Professional Documents
Culture Documents
Digital Signal Processing: Shenghua Zhou, Jing Lu, Pramod K. Varshney, Jianlai Wang, Hui Ma, Hongwei Liu
Digital Signal Processing: Shenghua Zhou, Jing Lu, Pramod K. Varshney, Jianlai Wang, Hui Ma, Hongwei Liu
Digital Signal Processing: Shenghua Zhou, Jing Lu, Pramod K. Varshney, Jianlai Wang, Hui Ma, Hongwei Liu
a r t i c l e i n f o a b s t r a c t
Article history: Transmit waveform diversity brings some advantages to colocated Multiple-Input Multiple-Output
Available online 16 December 2019 (MIMO) radars. Current waveform design techniques do not sufficiently exploit the degrees of freedom
from the receive end. In this paper, we devise waveform optimization criteria with the receive end
Keywords:
beamforming operation. With mutual correlation sidelobes suppressed at the receive end, some degrees
Colocated MIMO radar
of freedom can be released for a lower global sidelobe level. For transmit beampatterns with unequal
Receive beamforming
MIMO radar gains, two Peak Sidelobe Level (PSL) measures are discussed, one for identical range compression PSL
Waveform design and the other for the same distance to the noise level. Numerical results are given for beampatterns with
Peak sidelobe level mono-peak, two peaks with equal gains and two peaks with unequal gains, all indicating that the criteria
result in better sidelobe suppression.
© 2019 Elsevier Inc. All rights reserved.
https://doi.org/10.1016/j.dsp.2019.102635
1051-2004/© 2019 Elsevier Inc. All rights reserved.
2 S. Zhou et al. / Digital Signal Processing 98 (2020) 102635
shapes [13]. Third, good waveform optimization depends not only returns, [Jk ]i , j = δi +k, j , δk, j is the Kronecker delta symbol defined
on optimization algorithm in use but also on how to construct the by
waveform design criterion. A fair amount of work exists on wave-
form design for colocated MIMO radars with different optimiza- 0 k= j
tion tools, such as the mutual information based method [14], the
δk, j = (3)
1 else
quadratically constrained quadratic programming (QCQP) method
[15], the limited memory Broyden Fletcher Goldfarb and Shan- ar (·) ∈ C N r ×1 denotes the received steering vector, (·)T denotes
non (L-BFGS) algorithm [10], the sequential optimization method the transpose operation, Z = [z1 , · · · , z N c ]T ∈ C N c × N r denotes the
[16], the double least− pth minimax method [17], the sequential channel interference component, and zk , k = 1, · · · , N c stands for
quadratic programming (SQP) based method and the alternating the noise term at the kth snapshot.
direction method of multipliers (ADMM) based method [18]. The As [J−k ]i , j = δi −k, j = δ j ,i −k = δ j +k,i , we have
optimization tool in use is an important part of a good perfor-
mance but here we focus on how to construct better waveform JkT = J−k . (4)
optimization criteria.
In general, colocated MIMO radar waveform optimization in- With a range compression weight S/ N c , the output can be ex-
volves two tasks, namely, transmit beampattern matching and pressed in a vector form as
sidelobe suppression. It makes sense to exploit more degrees of
freedom for optimization. Transmit waveform optimization should x = vec SH X / N c = β av ( f a ) + I ⊗ SH / N c z (5)
suppress sidelobe outputs of range compressors at the receive end,
wherein the receive beamforming operation can suppress mutual where (·)H denotes the conjugate transpose operation, vec(·) de-
correlation sidelobes between angular waveforms. Therefore, an at- notes the vectorization operation,
tenuation factor is developed and imposed over mutual correlation
sidelobes such that some degrees of freedom can be exploited. av ( f a ) = ar ( f a ) ⊗ Rs (k) at ( f a ) (6)
Meanwhile, instead of square shaped desired transmit beampat-
terns, three control points are used for each peak of transmit denotes the compound transmit-receive steering vector, z = vec(Z),
beampattern, with two for the first null points and one for the ⊗ denotes the Kronecker-product operation, I denotes the identity
peak position, transmit antennas with limited elements. In our matrix,
waveform design criteria, three groups of spatial directions are
defined, one for transmit beampattern matching, one for spatial Rs (k) = SH Jk S/ N c (7)
receive channels and one for possible spatial directions. The op- denotes the shifted waveform covariance matrix, and (·) H
denotes
eration can reduce the number of elements to suppress. If the the conjugate transpose operation.
receive array has a much larger aperture than the transmit array
From (4), it can be shown that
[19], only auto correlation sidelobes should be suppressed. We also
discuss two different definitions of the final sidelobe level in cases RH H T H
s (k) = S Jk S/ N c =S J−k S/ N c = Rs (−k) . (8)
where a desired directional transmit beampattern has two peaks of
unequal heights. Numerical results are given for transmit beampat- If channel noise interference is white or wideband in contrast
terns with monopeak, two uniform peaks and two unequal peaks, to receive-end bandwidth, different rows of Z are often assumed
all indicating that the attenuation factor has a significant positive to be independent and identically distributed (i.i.d.). We assume
effect on sidelobe suppression. that the covariance matrix is Rz = E(zk zkH ), k = 1, · · · , N c , where
E(·) denotes the expectation operation. Then
2. Colocated MIMO radar receive end processing
E zzH = RTz ⊗ I (9)
2.1. Receive beamforming algorithms for colocated MIMO radar
and the interference component has the covariance matrix [20]
Consider a colocated MIMO radar with an N t -element linear H
transmit array and N r -element linear receive array. The waveform H H H
RZ = E I ⊗ S / N c zz I ⊗ S /Nc = Rz ⊗ Rs (0)/ N c . (10)
transmitted by the ith antennas is denoted by si ∈ C N c ×1 , i =
1, · · · , N t and denoted S = [s1 , · · · , s N t ] ∈ C N c × N t , where N c de-
notes the number of codes within each waveform, and C de- 2.2. Colocated MIMO radar receive beamforming
notes the complex plane. After coherent synthesis in space, trans-
mit waveforms will form different waveform signatures in space, A beamforming operation coherently combines signal x for a
termed as angular waveforms here and defined by high signal-to-interference-noise ratio (SINR) output. In practical
applications, receive beamforming is performed by a bank of re-
s ( f a ) = Sat ( f a ) (1) ceive beamforming filters [20]. Each filter is steered to a spatial
for spatial angular sector θ , where f a = 0.5 sin θ for linear ar- direction and is hence called a spatial receive channel here. The
ray, termed angular frequency, is defined to facilitate subse- weight of the spatial receive channel associated with angular fre-
quent derivations, and at (·) ∈ C N t ×1 denotes the transmit steering quency f a0 is denoted by ws ( f a0 ) and the output is an estimate of
vector. We can control the characteristic of angular waveforms the complex amplitude of target returns, i.e.,
through optimizing transmit waveforms S.
Signals received by the receive array can be written as β̂ ( f a0 ) = wH H
s ( f a0 ) x, s.t., ws ( f a0 ) av ( f a0 ) = 1. (11)
In practice, one may use another range compression weight,
X = β Jk s ( f a ) aTr ( f a ) + Z = β Jk Sat ( f a ) aTr ( f a ) + Z (2)
such as SRs−1 / N c if Rs is invertible, and then the constraint will
where β denotes the complex amplitude of target return, the pre- have a different form.
fix matrix Jk ∈ B N c × N c is a shift matrix, B = {0, 1} used to repre- The weights in [21,22] all have a Kronecker product structure
sent the signal mutual delay between range compressor and target as
S. Zhou et al. / Digital Signal Processing 98 (2020) 102635 3
R− T
z ar ( f a0 ) at ( f a0 ) If the window, denoted by wr , is imposed on ar ( f a0 ), the attenua-
ws ( f a0 ) = ⊗ . (12)
aTr ( f a0 ) R− T ∗
z ar ( f a0 ) aH
t ( f a0 ) Rs (0) at ( f a0 )
tion factor becomes
MIMO beamforming can form its weight with online data ρr ( f a0 ; f a ) = ar ( f a0 − f a )H wr / N r (23)
(adaptive) or without (quiescent). As Rz is generally
unknown,
the MIMO Capon algorithm estimate it by R̂z = E XH X / N c . As and wr = 1 becomes a special case.
waveform optimization is generally too time-consuming to be per- From the application perspective, ρk ( f a0 , f a ), standing for side-
lobe level in spatial receive channels, should have a low level.
formed online, we use the MIMO least square (LS), which simply
However, conventional MIMO radar waveform optimization algo-
assumes R̂z = I, and the weight is
rithms typically suppress only ck ( f a0 , f a ) [10]. It can of course
1 reduce ρk ( f a0 , f a ), but it is different from ck ( f a0 , f a ). In our for-
ws ( f a0 ) = ar ( f a0 ) ⊗ at ( f a0 ) (13) mulation, we introduce an additional term ρr ( f a0 , f a ), which is
N r N t p ( f a0 )
generally less than one for f a0 = f a .1 Therefore, we will suppress
where we used the fact N r = aH
r ( f a0 )ar ( f a0 ), and sidelobes represented directly by ρk ( f a0 , f a ), which equivalently
imposes a weight over conventional sidelobe measures ck ( f a0 , f a ).
p ( f a0 ) = aH
t ( f a0 )Rs (0)at ( f a0 )/ N t (14)
denotes the transmit beampattern, a measure of spatial distribu- 2.3. Quantization of angular frequency
tion of transmitted signal power. If Rs (0) = I, then p ( f a ) = 1 for
f a ∈ [−0.5, 0.5], corresponding to the uniform transmit beampat- The angular waveform s ( f a ) varies continuously along with the
tern. The transmit beampattern here means transmit gain in con- angular frequency f a = 0.5 sin θ , but we do not have optimization
trast to the uniform transmit beampattern. tools available that are efficient for optimization with continu-
ous f a . Therefore, we have to quantize the angular frequency by
In spatial receive channel identified by f a0 , the output of the
putting them in a number of bins and representing each bin by
target return coming from spatial direction f a can be expressed by
a discrete angular frequency. If too many angular frequencies are
used, the optimization time cost may be huge, whereas if the num-
ρk ( f a0 ; f a ) = wHs ( f a0 ) av ( f a ) (15)
ber of selected angular frequencies is insufficient, sidelobes over
where the common amplitude term β is ignored since it does not other angular frequencies may increase greatly. In this case, we
affect the sidelobe level. Hereafter ρk ( f a0 ; f a ) are termed as side- define three groups of angular frequencies in this paper.
lobe outputs of the spatial receive channels associated with f a0 The first group of angular frequencies is defined for transmit
and input angular waveforms from f a . beampattern matching, denoted by f0 ∈ C N 0 ×1 , where N 0 denotes
For the beamforming weight of form (12), we have the number of angular frequencies used for transmit beampattern
matching. One can control the accuracy of beampattern matching
ρk ( f a0 ; f a ) = ρr ( f a0 ; f a ) rk ( f a0 ; f a ) (16) by adjusting the number N 0 . There are many measures of transmit
beampattern matching performance [6], especially for omnidirec-
where tional transmit beampatterns. In radar applications, we are often
aH −T interested in transmit beampattern with one or more peaks. For
r ( f a0 ) R̂z ar ( f a )
ρr ( f a0 ; f a ) = (17) such desired beampatterns, we can use three angular frequencies
aTr ( f a0 ) R̂− T ∗
z ar ( f a0 ) to localize the position of a narrow transmit beam, one for the
ck ( f a0 ; f a ) peak and two for the first nulls. For wider beams, we can use more
rk ( f a0 ; f a ) = (18)
p ( f a0 ) angular frequencies for each peak.
The second group of angular frequencies, denoted by fa ∈
stands for the receive beamforming attenuation factor, and
C N a ×1 , contains the angular frequencies to which spatial receive
channels are steered, where N a denotes the number of spatial re-
ck ( f a0 ; f a ) = aH
t ( f a0 ) Rs (k) at ( f a )/ N t . (19)
ceive channels deployed at the receive end. Each filter of the bank
If f a0 = f a , ρk , rk and ck turn out to be auto correlation side- aims at a specific angular frequency (spatial direction). Each spa-
lobes or range sidelobes. From (8), we have tial receive channel deals with returns from a space volume, such
that target returns within the volume can pass the filter, whereas
ckH ( f a0 ; f a ) = c −k ( f a ; f a0 ) (20) target returns beyond will be attenuated significantly. In practice,
the number of spatial receive channels to deploy depends on many
so both auto and cross correlation sidelobes are conjugate sym- factors, such as the maximum SINR loss acceptable.
metric. It is expected that target returns from all spatial directions will
In general, ρr < 1, indicating that target returns from directions have low sidelobe outputs in the spatial receive channels. The third
that are not of interest will be attenuated. The MIMO LS algorithm group of angular frequencies, denoted by fm ∈ C N m ×1 , includes
will output sidelobes all possible angular waveforms from different spatial directions,
where N m denotes the number of real target returns to simulate.
1
ρk ( f a0 ; f a ) = [ar ( f a0 ) ⊗ at ( f a0 )]H av ( f a ) (21) In general, angular frequencies in fm have a smaller spacing than
N r B ( f a0 ) those in fa and then N m ≥ N a .
and the attenuation factor for the MIMO LS algorithm can be ex-
pressed by 2.4. Peak sidelobe level (PSL) measures
which, however, also has various definitions, different in the side- and RZ = Rz ⊗ Rs (0)/ N c , the noise covariance level output of the
lobes involved and the weights to combine them. spatial receive channel using weight (12) can be formulated as
We first define the PSL of transmit waveforms according to (7),
by σ 2 ( f a0 ) = E(wHs ( f a0 ) I ⊗ SH / N c zzH I ⊗ SH / N c ws ( f a0 ))
Fig. 1. Comparison of two PSL setting strategies. The PSL measures above were defined with continuous f a for
optimization, but no efficient optimization tool is available to solve
the channel noise level. It should be noted that we do not prefer this problem. Therefore, we formulate the waveform design criteria
one strategy to the other and will present results obtained under with f0 , fa and fm , all in the same form as
both strategies.
min max bd − diag(RA (0)), ea , em (39)
2.6. Transmit beampattern matching
where denotes the phase matrix of S, i.e., S = exp( j ), j de-
notes the imaginary part, ea denotes auto and mutual correlation
With angular frequencies in f0 , we construct a matrix
sidelobes with fa , em denotes mutual correlation sidelobes be-
RA (0) = AH tween fa and fm , and · denotes a kind of complex matrix norm.
t (f0 ) Rs (0) At (f0 ) (34)
Intended transmit waveforms are constant-modulus and are
where constructed through S = exp( j ). The sidelobes and beampattern
mismatches can be combined in various ways through the norm
At (f0 ) = [at ( f 0 (1)), · · · , at ( f 0 ( N 0 ))] (35) operation. If · stands for the 2−norm, ISL will be suppressed,
whereas we focus on the PSL, for which · stands for the infin-
is a matrix of transmitted steering vectors at angular frequencies
ity norm. One may also build other generalized norms to combine
in f0 . The diagonal elements of RA (0) stand for the responses of
those terms, which will not be discussed.
the transmit beampattern at f0 . Nondiagonal elements stand for
In order to suppress either PSLAG and PSLRG (no receive beam-
their mutual correlation without mutual shift.
forming attenuation) or PSLDG and PSLBG (with receive beamform-
With a desired transmit beampattern denoted by bd ∈ R N a ×1 ,
ing attenuation), we need to construct ea and em with fa and fm .
the transmit beampattern mismatch can be expressed by
Without loss of generality, we assume that fa and fm have no
elements in common. Filling in different ea and em , we obtain dif-
bd − diag(RA (0)) (36)
ferent waveform design criteria.
where diag(·) with a square matrix entry denotes a vector of its If PSLBG is intended to be suppressed, we have
diagonal elements.
ea = vec [Ra (k) (diag(1/p(fa ))Wa )]k∈ka
A desired transmit beampattern describes the shape of the (40)
transmit beampattern but its power may have a mismatch with em = vec [Rm (k) (diag(1/p(fa ))Wm )]k∈km
real signal power. One may introduce an optimization parameter
where auto and mutual correlation sidelobes are
to remove the ambiguity, but as transmit power satisfies the power
conservation law, we can build a constraint on bd by Ra (k) = AH
t (fa ) Rs (k) At (fa )
(41)
H
Rm (k) = AH
t (fa ) Rs (k) At (fm )
1= p ( f a )d f a = a ( f a )Rs (0)a( f a )d f a
diag(·) with a vector entry denotes a diagonal matrix with the vec-
1
N0 tor entry in the diagonal, vec(·) denotes the vectorization operator,
1
≈ aH (f0 (i ))Rs (0)a(f0 (i )) = tr(RA (0)) (37) p(fa ) denotes the transmit beampattern gains at fa , and
N0 N0
i =1
[Wa ]i ,n = ρr (fa (i ); fa (n)) = ρr (fa (i ) − fa (n), 0)
1 (42)
≈ bd
[Wm ]i ,n = ρr (fa (i ); fm (n)) = ρr (fa (i ) − fm (n), 0)
N0
Consequently, we eliminate the power ambiguity of the desired are matrices of receive beamforming attenuation factors.
transmit beampattern by normalizing bd so that If PSLDG are intended to be suppressed, we have
bd = N 0 . (38) ea = vec (Ra (k) Wr )k∈ka
(43)
em = vec (Rm (k) Wr )k∈km .
Normalization can avoid optimizing an additional power fitting
parameter and thus may lead to a better optimization result. For comparison purposes, we still consider two waveform op-
Nondiagonal elements of RA (0) stand for mutual correlation timization criteria without receive beamforming. In order to sup-
sidelobes between angular waveforms at k = 0. As sidelobes, they press PSLRG , we use
are also undesirable and should be suppressed. In [10], sidelobes
at k = 0 are taken into account in the waveform optimization cri- ea = vec [diag(1/p(fa ))Ra (k)]k∈ka
(44)
terion. If we are interested in PSL measures, sidelobes of angular em = vec [diag(1/p(fa ))Rm (k)]k∈km
6 S. Zhou et al. / Digital Signal Processing 98 (2020) 102635
and to order to suppress PSLAG , we use |sH (i )at ( f a0 )/ N t |2 ≤ s(i )2 · a( f a0 )/ N t 2 = N t (47)
ea = vec [Ra (k)]k∈ka where the equation holds if and only if s(i ) = exp( j ϕi )at ( f a0 ), the
(45) transmit beampattern
em = vec [Rm (k)]k∈km .
Fig. 2. Colocated MIMO radar with mono peak. (a) Desired and resulting transmit beampattern; (b) Angular waveforms at f a = 0 and f a = 0.15; (c) Range sidelobes of angular
waveforms at f a = 0; (d) Range sidelobes of angular waveforms from different directions; (e) Waveform signatures of phased-array radar in different spatial directions; (f)
Auto and mutual correlation sidelobes of the spatial receive channel at f a = 0 without Doppler concern.
Strictly speaking, a phased-array radar also transmits differ- Therefore, the performance loss caused by ignoring the difference
ent waveform signatures into different directions, but the differ- is trivial
ence is insignificant and can be omitted without performance loss Sidelobe outputs of all angular waveforms in the spatial receive
[26]. We go a step further to show the waveform signatures of a channel for f a = 0 Hz, corresponding to ck (0, f a ), fa ∈ [−0.5, 0.5],
phased-array radar transmitted into different directions in Fig. 2(e), are shown in Fig. 2(f). Although the angular waveforms in differ-
for f a = 0 Hz and f a = 0.2 Hz. From Fig. 2(e), there are blurs ent spatial directions have different signatures and range sidelobes,
near the code transition points. This is just a theoretical result their sidelobe outputs in the spatial receive channel have a low
and those blurs stand for high frequency components, which will level. This is because other spatial directions have a low transmit
be smoothed after passing band-pass circuits at the receive end. power gain that results in low sidelobes.
8 S. Zhou et al. / Digital Signal Processing 98 (2020) 102635
A heavier weight on the transmit beampattern matching term Results are shown in Fig. 3(e), wherein the transmit beampat-
can make the transmit power focused mainly into the peak di- tern is also presented. The sidelobe level depends heavily on the
rection. In this case, the rank of S is close to one and angular weight on the transmit beampattern matching. It can be seen that
waveforms transmitted into different spatial directions have sig- the fine spacing has sufficiently represented angular waveforms
natures close to each other. If the rank is one, all spatial directions from all spatial directions.
will have the same signature and the same range sidelobe level. Consider a receive array also with 10 half-wavelength spaced
Meanwhile, with close signatures, they have low mutual corre- elements and then calculate sidelobes after the receive beamform-
lation sidelobes at k = 0. Consequently, once the angular wave- ing operation with the MIMO LS weight as ar ( f a0 )/ N r , yielding
form at f a = 0 Hz has a low sidelobe level, other angular wave-
forms will inherent its low range sidelobes. This is the reason ρ ( f a0 , f a ) = max ρk ( f a0 , f a ). (50)
why we set fm = 0 but still obtain low mutual correlation side- k∈km
lobes.
The attenuation factor is calculated according to (17). The result is
In particular, if the weight imposed on transmit beampattern
shown in Fig. 3(f). It can be seen that the cross correlation side-
matching tapers off, the transmit gain toward the direction will
lobes are efficiently suppressed after receive beamforming. It is
decrease but the sidelobe level will improve significantly. For in-
obvious that the sidelobe level over the angular frequency dimen-
stance, if we impose a weight 0.5, we can obtain a PSLDG =
sion is nonuniform. In addition to receive beamforming, sidelobes
−36.24 dB. As an extreme, if we totally remove the transmit beam- can reach a lower level. From Figs. 3(e-f), the sidelobe level is
pattern matching pursuit, we can obtain an angular waveform with mainly determined by auto correlation sidelobes, instead of mutual
a range sidelobe level lower than −120 dB, which has been intro- correlation sidelobes. Consequently, using the same concept as in
duced in [26]. [27], we ignore cross correlation sidelobes of angular waveforms
but focus merely on suppressing sidelobes around two mainlobes,
3.3. Transmit beampattern with two uniform peaks through setting fm = [−0.35, −0.25, 0.25, 0.35]T .
With a lighter weight 0.02 on the transmit beampattern match-
The receive beamforming effect will be considered subse- ing term, we run the optimization process with (43) and obtain
quently. Consider a desired transmit beampattern with two peaks Fig. 4(a) that shows sidelobe outputs of two spatial receive chan-
at fa = {−0.3, 0.3} Hz with the same gain. Other spatial regions nels stacked together, where the receive beamforming effect is not
are considered as sidelobes. In this case, PSL measures, both chan- taken into account. The optimized PSL is −25.4 dB, with a 1.2 dB
nel PSL (PSLAC , PSLRC and PSLBC ) and global channel PSLs (PSLAG , decrease in contrast to waveform design without receive beam-
PSLRG and PSLBC ), are close to each other. Therefore, we run the formng. Fig. 4(b) shows PSL over range for sidelobes in two receive
optimization process with (44) that has a low computational cost. channels, without the receive beamforming effect. Fig. 4(b) clearly
Here a desired transmit beampattern is described by a gain and indicates that range sidelobes are still mainly determined by auto
a beam width. For the N t elements array, the first null width is correlation sidelobes, though cross correlation sidelobes are not
1/ N t . Therefore, we set f0 = [−0.4, −0.3, −0.2, 0.2, 0.3, 0.4]T Hz suppressed in the optimization criterion. Therefore, ignoring mu-
and bd = [0, 5, 0, 0, 5, 0]T . For the two peaks, the receive end de- tual correlation sidelobes does not lead to high-level mutual corre-
ploys two range compressors regarding fa = [−0.3, 0.3]T Hz and lation sidelobes and the setting is justified. Meanwhile, real range
real target returns are simulated by a 20-elements fm ranging from PSLs have two shallow notches around −0.3 Hz and 0.3 Hz, indi-
−0.475 Hz to 0.475 Hz, with spacing 0.05 Hz. cating that although quantization of the angular frequency through
With (44), we obtain a transmit beampattern shown in Fig. 3(a). a few angular waveforms will result in the sidelobe increase for
The resulting beampattern matches the desirable one well at the real target returns from all possible spatial directions, the degree
anchor points, i.e., f0 . The transmit power has been efficiently al- of increase can be controlled to an acceptable level by proper pa-
located into expected directions, resulting in low sidelobes of the rameter settings.
transmit beampattern. Waveform signatures in f0 are shown to- Attenuated by the MIMO LS algorithm, sidelobe outputs of more
gether with their sum in Fig. 3(b). It can be seen that their sum in angular waveforms in the two channels are stacked and shown in
the temporal dimension is close to a constant. That means that the Fig. 4(c), indicating that receive beamforming has made the mu-
transmit power switches between two directions, occupying nearly tual correlation sidelobes much lower than the auto correlation
all the time slots. sidelobes. The value of PSL over range can be found in Fig. 4(d).
Assume that two spatial receive channels are deployed to deal As receive beamforming does suppress auto correlation sidelobes,
with returns from two different spatial directions fa . Range com- the final PSL is identical to that before receive beamforming. In
pression sidelobes of angular waveforms at fm in two channels Fig. 4(b), mutual correlation between angular waveforms around
are stacked and shown together in Fig. 3(c). Two channel PSLs are two transmit beampattern peaks are comparable to the auto corre-
around −24.10 dB (three channel PSLs are close to each other). lation sidelobes, but in Fig. 4(d), it can be seen that mutual corre-
In practice, real target returns may come from any direction and lation sidelobes between returns in directions associated with the
Fig. 3(d) shows sidelobe outputs of angular waveforms from more two peaks have been efficiently suppressed, to an extent compara-
spatial directions, with a finer spacing of 0.01 Hz. It can be seen ble to mutual correlation sidelobes, i.e., around −45 dB. Therefore,
that sidelobes around fm are suppressed and the resulting PSLRC = the only obstacle for a lower global sidelobe level is auto correla-
−23.92 dB, with a 0.18 dB increase. Moreover, the final sidelobe tion sidelobes.
level is mainly determined by auto correlation sidelobes, which are Table 1 lists all sidelobe measures of designed waveforms in
much higher than cross correlation sidelobes. this section.
Sidelobe level in Fig. 3(d) is a bit higher than Fig. 3(c) of side-
lobes at optimization ticks. To show this fact more clearly, we 3.4. Transmit beampattern with two unequal peaks: Case I
calculate the PSL over the range for the sidelobes in Figs. 3(c-d),
excluding k = 0 parts, i.e., For directional transmit beampattern with identical peaks,
range sidelobe measures are close to each other. Noise levels in
two channels are close to each other as well. In the following, we
r ( f a0 , f a ) = max rk ( f a0 , f a ). (49) still consider desired transmit beampattern with two peaks but
k∈km
S. Zhou et al. / Digital Signal Processing 98 (2020) 102635 9
Fig. 3. Transmit beampattern with two uniform peaks. (a) Desired and resulting transmit beampattern; (b) Angular waveforms at f a = −0.3 and f a = 0.3; (c) Range sidelobes
at angular frequencies for optimization; (d) Range sidelobes at the angular frequencies; (e) PSL over the range without receive beamforming; (f) PSL over range after receive
beamforming.
allocate different powers to the two peaks. In this case, noise lev- bd = [0, 75, 0, 0, 25, 0]T . Real target returns are from −0.475 Hz to
els in two spatial receive channels become different and then one 0.475 Hz with spacing 0.05 Hz by assumption.
needs to choose from those PSL definitions. The resulting transmit beampattern is shown in Fig. 5(a), in-
Assume that the two peaks of the desired transmit beam- dicating that the resulting transmit beampattern agrees well with
pattern are located at fa = [−0.3, 0.3]T Hz, with a power ratio the desired one. The resulting power ratio of two peaks is 3.3 : 1,
3 : 1. We first suppress PSLBG and run the optimization process a bit mismatch with the desirable one. As most of the power has
with (40), still with f0 = [−0.4, −0.3, −0.2, 0.2, 0.3, 0.4]T Hz and been concentrated on two peaks, signal power in the sidelobe re-
10 S. Zhou et al. / Digital Signal Processing 98 (2020) 102635
Fig. 4. Transmit beampattern with two uniform peaks. (a) Optimized sidelobes in two spatial receive channels; (b) PSL over range for two channels without receive beam-
forming; (c) Sidelobe outputs in two channels with receive beamforming; (d) PSL over range for two channels with receive beamforming.
Table 1
PSL measures of optimization results.
gions is controlled to a low level. Therefore, our mechanism to
p ( f a0 )
form multiple peaks of transmit beampattern works well. ck ( f a0 , f a )/ p ( f a0 ) p ( f a ) = rk ( f a0 , f a ) (51)
p( f a )
For MIMO radar waveform optimization, the way to normal-
ize sidelobes ck ( f a0 ; f a ) is critical. Here PSLRC ( f a ) is normalized by with k ∈ km , f a0 ∈ f0 , and f a ∈ fm . It stems from a strict mathe-
p ( f a0 ) and PSLAC ( f a ) by p m . In practice, there is another common matical definition of waveform mutual correlation and will never
but unsuitable method to normalize sidelobes, given by exceed one.
S. Zhou et al. / Digital Signal Processing 98 (2020) 102635
Fig. 5. Transmit beampattern with two uniform peaks. (a) Desirable and resulting transmit beampatterns; (b) Mathematical definition of mutual correlation sidelobes ck ( f a0 , f a )/ p ( f a0 ) p ( f a ) for f a0 = 0 and f a0 = 0.15 stacked
together; (c) Range compression sidelobes rk ( f a0 , f a ) for f a0 = 0 and f a0 = 0.15 stacked together; (d) Angular waveform sidelobes ck ( f a0 , f a ), without receive beamforming; (e) Peak value over range excluding k = 0 for range
compression sidelobes, without receive beamforming; (f) Peak value over range excluding k = 0 for angular waveform sidelobes, without receive beamforming; (g) Range compression sidelobes with receive beamforming; (h)
Angular waveform sidelobes with receive beamforming; (i) Peak value over range excluding k = 0 for range compression sidelobes, with receive beamforming; (j) Peak value over range excluding k = 0 for angular waveform
sidelobes, with receive beamforming.
11
12 S. Zhou et al. / Digital Signal Processing 98 (2020) 102635
Fig. 6. Transmit beampattern with two uniform peaks. (a) Sidelobes of angular waveforms in two channels at optimization ticks; (b) Sidelobes of angular waveforms at more
angular frequencies; (c) Peak sidelobes over range in two channels; (d) Peak sidelobes over range for waveforms optimized with a Hamming window imposed on receive
beamforming.
For the waveforms at hand, the sidelobes in definition (51) are ures differ in the sidelobe outputs in the spatial receive channel re-
shown in Fig. 5(b), indicating that sidelobes are very high under garding f a = 0.3 Hz. The power ratio of the two peaks 3.3 : 1 here
this measure. This sidelobe definition overweights mutual correla- results in a 10.4 dB difference in the channel PSL at f a = 0.3 Hz.
tion sidelobes because transmit gains at fa , associated with spatial Auto correlation sidelobes in both figures are lower than mutual
receive channels, are always higher than those at fm that stands correlation sidelobes, because the waveform optimization criterion
for other spatial directions. In practice, targets from spatial regions has taken the receive beamforming effect into account but the op-
that are not of interest will also have low power returns in a sta- eration is not considered in Figs. 5(c-f).
tistical sense. That is the reason why we pay less attention on this In addition to receive beamforming with the 10-element re-
sidelobe definition. ceive array used in waveform optimization, sidelobes measured
Next, range compression sidelobes rk ( f a0 , f a ) are shown in
by ρk ( f a0 ; f a ) are shown in Fig. 5(g), and sidelobes measured by
Fig. 5(c) and angular waveform sidelobes ck ( f a0 , f a ) in Fig. 5(d),
ρr ( f a0 ; f a ) · ck ( f a0 ; f a ) in Fig. 5(h), where no window is imposed
both for k ∈ km , f a0 ∈ f0 , f a ∈ fm . For the spatial receive chan-
at receive beamforming.
nels regarding the lower peak, sidelobe level are different from
In Figs. 5(g-h), sidelobes at different angular frequencies are
Fig. 5(c) and Fig. 5(d), because two sidelobe measures impose
flatter than those in Figs. 5(c-d), as a result of waveform optimiza-
different weights over sidelobes outputs in the 0.3 Hz channel.
tion. The channel PSL is about −38 dB and the receive beamform-
From Figs. 5(b-d), it can be clearly seen that under different side-
ing results in a 12.7 dB drop of the PSL. Therefore, a well-designed
lobe measures, mutual correlation sidelobes are different. There-
fore, it is important to select reasonable sidelobe measures and it waveform optimization criterion can also result in a good wave-
is also the reason why we pay so much attention on PSL defini- form optimization result.
tions. Peak sidelobes over range excluding k = 0 for sidelobes in
Numerical values of sidelobe levels are shown in Figs. 5(e-f), Figs. 5(g-h) are shown in Figs. 5(i-j), respectively. The waveform
where Fig. 5(e) presents peak sidelobes over range excluding k = 0 optimization criterion with (40), tending to suppress PSLBG , gives
for range compression sidelobes in Fig. 5(c), and Fig. 5(f) presents rise to close sidelobe outputs in Fig. 5(i). If the sidelobes are nor-
peak sidelobes over range excluding k = 0 for angular waveform malized by peak gain as in Fig. 5(j), sidelobes in two different
sidelobes in Fig. 5(d). The resulting transmit beampattern is also spatial receive channels are not equivalent anymore. Specifically, at
presented for comparison purpose. It is obvious that the two fig- optimization points, the PSL in the 0.3 Hz channel is −47.59 dB,
S. Zhou et al. / Digital Signal Processing 98 (2020) 102635 13
about 10 dB lower than that in the −0.3 Hz channel. That is a attenuation factor results in a significant PSL drop to the optimiza-
significant difference. tion result. The Doppler problem is not considered but it is not
Waveform optimization with (40) tends to make different chan- difficult to generalize the criteria in case of Doppler mismatches.
nels yield the same channel PSL. The following part will discuss The SQP algorithm can be replaced by the L-BFGS algorithm and
waveform optimization results with (43), which tends to force PSLs the ADMM algorithm for a better performance output.
of different channels to have the same distance to the noise level.
Declaration of competing interest
3.5. For transmit beampattern with two unequal peaks: Case II
The authors declare that they have no known competing finan-
On the same transmit/array, we optimize with (43), under the cial interests or personal relationships that could have appeared to
same parameter settings as those used to obtain Fig. 5. The trans- influence the work reported in this paper.
mit beampattern is close to Fig. 5(a) and thus is omitted here.
According to Table 1, PSLDG is −42.23 dB, 10.7 dB lower than
References
PSLBG = −31.56 dB. Compared to Fig. 5, PSLBG rises by 3.1 dB but
PSLDG drops by 4.8 dB. Fig. 6(a) presents angular waveform side- [1] M. Friese, Polyphase barker sequences up to length 36, IEEE Trans. Inf. Theory
lobes attenuated by receive beamforming at angular frequencies 42 (4) (1996) 1248–1250.
fm . [2] N. Levanon, E. Mozeson, Radar Signals, J. Wiley, Hoboken, NJ, 2004.
From Table 1, sidelobes beyond selected angular frequencies [3] E. Fishler, A. Haimovich, R. Blum, D. Chizhik, L. Cimini, R. Valenzuela, Mimo
radar: an idea whose time has come, in: A. Haimovich (Ed.), Radar Conference,
may rise a bit. In Fig. 5, real PSLBG is 2.4 dB higher than the opti-
Proceedings of the IEEE, IEEE, 2004, pp. 71–78.
mization outcome. Fig. 6(b) shows sidelobes in two channels with [4] J. Li, P. Stoica, MIMO radar with colocated antennas, IEEE ASSP Mag. 24 (5)
a much finer sampling frequency of the angular frequency. Explicit (2007) 106–114.
values are shown in Fig. 6(c) for peak sidelobes over range exclud- [5] H. Deng, Polyphase code design for orthogonal netted radar systems, IEEE
ing k = 0. In Fig. 6(c), two PSL measures are equivalent for the Trans. Signal Process. 52 (11) (2004) 3126–3135.
−0.3 Hz channel, but are different for the 0.3 Hz channel. There- [6] P. Stoica, J. Li, Y. Xie, On probing signal design for MIMO radar, IEEE Trans.
Signal Process. 55 (8) (2007) 4151–4161.
fore, we plot them together in Fig. 6(c). [7] Y. Yang, R.S. Blum, Mimo radar waveform design based on mutual information
As mentioned above, PSLBG and PSLDG have different outcomes. and minimum mean-square error estimation, IEEE Trans. Aerosp. Electron. Syst.
If sufficiently suppressed, the former will force all spatial receive 43 (1) (2007) 330–343.
channels associated with transmit beampattern peaks to have the [8] A.M. Haimovich, R.S. Blum, L.J. Cimini, MIMO radar with widely separated an-
tennas, IEEE ASSP Mag. 25 (1) (2008) 116–129.
same channel PSL, but the latter will make channel PSLs have the
[9] E. Fishler, A.M. Haimovich, R.S. Blum, L.J. Cimini Jr., D. Chizhik, R.A. Valenzuela,
same distance to the noise distance. The former may have two Spatial diversity in radars – models and detection performance, IEEE Trans.
channels with one sidelobe level above the noise level but the Signal Process. 54 (3) (2006) 823–838.
other below the noise level. Therefore, the latter seems better, but [10] Y.C. Wang, X. Wang, H. Liu, Z.Q. Luo, On the design of constant modulus
the choice should be left for users in accordance to actual need in probing signals for mimo radar, IEEE Trans. Signal Process. 60 (8) (2012)
4432–4439.
the field. Moreover, users may suggest other waveform criteria and
[11] L. Xu, S. Zhou, H. Liu, J. Liu, Repeat radar jammer suppression for a colocated
this paper point out that one should note the rule of noise level in mimo radar, IET Radar Sonar Navig. 13 (9) (2019) 1448–1457, https://doi.org/
those criteria. 10.1049/iet-rsn.2018.5570.
In the receive beamforming process, we may impose a win- [12] H.A. Khan, Y. Zhang, C. Ji, C.J. Stevens, D.J. Edwards, D. O’Brien, Optimizing
dow to release more degrees of freedom for a lower global side- polyphase sequences for orthogonal netted radar, IEEE Signal Process. Lett.
13 (10) (2006) 589–592.
lobe level. When the Hamming window is imposed, we run the
[13] S. Zhou, H. Liu, X. Wang, Y. Cao, Mimo radar range-angular-Doppler sidelobe
optimization processing with (43) and then obtain a PSLDG = suppression using random space-time coding, IEEE Trans. Aerosp. Electron.
−42.49 dB, with a 0.26 dB drop from the optimization result with- Syst. 50 (3) (2014) 2047–2060.
out a window. At optimized angular frequencies, the PSL drops [14] M.R. Bell, Information theory and radar waveform design, IEEE Trans. Inf. The-
by 0.92 dB to −45.19 dB. Peak sidelobes over range are shown ory 39 (5) (1993) 1578–1597.
[15] O. Aldayel, V. Monga, M. Rangaswamy, Successive qcqp refinement for mimo
in Fig. 6(d). It can be seen that the imposed window helps to
radar waveform design under practical constraints, IEEE Trans. Signal Process.
suppress the sidelobe level further, but the PSL drop is not so 64 (14) (2016) 3760–3774.
significant, mainly because from Figs. 6(c-d), the sidelobe level is [16] G. Cui, H. Li, M. Rangaswamy, Mimo radar waveform design with constant
basically determined by auto correlation sidelobes. Therefore, the modulus and similarity constraints, IEEE Trans. Signal Process. 62 (2) (2014)
degrees of freedom that can be extracted from mutual correlation 343–353.
[17] L. Xu, H. Liu, S. Zhou, H. Ma, J. Zheng, J. Wang, Joint optimization of transmit
is limited in this case even for transmit/receive antenna arrays.
waveform and mismatched filter with expanded mainlobe for delay-Doppler
sidelobes suppression, in: 2018 International Conference on Radar (RADAR),
4. Conclusion and discussion 2018, pp. 1–6.
[18] J. Wang, Y. Wang, On the design of constant modulus probing waveforms with
This paper aimed at transmit waveform optimization for colo- good correlation properties for MIMO radar via consensus-ADMM approach,
IEEE Trans. Signal Process. 67 (16) (2019) 4317–4332, https://doi.org/10.1109/
cated MIMO radars in conjunction with receive array processing. TSP.2019.2928994.
Waveform design criteria were developed with implementation [19] H. Liu, S. Zhou, Y. Yu, H. Su, Detection performance of spatial-frequency diver-
concerns. An attenuation factor was introduced and incorporated sity mimo radar, IEEE Trans. Aerosp. Electron. Syst. 52 (5) (2014) 2091–2110.
into several waveform optimization criteria. We also defined three [20] S. Zhou, H. Liu, B. Liu, K. Yin, Adaptive mimo radar target parameter estima-
tion with Kronecker-product structured interference covariance matrix, Signal
groups of angular frequencies to develop the waveform design cri-
Process. 92 (5) (2012) 1177–1188.
teria, which also contributes to lower PSL outcomes. We defined [21] L. Xu, J. Li, P. Stoica, Target detection and parameter estimation for MIMO radar
several sidelobes measures for colocated MIMO radars and pointed systems, IEEE Trans. Aerosp. Electron. Syst. 44 (3) (2008) 927–939.
out that one should notice the distance from the sidelobe level to [22] D.Z. Feng, X.M. Li, H. Lv, H.W. Liu, Z. Bao, Two-sided minimum-variance dis-
the noise level when desired transmit beampattern have multiple tortionless response beamformer for mimo radar, Signal Process. 89 (3) (2009)
328–332.
peaks of different amplitudes. Two different PSL setting strategies
[23] J. Nocedal, S. Wright, Numerical Optimization, Springer-Verlag, NY, 2006.
were discussed, namely PSLBG and PSLDG . Several waveform opti- [24] J. Xu, J. Yu, Y. Peng, X. Xia, Radon-Fourier transform for radar target detection,
mization criteria were presented and numerical results were given I: generalized Doppler filter bank, IEEE Trans. Aerosp. Electron. Syst. 47 (2)
for different scenarios, indicating that the receive beamforming (2011) 1186–1202.
14 S. Zhou et al. / Digital Signal Processing 98 (2020) 102635
[25] J. Xu, J. Yu, Y. Peng, X. Xia, Radon-Fourier transform for radar target detection theory, wireless communications, image processing, radar signal process-
ii: blind speed sidelobe suppression, IEEE Trans. Aerosp. Electron. Syst. 47 (4) ing, and remote sensing. He was elected to the grade of fellow of the IEEE
(2011) 2473–2489. in 1997 for his contributions in the area of distributed detection and data
[26] S. Zhou, H. Liu, H. Zang, H. Su, Doppler sensitivity of MIMO radar waveforms, fusion. He was the President of the International Society of Information
IEEE Trans. Aerosp. Electron. Syst. 52 (5) (2016) 2091–2110.
Fusion in 2001. He is on the Editorial Board of the Journal on Advances in
[27] H. Liu, S. Zhou, H. Zang, Y. Cao, Two waveform design criteria for colocated
Information Fusion.
mimo radar, in: IEEE International Radar Conference, 2014, pp. 1–5.