E5.

21
HIGH-RESOLUTION DIRECTION FINDING USING CIRCULAR ARRAYS

M. Zeytznoilu’, J. Lztvat , J. Qzanr

‘Department of Electrical Engineering *Communications Research Laboratory

Ryerson Polytechnical Institute McMaster University
Toronto, Ontario Hamilton, Ontario
CANADA M5B 2K3 C.4NADA L8S 4L7

ABSTRACT algorithms are modified to compensate for strongly in-

terdependent behaviour of array sensors. We first dis-
This paper presents a discussion of high-resolution direc- cuss the array response models that will form the basis
tion finding algorithms applied to circular array struc- of signal processing algorithms for estimating DOA’s of
tures. Uncalibrated direction finding algorithms yield incident waves impinging upon the array. We demon-
highly biased direction-of-arrival estimates due to strong strate the highly biased performance of DF algorithms
mutual coupling effects among array elements. We de- based on the ideal array response model and suggest
rive alternate array response models based on array ge- modifications to existing DF algorithms.
ometry and electro-magnetic characteristics of sensors
and use the array output generated by these models to Let us assume that there are h’ narrow-band signals
calibrate direction finding algorithms. We apply this impinging upon an array of N identical sensors. Let
method to MUSIC algorithm and simulate its perfor- ai (t), as(t), . , UK(t) be the complex envelopes of these
mance. This approach exhibits superior estimation per- incident waves that arrive from directions 0r,&, . . ,BK
formance compared to the standard MUSIC algorithm. with respect to the line of reference of the array. Let
r,(t) be the output of the nth sensor. The array output
vector x(t) = [.cl(t),xz(t), . . ,z~(t)]~ is defined as

1. INTRODUCTION
x(t) = 2 s(8k)ak(t)+ v(t) (1)
k=l
This paper discusses signal processing algorithms for
high-resolution direction finding (DF) applied to circu-
where v(t) = [q(t), vz(t), . . . , up,(t)] T is the N x 1 vec-
lar array configurations. While the study of DF algo-
tor representing the noise processes at the array sensors.
rithms has been an active research topic for many years,
In Cl), s (0 k 1 re P resents the N x 1 dimensional array re-
most of the existing research body consists of applica-
sponse vector (steering vector) associated with Bk. By
tion of these algorithms to linear arrays. The underly-
introducing vector notation, (1) can be written as:
ing principles of such direction-of-arrival (DOA) estima-
tion algorithms are independent of the array geometry:
x(t) = Sa(t) + v(?) (2)
therefore, they can be equally well applied to circular ar-
rays. The principal advantage of circular arrays is their
where
ability to scan a beam through 360” with little change T
in either beamwidth or sidelobe level. These character- a(t) = [al(t),a2(tL.. ,aK(t)] , (3)
istics make circular arrays a highly desirable choice if
full peripheral coverage in DF is required.
s= [s(e,),s(e*),...,s(e,)l. (4)

Circular arrays present new challenges that mani- S is the array response matrix for an incident wave field
fest themselves in terms of strong mutual coupling ef- consisting of li waves at arrival angles {f?i, . , OK}.
fects between array sensors. In this paper we show The following assumptions are made on the signal model
that for circular arrays, DF algorithms based on ideai (2), and are considered valid throughout the paper.
response models exhibit significantly degraded perfor-
mance and highly biased DOA estimates, unless these Assumption 1: The sensor array is located at the
farfield such that planar wave expressions can be used
This work IS supported in part by Defense Research’Estab- for incident waves. Complex envelopes of the incident
Id~ment Ottawa, Canada waves are represented by a(t) which is modelled as a

- 3341 -
CH2977-7/91/0000-3341$1.00 Q 1991 IEEE
 zero-mean Gaussian vector with E[a(t)a+(.s)] = Z&b,,, the sensors as they respond to incident waves. Thus,
and correlation matrix C, such that [C,],, = P,,. the ideal response model that assumes the response of

Assumption 2: The array sensor noise vector v(t) is each sensor can be independently expressed, will fail to

;zy-mean Gaussian vector such that E[v(t)v+(s)] = yield accurate representation of the array output. An al-

v 15, with Xv = ~‘IN,N. The complex envelope vec- ternate response model for circular arrays where mutual

tor a(f) and the array sensor noise vector v(s) are un- coupling effects among sensors are incorporated into the

correlated for all t and s. model was derived in [2]. This analytic model is used
to simulate array response under the following two sce-
Assumption 3: The array response vectors {s(Bk),
narios:
Ic = 1,2,. , A’} corresponding to the columns of the
array response matrix S are linearly independent. l the array is operating in free space;
l the array is operating above lossy ground.
Assumption 4: All incident waves are assumed to im-
pinge upon the array at 0” elevation. Therefore, all sig- We refer to these models as the free space and the lossy

nal processing algorithms are constrained to estimate ground models. In particular, we consider two array

the DOA parameters in azimuth only. response matrices based on simulated data from the an-
alytic model representing the response of a uniform, 16

2. ARRAY RESPONSE MODELS element circular array operating in free space and above
lossy ground. The analytical model is used to simu-
The array response matrix determines how the incident late the normalized array response for an incident wave

signals interface with the sensor array. We consider impinging upon the array at incident angles that are

three array response models each described in terms of integer multiples of 0.5”‘.
a distinct array response matrix S. These models will
form the basis of our discussions where the performance For a uniform, N-sensor circular array one can sim-

of the DOA estimation algorithms in resolving incident ulate the array response at any incident angle, using

waves will be analyzed. steering vectors s(e) at 0 E [0’,360”/N) only. To con-

struct the array response matrix for an incident wave

2.1 Ideal Array Response Model field with K components, we extract normalized ar-
ray response vectors ~(0,) for IL = 1,. . ,I< from the set

The ideal array response matrix is determined by a {s(O),6’ E {0,0.5”, , (360”/N - 0.5”)}. The choice of

model that assumes differences in the response of indi- the underlying data set dictates the labelling of the re-

vidual sensors are due to propagation delays which are sponse matrix. Sf and S, refer to the response matrices

in turn determined by the array geometry. Let R repre- when the array is assumed to operate in free space and

sent the array radius and let q& be the angular position above lousy ground, respectively. Array output with re-
of the nth sensor measured from the array reference. If spect to an incident wave field is given by
there are K incident waves impinging upon the array at
angles 81,. ,6’~, the array response matrix of an ar- xj(t) = S,a(t) + v(t) (6)
ray with N identical sensors will be an N x K complex
valued matrix S,d with elements: where j E {f,g}. In our discussions to follow xf(t)
(xg(t)) represents the array output vector if the array
27r is operated in free space (above lossy ground).
b%d]nk = exp(J-RCOS(8k - &)),
x
Thus, &d(t), xf(t), xg(t) and corresponding array
for some k = 1,. . , K; n = 1,. . . , N. Here, we use “i8 response matrices Srd, Sf, S, will form the basis for
to indicate that S,d is the ideal array response with ele- the simulation studies. If xf(t) or xg(t) is compared to
ments that represent propagation delay among sensors x,d(t), one observes significant deviation from the ex-
measured with respect to the reference sensor at 4 = pected response based on the ideal model which may
0”. We use the notation x,d to indicate the array out- therefore degrade the performance of the DOA estima-
put vector based on: &d(r) = S,@(r) + v(t). tion algorithms based on the ideal model.

2.2 Other Array Response Models An equivalent way of representing the free space
and lossy ground array response models was discussed in
The ideal array response model serves well for establish- [2]-[4]. Let uon(t) be the open circuit voltage of the nth
ing a theoretical framework within which DF algorithms sensor which can be written as vO(t) = &a(t) where
can be analyzed; however, it fails to accurately model
the true array response, particularly for circular arrays. ’ The choice of 0.5” resolution is partially dictated by the
This is mainly due to significant interaction between computational requirements

- 3342 -
S,d is the ideal response matrix with elements given
B = [bh.+l, blc+z,. . . , bN] where b, represents the
in (5). Let Z, be the normalized impedance matrix
eigenvector corresponding to the rth eigenvalue of ft.
expressed in terms of the self and mutual impedance
between sensors and the load impedance. The open cir-
4. Determine the peak locations of the MUSIC spec-

cuit voltage vector is related to the array output vector

trum PM”(e) = [s+(e)BB+s(0)]-‘.

x(t), through the linear transformation Z,x(t) = v,,(t). To calibrate DF algorithms we incorporate simu-
Combining these last two equations and including the lated response data from alternate array models directly
array sensor noise vector we can write the array output into the algorithm by extracting the array steering vec-
vector tor s(e) from the array response data. Thus, m eval-
x(f) = z,‘s,d a(f) + v(t). (7) uating P&l,,(e) we use the noise-free, normalized, frw
In equation (7). we recognize S = Z;‘S,d. Here, the space (or lossy ground) array response at t? to deter-

array response matrix S equals S, or S, depending on mine s(0), i.e. we let s(0) = s,(@)a(t), with a(t) = 1.

the model the impedance matrix Z, has been derived Vt. This calibration procedure transforms calculation of
the algorithm’s objective function into a table look-up
from. Thus, the columns of the matrix S represent the
problem.
array response vectors that have been corrected for the
mutual coupling effects; as such they will serve as the
basis of the modified direction finding algorithms that 4. SIMULATION RESULTS
we discuss in the sequel.
Much of existmg work on the performance of DF algo-
3. SIGNAL PROCESSING rithms is based on an underlying model that assumes
ideal array response. However, observed performance
The problem of finding the DOA’s of I< plane waves im- of DF algorithms differ significantly from their nominal
pinging on a passive array of N sensors can be reduced levels if the array response is derived from other mod-
to that of estimating the parameters 81,. , of; in the els. In this section we present simulated performance of
signal model of (2) from the set of array output vectors the MUSIC algorithm in estimating the DOA’s of the
{x(ti), ,x(tbr)}. In this study we consider the MU- incident waves under various operating conditions.
SIC algorithm [5] based on the eigendecomposition of
the covariance matrix of the array output vectors. In the first simulation WC test the resolution ca-
pability and the estimation performance of the MUSIC
When some of the Incident waves are coherent as it algorithm based on the ideal response model. such that
happens with specular multipath components, the basic &,,(e) = [s;d(@)BBts,d(e)]-l. we generate array out-
MUSIC algorithm fails to resolve the incident wave field. put vectors using the free space array response model:
For uniform linear arrays spatial smoothing algorithms xf(t) = S,a(t) + v(t). These simulations are based on
have been proposed to overcome this inherent limita- a circular array with 16 uniformly spaced sensors and
tion of the MUSIC algorithm. However, due to the interelement spacing d = X/2. Other simulation param
circular geometry these spatial smoothing algorithms eter are given as follows.
cannot be directly applied to circular arrays. Extend-
ing these algorithms to circular arrays or creating new l Number of incident waves: K = 2.

pre-processing algorithms applicable to circular arrays l Incident angles: 0i = 0” and 02 = 5”;

is one of the current research activities. In our study of l Signal power: E[ar(t)al(t)] = E[az(t)a:(t)] = 1.

the MUSIC algorithm we restrict simulations to nearly l Incident waves are strongly correlated, pi2 = 0.98

coherent incident waves such that ]p,)] < 1, r # J for the l 64 snapshots are used to estimate the array output

elements of source signal correlation matrix. However, covariance matrix ir.

we addressed the question of coherent signals by using l Each Monte-Carlo experiment is based on 32 runs

maximum likelihood based DF algorithms, which are Fig. 1 depicts the DOA estimation performance of the
robust with respect to signal correlations. Results ob-
ideal response model based MUSIC algorithm as a func-
tained from these algorithms are reported in [l]. Here,
tion of SNR for -30 dB 5 SNR 5 50 dB.
we briefly summarize the standard MUSIC algorithm:

1. Given the array output vectors {x(ti), ,;x(rhr)} Here, we observe the DOA estimates converge to
estimate the array output covariance matrix R using e, = -3” and & = 8” as SNR is increased. Both
the relation & = & Et=, x(tm)xt(t,,,). DOA estimates display significant bias, and in [l] we
further observe that this residual bias disappears only
2. Determine and order the eigenvalues of H,.
for le, - 02 1 > 25” if all remaining simulation parame-
3. Using the eigenvectors corresponding to the N - Ii ters are held constant. Results of this simulation clearly
smallest eigenvalues of R form the projection matrix indicate the limited capability of the hiIUSIC algorithm

- 3343 -
 in estimatmg the DOA’s if t,he algorithm is based on [2] J. Litva and J. Qian, “Discussion of mutual coupling
the zdeal response model. Next, we calibrate the MU- in a VHF circular array antenna for application to
SIC algorithm using data derived from the same array radio direction finding,” CRL Tech. Rep., McMaster
response model as the output vectors were generated Umversity, Hamilton, Ontario, 1989.
from, such that PM,,(~) = (SUBBASE]-‘. Fig. 2 (31 I. J. Gupta and A. A. Ksienski, “Effect of mutual
displays the results from the calibrated MUSIC algo- coupling on the performance of adaptive arrays,”
rithm with all other parameters held constant. IEEE Trans. Antennas Propagat.. vol. AP-31, no.
5, pp. 785-791, 1983.
It is clearly seen that this approach exhibits supe-
rior estimat,ion performance when compared with the [4] C. C. Yeh, M. L. Leou and D. R. Ucci, “Bearing esti-

DOA estimates of Fig. 1. For this simulation i1 = 01 mations with mutual coupling present,” IEEE Trans.
and 02 = 82 at SNR > 20dB. As the free space array Antennas Propagat.,no. 10, pp. 133221335, 1989.

response data is only available at incident angles that [5] R. 0. Schmidt, “A signal subspace approach to mul-
are integer multiples of 0.5”, any estimation error whose tiple emitter location and spectral estimation,” PhD.
magnitude is less than 0.5” will be suppressed in the fi- dissertation, Stanford University, 1981.
nal results. Even when this quantization effect is taken
into consideration the estimates from the free space cal-
ibrated algorithm prove to be accurate to within 50.5”.

In further simulation studies reported in [I] we ob-

served similar estimation performance from the uncal-
ibrated MUSIC algorithm if the array output vectors
are generated with the lossy ground model. The resid-
ual bias in DOA estimates were more significant than
those depicted in Fig. 1 and they can be eliminated by
calibrating the algorithm using the lossy ground data.
To extend these results to incident wave fields with
fully correlated signals, we employed maximum likeli-
hood based DF algorithms; where we again observed SNR
that to eliminate residual bias in DO.4 estimates, the (dB)
DF algorithm has to be calibrated with data from the
appropriate array response model
-50’ 0’ 50’
5. CONCLUSIONS DOA

In this work our motivation has been to study the ef- Fig. 1 Peak spectrum of unmodified MUSIC algorithm
fects of mutual coupling on the performance of direction r
finding algorithms. The presence of mutual coupling e,,o*
distorts the structure of the array output covariance
matrix. As DF algorithms operate on the array out-
put covariance matrix this distortion results in severe
performance degradation. To compensate for non-ideal
array response we use simulated data from free space
response model that takes mutual coupling effects into
consideration. We modify DF algorithms by assembling
array steering vectors directly from simulated raw array
response data. In related simulation studies calibrated
DF algorithms are observed to exhibit exemplary per-
formance levels in resolving incident fields with closely
spaced source signals. SNR
(dB)
6. REFERENCES

[l] J Litva and M. Zeytinoglu, “Application of high-

resolution direction finding algorithms to circular ar- DOA
rays with mutual coupling present,” CRL Tech. Rep.,
McMaster University, Hamilton, Ontario, 1990. Fig. 2 Peak spectrum of calibrated MUSIC algorithm

- 3344 -