Вы находитесь на странице: 1из 8

APPLEBAUM ADAPTIVE BEAMFORMER WITH STEERING VECTOR ESTIMATION FOR CICLOSTATIONARY SIGNALS

Jos Ramn Cerquides Bueno - Juan Fernndez Rubio Signal Theory and Communications Department Polytechnic University of Catalonia E.T.S.I. Telecommunications Barcelona - SPAIN Apdo. 30002 Tel.(34) 3 401 64 58 Fax:(34) 3 401 64 47 E-mail: ramon@tsc.upc.es

ABSTRACT The Applebaum beamformer is well known in the field of array signal processing. It can be mathematically shown that, if the correct steering vector is known by the system, the Applebaum loop can achieve a maximum Signal to Interference-plus-Noise Ratio (SINR). However, little deviations between the values of the ideal and true steering vector result in severe penalizations in the output SINR. Our proposal is to estimate the real value of the steering vector directly from the received snapshots of data. To this subject, properties of cyclostationarity exhibited in the most commonly used communication signals are extracted and exploited, together with a direct iteration method in order to recover the eigenvector associated with the greatest eigenvalue without needing to compute the cyclic autocorrelation matrix. The result is an easy to implement, non-computationally complex system, which shows a fast convergence and low sensitivity to estimation errors.

I. Introduction An adaptive antenna is composed by several sensors located at different space positions. One signal is extracted from each one of these sensors. The output of the system is then constructed as a linear combination of these simple signals, as shown in figure 1. y(n) =
i=1 N

w* i x i (n) =

w x(n)

z
#1 #2

#1

x1 x2

* w1

* w2

where xi(n) are the samples of the signal received by the ith sensor, and wi is the complex weight assigned to this sensor. Its very useful to make use of a vectorial notation, where:
x

r 2 r N

r1 y

#2

#N #N xN
* wN

w(n)=[w1(n),w2(n),..,wN(n)]T is the complex weight vector and x(n)=[x1(n),x2(n),..,xN(n)]T

Fig. 1

is the vector of received data. From the antenna point of view, it's equivalent to have an antenna whose reception pattern can be easily changed by adjusting the values of the weights assigned to the sensors. The problem of beamforming can be stated as finding the optimal values of these weights in order to optimize (i.e. maximize or minimize) some previously defined cost function. Several solutions have been suggested, being able to establish a classification into two large groups, attending to the characteristics of the information provided to the system for rejecting undesired interference or noise. First, we can find the time reference beamformers (TRB), which employ a temporal reference to recover or distinguish between desired and interfering signals. The main problem that appears in this kind of beamformers is the need of transmit this reference. It doesnt bother the method you choose for the transmission, efficiency of the communications system is decreased. Another possibility is to include in the beamformer information related with the spatial environment, (i.e. Direction Of Arrival (DOA) of the signals). The main advantage is that no loss of efficiency results. However, as we shall discuss later, these kinds of beamformers are not free of drawbacks in practical implementations. Blind beamforming, on the other hand, works by restoring some property derived from the signal structure (i.e. Constant Modulus Algorithm). Usually, it implies a slower convergence, and gives place to the appearance of a new problem, known as the "capture problem" which happens when both signals (desired and interfering) exhibit the same properties. Most digital communications signals show a ciclostationary behavior, being able to exploit this fact in order to distinguish between desired signal and interference, thus reducing the capture problem. Our proposal is to take profit of these properties to determine spatial information for the desired signal, and then use this information into an Applebaum loop, the simplest beamformer driven by spatial conditions.

II. Signal model and generalized steering vector The signals collected by the sensors depend, in general, of its position, its reception pattern, the angle of incidence and the propagation conditions, as shown in figure two. In narrowband arrays, and assuming far field propagation, all the previously mentioned effects can be recollected in a single vector with complex elements. This vector is called the generalized steering vector and can be written as:
P

Fig. 2

v= ci a i
i=1

being P the number of paths, ci the ith path complex amplitude, and ai the D.O.A. vector associated to the ith path. Each one of the impinging signals has its own generalized steering vector. The model used for the snapshots is then given by: x(n) =vs(n) +Ai(n) +r(n)

where: v Generalized steering vector for the desired signal s(n) Desired signal A Matrix of generalized steering vectors for the interfering signals Interference's vector r(n) Additive gaussian noise

i(n)

III. The Applebaum beamformer The Applebaum beamformer (fig. 3) is the simplest structure that makes use of spatial information about the desired signal. To lead the weights to his optimal value, the objective taken is to minimize the power of the output signal, while the gain of the antenna is fixed for some generalized steering vector, which may be known in advance, at the design stage. The mathematical formulation is very popular, and we will use here only the final equation, which shows how the weights must be updated:
Signals from other sensors

x i (n) w* i (n) 1-z


-1

y(n)

vi

Fig. 3 where v is the generalized steering vector, i = 1,2 N and is a parameter that controls the convergence of the algorithm. To ensure it, is usually taken as (n)=/Pt(n) where H Pt(n)=x (n)x(n) is a estimation of the total power received, and is a normalized factor which must be chosen in the range 0<<2.
* * w* i (n +1) = wi (n) + vi - xi (n) y (n) ;

The main problem of the Applebaum beamformer is that the generalized steering vector must be known. As mentioned in the introduction, this is a serious drawback. In first place, the phenomenon of multipath propagation is very difficult to control, because of its dependence with the atmospheric conditions, ground reflection coefficients, etc. Another causes of indetermination are the reception patterns of the sensors, usually assumed identical. The same assumption is taken with the splitters and preamplifiers that precede the beamformer. Differences of gain between channels (in modulus or phase) are difficult to avoid or contemplate in a "a priori" design. Putting together all these effects leads to differences between the ideal value and the real value of the generalized steering vector. The problem is aggravated if we take into account the high sensitivity of the Applebaum loop to deviations in the steering vector.

IV. Cyclic autocorrelation matrix and steering vector It's said that a signal x(n) is ciclostationary (in a wide sense) if its mean x(n) and its autocorrelation are both periodic functions in n with period M, that is: x (n) = E{x(n)} = E{x(n + lM)} = x (n + lM) R xx (n, m) = E{x(n) x* (n - m)} = E{x(n + lM) x* (n + lM - m)} = R xx (n + lM, m) These properties hold for many communications signals, especially those used in digital communications. However, an important requirement must be accomplished: the signal must be oversampled (that is, more than a sample per symbol is needed). Our interest is essentially put in the periodicity shown by the autocorrelation function, because communication signals usually show zero mean. Thus, given a periodic autocorrelation function Rxx(n,m), it's possible to express it as a Fourier series development. That is: R xx (n, m) =
2 kn 1 N-1 k R xx (m) e j M N k=0

where the coefficients of the development are the so-called cyclic autocorrelation functions while the parameter k plays the role of a digital frequency that we will call cycle frequency. The function Rx(m) can be obtained as:
N-1 k (m) = R xx

n=0

Rxx (n,m)

e-j

2 kn M

N-1 2 kn = E x(n)x*(n-m) e-j M n=0

Plots of the magnitude of Rxx(n,m) and Rx(m) for a BPSK baseband signal with nine samples per period are shown in figures 4 and 5 to help clarify concepts. It's interesting to note that, when k=0, the cyclic autocorrelation function is equal to the classical autocorrelation. For other values of k, the function has valuable information that does not appear in the classical autocorrelation function. Another interesting consideration is that the cyclic autocorrelation function of a stationary signal will only take value other than zero when k equals 0. The same happens when the function Rx(m) is computed for a ciclostationary signal whose period (say L) is different from M (exceptions must be done for some values of k in the special case of M being an integer multiple of L).

Fig. 4. Plot of Rxx(n,m)

Thus, returning to the field of arrays, the cyclic autocorrelation function can be used as a powerful tool to difference between desired signal and interfering ones. Let's continue with our previous model for the received signals. If we make the next set of assumptions: 1.- The cyclic autocorrelation function of s(n) for some set of values of M, k and m is not zero (some of these values will be chosen by design, while others may be available at the synchronization circuits).

Fig. 5. Plot of Rx(m)

2.- The cyclic autocorrelation function for the interfering signals and the noise is zero. These assumptions can be expressed mathematically in the following way: k (m) = k (m)Ps 0 k0 Rss ss m, M fixed R ikq i r (m) = 0 q, r = 1 I Rk n q n r (m) = 0 q, r = 1 N where I is the total number of interferences, N is the number of sensors, Ps is the mean power of the desired signal, and s(m) is a normalization parameter that measures the quantity of cyclostationarity of the desired signal. We will call it cyclic autocorrelation coefficient. The cyclic autocorrelation matrix can be defined as:
N-1 2 kn k R xx (m) = E x(n) x H (n - m) e-j M n=0

Substituting the signal model previously stated in this equation and making use of the established assumptions leads to:
k k R xx (m) =vvH ss (m) Ps

showing that the cyclic autocorrelation matrix is a rank-one matrix if k,m and M are conveniently chosen. Another important fact to remark is that the eigenvector associated with the only non-zero eigenvalue is exactly the generalized steering vector of the desired signal.

V. Estimation of the generalized steering vector Unfortunately, it's impossible to dispose of the cyclic autocorrelation matrix in the receiver, because it has been defined as the expected value, that is, taking a average over the realizations. However, if the signals involved are cycloergodic (and this condition holds for

all the communications signals), it's possible to estimate this matrix from the data that is being received. One possibility (although it does not correspond to the maximum likelihood criterion) that allows to take into account only the most recent samples is to take an iterative estimator of the form:

[R

k xx (m) n+1 = _

[R

2 kn k xH (n - m) e-j M xx (m) n + (1- )x(n)

where is called the forgetting parameter. An easy way of obtaining the value of the generalized steering vector v from the estimated cyclic autocorrelation matrix is given by the direct iteration method. In order to reduce the computational load of the algorithm the next solution is proposed: 1. Accept that [Rx(m)]n can be decomposed in the following form:

[R

k v vH xx (m) n _ n n n

where the indexes k and m have been removed on the right hand of the equation only for notational simplicity purposes. Vector vn is assumed to be normalized in some form (i.e. vvn=1 ) 2. We do now compute the matrix [Rx(m)]n+1, and project the vector vn over this matrix. The new vector obtained is called un+1: u n+1 = n vn + (1- )x(n) x H (n - m) e-j
2 kn M

vn

3. Normalization of un+1 yields the values of the new steering vector vn+1 and the associated eigenvalue n+1: n+1 = _ u n+1 _ vn+1 = u n+1 n+1

Under the assumptions taken in the model, the so constructed succession (vn,n) converges to (v,) as n grows. The final weight adaptation equation can be definitely written as: wH (n +1) = wH (n) + vn -x(n) y* (n) (n) Pt

VI. Simulations and results Array geometry: Linear equally spaced array with 6 sensors.

Signal scenario:

The desired signal is a BPSK signal, which is received in presence of a interfering unmodulated carrier and additive white gaussian noise.

Signals parameters: Desired signal: BPSK signal with 3 paths. PATH SNR D.O.A. 1 5 dB 0 2 -5 dB 30 3 0 dB 20

Number of samples per symbol: M=2 Cyclic parameters chosen: k=m=1 Interfering signal: Unmodulated carrier. Angle of arrival: -20 dB INR=5 dB

Other parameters of the simulation: =0.99 =0.1 v0=[1,0,0,0,0,0]T 0=1; w(0)=[1,0,0,0,0,0]T

Initial estimated steering vector Initial estimated eigenvalue Initial weight vector

The results of the simulation are shown in figs. 6 and 7. Graph 6 shows the evolution of the output SINR, while graph 7 corresponds to the reception pattern after 500 samples. Fig. 6. Evolution of the output SINR Fig. 7. Reception pattern after 500 samples

VI. Conclusions

A new method of beamforming has been developed. It exploits the fact that, when the signal subspace has rank one, the only existing signal eigenvector coincides with the desired signal generalized steering vector. Cyclostationary properties have been used to construct the cyclic autocorrelation matrix, allowing us to make disappear information related with the interfering signals and noise. The direct iteration method is used to determine the eigenvector of interest, and has been formulated in a such a way that avoids the need of computing the cyclic autocorrelation matrix, with the consequent reduction of computational load. The algorithm does not depend on the geometry of the array or the sameness of the reception patterns, and its independent of the phase and magnitude deviations between different channels. Finally, in the simulations, the algorithm has proven to have a fast convergence towards the optimal value of the weights and very low sensitivity to instantaneous estimation errors of the steering vector.

Вам также может понравиться