22.01.2014 Views

DOA Estimation of Coherent Signals in Heavy Noise ... - IERI

DOA Estimation of Coherent Signals in Heavy Noise ... - IERI

DOA Estimation of Coherent Signals in Heavy Noise ... - IERI

SHOW MORE
SHOW LESS

You also want an ePaper? Increase the reach of your titles

YUMPU automatically turns print PDFs into web optimized ePapers that Google loves.

2012 International Conference on Environmental Eng<strong>in</strong>eer<strong>in</strong>g and Technology<br />

Advances <strong>in</strong> Biomedical Eng<strong>in</strong>eer<strong>in</strong>g, Vol.8<br />

<strong>DOA</strong> <strong>Estimation</strong> <strong>of</strong> <strong>Coherent</strong> <strong>Signals</strong> <strong>in</strong> <strong>Heavy</strong> <strong>Noise</strong> Environment<br />

Shaowei Chen, Y<strong>in</strong>gy<strong>in</strong>g Wei and X<strong>in</strong>jian Shi<br />

Department <strong>of</strong> Electronisc and Information<br />

Northwestern Polytechnical University<br />

Xi’an, Shaanxi, Ch<strong>in</strong>a<br />

Y<strong>in</strong>gy<strong>in</strong>g-1986523@163.com<br />

Keywords: Direction <strong>of</strong> arrival; coherent signal; s<strong>in</strong>gular value decomposition; SNR.<br />

Abstract. After research<strong>in</strong>g s<strong>in</strong>gular value decomposition (SVD) algorithm and Toeplitz algorithm,<br />

a novel high efficient algorithm <strong>in</strong> multipath environment is proposed <strong>in</strong> the paper. First a new<br />

Toeplitz matrix is reconstructed us<strong>in</strong>g the maximum eigenvectors <strong>of</strong> signal covariance matrix. Then<br />

the <strong>DOA</strong> (Direction <strong>of</strong> Array) <strong>of</strong> <strong>in</strong>cident source can be resolved by the way <strong>of</strong> matrix SVD<br />

algorithm. Simulations results, presented <strong>in</strong> section IV, show that the new algorithm has the better<br />

performance than the conventional SVD and Toeplitz algorithm <strong>in</strong> accuracy and stability at low<br />

signal to noise ratio (SNR). When SNR is -7, the identification error <strong>of</strong> Toeplitz algorithm is up to<br />

2. 50, and SVD algorithm lose effectiveness. But the identification error <strong>of</strong> new TSVD algorithm is<br />

near to 0.<br />

I. Introduction<br />

<strong>DOA</strong> is one <strong>of</strong> key research issues <strong>in</strong> spatial spectrum estimation. A lot <strong>of</strong> mature algorithms have<br />

been developed up to now, such as MEM、MVM、ESPRIT and MUSIC etc. Influenced by<br />

multipath propagation exist<strong>in</strong>g <strong>in</strong> radar, received signal under actual circumstances is always<br />

coherent. To address the issue, now there are two types <strong>of</strong> algorithms: First, Dimension-reduc<strong>in</strong>g<br />

Way. It can be further classified <strong>in</strong>to spatial smooth<strong>in</strong>g algorithm and matrix reconstruction<br />

algorithm, such as modified spatial smooth<strong>in</strong>g [2] (MSS), s<strong>in</strong>gular value decomposition [3] (SVD),<br />

etc; Second, non Dimension-reduc<strong>in</strong>g Way. There are Toeplitz method, dummy array<br />

transformation method and some other new de-correlated methods [5-7]. Research<strong>in</strong>g on these<br />

algorithms, performance <strong>of</strong> Dimension -reduc<strong>in</strong>g Method will drop because <strong>of</strong> the disadvantage <strong>of</strong><br />

array aperture loss. And non Dimension-reduc<strong>in</strong>g Method needs to amend and judge the data <strong>in</strong><br />

advance, which <strong>of</strong>ten lead to great estimation error.<br />

A novel <strong>DOA</strong> algorithm, based on SVD and Toeplitz, is proposed <strong>in</strong> this paper. It comb<strong>in</strong>es the<br />

advantages <strong>of</strong> two algorithms, high resolution and stability. At last a large number <strong>of</strong> simulations<br />

show the good performance <strong>of</strong> new algorithm <strong>in</strong> this paper.<br />

2. Mathematical model<br />

Consider a l<strong>in</strong>ear receiv<strong>in</strong>g array with M equally spaced elements and direct sources with a total <strong>of</strong><br />

N paths. At the array, source wave front propagation is assumed planar and source energy is<br />

<strong>in</strong>cident on the array a dist<strong>in</strong>ct angles {θ 1 , θ 2 , …. , θ N }. The propagation channel is<br />

homogeneous and the signals are narrowband (i.e. source bandwidth is much smaller than the<br />

978-1-61275-026-2/10/$25.00 ©2012 <strong>IERI</strong> ICEET 2012<br />

170


eciprocal <strong>of</strong> propagation time <strong>of</strong> the signal across the array). Then the received signal at the kth<br />

time is given by<br />

X( k) = AS( k) + N ( k), k = 1,2,..., K<br />

(1)<br />

Where,<br />

X ( k) = [ x1<br />

( k) … x ( )]<br />

T<br />

M<br />

k<br />

T<br />

jw0<br />

k<br />

Sk ( ) = [ s1<br />

( k) … sN( k) ] , si( k) = ai( ke )<br />

A= [ a( θ ) a( θ ) … a( θ )]<br />

1 2<br />

a θ β π θ λ<br />

− jβ<br />

− j( M −1)<br />

i<br />

(<br />

i) = [1 e e β ],<br />

i<br />

= 2 ds<strong>in</strong> i<br />

/<br />

Nk ( ) = [ n1<br />

( k) n ( )]<br />

T<br />

M<br />

k<br />

Where, d is the sensor spac<strong>in</strong>g, λ is the process<strong>in</strong>g center wave length, ω 0 is the process<strong>in</strong>g center<br />

angular frequency<br />

Therefore, the spatial covariance matrix is:<br />

2<br />

R = [ XX H ] = A [ SS H ] A H + I<br />

H<br />

H<br />

= USΣSUS + UNΣ NU<br />

N<br />

E E σ<br />

Where, ‘H’ denotes the complex conjugate transpose; ‘E’ stands for expectation, an ensemble<br />

average operator; I is M×M identity matrix; U S =[e 1 e 2 … e N ] is signal subspace, consist<strong>in</strong>g by<br />

eigenvectors correspond<strong>in</strong>g to largest N eigenvalue; U N =[e N+1 e N+2 … e M ] is noise subspace,<br />

consist<strong>in</strong>g by eigenvectors correspond<strong>in</strong>g to smallest (M-N) eigenvalues; Σ S is diagonal matrix<br />

consist<strong>in</strong>g <strong>of</strong> large eigenvalue; Σ N is diagonal matrix consist<strong>in</strong>g <strong>of</strong> smallest N eigenvalues; σ 2<br />

denote ideal white noise power.<br />

In direction f<strong>in</strong>d<strong>in</strong>g problems, eigenstructure methods <strong>of</strong> estimat<strong>in</strong>g the directions <strong>of</strong> arrival {θ i }<br />

are based on exploit<strong>in</strong>g this structure <strong>of</strong> R. That is, R has a range basis consist<strong>in</strong>g <strong>of</strong> the direction<br />

vectors <strong>of</strong> source plane waves, and R is the sum <strong>of</strong> an N-rank matrix U S Σ S U H S and σ 2 I.<br />

The noise subspace vectors are orthogonal to all the steer<strong>in</strong>g vectors. The MUSIC spectrum is a<br />

cont<strong>in</strong>uous spectrum formed as<br />

1<br />

PMUSIC<br />

( θ ) =<br />

(3)<br />

H<br />

2<br />

a ( θ ) U<br />

The source directions can be estimated as the peaks <strong>of</strong> the spatial spectrum.<br />

Conventional eigen-structure methods, however, are applicable only when the path signal<br />

covariance matrix is non-s<strong>in</strong>gular. When some <strong>of</strong> the path signals are coherent (for examples, if<br />

they orig<strong>in</strong>ate from the same source), the path signal covariance matrix will be s<strong>in</strong>gular and<br />

properties noted do not hold. Up to now, to solve this problem there are many methods for coherent<br />

signal. SVD and Toeplitz algorithm are two k<strong>in</strong>ds <strong>of</strong> typical methods.<br />

2.1 SVD method<br />

SVD (s<strong>in</strong>gular value decomposition) algorithm has been proposed as Reference [3]. The<br />

[3]<br />

largest eigenvector e 1 is formed as<br />

T<br />

e1 = [ e11, e12, e13...... e1M<br />

]<br />

k=<br />

N<br />

(4)<br />

T<br />

= α a( θ ) = [ x(1) x(2)... x( M)]<br />

∑<br />

k = 1<br />

1<br />

k<br />

Therefore, the largest eigenvector e 1 conta<strong>in</strong>s all the source direction <strong>in</strong>formation. Then a new<br />

matrix Y can be formed by utiliz<strong>in</strong>g e 1 .<br />

⎡ e1,1 e1,2 e1,<br />

p ⎤<br />

⎢<br />

⎥<br />

⎢<br />

e1,2 e1,3 e1, s+<br />

1<br />

Y = ⎥<br />

(5)<br />

⎢ ⎥<br />

⎢<br />

⎥<br />

⎢⎣<br />

e1, m<br />

e1, r+<br />

1<br />

e1,<br />

M ⎥⎦<br />

N<br />

N<br />

(2)<br />

171


Where, m+p-1=M. The noise subspace can be obta<strong>in</strong>ed by s<strong>in</strong>gular value decomposition <strong>of</strong> the<br />

matrix Y. F<strong>in</strong>ally, <strong>DOA</strong> can be estimated accord<strong>in</strong>g to formula 3.<br />

2.2 Toeplitz approximation method<br />

S. Y. Kung et al.[4] proposed a Toeplitz Approximation Method (TAM) which is based on a<br />

reduced order Toeplitz approximation <strong>of</strong> an estimated spatial covariance matrix. The estimated<br />

covariance matrix, <strong>in</strong> the case <strong>in</strong> which sources are uncorrelated and statistically stationary, is<br />

Toeplitz. In a multipath environment, however, the source paths are fully correlated, and this<br />

covariance matrix is not Toeplitz. The Toeplitz structure can he guaranteed by employ<strong>in</strong>g spatial<br />

smooth<strong>in</strong>g, which destroys cross correlation between directional components. The TAM is designed<br />

for robustness <strong>in</strong> an arbitrary ambient noise environment.<br />

3. Tsvd algorithm<br />

In accordance with the forego<strong>in</strong>g analysis, a novel direction f<strong>in</strong>d<strong>in</strong>g algorithm, based on the<br />

advantages <strong>of</strong> SVD and TAM, is proposed <strong>in</strong> this paper. First, eigenvalue decomposition <strong>of</strong> the<br />

spatial covariance matrix is [U, S]= eig (R), where, U denotes eigenvectors matrix; S represents<br />

egenvaluses matrix. As previous analysis, the largest eigenvector e 1 associated with the matrix U<br />

completely <strong>in</strong>clude <strong>in</strong>formation <strong>of</strong> received source paths. Consider<strong>in</strong>g correlated source paths,<br />

decoherent e 1<br />

'<br />

rm<br />

= e11 × e1<br />

m<br />

(6)<br />

Obviously, the angles element still exists <strong>in</strong> the vector r m .<br />

As we all known like given <strong>in</strong> the TAM algorithm, the Toeplitz matrix can destroy cross<br />

correlation between directional components. Now, we form the new covariance matrix as follow:<br />

⎡ r1 r2<br />

rM<br />

⎤<br />

⎢ '<br />

r2 r1 r ⎥<br />

M −1<br />

Y = ⎢<br />

⎥<br />

⎢ ⎥<br />

(7)<br />

⎢ ' '<br />

⎥<br />

⎣rM<br />

rM−<br />

1<br />

r1<br />

⎦<br />

We can easily verify the new matrix Y is Toeplitz structure (conjugate symmetric matrix).<br />

Then, perform an SVD on Y, [U, S, V]=svd(Y) and arrange the s<strong>in</strong>gular values <strong>in</strong> decreas<strong>in</strong>g.<br />

The SVD is<br />

⎡S1 0 ⎤⎡V′<br />

1⎤<br />

Y = USV′<br />

= [ U1 U2]<br />

⎢<br />

0 S<br />

⎥⎢<br />

2<br />

V<br />

⎥<br />

(8)<br />

⎣ ⎦⎣ 2 ′ ⎦<br />

Where, S should ideally be a diagonal matrix <strong>of</strong> N eigenvalues; the N×N diagonal matrix S 1<br />

conta<strong>in</strong>s the N largest s<strong>in</strong>gular values; the (M-N)×(M-N) diagonal matrix S 2 conta<strong>in</strong>s the (M-N)<br />

smallest s<strong>in</strong>gular values. U stands for left s<strong>in</strong>gular matrix; V stands for right s<strong>in</strong>gular matrix.<br />

Similarly, the N columns eigenvectors correspond<strong>in</strong>g to S 1 <strong>of</strong> matrix U is signal subspace. The (M-<br />

N) columns eigenvectors correspond<strong>in</strong>g to S 1 <strong>of</strong> matrix U is noise subspace Un=U (:, (N+1):M).<br />

With the equation 3, the source directions can be estimated quickly.<br />

In summary, the five steps <strong>of</strong> the novel approach are:<br />

Step1<br />

Calculate the spatial covariance matrix R;<br />

Step2<br />

Perform eigenvalue decomposition on R. Extract the largest vector e 1 accord<strong>in</strong>g to the largest<br />

eigenvalue;<br />

Step3<br />

Reconstruct the new Toeplitz covariance matrix Y accord<strong>in</strong>g to equation 7;<br />

Step4<br />

172


Perform an SVD on Y, so we can easily obta<strong>in</strong> noise subspace Un;<br />

Stpe5<br />

Search the peak <strong>of</strong> the cont<strong>in</strong>uous spectrum as the <strong>DOA</strong> <strong>of</strong> source accord<strong>in</strong>g to formula 3.<br />

4. Simulations<br />

The snapshot model consists <strong>of</strong> a l<strong>in</strong>ear array <strong>of</strong> 8 equispaced with <strong>in</strong>ter-element spac<strong>in</strong>g <strong>of</strong> one half<br />

wavelengths. Two equal-powered emitters at angular locations 15° and 22°. The number <strong>of</strong><br />

snapshots sets to 100. In order to perform SVD 、Toeplitz and TSVD applied to the direction<br />

f<strong>in</strong>d<strong>in</strong>g, the temporal snapshot data and Gaussian noise data were generated artificially by a<br />

computer. The simulation results is given as Fig.1-4.<br />

Fig. 1 shows simulation results for SVD、Toeplitz and TSVD at SNR <strong>of</strong> 7dB. Three methods<br />

all can resolve the emitters, but TSVD yields nearer optimal angle estimates than other two<br />

algorithms. But SNR decrease to -7dB, Fig. 2 shows that TSVD has the accurate performance. The<br />

SVD cannot resolve the angles, while the error <strong>of</strong> the Toeplitz is up to 2.5°.<br />

150<br />

Spatial spectrum (dB)<br />

100<br />

50<br />

0<br />

TSVD<br />

SVD<br />

Toeplitz<br />

-50<br />

0 10 20 30 40<br />

Angle (°)<br />

Figure.1 The three arrows <strong>in</strong>dicate the orig<strong>in</strong>al spectral peaks at 7dB SNR.<br />

100<br />

Spatial spectrum (dB)<br />

80<br />

60<br />

40<br />

20<br />

0<br />

-20<br />

TSVD<br />

SVD<br />

Toeplitz<br />

0 10 20 30 40<br />

Angle (°)<br />

Figure.2 The three arrows <strong>in</strong>dicate the orig<strong>in</strong>al spectral peaks at -7dB SNR.<br />

In the same environment condition we carry 100 Monte Carlo trials. Fig. 3-4 respectively show<br />

Monte Carlo simulation results <strong>of</strong> the error and success probability for SVD、Toeplitz and TSVD.<br />

Obviously, the success probability <strong>of</strong> TSVD is near to 100% at a SNR <strong>of</strong> -10dB, <strong>in</strong> contrary SVD is<br />

down to 0. Although Toeplitz can achieve high success probability, the error is greater than new<br />

method. So we can summarize TSVD has optimal resolution and stability compar<strong>in</strong>g to<br />

conventional SVD and Toeplitz algorithms particularly at low SNR.<br />

173


<strong>Estimation</strong> error (°)<br />

1.4<br />

1.2<br />

1<br />

0.8<br />

0.6<br />

0.4<br />

0.2<br />

TSVD<br />

SVD<br />

Toeplitz<br />

0<br />

0 5 10 15 20<br />

SNR (dB)<br />

Figure.3 <strong>Estimation</strong> error for three algorithms at SNR <strong>of</strong> 0-20.1<br />

1<br />

0.8<br />

TSVD<br />

SVD<br />

Toeplitz<br />

Success probability<br />

0.6<br />

0.4<br />

0.2<br />

0<br />

-10 -5 0 5 10 15 20<br />

SNR (dB)<br />

Figure.4 Success probability for SVD、Toeplitz and TSVD at SNR <strong>of</strong> -10 to 20.<br />

5. Conclusion<br />

TSVD algorithm for <strong>DOA</strong> f<strong>in</strong>d<strong>in</strong>g problems, with application <strong>in</strong> multipath environment with l<strong>in</strong>ear<br />

equispaced arrays, has been presented. The algorithm comb<strong>in</strong>es the advantages <strong>of</strong> two classical<br />

algorithms, has excellent performance. At last simulations results <strong>in</strong>dicate that it can be optimistic<br />

<strong>in</strong> expect<strong>in</strong>g high resolution and good accuracy when a greatly low SNR is used.<br />

References<br />

[1] H. Krim, M. Viberg. “Two Decades <strong>of</strong> Array Signal Process<strong>in</strong>g Research,” IEEE Signal<br />

Process<strong>in</strong>g Magaz<strong>in</strong>e, vol. 7, pp. 67-94, 1996..<br />

[2] R. T. Williams, S. Prasad, A. K. Mahalanabis and L. H. sibul. “An Improved Spatial Smooth<strong>in</strong>g<br />

Technique for Bear<strong>in</strong>g <strong>Estimation</strong> <strong>in</strong> a Multipath Environment,” IEEE Trans on ASSP, vol.<br />

36,no. 4, pp. 425-432, 1998.<br />

[3] D. W. Tufts, A. C. Kot and R. J. Vacco. “The Threshold Analysis <strong>of</strong> SVD-based Algorithms,”<br />

IEEE <strong>in</strong>ternational conference ICASSP-88, pp. 2416-2419. 1988.<br />

[4] Kung S Y, Lo C K, Foka R. “A toeplitz approximation approach to coherent source direction<br />

f<strong>in</strong>d<strong>in</strong>g,” IEEE <strong>in</strong>ternational conference ICASSP-86, vol. 11, pp. 193-196, 1986.<br />

[5] Kaneko, Kazumasa Sano, Akira. “Music-like Iterative <strong>DOA</strong> <strong>Estimation</strong> <strong>in</strong> Multipath<br />

Environments,” IEEE Sensor Array and Multichannel Signal Process<strong>in</strong>g Workshop. pp. 212-<br />

215, 2008.<br />

174


[6] T<strong>in</strong>g D<strong>in</strong>g. “Modified <strong>DOA</strong> <strong>Estimation</strong> Algorithm For Reduc<strong>in</strong>g Correlation <strong>of</strong> Multi-path<br />

<strong>Signals</strong>,” Information Security and Communications Privacy, vol. 2, pp. 53-55, February. 2008.<br />

[7] X<strong>in</strong> Xie, Guol<strong>in</strong> Li and Huawen Liu. “<strong>DOA</strong> <strong>Estimation</strong> <strong>of</strong> <strong>Coherent</strong> <strong>Signals</strong> Us<strong>in</strong>g one<br />

Snapshot,” Journal <strong>of</strong> Electronics and Information Technology. Vol. 32, no. 3, pp. 604-608,<br />

March 2010.<br />

175

Hooray! Your file is uploaded and ready to be published.

Saved successfully!

Ooh no, something went wrong!