Вы находитесь на странице: 1из 4

Spiral Array Design with Particle Swarm

Optimization
Rilin Chen, Pengxiao Teng, Yichun Yang
Key laboratory of noise and vibration research
Institute of Acoustics, Chinese Academy of Sciences
Beijing, China
chenrl@mail.ioa.ac.cn
AbstractIn this paper, a novel optimization method based on
the particle swarm optimization (PSO) algorithm for the array
design is proposed. The PSO is introduced to search for the
optimal microphone distribution on the Archimedes spiral line
with the minimum sidelobe level and the narrow mainlobe width.
Numerical simulation show the better performance of the array
reconfiguration designed by means of PSO algorithm. Experiments in the semi-anechoic room demonstrate the effectiveness of
the proposed scheme.
Keywords- spiral Array;PSO; microphone distribution;

I.

INTRODUCTION

Recent studies of microphone arrays embraced a wide


range of applications, including multimedia communication,
speech recognition, video conference system, mobile robot and
falut source localization [1]. Beamforming of planar microphone array overlapped with the video image, which is regarded
as acoustic imaging or acoustic camera [2], is an established
technique for efficient and accurate noise source localization
and measurement. Performance of planar arrays is determined
by many parameters, such as the number of microphones, array
design (microphone distribution), and the aperture of array. It is
desired to optimize microphone distribution that fulfill specific
requirements, like the mainlobe width (MLW) and the sidelobe
level (SLL) which will affect the accuracy of localization and
measurement by the arrays. On the condition of the same
number of microphones and the same aperture of arrays, array
configuration design is the dominant factor for the
performance. Therefore, searching for the best microphone
distribution is an important work to acquire optimal
performance.
In array design process, the objective function and
constraint condition are always nonlinear and nondifferentiable. Therefore, heuristic methods, like genetic algorithm (GA)
[3] and particle swarm optimization (PSO) [4, 5] are often
adopted to solve these problems. GA is very efficient at
exploring the entire search space, but it is relatively poor in
finding the precise local optimal solution in the region where
the algorithm converges. Particle swarm optimization is similar
in some ways to genetic algorithms, but is much easier to
understand and implement and can find much precise solution.
Particle swarm optimization which was developed by Kenney
and Eberhart originated in studies of the social behavior of
birds flocking and fish schooling in their search for food [6]
when the researchers realized that their simulation algorithms

possessed an optimizing process. During the optimization


process of PSO, members of this population are flying
according to their previous flying experience, including their
individual and the swarm experience. It is accomplished by
adjusting their decision parameters toward the two best
solution Pbest that is the best individual found so far by that
individual, and Gbest that is the best individual previously by the
population.
So, particle swarm optimization is desired to optimize the
array design. In common, many researches are based on the
rectangular grid by making some grid point on or off to
accomplish a new reconfiguration of microphones. The present
work is to find the best array design on the Archimedes spiral
line, which clearly showed its advantage that is able to keep
much greater performance [7].
The structure of this paper is as follows: In Section , the
beampattern formula is described. In Section , the particle
swarm optimization approach for array design is presented in
detail. In Section , data simulations and results are presented
and discussed. Finally, conclusions are given in Section .
II.

BEAMPATTERN FORMULA

Beampattern reflects the spatial localization accuracy and


separation ability of a microphone planar array. The mainlobe
width shows its separation ability for different sources. The
narrower the mainlobe is, the stronger the spatial separation
ability is. If the mainlobe width is too wide, the array will not
be able to separate multi-sources. The sidelobe level indicates
the interference rejection capability of arrays. The sidelobe of
lower level will increase its anti-interference ability.
Z

Plane wave

P2

P1

P3
P4

P5

Figure 1. Model of incident plane wave

In Fig. 1, there is incident plane wave to the spiral array (z=0),


so the time delay relative to the physical center of the array is
as follows:

mt =

kpm

(1)

where
c is the sound speed;
pm is the position vector of the microphone m;
k is the wavenumber.

u x sin cos

k = u y = sin sin

u cos
z

(2)

where (, ) is the direction of the incident plane wave. The


beampattern of the spiral array can be expressed as:

B (k, f ) =

e
M

j 2 fpm k c

(3)

i =1

where f is the frequency of the incident plane wave and M is


the number of microphones. Without loss of generality, it is
assumed that =90 which make uz=0. The wavenumber (2)
will be simplified to

ux

uy

k =

(4)

where -1ux1, -1uy1. The mainlobe width of the


beampattern is expressed as:

MLW ( k , f ) = 2

(5)

where
is the wavelength of incident plane wave;
d is the aperture of the spiral array;
is the correction factor of aperture equivalent.
The sidelobe level can be written as:
SSL ( k , f ) =
max
B ( k, f ) .
MLW ( k , f

2< k <

vid = w vid + c1r1 ( pid xid ) + c2 r2 ( g d xid )


xid = xid + vid

(7)
(8)

where c1 and c2 are two positive constants (usually c1 = c2 = 2),


r1 and r2 are two random functions in the range [0, 1], and w is
the inertia weight which is usually a constant between 0.4 to
0.9. Obviously, (7) is used to calculate a new particle velocity
which is determined by its previous one and the distance of its
current position from the personal best position and the
distance of its current position from the global best position.
Then the particle changes to the new position according to (8).
The inertia weight w which is just proportional to the old
velocity and keeps the particle traveling in the same direction
is employed to influence the trade-off between global and
local exploration abilities. A large inertia weight tends to
explore the global search space while a small inertia weight
tends to search in the current area. A suitable inertia weight
can make a balance between global and local area. Some
researchers find advantage in choosing w a variable decreasing
with the optimization process which will encourage local
searching at the end of the optimization process [8].
In this paper, array design on the Archimedes line of good
performance is searched by PSO technique. So the fitness
function, which is computed in each iteration and is used for
exploring for the minimum value, is written as [11]:

fitness = SLL ( k , f ) MLW ( k , f )


p1

p2

(9)

where p1, p2 are randomly chosen exponents.


Steps of particle swarm optimization approach are detailed
as follows:

(6)

The mainlobe width (5) and the sidelobe level (6) are two
conflicting objects so that it is impossible to obtained the
design that both possesses narrowest mainlobe width and
lowest sidelobe level. Therefore, the design of reasonable
mainlobe width and sidelobe level becomes a trade-off choice.
III.

and get the best solution. The particles are manipulated


according to the following equations:

ARRAY DESIGN WITH PSO

James Kennedy and Russell C. Eberhart [6] originally


proposed the PSO algorithm for optimization. PSO is operated
on a model of social interaction between particle and swarm
intelligence to achieve optimization of a problem-specific
fitness function. Consider a set of swarm of N particles with
the position vector Xi=(xi1, xi2, xiD) which is a possible
solution of the problem and the velocity vector Vi=(vi1, vi2,
viD). Every time, a new fitness value can be obtained according
to the position vector, and Pibest=( pi1, pi2, piD) which is the
best previous position of particles and Gbest=( g1, g2, gD)
which is the global best position is updated. As can be noted,
gbest and Pbest will attract particles traveling to the best particle

Step 1. Initialization:
The first step is to select the suitable parameters for PSO,
including population size, particle dimension etc. which will
affect the convergence speed and their reasonable ranges
which are used for judging whether they are cross-border.
Step 2. Initialize Xi, Vi and Pibest, Gbest:
Each particle is initialized by its own randomly location with a
random velocity to start their optimization process. To start
with, all particles' positions (microphone distribution) become
their respective individual best and then are made a compareison to select the first global best since they don't even have
their experience.
Step 3. Evaluate particles' fitness and update Pibest and Gbest:
The fitness function, which is the product of sidelobe level and
mainlobe width of the array expressed as (9) in this paper,
returns a fitness value to be assigned to the current location. If
that value is greater than the value at the respective individual
best for that particle, then the appropriate locations are
replaced with the current location and update the respective
individual best and its fitness value. Compare all of the individual best and get the greatest one as the global best encountered so far.
Step 4. Update the velocity and position:

The velocity of the moving particles is a real-value vector,


which will control the particles' direction and speed. The
velocity affected by the individual best and global best will
determine the particles' next location.
Step 5. Termination criteria:
This is the condition under which the search process will
terminate. In this study, the search will terminate if the
following criteria is satisfied: The number of iterations reaches
the maximum allowable number or the fitness value meet a
certain condition. If the termination criteria is satisfied, then
stop, or else go to step 3.
IV.

while the ring array and the double ring array have the high
sidelobe level. The wheel array show great performance of
narrow mainlobe width and low sidelobe level compared with
other designs. So the compare between the design by means of
PSO algorithm (Described as PSO design in the following) and
the wheel array is exhibited in Fig. 2. It indicates that the PSO
design performs better than the wheel array. The PSO design
shows that the sidelobe level is much lower although the
mainlobe is slightly wider, which benefits accurately localization and anti-interference ability.

SIMULATION AND EXPERIMENT

The Archimedes spiral line is express as:

xi = ri cos ( i ) , yi = ri sin ( i )

(10)

ri = a i

(11)

where (xi, yi) is the position of the microphone i, i is the angle


between the one and x axis, ri is the distance from origin to the
microphone i. PSO algorithm is used to search for the best
microphone distribution which gives consideration to both the
low sidelobe level and narrow mainlobe width compared with
other distributions on the Archimedes spiral line.
A. Numerical simulation
In the following simulation, the aperture of the array is
constrained to 1m, so it is required rmax0.5m. The parameter
rmax is the distance from origin to the farthest microphone. To
meet this requirement, the parameter a should be limited as
a=1/(2max) where max is the angle of the farthest microphone
position. Also, it is assumed that the number of microphone is
64 and the frequency f=2000Hz is selected to simplify the
simulation to be the one as a single frequency problem. In
practical application, a relative regular distribution is desired to
make it much easier to design an available system. Based on
the consideration above, it is assumed that the element is
uniformly equal-angle distributed on the spiral line which
satisfied i= i and [0, 2]. In practice, there is always a
camera for video image in the center of the array. So it is
required that there is no element in the center and usually rmin
0.1m which equals to 10.1/ a is limited. In the start of
PSO process, the initialization of parameters is employed as:
N=20, D=1, c1 = c2 = 2, v[0, 2], [0, 2]. The termination
criterion is that the number of iterations reaches gen=100. The
inertia weight is as follows:

wi = wmax

wmax wmin
gen

(a)

(c)

(b)

(d)

Figure 2. Array designs (top) and their respective beampatterns (bottom)

B. Experiments in the semi-anechoic room

Figure 3. Acoustic Imaging System

Fig. 3 exhibits a real-time acoustic imaging system (left)


with 64 microphones (top right) and a camera (bottom right) in
the center of the array. The system is used to make some sound
localization experiments and sound field measurement.

(12)

where gen is the number of iterations and t is the time index of


iteration.
Adopting the PSO algorithm, we finally get the array
design just like Fig. 2(a) while Fig. 2 (c) shows its beampattern.
In the real application, there are a lot of different array designs,
including the rectangular array, the single ring array, the double
ring array, the wheel array etc. [9, 10]. In these designs, the
performance of the wheel array is greater than others. There is
a problem of many grating lobes about the rectangular array

Figure 4. Experiments with a real-time system in the semi-anechoic room

On the left side of Fig. 4, there is a sound field measurement for the automobile engine of a car in the semi-anechoic room.
It revealed the main sound sources of the engine with the
frequency ranged from 2000Hz to 2200Hz for band noise measurement while the distance is 0.75m from the engine to the
array. The top right is a sound field distribution while the fan
was running. The sound radiating became a stable sound field.
The bottom right is the sound localization result of two
speakers with the distance from each other about 1m. The
system located the accuracy position displayed as two reddest
points. All localization experiments and measurement showed
that the PSO design is an available one for the practical
applications.
CONCLUSION

This research has taken advantage of the flexibility of the


particle swarm optimization technique by applying it to the
problem of microphone reconfiguration of the array. In order to
be able to adopt the PSO algorithm to search for the optimal
array distribution, the fitness function which is the product of
sidelobe level and mainlobe width of the array is deduced to
search for the best design. The primary numerical results as
reported are solid validation of the PSO algorithm used in array
design. Experimental results have shown that the PSO
technique is able to get a validate array design used in the
practical applications. Future work with the particle swarm
might extend to many different areas of array processing and
analysis.
REFERENCES

[1]

Y.Huang and J.Benesty,Eds. Audio Signal Processing For NextGeneration


Multimedia
Communication
Systems,
Norwell,
MA:Kluwer,2004.
[2] Harry L.Van Trees. Optimum Array Processing Part IV of Detection,
Estimation, and Modulation Theory, New York, John Wiley &Sons,
2002.
[3] Mahanti, G. K., N. Pathak, and P. Mahanti, Synthesis of thinned linear
antenna arrays with fixed sidelobe level using real code dgenetic
algorithm, Progress In Electromagnetics Research,PIER 75, 319328,
2007.
[4] E. C. Laskari, K. E. Parsopoulos, and M. N. Vrahatis, Particle swarm
optimization for minimax problems, in Proc. Congr. Evolutionary
Computation (CEC 02), vol. 2, 2002, pp. 15761581.
[5] Chen, T. B., Y. L. Dong, Y. C. Jiao, et al., Synthesis of circular antenna
array using crossed particle swarm optimization algorithm, Journal
ofEle ctromagnetic Waves and Applications, Vol. 20, 17851795, 2006.
[6] J. Kennedy, The PSO: social adaptation of knowledge, in Proc IEEE
Int. Conf. on Evolutionary Computation, 1997, pp. 303308.
[7] Van Tonder, J.J. Cloete, J.K. , A Study of an Archimedes Spiral
Antenna, IEEE Antennas and Propagation Society International
Symposium, 20-24 Jun ,1994.
[8] R. C. Eberhart and Y. Shi, Comparing inertia weights and constriction
factors in particle swarm optimization, in Proc. Congr. Evolutionary
Computation, vol. 1, 2000, pp. 8488.
[9] J. Hald and J. J. Christensen, A class of optimal broadband phased
array geometries designed for easy construction, Proceedings of Inter
Noise 2002.
[10] J. Hald and J. J. Christensen, A novel beamformer array design for
noise source location from intermediate measurement distances,
Proceedings of Inter-Noise 2002.
[11] Palermo G., Silvano C. Discrete Particle Swarm Optimization for
Multi-objective Design Space Exploration, Digital System Design
Architectures, Methods and Tools, 2008, pp. 641-644.

Вам также может понравиться