Applied Acoustics 120 (2017) 75–84

Contents lists available at ScienceDirect

Applied Acoustics
journal homepage: www.elsevier.com/locate/apacoust

On the design and optimization of acoustic network resonators for

tire/road noise reduction
Bin Wang, Denis Duhamel ⇑
Université Paris-Est, Laboratoire Navier, ENPC-IFSTTAR-CNRS, UMR 8205, Ecole des Ponts ParisTech, 6-8 Avenue Blaise Pascal, 77455 Marne-la-Vallée, France

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

Article history: In this work we propose a numerical method for the calculation of resonant frequencies of network res-
Received 18 January 2016 onators and we also present an optimization method based on genetic algorithms to get targeted reso-
Received in revised form 14 October 2016 nant frequencies of the network resonators. We can optimize parameters of the network structure
Accepted 11 January 2017
such as junction types and end positions. Experiments are conducted on optimized wooden network res-
onators to validate the method. Good agreement is found between the measured and targeted resonant
frequencies. Applications to tire/road noise are considered.
Keywords:
Ó 2017 Published by Elsevier Ltd.
Acoustic network resonator
End correction
Resonant frequency
Optimization

1. Introduction quencies, we can see a maximum sound absorption. The theory

and applications of quarter-wave resonators are summarized in
Tire treads and road textures in the contact zones between tires [5].
and roads can be considered as acoustic network resonators. Con- The quarter-wave tube has an open and a closed end, but res-
sequently, the acoustic fields around the tire/road systems are onators with two open ends can also be used for the sound absorp-
influenced by network resonances. These network resonances in tion, especially for the case where air needs to be transported
the contact zone in Fig. 1 are seen as one of the noise enhancement through walls or one needs to see through the wall. Eerden [4]
mechanisms in [1]. Since the network resonators in the contact studied this case, and concluded that at low frequencies
zone have large influence on the acoustic fields around their reso- ðf < 2000 HzÞ the waves propagating in the resonator are not
nant frequencies, the network resonators should be investigated in absorbed at the end but are reflected back into the resonator due
detail. First studies of some simple acoustic resonators are to the mass reactance at the free end. For higher frequencies
reviewed as follows. (2000–10,000 Hz) the waves are absorbed due to radiation into
Besides porous materials and perforated panels, narrow infinity. In order to create broadband sound absorption, coupled
quarter-wave tube resonators are also widely used for the sound tube resonators with different cross-sectional areas and lengths
absorption in a wall or panel for a narrow frequency band based can be applied. The mechanism for the broadband absorption is
on the resonance of air inside the tube and the viscous shear and that the sound energy is dissipated by the viscothermal effects
thermal conductivity losses on the tube walls. The model by Zwik- and the incident waves are cancelled due to the broadband reso-
ker and Kosten [2] for wave propagation in cylindrical tubes nance of air in the coupled resonators.
included the viscosity and thermal conductivity. Tijdeman [3] Helmholtz resonators (HRs) are also used to control the noise
proved that this model is complete and accurate for both narrow inside enclosures in many studies. Helmholtz resonators can be
and wide tubes. Eerden [4] studied the influence of the viscous considered as a mass-spring system. The spring stiffness is repre-
and thermal conductivity losses on the absorption coefficient and sented by the volume of air and the mass is given by the small col-
concluded that the viscothermal effects cannot be neglected if umn of vibrating air in a perforation of the panel. The energy can be
the resonators are used for sound absorption because they result dissipated by the vibrating air and the porous material placed in
in energy being dissipated and the effective speed of sound inside the volume. See [6–10] for different applications of these Helm-
the tube can be considerably reduced. Around the resonant fre- holtz resonators to noise reduction. More specifically T-shaped
acoustic resonators can be seen in many studies for noise control
⇑ Corresponding author. in small enclosures, see for instance [11–16] for models and exper-
E-mail address: [email protected] (D. Duhamel). imental results on these resonators.

http://dx.doi.org/10.1016/j.apacoust.2017.01.017
0003-682X/Ó 2017 Published by Elsevier Ltd.
76 B. Wang, D. Duhamel / Applied Acoustics 120 (2017) 75–84

or closed ends. For the open ends, end corrections depend on flange
shapes. So, first, methods for the calculation of end corrections will
be introduced in Section 2. Next, in Section 3, a numerical method
for the calculation of resonant frequencies of network resonators
will be developed. Then, in Section 4 an optimization method will
be proposed to get the targeted resonant frequencies. Some exam-
ples of the application of the optimization method are shown in
Section 5 while comparisons with experimental measurements
Fig. 1. Network resonators in the contact zone between a tire and a road. are given in Section 6. Last some conclusions will be given.

If we are interested more specifically in tire noise, today the 2. Determination of end corrections
noise due to the vibrations of a rolling tire can be calculated with
convincing accuracy. However, air pumping is not understood very The length is an important parameter for calculating the reso-
well. In [17] Hayden described the air movement in the contact nant frequencies of a network with open ends. A short distance
zone between a rolling tire and a road. Air is squeezed out when should be added to each end of the network to get precise results.
the treads at the entrance of the contact zone are compressed on This short distance is called the end correction, which makes each
the road surface, and flows into the voids when the treads lift up straight part of the network a little longer than its physical length.
from the road surface. Daffayet et al. [18] measured the pressure From the perspective of waves, standing waves occur during the
in cylindrical cavities over which a smooth tire rolled. They network resonances. The sound waves are reflecting at open ends,
assumed that the noise is generated by opening and closing the which are not perfectly at the end sections of the network, but at
cavities in the contact zone. Ronneberger [19] thought that air small distances (end corrections) outside the network.
was displaced by the changing gaps between the tire and road sur- The end corrections of the network open ends can be obtained
faces, because the treads are deformed by road roughness. These from the radiation impedances which have small but finite values
sources are located in the contact zone between the tire and the by (1) from [27].
road and the sound is modified by the horn effect. 

Horn effect is an essential noise enhancement mechanism. The ~d ¼ Re k1 arctan Z r ð1Þ
iqc
tire/road system can be seen as a horn-like structure. The surfaces
of the tire and the road constitute horns in front of and behind the The upper script  means that it is a frequency-dependent
contact zone. The noise generated in the contact zone is amplified quantity. Here only the real part of the end correction is consid-
by the multiple reflections between the tire surface and the road ered, which is the most useful in the present study. To estimate
surface which are acoustically reflecting surfaces. The amplifica- the end corrections of the network, the radiation impedances of
tion of the horn effect reaches up to 10–20 dB in the results of pre- the open ends should be calculated first by the impedance transfer
vious studies, where the road and the tire are modeled with equation of an acoustic transmission line (2) from [27], because Z r
smooth surfaces. A first attempt at an analytical description of cannot be calculated or measured directly at the pipe end.
the horn effect was made by Ronneberger [20]. Kropp et al. [21] 
 
suggested a theoretical model based on multipole synthesis. The Z l
Z r ¼ iqc tan arctan  kl ð2Þ
model can provide a reasonable prediction of noise levels at mid iqc
and high frequencies for a tire placed on a hard surface. However,
it overestimates the horn amplification effect at low frequencies. Z l is the impedance at an abscissa x ¼ l, i.e., at a distance l from the
Graf et al. [22,23] first investigated experimentally the horn ampli- open end. It can be calculated as Dalmont did using a BEM numer-
fication of sound generated by a simple acoustic source. The ical method [27].
boundary element method is then shown to give predictions. The Eq. (2) means that the radiation impedance can be obtained
dependence of the horn-effect on different geometrical parameters from the case where Z l is an input impedance of a pipe of length
is also investigated both through experiments and boundary ele- l. Then the end corrections can be obtained by (1). The values of
ment calculations. It shows that for the intermediate frequency end corrections depend on flange shapes. An open pipe end with
range the BEM provides an excellent tool to calculate the horn different flange shapes has different end corrections. In this
effect for practical geometries. The aim of the work by Anfosso section, end corrections of a network with complex flanges are
et al. [24,25] is also to predict the amplification due to horn effect. calculated by Dalmont’s methods or by the BEM. The flanges are
Sound pressure amplification of a 2D infinite rigid cylinder is a round surface and a plane surface (see Fig. 2), which will be used
obtained using the analytical approach based on modal decompo- in Section 4. The network is shown in red. Some pipes are identi-
sition of sound pressure. It gives quick and accurate results, but is fied by the numbers one to three shown in green.
limited to simple geometrical configurations and purely reflecting Only the end corrections of longitudinal pipes will be discussed.
properties of boundaries. In [26] Fadavi et al. deal with the horn For transverse pipes they can be calculated approximately by Dal-
effect using a 3D cylinder tire model. The sound pressure and mont’s fit formula (3) for rectangular flanges because the flange of
sound amplification are calculated in the space around the 3D tire transverse ends is flat.
model using the Boundary Element Method. The influence of dif- "
5 #
ferent parameters such as the position and size of the source are asq asq asq
dsq ¼ dsq1 þ ðdsq0  dsq1 Þ þ 0:057 1 aeff ð3Þ
studied in terms of amplification and sound pressure spectrums. bsq bsq bsq
All these studies are made for smooth roads and tires and do not pffiffiffiffi
take into account the real geometry of the tire or the road. with dsq1 ¼ 0:811aeff and dsq0 ¼ 0:597aeff , in which aeff ¼ 2asq = p.
In this work, we want to estimate the influence of non smooth Here, 2bsq is the flange width, respectively. bsq is the shorter one
geometries on the horn effet. For this, we focus on network res- of the two distances between the transverse end and the sides AD
onators and use several assumptions for the networks. There are or BC in Fig. 2 for which the value of bsq for pipe number 2 is shown.
only right-angled junctions in the networks. The pipes in the net- 2asq ¼ 0:009 m is the pipe width which is small and not shown in
works have the same cross-section. The networks could have open the figure.
 B. Wang, D. Duhamel / Applied Acoustics 120 (2017) 75–84 77

The acoustic velocity at the closed end of the pipe is unity, so

the impedance of the piston is obtained. Then the radiation impe-
dance is calculated from this impedance by Eq. (2).
The end corrections of the longitudinal pipe for cylinders of dif-
ferent widths are shown in Fig. 4 for ka < 0:23, and the frequency
interval is 200 Hz. They are obtained by BEM as no analytical for-
mula can be used in this case. For each width of the cylinder, the
end corrections are frequency-dependent, but their standard
deviations are very small (see Fig. 5), which indicates that they
tend to be very close to the mean values. The standard deviation
is calculated by Eq. (4). So the mean values of end corrections for
each width in Fig. 5 are used to get a fit formula of end corrections
(5), in which w is half of the width.
rffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi
Fig. 2. A network (in red) between a round surface and a rigid plane surface. (For 1 XN
interpretation of the references to colour in this figure legend, the reader is referred r¼ i¼1 i
ðx  lÞ2 ð4Þ
N
to the web version of this article.)
P
with l ¼ N1 Ni¼1 xi , where xi is the end correction at each frequency
The longitudinal pipes at different positions have different
shapes and dimensions of the flanges. Although their ends may and N is the number of frequencies calculated.
not be in the middle of the flanges, in the interest of simplicity they d ¼ 0:00808 þ 0:22128w  3:72112w2 þ 19:80897w3 ð5Þ
are approximately assumed in the middle of the flanges. The round
surface in Fig. 2 is a cylinder. In the contact zone, the cylinder sur- This fit formula will be used in Section 4. The same procedure
face is flat. can be used to get fit formulas for other pipes and flanges of differ-
In order to calculate the end corrections of longitudinal pipes, ent shapes and dimensions.
half a cylinder is meshed for BEM, which is shown in Fig. 3a and
b. This is a longitudinal pipe, like pipe 1 in Fig. 2, but located at 3. Calculation of resonant frequencies
the center of the contact zone. Since the plane surface is rigid, it
can be considered as a symmetric plane. A symmetric plane of In the network, there are columns of air. Waves traveling in air
the cylinder exists, so only half of the cylinder is meshed. The will reflect back when they reach the end. A standing wave is cre-
radius of the cylinder is 0.27 m. The half pipe has rectangular ated at the resonant frequency of the network. An open network
cross-section with area of 0:0045 m  0:009 m. The length is will resonate if there is an anti-node at each open end. These
0.1 m. It has an open end and a closed end which has a unit velocity anti-nodes are places where there are maximum velocities and
as boundary condition. The base surface behind the piston is mod- minimum pressures (p ¼ 0). However, these zero pressures are
eled as flat to reduce the elements because the flange shape behind not at the physical ends of the network as they are altered by their
the piston has no influence on the end correction of the open end of contacts with air. In fact they are at the end correction positions of
this pipe. It has a very small impedance which is used to minimize the network.
the reflections. Other parts of the model are rigid. Knowing the resonant frequencies of the network is essential.
The acoustic pressure at the central node of the closed end is The network has a great influence on the exterior acoustic field
calculated by a BEM software. This software was developed by around the resonant frequencies, and at other frequencies there
the authors and some details on the formulation which has been is no change of acoustic pressure. In order to calculate this influ-
used can be found in [28]. For the computations in this article, first ence the resonant frequencies should be estimated first, otherwise
order elements of triangular or quadrangular shapes were used. all the frequencies should be calculated and it leads to large
The number of nodes was approximately 1600 and the number computations.
of elements 2900. This was found sufficient for the frequency band There are analytical solutions for simple pipes such as a straight
which is studied in the following. pipe, a T-shaped pipe and so on. However, numerical methods

Fig. 3. (a) Half a mesh of the boundary element model of a rectangular tube with a cylindrical flange; (b) details of the mesh of the tube.
78 B. Wang, D. Duhamel / Applied Acoustics 120 (2017) 75–84

avoid using these interior corrections, numerical methods should

be adopted for the calculation of the resonant frequencies.
The resonant frequencies can be obtained by solving an eigen-
value problem (6) of an undamped finite element model.

ðK  x2 MÞ/ ¼ 0 ð6Þ
K is the stiffness matrix, M is the mass matrix, x ¼ 2pf and / is
the eigenvector.
Finite element software Abaqus is chosen to perform the reso-
nant frequencies extraction procedure, which is a linear perturba-
tion procedure. The networks have open ends. The impedance
boundary conditions cannot be applied to these ends in Abaqus.
So their end corrections should be taken into account in the mod-
els. The branches with open ends in finite element models are cre-
ated longer than their physical lengths. Since at the resonant
frequencies the pressures are zero at the end correction positions,
zero pressure boundary conditions are applied. The other boundary
Fig. 4. End corrections of longitudinal pipes with cylindrical flanges of different conditions are rigid. The values of these end corrections for open
widths within 2000 Hz (ka < 0:23). ends are computed by relation (5). Then the network with cor-
rected lengths for the open ends is meshed.
For a 2D network whose junctions and branches are in the same
plane, an automatic calculation procedure can be implemented to
get the resonant frequencies. Abaqus GUI will generate an input
file ‘⁄.inp’ which is sent to the processor while running a simula-
tion. A simple case of a straight pipe is solved to get the ‘⁄.inp’ file
which can be modified for complex networks. In this file, the
straight pipe mesh information is deleted, then mesh information
of the network to be solved is added. New node and element sets
are defined for the zero pressure boundary conditions. The other
parts of this file keep the same data as before.
Since 2D meshes are quite easy to get, Matlab is used to create
the mesh information for the network, and the node and element
sets for the boundary conditions. The other information from the
old file of straight pipe is written into the new ‘⁄.inp’ file by Matlab
as well.
The automatic procedure for calculating the resonant frequen-
cies of 2D networks is: create the ‘⁄.inp’ by Matlab, submit it to
Abaqus by commands in Matlab, and then post-process the Abaqus
Fig. 5. Standard deviations and mean values of end corrections of longitudinal output script in Matlab. This Matlab programme is named as
pipes with cylindrical flanges of different widths.
‘2DNRF’ (for 2D Network Resonance Frequency), which will be
used in Section 4.

should be adopted in order to get the resonant frequencies of

complex networks. It is assumed that there is a plane wave in 4. Optimization of network structures
the straight parts of the network for the cross-sections are quite
small compared to the wavelengths. The value and number of resonant frequencies vary according to
Merkli [11] proposed a simple model to calculate the resonant the network structures. Because of this dependence, it should be
frequencies of a T-shaped resonator by using the wave propagation possible to find the wanted resonant frequencies or a maximum
theory. In this model, end corrections for open ends should be con- number of resonant frequencies by optimizing the network
sidered. Besides, the end correction of each branch at the junction structure.
position should also be included in the model. So for each open The parameters that describe a network are the number and
branch, the effective length includes the physical length of the positions of ends, junction types, junction positions and pipe
branch, the end correction at the junction position and the end cor- cross-sections. In this work there are only longitudinal and trans-
rection of the open end. Dalmont’s [27] methods and results can be verse pipes in a network, so the junction positions depend on the
used for the calculations of the open end corrections. With his end positions. If the end positions are known, the junction posi-
methods, the open end flanges could have any shape and dimen- tions are known. The 2D network considered in this work consists
sion. The end corrections at the junction position can be calculated of the three types of junctions shown in Fig. 6a–c.
with the model of Li [12]. Different networks can be formed by using different parame-
Theoretically, Merkli’s model can be applied to other cases such ters. Their resonant frequencies can be estimated by the finite
as a L-shaped junction pipe, a cross junction pipe or a network. element method 2DNRF proposed before. By changing the
However before estimating the resonant frequencies, the end cor- parameters, networks with the wanted resonant frequencies or a
rections of branches at the junction positions should be calculated maximum number of resonant frequencies can be found. Opti-
precisely. Existing models can only give approximate estimations mization methods should be adopted to optimize the parameters.
of these interior end corrections. If we use Merkli’s model to Genetic algorithm (GA) is used for this purpose in this article but
calculate the resonant frequencies of a network, there will be large other optimization methods such as Simulated Annealing could
errors due to the approximate interior end corrections. In order to also be used.
 B. Wang, D. Duhamel / Applied Acoustics 120 (2017) 75–84 79

Fig. 6. (a) T junction with left branch; (b) T junction with right branch; (c) cross junction.

Table 2
Results for the targeted resonant frequencies.

Case Targeted (Hz) Obtained (Hz) Error (%)

1 1250 1254 0:32
2 1400 1400 0
3 1600 1590 0:625

used. For a network each parameter of junction type is encoded

as an integer in 0, 1 or 2 which represent the junction types in
Fig. 6a–c, and the other parameters of network dimensions (such
as end positions and cross-sections) are encoded as binary strings.
Then the initial population can be generated by using a random
number generator. There are two parts in a population, a binary
Fig. 7. An example of network for the chromosome representation.
part of network dimensions and an integer part of junction types.
We can create a random binary matrix of size Nind  Lindfor the
parameters of network dimensions, where Nind specifies the num-
ber of individuals and Lind the length of the individuals. For the
parameters of junction types, we can give an integer (0, 1 and 2)
jun jun
matrix of size Nind  Lind where Lind is the number of junc-
tions. By combining these two matrices, the initial population is
obtained.
Here is an example of an individual in a population of networks

m n
zfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflffl}|fflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflffl{ zfflfflfflfflfflfflfflfflfflfflfflfflfflffl}|fflfflfflfflfflfflfflfflfflfflfflfflfflffl{
½0100...1111;0101...0101 |fflfflfflfflfflfflfflfflfflffl{zfflfflfflfflfflfflfflfflfflffl}l ;......;0101...:0111;2;1;0;......;1;0;2 network i
ð7Þ

This is the chromosome representation of the network in Fig. 7.

The network has 3 rows and 4 columns. In the first part of the chro-
mosome, there are m dimension parameters and each one has l
Fig. 8. An example of networks generated randomly in the first generation of the binary strings which can be converted to a real value. So there
GA procedure. are m real values, and for this network m is 8. We specify the
boundary of the network, so the open ends move only on this
boundary and along with the junctions. Therefore, we only need
The chromosome representations should be decided first. GAs to generate the real values for the coordinates of the junctions
operate on a number of potential solutions, called a population of and the cross sectional area of the network. Since we use the same
chromosomes, consisting of some encoding of the network param- cross section for the network, only one real value is needed for the
eter set. The binary and integer chromosome representations are area. The other seven real values represent the positions of the

Table 1
Coordinates of central points of open ends of a unflanged network.

End x (m) y (m) End x (m) y (m)

1 0:0475 0:051684 8 0:0475 0:051684
2 0:0325 0:051684 9 0:0325 0:051684
3 0:0325 0:051684 10 0:0325 0:051684
4 0:0475 0:051684 11 0:0475 0:051684
5 0:076684 ½0:0475; 0:015 12 0:076684 ½0:015; 0:0475
6 0:076684 ½0:01; 0:01 13 0:076684 ½0:01; 0:01
7 0:076684 ½0:015; 0:0475 14 0:076684 ½0:0475; 0:015
80 B. Wang, D. Duhamel / Applied Acoustics 120 (2017) 75–84

three rows and four columns which determine the coordinates of If the purpose is to find as many resonant frequencies as
the junctions. In the second part of the individual there are possible, the objective function is the number of resonant
n ¼ 12 junctions and each integer represents the junction type. If frequencies. A minus sign is added to get a minimization prob-
this population consists of k networks, its chromosome representa- lem. One has
tion is a k  ðm þ nÞ matrix.
The purpose of this work is to find the wanted resonant fre- ObjV 2 ðxi Þ ¼ countðf r1 ; f r2 ; . . .Þ ð9Þ
quencies or as many resonant frequencies as possible by designing
and optimizing the network. The 2D FEM, implemented in the We can minimize the two objective functions, which only
MATLAB code 2DNRF proposed before, has been used for estimat- depends on the parameters of the network (dimensions and junc-
ing values of f r and the number of resonant frequencies of a net- tion types).
work. If f w is the wanted resonant frequency and there are For the minimization problem, the most fit individuals have the
several resonant frequencies f r1 ; f r2 ; . . . for the network i, the objec- lowest numerical values of the associated objective function. The
tive function for the first purpose is given by raw objective values in Eq. (8) or (9) are usually only used as an
intermediate stage in determining the relative performance of
ObjV 1 ðxi Þ ¼ minðjf r1  f w j; jf r2  f w j; . . .Þ ð8Þ individuals. The fitness function is normally used to transform
the objective function value into a measure of relative fitness. Here,
where xi is the chromosome of network i. individuals are assigned a fitness according to the rank of their

Fig. 9. (a) Network with the resonant frequency 1254 Hz; (b) network with the resonant frequency 1400 Hz; (c) network with the resonant frequency 1590 Hz.
 B. Wang, D. Duhamel / Applied Acoustics 120 (2017) 75–84 81

5.1. Unflanged networks

In this example there is no flange, so the end corrections,

0.001684 m, can be calculated by Eq. (10) which is from paper [27].

dsq0 ¼ 0:597aeff ð10Þ

pffiffiffiffi
with aeff ¼ 2asq = p where 2asq is the pipe width.
Transverse (x axis) and some of longitudinal coordinates of the
ends are given in Table 1. The other longitudinal coordinates of the
ends vary within the ranges which are shown in Table 1 as well.
The pipe cross-sections are the same, 0:005 m  0:005 m. Bound-
ary conditions p ¼ 0 should be applied to the open ends, where p
is the acoustic pressure.
By giving the GA parameters, such as the number of individuals,
the generation gap, the maximum number of generations and the
wanted resonant frequencies, and running the GA program once,
an optimal network can be found. This network has the closest res-
onant frequency to the targeted resonant frequency. The GA
Fig. 10. The minimum difference between the targeted frequency 1400 Hz and the parameters should be adjusted for each case to obtain acceptable
resonant frequencies of individuals in each generation in the GA procedure.
results.
In Table 2, the wanted resonant frequencies are obtained by
optimizing the network structure and the maximum error is less
than 1%. The network structures can be seen in Fig. 9a–c. Besides
raw objective values in the population rather than the raw
the resonant frequencies shown in Table 2, these three networks
performance.
have other resonant frequencies within 2000 Hz. They are
A population is first initialized and then improved through
1926 Hz in case 1, 1976 Hz in case 2 and 1947 Hz in case 3.
repetitive application of the mutation, crossover and selection
In Fig. 10 we can see the minimum difference between the tar-
operators. This process is repeated until a fixed number of genera-
geted frequency 1400 Hz and the resonant frequencies of individ-
tions has been reached. If we are not satisfied with the results, we
uals in each generation in GA procedure. In the first generation
can adjust the specified parameters of GA or networks and then
the difference is 259.52 Hz. From the second generation it becomes
run the optimization program again until we get targeted resonant
6.93 Hz. In the 7th generation a very good result is found. The dif-
frequencies.
ference between its resonant frequency and the targeted frequency
1400 Hz is 0.22 Hz. The number of individuals in each generation
5. Design examples used in the GA procedure is 1000. This leads to a computing time
of about two hours to get the final result by the GA algorithm.
In order to illustrate that GA works for the optimization of 2D
networks and that the resonant frequencies can be obtained by 5.2. Flanged networks
the GA process, some examples are given.
All the dimension parameters of the network can be optimized In this example the network is between the cylinder and the
within given ranges during the GA process, but for the sake of plane surface used in Section 2. The calculations of open end cor-
simplicity some of them are specified. The network to be optimized rections are presented before. The end corrections of longitudinal
in this section has 14 open ends and 12 junctions. The junction open ends can be calculated by formula (5). For the transverse ends
positions are arranged in four columns and three rows, and their their corrections, 0.0054 m, 0.0057 m, and 0.0054 m, are given by
coordinates depend on end coordinates. In Fig. 8 there is an exam- formula (3).
ple of networks generated randomly in the first generation of the Longitudinal and some transverse coordinates of the ends are
GA procedure. In Fig. 8 we can see 14 positions for the open ends, given in Table 3. The other transverse coordinates vary within
but the network has 12 open ends due to the junction types. The the ranges which are shown in Table 3. The pipe cross-sections
circles represent the junction positions. Three types of junctions are the same, 0:009 m  0:009 m. Boundary conditions p ¼ 0
introduced before can be seen in this example. They are T junction should be applied to the open ends, where p is the acoustic pres-
with left branch, T junction with right branch and cross junction. sure. In Table 3, EC means the value of longitudinal end correction,
The corrections of open ends can be obtained by the methods which can be calculated by formula (5).
introduced before. If the ends are closed, there is no need to add In this section we try to get networks with the maximum
end corrections to the original pipe lengths. number of resonant frequencies within the frequency range

Table 3
Coordinates of central points of open ends of the network between a cylinder and a plane surface.

End x (m) y (m) End x (m) y (m)

1 ½0:042; 0:0705 0:05  EC 8 ½0:0705; 0:042 0:05 þ EC
2 ½0:0045; 0:033 0:05  EC 9 ½0:033; 0:0045 0:05 þ EC
3 ½0:033; 0:0045 0:05  EC 10 ½0:0045; 0:033 0:05 þ EC
4 ½0:0705; 0:042 0:05  EC 11 ½0:042; 0:0705 0:05 þ EC
5 0:0804 0:03333 12 0:0804 0:03333
6 0:0807 0 13 0:0807 0
7 0:0804 0:03333 14 0:0804 0:03333
82 B. Wang, D. Duhamel / Applied Acoustics 120 (2017) 75–84

Fig. 11. (a) Network 1; (b) Network 2.

Table 4
Networks with four resonant frequencies.

Resonant frequency 1st (Hz) 2nd (Hz) 3rd (Hz) 4th (Hz)
Network 1 1055 1797 1799 1929
Network 2 1201 1799 1890 1989

0–2000 Hz. With the dimension parameters of networks given

before, networks with as many resonant frequencies as possible
will be searched by GA. The networks of interest have repeated
junctions in the longitudinal direction (y axis in Fig. 8). It means
that in each column of the network in Fig. 8 the junctions should
be the same.
By changing the GA parameters and running the GA program
several times, four resonant frequencies at most within
0–2000 Hz are found. There are 2 networks with 4 resonant
frequencies, and they are shown in Fig. 11a and b. The resonant
frequencies are shown in Table 4. It was found that different runs
of the software lead to networks with quite similar structures but
with slightly different coordinates of the junctions.
In Fig. 12, the maximum number of resonant frequencies of a
network in the first generation is 3. It changes to 4 in the 8th
generation and then keeps the same value until the end of the
GA procedure. Since we are interested in periodic networks, only
networks with repeated junctions could be generated in each gen-
eration. So the number of individuals in each generation is just 30. Fig. 12. The maximum number of resonant frequencies for each generation in the
From the above examples one can conclude that the GA process GA procedure.
developed before is suitable for the optimization and design of net-
works. Once the ranges of dimension parameters of networks are 6.1. Unflanged network
given, networks with targeted resonant frequencies can be found.
The design examples will be validated by measurements in An unflanged network in Fig. 13a is measured in order to vali-
Section 6. date the optimization. The network in Fig. 13a is built with wood.
A speaker and a microphone are put near the open ends in Fig. 13a.
6. Experimental validation The optimized network is a 2D network, so the depth is not given
in the optimization. In the experiments, the depth that we use is
Wooden networks are built and measured in order to validate 0.01 m.
the optimization in Section 5. First a network without flange is The SPL obtained by the microphone is shown in Fig. 13b. We
measured. This network is optimized and built to target the reso- can see a deep valley at frequency 1370 Hz which is close to the
nant frequency 1400 Hz. Then a network between a cylinder and targeted resonant frequency 1400 Hz. So the GA optimization pro-
a plane surface is built and measured to search 4 resonant frequen- cedure developed before can give the resonant frequencies that we
cies predicted in Section 4. want.
 B. Wang, D. Duhamel / Applied Acoustics 120 (2017) 75–84 83

Fig. 13. (a) Optimized wooden network with the resonant frequency 1400 Hz; (b) measured SPL for the optimized wooden network with the resonant frequency 1400 Hz.

Fig. 14. (a) Sketch of experimental setup; (b) an optimized network between a cylinder and a plane surface.

Fig. 15. (a) Predicted results of an optimized network between a cylinder and a plane surface; (b) measured results of an optimized network between a cylinder and a plane
surface.
84 B. Wang, D. Duhamel / Applied Acoustics 120 (2017) 75–84

6.2. Flanged network to describe the road asperities and to mesh a part of the road in
the boundary element model.
The network between the cylinder and the plane surface is opti-
mized in Section 5.2. The acoustic fields of network 1 in Fig. 11a are Acknowledgment
calculated and measured in order to validate the GA optimization
methods. When we measure the case without network, we close This work was funded by the China Scholarship Council (CSC)
the pipe ends in Fig. 14b with woods. and by the French Environment and Energy Management Agency
In our calculations, a dimensionless point source is used, and it (ADEME).
has been assumed omni-directional. The general tendencies and
frequency-dependant variations of the sound pressure will be com- References
pared between the predicted and measured results. Such compar-
isons are very common and widely used in previous studies such as [1] Sandberg U, Ejsmont JA. Tyre/road noise reference book, INFORMEX; 2002.
[2] Zwikker C, Kosten C. In: Zwikker C, Kosten CW, editors. Sound absorbing
in [29,14]. So it is not necessary to use a omni-directional source in materials. Elsevier Publishing Company; 1949.
our experiments. A numerical signal is generated and then sent to a [3] Tijdeman H. On the propagation of sound waves in cylindrical tubes. J Sound
real loudspeaker. This is the source used in these experiments. Its Vib 1975;39(1):1–33.
[4] van der Eerden FJM. Noise reduction with coupled prismatic tubes. PhD thesis,
dimensions are 0:12 m  0:12 m  0:12 m. University of Twente, Enschede, the Netherlands; November 2000.
The radius and width of the cylinder are 0.27 m and 0.15 m. Its [5] Field C, Fricke F. Theory and applications of quarter-wave resonators: a prelude
location can be seen in the sketch of the experimental setup to their use for attenuating noise entering buildings through ventilation
openings. Appl Acoust 1998;53(1):117–32.
Fig. 14a. Sweep signals within 2000 Hz are generated by a genera- [6] Fahy F, Schofield C. A note on the interaction between a helmholtz resonator
tor. The center of the speaker is at (1 m, 0 m, 0.06 m). The micro- and an acoustic mode of an enclosure. J Sound Vib 1980;72(3):365–78.
phone is at (0.1 m, 0 m, 0.005 m). [7] Slaton W, Zeegers J. Acoustic power measurements of a damped
aeroacoustically driven resonatora). J Acoust Soc Am 2005;118(1):83–91.
The predicted results are shown in Fig. 15a. The resonant fre-
[8] Cummings A. The effects of a resonator array on the sound field in a cavity. J
quencies of this optimized network are given in Section 5.2, which Sound Vib 1992;154(1):25–44.
are 1055 Hz, 1797 Hz, 1799 Hz and 1929 Hz. The second and the [9] Doria A. Control of acoustic vibrations of an enclosure by means of multiple
resonators. J Sound Vib 1995;181(4):673–85.
third resonant frequencies are very close, so just one extremum
[10] Inácio O, Henrique L, Antunes J, et al. Design of duct cross sectional areas in
can be seen around 1800 Hz in Fig. 15a, where we can see two bass-trapping resonators for control rooms. Noise Control Eng J 2007;55
other extrema around 1055 Hz and 1929 Hz. (2):172–82.
The measured results are given in Fig. 15b, where the variations [11] Merkli P. Acoustic resonance frequencies for a t-tube. Zeitschrift für
angewandte Mathematik und Physik ZAMP 1978;29(3):486–98.
of SPL are similar to the predicted results in Fig. 15a. Around the [12] Li D, Vipperman JS. On the design of long t-shaped acoustic resonators. J
wanted resonant frequencies, reductions of the measured SPL are Acoust Soc Am 2004;116(5):2785–92.
close to reductions of the predicted SPL. The lower dip in the exper- [13] Li D. Vibroacoustic behavior and noise control studies of advanced composite
structures. PhD thesis, University of Pittsburgh; 2003.
imental curve compared to the computation can be explained by [14] Yu G. Acoustic resonators for noise control in enclosures: modelling, design
the damping in the system coming for instance from the wood and optimization. PhD thesis, The Hong Kong Polytechnic University; 2009.
material. From the predicted and measured SPL we can see that [15] Li D, Vipperman JS. Noise control of mock-scale chambercore payload fairing
using integrated acoustic resonators. J Spacecraft Rock 2006;43(4):877–82.
the optimized wooden network has the resonant frequencies that [16] Lane SA, Richard RE, Kennedy SJ. Fairing noise control using tube-shaped
we want. resonators. J Spacecraft Rock 2005;42(4):640–6.
[17] Hayden RE. Roadside noise from the interaction of a rolling tire with the road
surface. J Acoust Soc Am 1971;50(1A) [113–113].
7. Conclusions [18] Deffayet C et al. Air-pumping phenomena in road cavities. In: INTROC 90-
international tire/road noise conference 1990, Gothenburg, Sweden; 1990.
A numerical method for the calculation of resonant frequencies [19] Ronneberger D. Experimentelle und theoretische untersuchungen spezieller
mechanismen der rollgeräusche. Mitteilungen des Instituts für Straßen-,
of network resonators has been proposed. Based on this numerical Eisenbahn-und Felsbau an der ETH Zürich 1984;57:55.
method, an optimization method using genetic algorithm is devel- [20] Ronneberger D. Towards quantitative prediction of tyre/road noise. In:
oped in order to get the targeted resonant frequencies of the net- workshop on rolling noise generation, Institut fur Technische Akustik,
Technische Universitat Berlin; 1989.
work resonators. The optimization method is validated by [21] Kropp W, Becot F, Barrelet S. On the sound radiation from tyres. Acta Acust
experiments of design examples of wooden network resonators. United Acust 2000;86(5):769–79.
The measured resonant frequencies of the optimized network [22] Graf R, Kuo C, Dowling A, Graham W. Horn amplification at a tyre/road
interface-part i: experiment and computation. In: Proceedings of internoise
resonators agree well with the targeted resonant frequencies. The 99-the 1999 international congress on noise control engineering held in fort
good agreements also prove that the end corrections are estimated lauderdale, Florida, USA, 6–8 December 1999, vol. 1; 1999.
correctly. [23] Graf R, Kuo C, Dowling A, Graham W. On the horn effect of a tyre/road
interface, part I: experiment and computation. J Sound Vib 2002;256
Tire treads and road textures in the contact zones between tires
(3):417–31.
and roads can be considered as networks. The network resonance [24] Anfosso-Ledee F. Tyre/road noise horn effect: 2D modelling by analytical and
is an important mechanism of the tire/road noise. With the meth- numerical models. Tech rep, national PREDIT project; 2000.
ods proposed in this work, it is possible to investigate the influence [25] Anfosso-Ledee F, Klein P, Fadavi A, Duhamel D. Tire/road noise: comparison of
2d and 3d models for horn effect. In: Internoise 2000. Proceedings of the 29th
of these networks on the tire/road noise and to design tire treads international congress on noise control engineering, held 27-31 August 2000,
and road textures. Nice, France., vol. 1; 2000.
For a real tire the method could be applied by first estimating [26] Fadavi A. Tire/road noise: 3d model for horn effect. In: The 29th international
congress and exhibition on noise control engineering; 2000.
the end corrections of the pipe. This will need a fine mesh of the [27] Dalmont J-P, Nederveen C, Joly N. Radiation impedance of tubes with different
tire surface and potentially large computations but efficient meth- flanges: numerical and experimental investigations. J Sound Vib 2001;244
ods such as BEM based on Fast Multipole Methods could help to get (3):505–34.
[28] Duhamel D. L’acoustique des problèmes couplés fluide-structure - application
results in a reasonable time. Computations of resonant frequencies au contrôle actif du son. PhD thesis, Ecole Nationale des Ponts et Chaussées;
of the network in the contact zone should be done without too 1994.
much difficulties. The method can also be applied to estimate the [29] Cummings A. The effects of a resonator array on the sound field in a cavity. J
Sound Vib 1992;154(1):25–44.
influence of the road surface of the acoustic radiation. This needs