Professional Documents
Culture Documents
INTRODUCTION
1052 IEEE TRANSACTIONS ON AEROSPACE AND ELECTRONIC SYSTEMS VOL. 46, NO. 3 JULY 2010
jamming signal is present, ESPRIT would not be TABLE I
able to estimate the DOA of any additional desirable Algorithm Overview
signal. Thus, it is not usable in the problem we are
Step 1: Find beamforming weights that minimize the jammer’s
trying to solve [9]. power.
Despite the stated disadvantages, MUSIC is the Step 2: Apply threshold detection to the beamformer outputs
existing algorithm that seems best suited for our of each range-Doppler bin of interest.
system. Since it is usable with our simple hardware Step 2a: If a target is detected, record its range and Doppler
structure, MUSIC is more statistically efficient than and proceed to Step 3.
Step 2b: If no target is detected, start over with the next
Capon and offers a much reduced computational coherent processing interval (CPI).
burden as compared with ML estimation [6]. Because Step 3: Estimate relative phase information for each detected
of its popularity in literature, an analytical and target.
numerical comparison of MUSIC and our algorithm Step 4: Calculate DOAs from the phase information.
will be given later.
The rest of the paper is organized as follows.
In Section I, first, we give a full description of The variables ®(n) and Á(n) respectively denote
our algorithm, starting with the system model and the amplitude and the time-varying phase of the target
continuing with a tabular list of algorithm steps. at antenna 1, while μ and ± denote the relative phase
Next, we proceed with a detailed description on our angles at antenna 2 and 3. In a similar manner, the
methodology for interference cancellation, target parameters ¯(n), ¸(n), ², and ´ denote the amplitude,
detection, and phase angle estimation. Afterwards, we time-varying phase, and electrical phase angles of the
analytically identify the spatial scenarios of a jammer jamming signal. The noise, vi (n) is a white zero-mean
and a target in which the proposed technique will complex random variable with variance ¾ 2 and is
reliably estimate a target’s DOA. A similar analysis uncorrelated with vm (n) for i 6= m. All Greek letter
of MUSIC is given that suggests the two algorithms variables represent real numbers.
will exhibit the same spatial dependency. Next, in We now give an overview of our algorithm which
Section III the statistical performance of the algorithm does not fit either of the paradigms introduced above,
is examined through a collection of simulations. In i.e., we do not scan a narrow beam nor do we use a
Section IV, we present a brief radar signal processing parametric method to estimate the steering vectors
example to demonstrate how our proposed algorithm of all present source signals. Throughout the rest of
might be used in practice. Finally, Section V contains this paper, we refer to the desired signal as a target
the conclusions of this work. signal because this approach has been motivated
from the signal processing needs of a radar system.
We have also chosen to use a noise jammer for the
II. SYSTEM MODEL interference source because of the ease at which one
can be simulated, but application need not be limited
Three antennas in an arbitrary geometry make up to this case. The algorithm steps are enumerated in
our receiver structure. The received signal at the ith Table I.
element at time n, is denoted by xi (n) and is formed Before proceeding to the algorithm details,
from the coherent addition of the target signal ti (n), we make some additional comments. In regards
the jammer signal ui (n), and the noise vi (n). Therefore, to Step 3, in the weighted sum, the jammer’s
xi (n) = ti (n) + ui (n) + vi (n), i = 1, 2, 3: (1) power is negligible, but the target’s directional
phase information is distorted. For this reason
Assuming point sources and equal gains for the three the phase information has to be recovered in a
receivers, the target and interfering signal at each novel manner. The recovery is accomplished by
sensor will be phased replicas [10]. We also assume arranging those sums in a phasor diagram and
narrowband signals, which means that the relative exploiting the problem’s natural symmetry. At Step
phases of the received signals will be constant across 4, the DOAs of multiple targets can be estimated,
the entire band. The target signals are modeled as sequentially, if those targets are resolvable in range
or Doppler. Our approach is philosophically similar
t2 (n) = t1 (n)ejμ , t3 (n) = t1 (n)ej± (2)
to [11] the Applebaum and Wasiewicz algorithm
where for jammer suppression and target DOA estimation
t1 (n) = ®(n)ejÁ(n) (3) in monopulse systems. Like [11], instead of using
beamforming in the traditional fashion, to maximize
and the interfering signals are the signal-to-interference-plus-noise ratio (SINR),
u2 (n) = u1 (n)ej² , u3 (n) = u1 (n)ej´ (4) beamforming is used to null a jamming signal.
Nulling the jammer enables a reduced-complexity
where mathematical technique for estimating target
u1 (n) = ¯(n)ej¸(n) : (5) signal parameters. Unlike [11], we employ phase
xi,m = [xi , xm ]> , i = 1, 2, 3 and i 6= m (6) and the above ensemble average is estimated with the
average of Nt noisy snapshots by the calculation
1
wi,m = p [1, ej!i,m ]> : (7) N
2 1 Xt
1054 IEEE TRANSACTIONS ON AEROSPACE AND ELECTRONIC SYSTEMS VOL. 46, NO. 3 JULY 2010
For our three-antenna receiver, the process
of averaging jammer phase angles or performing
eigenvalue decomposition (EVD) only needs to be
carried out twice because any third weight vector can
be calculated from the first two. From (4) and (9), we
know that the optimal !1,2 = ¼ ¡ ², and the optimal
!1,3 = ¼ ¡ ´. It can also be shown that, in order to
cancel the jammer from the beamformer output of
antennas 2 and 3,
!2,3 = ¼ ¡ ´ + ²: (17)
Fig. 1. Vector diagram showing derivation of angle of r1,2
By combining the equations for each of the weight (i.e., 6 r1,2 ) in (24).
phase angles we arrive at the relationship
!2,3 = !1,3 ¡ !1,2 + ¼: (18) over any map region of interest. Using (19) an entire
“cleaned” range-Doppler map can be created from
B. Signal Detection values of p. An example on target detection is given
in Section IV.
After the training snapshots are collected and Depending on antenna geometry or system
the weights have been estimated, they are used to requirements, alternate definitions for a detection
search for a target signal within the jammer-dominated variable might be useful. One example would be to
received signal. Multiplying the weighting vectors use p = 13 (jr1,2 j2 + jr1,3 j2 + jr2,3 j2 ). Another possible
with their respective received signal vectors results detection criteria would be to use an m-of-n threshold
in the residue power detection variable detection on the three beamformer outputs, e.g.,
3 3 2-of-3 jri,m j2 should exceed the threshold to declare
1X > 1X a detection.
p= jw1,m x1,m j2 = jr1,m j2 : (19)
2 2
m=2 m=2
1056 IEEE TRANSACTIONS ON AEROSPACE AND ELECTRONIC SYSTEMS VOL. 46, NO. 3 JULY 2010
this occurs by inspecting (32). If an entry of EN is
zero, any particular inner product with that vector can
be achieved by an infinite number of steering vectors.
Therefore, even at MUSIC’s asymptotic limit, if
μ = ², an EL estimate cannot be made; if ± = ´, an AZ
estimate cannot be made; and if ¡μ + ² + ± ¡ ´ = 0,
neither an AZ nor an EL estimate can be made. Fig. 2. Array geometry used in simulations.
From the above analysis, one would expect the
proposed algorithm and MUSIC to behave similarly case is more complex than our phase angle estimation
for large numbers of snapshots. Unfortunately, it method by a factor of 1,348.
is generally very difficult to analytically trace the
performance of a DOA estimation algorithm for F. Application to an L-Shaped Array
multi-source signal cases [10, 16, 17]. In order to
compare the performance of our proposed algorithm Although our algorithm can be applied to general
and MUSIC as a function of snapshots, we have array shapes, to further analyze its properties we now
instead conducted a simulation which is described in consider a highly utilized array shape, the L-shaped
Section III. array. The L-shaped array is popular because it can be
It is difficult to directly compare the computational used to estimate DOA angles in two dimensions. Also,
complexity of our algorithm and MUSIC because AZ and EL angles can be defined such that spatial
of the various engineering decisions that must be angle only depends on a single electrical phase angle.
made before applying either. Consequently, from Here we designate the relative signal phases at sensor
our algorithm we omit from this comparison the 1 and 2 to measure AZ, and the relative signal phases
computations required to estimate the beamforming at sensor 1 and 3 to measure EL, i.e.,
weights, and from MUSIC, we omit the computations μ ¶ μ ¶
μ ±
required to estimate the 3 £ 3 cross-correlation AZ = arcsin , EL = arcsin (36)
¼ ¼
matrix, the subsequent EVD, and the selection
of the correct Pmu peak. In other words, for our where the distance from antennas 1 to 2 and from 1
algorithm we consider here only the computations to 3 is a half-wavelength. Thus, antenna 1 is at the
required in calculating the beamformer outputs and corner of the L-shape and the point 0± AZ and 0± EL
the phase angle estimation. For MUSIC we consider corresponds to the boresight of the receiver array. The
only the grid search of (32). In the description array structure is depicted in Fig. 2.
that follows the multiplication of two complex With a half-wavelength spacing and traditional
numbers involves four “real multiplications,” and interferometry, relative signal phases would measure
the addition of two complex numbers two “real between §¼. However, when using the proposed
additions.” Divisions are counted as multiplications, algorithm, μ and ± will not fall within this range in
and subtractions are counted as additions. We general. Therefore, before applying (36), μ and ± must
assume that measuring the phase angle of a first be appropriately unwrapped.
complex number can be done inexpensively
through the use of a look-up table and that the III. NUMERICAL RESULTS
complexity involved in multiplying a number by 2
is negligible. Now, we present a set of numerical results which
Averaging our phase angle estimates from are obtained from simulations. All of the simulations
N jammer-plus-target-plus-noise snapshots, our were performed with the L-shaped array described
algorithm requires 24N + 1 real multiplications, above. First, the algorithm’s statistical performance
12N ¡ 1 real additions. For MUSIC, each grid point for estimating target DOA is compared with the
requires 16 real multiplications and 4 real additions. baseline interferometer performance when only
We now compare the complexity of the phase angle a target signal is present, i.e., no jamming. Next,
estimation in our algorithm with the complexity of we revisit the issue of jammer-to-target angular
a MUSIC grid search in an example. If we use a separation and demonstrate how the separation can
typical N value of 5, our algorithm requires 121 affect performance. Last, our algorithm’s performance
real multiplications and 59 real additions. For the is compared with that of the well-known MUSIC
MUSIC grid search we use points with 1± of precision algorithm.
from ¡50± to 50± AZ and ¡50± to 50± EL. This
example grid has 10,201 points, resulting in a total of A. Simulation 1
163,216 real multiplications and 40,804 real additions.
Therefore, even though we limited the search span to The target was placed at 0± AZ, 0± EL. The
be much smaller than that of a full hemisphere and jammer signal, simulated as a white noise signal,
used a coarse step size of 1± , MUSIC’s search in this was placed at 22± AZ, 55± EL. The jammer-to-noise
1058 IEEE TRANSACTIONS ON AEROSPACE AND ELECTRONIC SYSTEMS VOL. 46, NO. 3 JULY 2010
Fig. 6. Estimation variances for simulation 3. Fig. 7. Estimation variances for simulation 4.
signal was placed at 46± AZ, 15± EL with a JNR of again used to estimate the beamforming weights.
50 dB. Twenty snapshots were utilized to estimate In Simulation 3, 15 target-plus-jammer-plus-noise
the jammer covariance matrices and the canceling snapshots resulted in nearly equivalent average
weights. The target’s DOA was then estimated using performance of the two algorithms, so 15 snapshots
target-plus-jammer-plus-noise snapshots, while the were used here as well. The estimation variances
number of these snapshots was varied from 1 to 15. are plotted in Fig. 7 after 1000 iterations. The
For the cases when the number of snapshots was performance of the two algorithms is quite similar.
greater than 1, we averaged the estimates for the For the scenarios where our algorithm is relatively less
relative phases, μ and ±, after first properly wrapping reliable, MUSIC is also. The results of Simulation 3
those values into the §¼ range. Then the averages and Simulation 4 suggest that, while being a powerful
for μ and ± were used to calculate AZ and EL. algorithm and useful for complex multi-signal
The same weighting was used for each snapshot of scenarios, MUSIC’s computational complexity is
that particular simulation iteration. The simulation unwarranted in the problem on which we are focused.
was carried out 1000 times, and the resulting DOA
estimation variances are plotted in Fig. 6. An IV. RADAR EXAMPLE
automatic minimization routine was used to obtain
the MUSIC estimates instead of a grid search so that For this example, a CW radar is simulated with
the grid step size would not affect the simulation’s a 10 GHz center frequency. The radar samples 256
statistics. Our algorithm outperforms MUSIC when times per CPI at a rate of 51.2 kHz, which results in a
a small number of snapshots containing the target CPI of 5 ms. CW was chosen for convenience, but all
signal are available. Furthermore, MUSIC cannot be of the below steps could be used with a pulse-Doppler
used with only a single snapshot. As expected, the system. Using standard equations from [18], we find
performance of the two algorithms begins to converge that the maximum unambiguous range-rate, or relative
at larger numbers of snapshots. velocity, is 384 m/s, and the range-rate resolution is
3 m/s.
D. Simulation 4 Reflections from two targets are simulated. The
first target is closing at rate of 121 m/s (or ¡121 m/s)
In the analysis of the preceding section, we from the direction of ¡51± AZ and 8± EL, and the
showed that the relative AZ and EL of the jammer second target is opening at a rate of 86 m/s (or
and target incident signals affects the DOA estimation +86 m/s) from the direction of 19± AZ and ¡30± EL.
performance of MUSIC in a similar way to how it White Gaussian noise (WGN), uncorrelated from
affects our algorithm. Although performance trends antenna to antenna, is added to the received signal
could be identified, a full analytical description such that the noise floor will correspond to 0 dB on
of MUSIC’s performance cannot be found [10]. the generated radar maps. With a JNR of 40 dB, the
As a result, it seems useful to further the MUSIC jamming signal is simulated at a location of 46± AZ
comparison with another simulation. The target is and 37± EL relative to the receiver. The resulting
kept at 7± AZ, ¡25± EL with an SNR of 20 dB. map from a single CPI is shown after Doppler
The jammer signal was simulated with a constant processing in Fig. 8, and we notice that without
15± EL, while its AZ was swept across space. further work, only the jamming signal can be seen.
Twenty snapshots of the jammer at 50 dB JNR were The target signals we are trying to detect and track are
Fig. 9. Target detections after jammer cancellation. Fig. 11. Beamformer output jr1,3 j=® as function of signal EL.
1060 IEEE TRANSACTIONS ON AEROSPACE AND ELECTRONIC SYSTEMS VOL. 46, NO. 3 JULY 2010
well even when the number of target-containing
snapshots available is small. This property makes it
attractive for use in post-Doppler processing where it
is common for a target signal to straddle only a few
range-Doppler bins. The DOAs of multiple targets
can be estimated from one CPI as long as those target
signals are resolvable in range or Doppler.
REFERENCES
1062 IEEE TRANSACTIONS ON AEROSPACE AND ELECTRONIC SYSTEMS VOL. 46, NO. 3 JULY 2010
Lance Schmieder received a B.S.E.E. from the University of Tennessee in
Knoxville in 2005 and was recognized as a Top Collegiate Scholar for the
College of Engineering. In 2008, he completed his M.S.E.E. at the University of
Texas at Dallas (UTD).
He was a Jonsson Fellowship recipient at UTD and a researcher in the
Wireless Communications Research Laboratory. Currently, he is an analyst with
the Mustang Technology Group in Allen, TX, where his primary duties involve
signal processing, system simulation, and algorithm development.
Mr. Schmieder is a member of the Tau Beta Pi and Eta Kappa Nu honor
societies.
Donald Mellon received an advanced degree is in solid state physics from Iowa
State University in 1970.
He is currently an independent contractor supporting radar development
programs at Mustang Technology Group in Allen, TX. Dr. Mellon has over
30 years experience as an analyst and systems engineer on advanced radar
and communication systems obtained primarily in Texas Instrument’s Defense
Electronics organization.