obtained through the above process based
Figure 19. The change in the combined pitch sequences in 10 on the genetic algorithm were compared using a rando-
chromosomes through the evolutionary convergence: (a) initial mized algorithm regarding the relative location in the
chromosomes; (b) converged chromosomes. two-dimensional normal distribution. As a result, the
centre of the two-dimensional normal distribution was
chromosome was used as the decision reference to iden- moved in the direction of improvement and the cohe-
tify whether or not the tyre pitch sequence has the siveness of the normal distribution increased. The r.m.s.
dominant genetic information; this was calculated on value was reduced from 0.3166 to 0.3159 at the level of
1184 Proc IMechE Part D: J Automobile Engineering 226(9)

99.78%. The peak value was reduced from 3.4588 to 9. Hayden RE. Roadside noise from the interaction of a
1.7239 at the level of 49.84%. From these results, it can rolling tire with road surface. In: Purdue noise control
be concluded that the genetic algorithm is more effec- conference, West Lafayette, IN, USA, 1971, pp.62–67.
tive at reducing the peak value than the r.m.s. value in West Lafayette, IN: Purdue University.
the harmonic spectrum of the predicted tyre air- 10. Ejsmont JA, Sandberg U and Taryma S. Influence of
pumping noise; moreover, the proposed method has a tread pattern on tire/road noise. SAE paper 841238,
1984.
sufficient level of usefulness to serve as part of the pro-
11. Jennewein M and Bergmann M. Investigations concern-
cess of optimizing the tyre pitch sequence to reduce the ing tyre/road noise sources and possibilities of noise
tyre air-pumping noise. reduction. Proc IMechE Part D: J Automobile Engineer-
ing 1985; 199(3): 199–206.
Funding 12. Kim S-T, Jeong W-T, Park Y-H and Lee S-G. Prediction
method for tire air-pumping noise using a hybrid tech-
This work was supported by an Inha University nique. J Acoust Soc Am 2006; 119(6): 3799–3812.
Research Grant and by the Mid-career Researcher 13. Becker M. Genetic algorithms for noise reduction in
Program through a National Research Foundation tire design. In: 2006 IEEE international conference on
of Korea grant (grant no. 2010-0014260) funded by systems, man, and cybernetics, Taipei, Republic of China,
the Korean Ministry of Education, Science and 8–11 October 2006, Vol 6, pp.5304–5308. New York:
Technology. IEEE.
14. Becker M, Jaschke S and Szczerbicka H. Tread profile
optimization for tires with multiple pitch tracks. In:
References IEEE 13th international conference on intelligent engi-
1. Kim H-W, Lee S-K and Na E-W. Sound quality evalua- neering systems, Barbados, 16–18 April 2009, pp.125–
tion of the impact noise induced by road courses having 130. New York: IEEE.
an impact bar and speed bumps in a passenger car. Proc 15. Mundl R, Fischer M, Wiese K et al. Virtual pattern opti-
IMechE Part D: J Automobile Engineering 2010; 224(6): mization based on performance prediction tools. Tire Sci
735–747. Technol 2008; 36(3): 192–210.
2. Lee H-H and Lee S-K. Objective evaluation of interior 16. Chen X, Chen L, Chen Y and Xiao W. Adaptive immune
noise booming in a passenger car based on sound metrics genetic algorithm for tire tread pattern pitch parameters
and artificial neural networks. Appl Ergonomics 2009; optimization. In: 3rd international symposium of intelligent
40(5): 860–869. information technology application, Nanchang, People’s
3. Lee H-H, Kim S-J and Lee S-K. Design of new sound Republic of China, 21–22 November 2009, vol. 1, pp.52–
metric and its application for quantification of an axle 55. New York: IEEE.
gear whine sound by utilizing an artificial neural net- 17. Che Y and Chen L. Method of summation and intelligent
work. J Mech Sci Technol 2009; 23(4): 1182–1193. optimization for tire tread noise based on adaptive
4. Kim T-G, Lee S-K and Lee H-H. Characterization and immune genetic algorithm. In: 2009 international confer-
quantification of luxury sound quality in premium-class ence on energy and environment technology, Guilin,
passenger cars. Proc IMechE Part D: J Automobile Engi- Guangxi, People’s Republic of China, 16–18 October
neering 2009; 223(3): 343–353. 2009, pp.392–395. New York: IEEE.
5. Zwicker E and Fastl H. Psychoacoustics: facts and mod- 18. Li X-H, Guo B and Yang H-Y. Application of tread pat-
els, 2nd edition. Berlin: Springer, 1999. terns noise-reduction based on fuzzy genetic algorithm.
6. La C, Poggi M, Murphy P and Zitko O. NVH considera- Fuzzy Inf Engng 2009; 62: 1141–1148.
tions for zero emissions vehicle driveline design. SAE 19. Davis L. Handbook of genetic algorithms. New York: Van
paper 2011-01-1545, 2011. Nostrand Reinhold, 1991.
7. Rust A and Graf BJ. NVH of electric vehicles with range 20. Goldberg DE. Genetic algorithms in search, optimization
extender. SAE paper 2010-01-1404, 2010. and machine learning. Reading, MA: Addison-Wesley,
8. Genuit K. The future of NVH research – a challenge by 1989.
new powertrains. SAE paper 2010-36-0515, 2010.

You might also like