Anda di halaman 1dari 11

Applied Acoustics 72 (2011) 873883

Contents lists available at ScienceDirect

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

Spatial smoothing for localized correlated sources Its effect on different localization methods in the neareld
Wolfram Pannert
University of Applied Sciences Aalen, Department of Mechanical Engineering, Beethovenstr. 1, 73430 Aalen, Germany

a r t i c l e

i n f o

a b s t r a c t
The spatial smoothing (SS) technique has been proved to be effective in decorrelating coherent signals by restoring the rank of the signal covariance matrix R. Averaging the covariance matrices of subarrays of the original array, is a technique which increases the rank of the smoothed matrix RSS. Algorithms like MUSIC or Capon, which rely on the use of the signal covariance matrix R and fail in the case of correlated sources, can be applied to scenarios with correlated sources after spatial smoothing. However, SS is most practically applied to uniformly spaced arrays or to arrays which have a translational symmetry. In addition the formulation is strictly applicable only to such fareld conditions, where the incoming waves are plane waves and the steering vectors to the sources of the different subarrays are identical. These conditions are not fullled in the neareld. Spatial smoothing is now applied with an acoustic camera in the neareld and it is shown that up to some limits this technique is applicable. Effects/limitations are studied using simulation and measurements with several Beamforming algorithms (MUSIC, Capon and Orthogonal Beamforming) are carried out. The results demonstrate the benets of SS even in the neareld up to some limits, which are given through the distance of the different subarrays in comparison to the spatial resolution of the Beamforming algorithm. Especially at lower frequencies SS in connection with MUSIC- or Capon-Beamforming give better resolution in comparison to D + S Beamforming. 2011 Elsevier Ltd. All rights reserved.

Article history: Received 2 February 2011 Received in revised form 10 May 2011 Accepted 19 May 2011 Available online 15 June 2011 Keywords: Spatial smoothing Beamforming Coherent signals Covariance matrix

1. Introduction Array processing as a eld is well developed for signal processing in acoustics and radar applications. It combines the signals of the sensors in the array into a single output. A summary of standard and advanced Beamforming methods in acoustics was written by Van Trees [1]. Much research has been focused on high-resolution Beamformers. A high-resolution Beamformer was developed in 1969 by Capon, known as Minimum Variance Distortionless Response (MVDR) Beamformer [2]. One of the largest problems with high resolution methods is performance degradations with correlated sources. Historically, the problem of correlated sources has been studied exclusively in the fareld. Only recently the problem of correlated sources has been analyzed in the neareld [3,4]. The original use of spatial smoothing was introduced by Evans et al. [5]. The modication to the Beamforming problem was done by Shan and Kailath [6]. The performance of the method for direction of arrival estimation (DOA) was also studied by Shan et al. [7]. Spatial smoothing remained the most widely used method for separating correlated sources. Reddy et al. in 1987 studied the
E-mail address: wolfram.pannert@htw-aalen.de 0003-682X/$ - see front matter 2011 Elsevier Ltd. All rights reserved. doi:10.1016/j.apacoust.2011.05.011

performance of MVDR and how spatial smoothing improved the method. A representation of the output power of MVDR and the specic effects of correlation were mathematically proven [8]. Spatial smoothing was also shown to be a toeplitzization of the covariance matrix by Takao et al. [9]. The covariance matrix becomes more diagonal as signals are decorrelated. The work was extended to a more general form by Tsai in 1995 [10]. Bresler et al. [8] provided a proof of the effectiveness of spatial smoothing through signal subspace and parameter estimation. The proof allowed iterative methods to be used for estimating signal parameters [11]. Another high-resolution Beamformer is the MUSIC Beamformer [12]. The MUSIC algorithm is an example of a subspace-algorithm, which divides the vector space of the covariance matrix in a signalsubspace and a noise-subspace. For the MUSIC-algorithm correlated sources have the effect that part of the signal subspace is indistinguishable from the noise subspace and results in a dshift of signal eigenvectors into the noise subspace. As a result, the observed noise subspace is no longer orthogonal to the steering vectors and the MUSIC algorithm fails. Similar effects make problems with the Capon algorithm, where an inversion of the covariance matrix is necessary. In the case of correlated sources

874

W. Pannert / Applied Acoustics 72 (2011) 873883

Fig. 1.1. Neareld geometry: ~ r q = vector from origin to source, ~ r n = vector from sensor to source, h = angle of source, DF = distance to focus-plane xn = n d position of the microfones, d = distance of the sensors.

role in reconstructing the soundeld. The evanescent waves decay exponentially within several wavelengths which is the present consideration smaller than about 0.34 m (for f P 1000 Hz). The simulation and measurements are carried out at a distance of 1 m or 2 m away from the sources, where the evanescent waves have decayed, but the soundeld is assumed to be well decribed by a spherical wave. A neareld representation of a narrow-band signal representing a point source is a spherical wave. The free-space wave s(rq) for a point source with the power r2 located at rq can be represented by Eq. (1.1). For a uniform linear array (ULA) the distance from the source to the sensors is modeled by rn(h, rq) as shown in Eq. (1.2). The range from the source to one of the n sensors, given by rn, is determined by the distance from the source to the phase center of the array rq, and the distance of the sensor from the phase center of the array xn = n d where d is the distance between two sensors. A single range bin, which is a circle from the phase center of the array, can be selected by holding rq constant and varying h (see Fig. 1.1).

ejkrn sn r q r rn q rn rq cosh2 r q sinh n d2

1:1 1:2

For fareld conditions (rq ? 1), the spherical wave becomes a plane wave and the angle h for a single source is the same for every sensor, whereas it varies in the neareld. The signal for a source located at (rq, h) which impinges on the array with M microphones, can be represented by the vector s.
Fig. 1.2. Forming the smoothed covariance matrix RSS.

s s1 rq ; s2 r q ; . . . ; sM r q T

1:3

the covariance matrix is bad conditioned. (For details to the Delay and Sum Beamforming, MUSIC-Beamforming and Capon-Beamforming see Appendix A.) As a third example Orthogonal Beamforming uses directly the eigenvalues of the covariance matrix and in the sense of a principal component analysis, the P largest eigenvalues can approximately be assigned to the P sources (see Sarradj [13]). The small eigenvalues correspond to noise. If now two sources are correlated one of the associated eigenvalues is approximately zero and the algorithm fails in separating the two sources. The Beamformers considered in this article do not reconstruct the complete soundeld like holographic methods, but they estimate the direction and the power of point sources. 1.1. Neareld model The term neareld is used in this article not in the sense of acoustical holography, where evanescent waves play an important

where si(rq) is the complex amplitude, at a given frequency, of signal of the ith microphone. 1.2. Conventional Beamforming The simplest form of Beamforming is the conventional Delayand-Sum Beamforming (D + S) which uses the time delay between the sensor signals for an assumed source position and sums the channels up after correcting for the time delay. The time delays are contained in a so call steering vector v. To achieve unity gain, the steering vectors v for a steering location rs are scaled by the number of sensors M.

v s1 rs ; s2 rs ; . . . ; sM rs T

1:4

The array-pattern is now given through the scalar-product between the array-vector s and the steering-vector v.

Bh

1 H v s M

1:5

Fig. 2.1. Subarrays in the case of an ULA.

W. Pannert / Applied Acoustics 72 (2011) 873883

875

where the superscript H means the hermitian-conjugate of a vector or matrix. The array pattern is the output of a Beamformer with xed steering position rs, while varying the source location rq. This describes how the output is affected by signals from different loca-

tions. A null in a direction means that the power is low for signals coming from that direction. A peak in a direction means the signals from that direction are passed with up to unity gain. Delay-and-Sum Beamformers have 13 dB sidelobes, if equal weight is given to the signals from all microphones.

Fig. 2.2. Two sources in the fareld with/without spatial smoothing: f = 2000 Hz (frequency of the sources), theta1 = +10 (angle of the 1st source), theta2 = 10 (angle of the 2nd source).

Fig. 2.3. One source in the neareld with/without spatial smoothing: f = 4000 Hz (frequency of the source), DF = 1.0 m (distance array-source), d = 0.05 (sensorspacing), L = M d (size of the ULA).

876

W. Pannert / Applied Acoustics 72 (2011) 873883

1.3. Covariance matrix High resolution Beamformers use the covariance matrix R (R = s sH) of the sensor signals, since it contains all the data needed for each source: power, direction and correlation. In order to understand the effects of correlated signals, a covariance matrix model for a two-source-situation is used. The two-signal model is specied by Eqs. (1.6)(1.8), where v1 and v2 refer to the steering vector to the 1st and 2nd source. P is the correlation matrix of 2 the two sources, r2 i (i = 1, 2) is the power of the signals and rn the power of uncorrelated noise. The statistical correlation be-

tween two signals is p and increases the off-diagonal elements. p has a range from 0 (no correlation) to 1 (perfect correlation).

Delay+Sum - and Capon- Beamforming


0

-5

-10

dB
-15 -20 -25 -1

-0.8

-0.6

-0.4

-0.2

0 x[m]

0.2

0.4

0.6

0.8

Delay+Sum - and Music- Beamforming


0

-5

-10

Fig. 2.5. Two sources in the neareld with/without spatial smoothing: f = 1000 Hz (frequency of the sources), X1 = 0.3 m (position 1st source), X2 = 0.3 m (position 2nd source), DF = 1.0 m (distance array-source), L = 0.8 m (length of the array).

dB
-15 -20 -25 -1

-0.8

-0.6

-0.4

-0.2

0 x[m]

0.2

0.4

0.6

0.8

Fig. 2.4. Two sources in the neareld with/without spatial smoothing: f = 2500 Hz (frequency of the sources), X1 = 0.3 m (position 1st source), X2 = 0.3 m (position 2nd source), DF = 1.0 m (distance array-source), L = 0.8 m (length of the array).

Fig. 3.1. Two-dim. subarrays in a 7 5 regular array; red (full line) = position of the central array. (For interpretation of the references to colour in this gure legend, the reader is referred to the web version of this article.)

W. Pannert / Applied Acoustics 72 (2011) 873883

877

The eigenvectors of covariance matrix R can now be investigated. The eigenvectors which belong to the large eigenvalues of R (signal subspace) span the same subspace as the steering vectors pointing to the targets and the magnitude of the eigenvalues correspond approximately to the power of the signals.

A v 1 " P

v2
#

1:6 1:7

The correlation p from Eq. (1.7) causes many problems for Beamformers based on covariance matrix inversion. In the uncorrelated case, the number of sources is the same as the number of dominant eigenvalues of the covariance matrix. Johnson and Dudgeon [14, p. 386] show that two perfectly correlated signals with |p| = 1 and the same power r2, result in a R of the form shown in Eq. (1.9). The result are M 1 eigenvectors with eigenvalues of r2 n due to noise and a single eigenvector with an eigenvalue of M r2 + r2 n due to signal plus noise.
H 2 R r2 n I r v 1 v 2 v 1 v 2

r2 r1 r2 p 1 r1 r2 p r2 2
H 2 nI

1:9

R APA r

1:8

The two-dimensional signal subspace of two uncorrelated signals has collapsed to an one-dimensional space, where the two correlated signals are mixed into an indistinguishable signal.

Fig. 3.2. Capon-Beamforming: f = 25892690 Hz (frequency range), DF = 2 m (distance array-source), distance between speakers = 0.4 m.

878

W. Pannert / Applied Acoustics 72 (2011) 873883

The solution for this problem in the far-eld case is spatial smoothing unless a better method is found. For an ULA it averages along the diagonal of the covariance matrix as shown in Fig. 1.2. For a two-dimensional regular array it is a little more complicated. The averaging process assumes a plane-wave because only now a source signal will have the same direction of arrival for any section of the array. However this fact is not true in the neareld. Looking to the mechanism how spatial smoothing works, the averaging process reduces the correlation by combining spatially separated parts of the array (subarrays). The correlation of the sources has different effects on different subarrays. This can be seen in the terms proportional to p in the expression for the nth and mth sensor in Rn,m in Eq. (1.10) which describes the two-signal model used in Eq. (1.9). Note that rn(hj) represents the distance from the nth sensor to the source located at an angle hj, rm(hj) the same for the mth sensor. Under fareld conditions the rst two terms depend only on the difference (n m) in an ULA and so the corresponding elements in the covariance matrices Ri of the subarrays are the same. The terms proportional to p vary and are reduced by averaging.
jkr n h1 rm h1 jkrn h2 r m h2 Rn;m r2 r2 p 1e 2e

r1 r2 ejkrn h2 rm h1 pr1 r2 ejkrn h1 rm h2

1:10

In the neareld the rst two terms will also vary from subarray to subarray. This results in a widening of sources the amount of spread depends on the spatial shift of subarrays against each other. The sources are still decorrelated since the p terms still vary more in the neareld. A further effect of spatial smoothing is a loss of spatial resolution through reduction of aperture size due to the forming of smaller subarrays out of the original array. 2. One-dimensional spatial smoothing In this part of the work the effect of spatial smoothing for an ULA under fareld conditions and neareld conditions is studied with simulations. The set-up was an array with M equidistant

sensors and d = 0.05 m spacing. Three subarrays were built and averaged (Fig. 2.1). The 2nd array is the central array which is used for focusing and the 1st and the 3rd array is shifted for 0.05 m compared to the central array. The simulated data had a SNR of 40 dB. Fig. 2.2 presents the effect of spatial smoothing on MUSIC-, Capon-, and D + S Beamforming with an ULA (L = 0.8 m) and two correlated sources with same amplitude and position at 10 in the fareld. The spatial smoothing has decorrelated the two sources which has also an effect on the D + S Beamforming. The number of dominant eigenvalues of the correlation matrix R has changed from one to two. Without spatial smoothing there is only one virtual source in the middle between the real sources. The next series in Fig. 2.3 show the result for Capon and D + SBeamforming for one source under neareld conditions at different values of the array length L. As the spatial resolution DX becomes better (smaller DX) with increasing size of the array one can notice the effect, that the three subarrays see the source under different angles. So the averaged covariance matrix RSS contain three sources. From the result it is clear, that there is only a neareld-effect if the spatial-resolution DX is smaller than the shift of the subarrays in comparison to the central array, which is used for focusing. Similar results are known for Synthetic Apertur Radar (SAR), where range migration effects need not to be compensated, when they are less than the size of the range-resolution. In Fig. 2.4a similar scenario is studied now with two sources in the neareld. The results in the neareld show that spatial smoothing has not the potential as in the fareld situation nevertheless, the dynamic is improved for both algorithms MUSIC and CaponBeamforming as compared to the situation without SS. D + S Beamforming is nearly not inuenced by SS. Further simulations are carried out for different frequencies, source separations and correlation coefcients. The results show that especially in situation, where the resolution of D + S Beamforming, Capon- and MUSIC-Beamforming without SS is insufcient to separate the sources, the application of SS improves the effectiveness of Capon- and MUSIC-Beamforming.

Fig. 3.3. Capon-Beamforming with two uncorrelated sources; same parameters as in Fig. 3.2.

W. Pannert / Applied Acoustics 72 (2011) 873883

879

One example is given in Fig. 2.5. Capon- and MUSIC-Beamforming with SS are able to separate two sources where D + S Beamforming and Capon- and MUSIC-Beamforming without SS clearly fails. 2.1. Summary Especially at low frequencies where the spatial-resolution is poor, Capon and MUSIC Beamforming show their potential of separating two sources after application of SS. The resolution is even better as with the D + S Beamforming. So one can draw the conclusion that mainly at low frequencies, one can apply spatial smoothing without having the effect that a source is split up into several parts.

3. Two-dimensional formulation and experimental setup In the two-dimensional case the array must have a translational symmetry like for example the regular 7 5 array shown in Fig. 3.1. The sensors are clustered into overlapping 6 4 subarrays. One possibility with four subarrays is shown in Fig. 3.1. Processing a scenario with the smoothed covariance matrix, RSS is calculated with a central 6 4 array (red in Fig. 3.1), which only exists virtually. For the experiments a commercial system was used (CAE Noise Inspector) with a 8 6 regular array with 0.1 m distance between the sensors. The system has the benet, that it offers an interface for user-dened algorithms in Matlab-Code. From

Fig. 3.4. Music-Beamforming: f = 25892690 Hz (frequency range), DF = 2 m (distance array-source), distance between speakers = 0.4 m.

880

W. Pannert / Applied Acoustics 72 (2011) 873883

the simulation results in the last section one can deduce, that the shift of the different subarrays compared to the central array (which is virtual and only used for focusing) should not be greater than the spatial resolution of the high-resolution-algorithm. This is a limitation for the number of subarrays and therefore like in Fig. 3.1 four subarrays are used. So the shift of a source seen from a subarray compared to the central array is 0.05 m and one should expect only a splitting of sources, if the spatialresolution is better than 0.05 m. 3.1. Experimental results The following series in Figs. 3.23.6 show the results for two loudspeaker signals and different Beamforming algorithms. In Figs. 3.2 and 3.4 the signals are perfectly correlated (white noise monosignal), whereas in Figs. 3.3 and 3.5 uncorrelated signals are used for comparison. The potential of spatial smoothing is tested with three different Beamforming algorithms high resolution CaponBeamforming, high resolution Music-Beamforming and Orthogonal Beamforming which can display different uncorrelated sources separately. The results are displayed with 6 dB dynamic. The bandwidth of approximately 100 Hz contains the summed results of a narrowband FFT-processing. In Fig. 3.2 the result for Capon Beamforming is presented and the effect of spatial smoothing is compared. In Fig. 3.3 for comparison there is the result for Capon-Beamforming applied to a similar scenario with two completely uncorrelated sources. From the result it is clear, that spatial smoothing in the neareld does not work perfectly, but the potential of resolving correlated sources is given. Next there is the MUSIC-algorithm applied to the same scenario. In Fig. 3.5 for comparison there is the result for MUSIC-Beamforming applied to a similar scenario with two completely uncorrelated sources as in Fig. 3.3. Fig. 3.6 shows the result for Orthogonal Beamforming. Orthogonal Beamforming is able to separate uncorrelated sources approx-

imately. Without spatial smoothing the scenario is resolved in two peaks but they belong to one eigenvalue of the cross spectral matrix. The following pictures show the result after application of spatial smoothing. The signal of the two speakers has been decorrelated and Orthogonal Beamforming can separate the scenario in source 1 and source 2. The number of dominant eigenvalues in the cross spectral matrix has changed from one to two. In Fig. 3.7 for comparison there is the result for Orthogonal Beamforming applied to a similar scenario with two completely uncorrelated sources as in Fig. 3.6.

4. Conclusion The spatial smoothing technique allows the use of Beamforming algorithm which usually are applicable only for scenarios with decorrelated sources. In the present investigation three Beamforming algorithms are compared Capon, MUSIC, and Orthogonal Beamforming. Whereas spatial smoothing in the fareld with plane waves is known to work properly, problems come up in the neareld, because sources appear under different look-angles which correspond to the different subarrays. With the simulation and measurements its shown that under certain conditions this technique is also applicable in the neareld. The conditions are the distance of the centre of the subarrays from the centre of the central array (which is used for focusing) should be less than the spatial resolution of the Beamforming algorithm. Otherwise single point sources are split up. The experimental results show that the potential of the different algorithms is worse compared to the ideal case of not correlated sources, but there is great improvement in resolution compared to the processing without the spatial smoothing technique. Especially at low frequencies, Capon and Music Beamforming provide better resolution than standard D + S Beamforming. Whereas SS has less inuence on D + S Beamforming, the is a great benet of SS for the high resolution algorithms Caponand MUSIC-Beamforming. In the case of Orthogonal Beamforming SS also shows its potential to decorrelate two previously correlated sources.

Fig. 3.5. Music-Beamforming with two uncorrelated sources; same parameters as in Fig. 3.4.

W. Pannert / Applied Acoustics 72 (2011) 873883

881

Fig. 3.6. Orthogonal Beamforming: Df = 1/12 octave at 3140 Hz (frequency range), DF = 2 m (distance array-source), distance between speakers = 0.4 m.

882

W. Pannert / Applied Acoustics 72 (2011) 873883

Fig. 3.7. Orthogonal Beamforming with two uncorrelated sources; same parameters as in Fig. 3.6.

Appendix A. Appendix Expressions for the Beampattern with different Beamforming algorithms. ~ V hi is the steering vector to a location hi.  R is the M M cross spectral matrix of the M microphone signals. The superscript H means the hermitian conjugation (transpose and complex conjugation). A.1. Delay and sum Beamforming

A variational approach leads to:

BCapon hi

1 ~ V H hi R1 ~ V hi

The cross-spectral matrix R is bad conditioned in the case of correlated sources and a regularization is necessary (diagonal loading).

A.3. MUSIC-Beamforming The MUSIC (Multiple Signal Classication) algorithm belongs to the class of subspace algorithms (see Ref. [12]). The term subspace refers to the space of eigenvectors of the cross-spectral matrix R. In the presence of P sources R has P large eigenvalues and M-P small eigenvalues which belong to the noise in the system. ~S and U ~N build the signal-subspace The related eigenvectors U k k and the noise-subspace. The separation between the two spaces occurs through comparison with a threshold.

BDS hi ~ V H hi R ~ V hi
A.2. Capon Beamforming Capon Beamforming tries to minimize the inuence of sources of interference. Ideally it results in a very sharp Beampattern (see Ref. [2]).

W. Pannert / Applied Acoustics 72 (2011) 873883

883

Further on one can show that the steering vectors ~ V hi pointing to the P sources in the scenario, are orthogonal to the noise eigen~N with k = P + 1. . .M. vectors U k

~ ~N 0 V H hi U k

for i 1 . . . P and k P 1 . . . M

With this feature one can form a pseudo-Beampattern which has very distinct maxima in the direction of the P sources (hm = hi)

BMUSIC hm

1 ~N j2 j~ V H hm U k

The height of the maxima has no relation to the signal power of the corresponding source. References
[1] Van Trees HL. Optimum array processing. Wiley Interscience. ISBN 0-47109390-4. [2] Capon J. High-resolution frequency-wavenumber spectrum analysis. In: Proc IEEE, vol. 57; August 1969. p. 140818. [3] Agrawal M, Abrahamsson R, Ahgren P. Optimum Beamforming for a neareld source in signal-correlated interferences. Signal Process 2006;86(5):91523. [4] Lee Ju-Hong, Chen Yih-Min, Yeh Chien-Chung. A covariance approximation method for near-eld direction-nding using a uniform linear array. IEEE Trans Signal Process 1995;43(5):12938.

[5] Evans JE, Johnson JR, Sun DF. Application of advanced signal processing techniques to angle of arrival estimation in ATC navigation and surveillance systems technical report. M.I.T. Lincoln Laboratory, Lexington, Massacusetts; June 1982. [6] Shan Tie-Jun, Kailath T. Adaptive Beamforming for coherent signals and interference. IEEE Trans Acoust Speech Signal Process 1985;33(3):52736. [7] Shan Tie-Jun, Wax M, Kailath T. On spatial smoothing for direction-of-arrival estimation of coherent signals. IEEE Trans Acoust Speech Signal Process 1985;33(4):80611. [8] Reddy V, Paulraj A, Kailath T. Performance analysis of the optimum Beamformer in the presence of correlated sources and its behavior under spatial smoothing. IEEE Trans Acoust Speech Signal Process 1987;35(7): 92736. [9] Takao K, Kikuma N, Yano T. Toeplitzization of correlation matrix in multipath environment, vol. 11; April 1986. p. 18736. [10] Bresler Y, Reddy VU, Kailath T. Optimum Beamforming for coherent signal and interferences. IEEE Trans Acoust Speech Signal Process 1988;36(6):83343. [11] Tsai Churng-Jou, Yang Jar-Ferr, Shiu Tsung-Hau. Performance analyses of Beamformers using effective SINR on array parameters. IEEE Trans Signal Process 1995;43(1):3003. [12] Schmidt RO. Multiple emitter location and signal parameter estimation. In: Proc RADC spectrum estimation workshop. Grifths AFB, Rome, New York; 1979. p. 24358. [13] Sarradj E. A fast signal subspace approach for the determination of absolute levels from phased microphone array measurements. J Sound Vibr 2010;329:155369. [14] Johnson Don H, Dudgeon Dan E. Array signal processing: concepts and techniques. Englewood Cliffs (NJ): PTR Prentice Hall; 1993.

Anda mungkin juga menyukai