A Maximum a Posteriori Approach to Beamforming in the Presence of Calibration Errors A. Swindlehurst Dept. of Elec. & Comp. Engineering Brigham Young University Provo, UT 846 Abstract The performance of DF-based beamformers is seriously degraded in situations the array is imprecisely calibrated, or when the spatial coherence of the signal wavefronts is perturbed. When the calibration errors or perturbation may be characterized by a set of parameters drawn from a known Gaussian distribution, a maximum a posteriori () estimator may be used to separately estimate the directions of arrival and the perturbation parameters, resulting in essentially an on-line auto-calibration. This paper examines the improvement that results from using the auto-calibrated steering vectors in standard DF-based beamformers to estimate the received signal waveforms and suppress unwanted interference. For the special case of additive unstructured calibration errors and uncorrelated signals, it is shown that the beamformer is similar in form to so-called \subspace corrected" approaches.. Introduction All methods for direction-nding (DF) and DF-based beamforming rely on the availability of information about the array response, and assume the signal wavefronts have perfect spatial coherence. The assumption of a known array response (e.g., from calibration data or a known array geometry, etc.) and idealized wave propagation are never satised in practice. Due to changes in antenna location, temperature, and the surrounding environment, the response of the array may be signicantly dierent than when it was last calibrated. Furthermore, the calibration measurements themselves are subject to gain and phase errors, and they can only be obtained at discrete points in the parameter space (thus necessitating interpolation techniques for uncalibrated directions). For the case of analytically calibrated arrays of nominally identical, identically oriented elements, errors result since the elements are not really identical and their locations are not precisely known. Furthermore, even if the calibration data were precisely known, inhomogeneous propagation eects cause the actual array response to a given signal to be dierent from the response at the time of calibration. Depending on the degree to which the actual antenna response diers from its nominal value, DF and beamformer performance may be signicantly degraded. To account for the eects described above, a slightly generalized model for the array response will be considered in this paper. The response will be parameterized not only by the directions of arrival (DOAs) of the signals, but also by a vector of perturbation or \nuisance" parameters that describe deviations of the response from its nominal value. These parameters can include, for example, displacements of the antenna elements from their nominal positions, uncalibrated receiver gain and phase osets, etc.. With such a model, a natural approach is to attempt to estimate the unknown nuisance parameters simultaneously with the signal parameters. Such methods are referred to as auto-calibration techniques, and have been proposed by a number of authors, including [,,, 4, 5] among many others. When auto-calibration techniques are employed, it is critical to determine whether both the signal and nuisance parameters are identiable. In certain cases they are not; for example, one cannot uniquely estimate both DOAs and sensor phase characteristics (unless of course additional information is available, such as sources in known locations [6], etc.). The identiability problem can be alleviated if the perturbation parameters are assumed to be drawn from some known a priori distribution. While this itself represents a form of additional information, it has the advantage of allowing an optimal maximum a posteriori () solution to the problem to be formulated [7, 5]. In [5] it is shown that, by using an asymptotically equivalent approximation to the resulting criterion, the estimation of the signal and nuisance parameters can be decoupled, leading to a signicant simplication of the problem. Presumably, any of the above auto-calibration methods would provide not only improved DOA estimates, but also calibration information that would be useful in beamformer implementation. In this paper, beamformer performance is investigated for the case the optimal perturbation parameter estimates of [5] are used to update the array calibration. Simulations demonstrate that such an approach can result in a signicant performance improvement, measured using either interference rejection capability or mean-squared error. In addition, for simple additive unstructured cal-
ibration errors, the approach is shown in certain cases to yield a beamformer similar to the subspace corrected algorithms described in [8, 9, ].. Mathematical Model and Algorithms The response of an arbitrary array of m sensors for a given DOA will be denoted by the m-vector a(; ), which is parameterized by a vector IR p that describes the array perturbation. The array output is then modeled by the following familiar equation: x(t) [a( ; ) a( d ; )] 6 4 s (t). s d (t) 7 5 + n(t)() A(; )s(t) + n(t) ; () s(t) and n(t) represent the received signals and noise, respectively. Assuming samples are taken from the array, the following data matrix may be dened: X [x(t ) x(t )] A(; )S + ; () S and are dened similarly to X. Both s(t) and n(t) are assumed to be temporally white zero-mean complex Gaussian random processes, with covariances given by I and P lim! X t s(t)s (t) ; (4) respectively ( and P are unknown and must be estimated). The perturbation vector is also assumed to be drawn from a Gaussian distribution with known mean (corresponding to the nominal, unperturbed array response) and covariance. Given the above, the covariance of the array output and its eigendecomposition may be written as R lim! X t E[x(t)x (t)] (5) A(; )PA (; ) + I (6) E s s E s + E n E n ; (7) s is a diagonal matrix containing the d largest eigenvalues, and the columns of the md matrix E s are the corresponding unit-norm eigenvectors. Similarly, the columns of E n are the m? d eigenvectors corresponding to. Using well known arguments, one can show that E n is orthogonal to A(; ), while the range space of E s coincides with that of A(; ) when P is full rank... An Asymptotic Estimator In [5], it is shown that estimates of and asymptotically equivalent to those from the exact estimator may be obtained by setting ^ arg min a ^Ma? ^f T ^??^f (8) ^??? f ; (9) a vec(a(; )) () ^M ^U T ( ^E n ^E n) () ^U ^? ^A y ^E s ~ ^? s ^? Re ^D ^M ^D +? ^E s ^A y () () ^f Ref ^D ^Ma g (4) D @a(; ) @a(; ) ; : : : ; ; (5) @ @ p ; and ^ and ^A are \consistent" estimates determined from some initial estimation step. The above approach is quite general in that, by proper choice of, it can be applied to arbitrary types of model errors. Another key advantage is that the estimation of and is decoupled; a search is required only for the d DOA parameters in, and not for (which is calculated directly given ^). Other properties and uses of the algorithm are outlined in [5]... Optimal Beamformers The minimum mean squared error (MSE) beamformer weights are given by W MSE arg min W kw X? Sk F (6) R xs lim! R? R xs R? A(; )P ; (7) X t x(t)s (t) A(; )P : When the desired signal is uncorrelated with the interference, P is diagonal and the minimum MSE solution is Strictly speaking, the equivalence of the above estimator and the optimal approach only holds for rst order errors? that are \of the same order" as the nite sample eects of the noise. In other cases (particularly those model errors are dominant), a dierent approach should be used. For more details, see [5, ].
just a scaled version of the so-called minimum variance distortionless response (MVDR) beamformer: W R? a(^) a (^)R? a(^) : (8) In the general case the signal and interference are correlated, the optimal weights depend on the signals themselves through R xs or P, and thus they cannot be used directly (i.e., without a training sequence, for example). In the approach of [], the quantities P and R in (7) are replaced by their structured ML estimates: ^P s A y ( ^R? ^ I)A y ^R s A ^Ps A + ^ I ; A A(^; ), () y denotes a (left) pseudoinverse, and ^R is a sample estimate of R. Since calibration errors were not addressed in [], the nominal model was used to calculate the beamformer weights. evertheless, the method performs well when calibration errors are present, as recently demonstrated in []. On the other hand, the MVDR approach is well known to be hyper-sensitive to array perturbations, especially at high SR. While ad hoc methods employing articial noise injection have been used to combat this problem, other techniques based on subspace corrected (SC) weights have found success in experimental systems [9, ]. In these approaches, the R? term in (8) is replaced by E s s? E s. This is equivalent to projecting a(^) onto the signal subspace prior to forming the MVDR weights. One of the goals of this paper is to study the improvement that results from using the method of [] with A(^; ^) rather than A(^; ), ^ is obtained from the estimator in (9). This approach will be referred to as the beamformer in the sequel. In the next section, an interesting connection is made between the beamformer and the SC-MVDR method. In particular, it is shown that for simple unstructured array errors and uncorrelated signals, the SC-MVDR and weights have a very similar form.. Some Special Cases For the moment, consider the following simple unstructured model for the perturbed array response: A(; ) A() + A ~ ; (9) Refvec( ~A)g Imfvec( A)g ~ ; () and the columns of ~A, denoted ~a i, are modeled as zero mean Gaussian random vectors with moments E[~a i ~a k] ik I i; k ; ; d () E[~a i ~a T k ] i; k ; ; d: () This model corresponds to an additive, circularly symmetric complex array perturbation that is uncorrelated from sensor to sensor, but possibly -dependent. It is easy to verify that under these assumptions, the covariance of is given by Refg I? Imfg I ; () Imfg I Refg I the i; k th element of the matrix is ik. It is interesting to examine the form of the estimate ^ for this case. To begin with, note that for the above model and ^D [I ji], I is md md. Thus, ^??? f, and " Re( ^M +?? I)? Im( ^M + #? I) Im( ^M +? I) Re( ^M +? I) f Ref ^Ma g : Imf ^Ma g Using the fact that, for any invertible matrix Z,? RefZg? ImfZg RefZ? g? ImfZ? g ImfZg RefZg ImfZ? g RefZ? ; g (4) it is easy to show that Re? ^M + ^? 6? I ^Ma 4 7 5 ^Ma : (5) Im ^M +? I? A further simplication of (5) is possible that is quite revealing. Using the denition of ^M in (), note that ^M +?? I ( ^U T +? ) ( ^E n ^E n) +?? ( ^E s ^E s) ^U T +?? ( ^E n ^E n) + ( ^E s ^E s) : Multiplying the last equation above on the right by ^Ma and simplifying then yields ^? nh? 4 Re I + (U T )? i o? ( ^E n ^E n) a Imnh? I + (UT )? i o? ( ^E n ^E n) a 5 : (6)
Finally, using (9)-() and properties of the Kronecker product, the estimate of the array response becomes A(^; ^) A(^)? ^E n ^E na(^) I + (UT )?? : (7) The key point of interest is that, if?!, then the estimate of the array response converges to a subspace corrected version of the nominal response: lim A(^; ^) ^E s ^E sa(^) :?! Furthermore, if the estimated array response is used in (8), the MVDR beamformer (8) will converge to the SC-MVDR approach. The condition?! occurs either with a large data sample, or when the array perturbation is large. In either case, the information provided by the prior distribution of is of little value, and is essentially ignored by the criterion. This observation provides some theoretical justication for the SC-MVDR technique, which previously had been derived using ad hoc (but well motivated) reasoning. However, in cases the prior cannot be neglected, using SC response vectors for beamforming will not be optimal and signicant degradation can result. This is seen in the simulation examples described later... Gain and Phase Errors For arrays composed of nominally identical elements, a common approach used to describe deviations in the array response attempts to model the non-uniform gain and phase eects of the receiver electronics behind each antenna element. In this model, the nominal response is perturbed by an unknown complex diagonal matrix: and A(; ) GA() ; (8) Refgg ; (9) Imfgg g diagfgg. The mean of the distribution for in this case is given by [e T ] T, e is an m vector of ones. For simplicity, in this discussion the covariance of will be assumed to be (a )I, which implies that the individual gain and phase errors are all mutually independent and identically distributed. The derivation of the estimate of and hence g is straightforward but somewhat cumbersome, and thus will not be presented here. However, the result is quite simple, and is given by Z 4 d X i;k ^g? I + a Z? e ; () u ki a(^ i )a T (^ k ) 5 ( ^E n ^E n) ; () Interference Gain (db) 5 5 5 MVDR 5 5 5 5 umber of Sensors Figure : A Comparison of Beamformer Performance, Unstructured Calibration Errors u ki is the k; i th element of ^U, () denotes conjugation, and an element-wise (Hadamard) product. ote that for very small gain/phase errors a!, ^g! e and hence ^G! I as expected. 4. Simulation Results In this section, the performance of the beamformer is studied by means of a number of simulation examples. The rst example involves a nominally unitgain uniform linear array perturbed by an unstructured calibration error in the form of equation (9)-() with ai and a :. The array receives samples of two db SR uncorrelated Gaussian signals with arrival angles of 5 and 5. Using DOA estimates from the optimal estimator, the relative interference rejection capability of the MVDR, SC-MVDR, and beamformers was calculated for various array sizes. The results are plotted in Figure based on 5 independent trials. The plot shows the gain of the beamformer weights for the 5 source in the direction of the 5 interferer (normalized for a unit gain response at 5 ). The subspace correction eliminates the signal cancelation effect of the MVDR approach, but the beamformer provides a signicant advantage, especially for larger arrays. The above simulation was repeated assuming receiver gain/phase errors as described by (8)-(9), also with a :, and the results are plotted in Figure. Algorithm performance is seen in this case to depend very little on the type of calibration error encountered. When the signals arriving at the array are highly correlated, interference rejection is no longer an appropriate performance criterion. In such cases, an optimal beamformer will attempt to combine correlated arrivals with the desired signal to improve the quality of the resulting estimate, as measured using (for example) mean-squared error. To examine beamformer
Interference Gain (db) 5 5 MVDR ormalized RMS Error.8.7.6.5 Uncompensated MMSE (Training Signal) 5.4 5 5 5 umber of Sensors Figure : A Comparison of Beamformer Performance, Gain- Phase Calibration Errors..5..5..5..5.4 Multipath Time Delay in Symbol Periods Figure : Root MSE Performance of Various Beamformers for a Multipath Channel performance for the case of correlated signals, a tworay multipath channel was simulated for various relative delays. A miscalibrated 5-element linear array was assumed to receive a random QPSK signal from?6, as well as a slightly delayed copy of the signal from 6. Both arrivals had an SR of db, and the array was again perturbed according to (9)-() with ai and a :5. For each trial, DOA estimates were obtained based on 75 samples from the array, and normalized RMS signal errors were computed. The results are plotted in Figure for various relative delays between the two arrivals. The \uncompensated" approach corresponds to the method of [] implemented with A(^; ) rather than A(^; ^) as in the beamformer. The minimum MSE curve was obtained using a known 75-sample training sequence to compute the optimal weights, and was included to give an idea of the \best possible" performance. While the SC-MVDR approach can to some degree compensate for array perturbations, it cannot eliminate signal cancelation due to the presence of a correlated arrival, and its performance in this case is quite poor. For small delays, correcting for calibration errors yields a 5-% improvement in RMS error, which translates into a reduction in symbol error rate of approximately a factor of 6 (from.4 to.7) for this example. References [] A. Paulraj and T. Kailath, \Direction-of-Arrival Estimation by Eigenstructure Methods with Unknown Sensor Gain and Phase", In Proc. IEEE ICASSP, pages 7.7.{7.7.4, Tampa, Fla., March 985. [] Y. Rockah and P. M. Schultheiss, \Array Shape Calibration Using Sources in Unknown Locations { Part I: Far-Field Sources", IEEE Trans. on ASSP, 5:86{99, March 987. [] A. J. Weiss and B. Friedlander, \Array Shape Calibration Using Sources in Unknown Locations - A Maximum Likeli- hood Approach", IEEE Trans. on ASSP, 7():958{966, Dec. 989. [4] M. Wylie, S. Roy, and H. Messer, \Joint DOA Estimation and Phase Calibration of Linear Equispaced (LES) Arrays", IEEE Trans. on Sig. Proc., 4():449{459, Dec. 994. [5] M. Viberg and A. Swindlehurst, \A Bayesian Approach to Auto-Calibration for Parametric Array Signal Processing", IEEE Trans. on Sig. Proc., SP-4():495{57, December 994. [6] J. Lo and S. Marple, \Eigenstructure Methods for Array Sensor Localization", In Proc. ICASSP, pages 6{6, Dallas, TX, 987. [7] B. Wahlberg, B. Ottersten, and M. Viberg, \Robust Signal Parameter Estimation in the Presence of Array Perturbations", In Proc. IEEE ICASSP, pages 77{8, Toronto, Canada, 99. [8] D. Tufts, I. Kirsteins, and R. Kumaresan, \Data Adaptive Detection of a Weak Signal", IEEE Trans. on Aero. and Elec. Sys., 9():{6, Mar. 98. [9] R. Schmidt and R. Franks, \Multiple Source DF Signal Processing: An Experimental System", IEEE Trans. on Antennas and Propagation, 4():8{9, March 986. [] B. Wahlberg, I. Mareels, and I. Webster, \Experimental and Theoretical Comparison of some Algorithms for Beamforming in Single Receiver Adaptive Arrays", IEEE Trans. on Antennas and Propagation, 9():{8, Jan. 99. [] M. Viberg and A. Swindlehurst, \Analysis of the Combined Eects of Finite Samples and Model Errors on Array Processing Performance", IEEE Trans. on Sig. Proc., 4():7{8, ovember 994. [] B. Ottersten, R. Roy, and T. Kailath, \Signal Waveform Estimation in Sensor Array Processing", In Proc. rd Asilomar Conference on Signals, Systems, and Computers, pages 787{79, Asilomar, CA., ovember 989. [] J. Yang and A. Swindlehurst, \The Eects of Array Calibration Errors on DF-Based Signal Copy Performance", IEEE Trans. on Sig. Proc., 4():74{7, ovember 995.