Cognitive MIMO Radar

Joseph Tabrikian
Signal Processing Laboratory
Department of Electrical and Computer Engineering
Ben-Gurion University of the Negev

Involved collaborators and Research Assistants:

Prof. R. Shavit, Prof. H. Messer, Dr. I. Bilik, I. Bekkerman, W.
Huleihel, M. Teitel, N. Sharaga, O. Isaacs

BGU Radar Symposium 2016

Outline

 MIMO radar at a glance

 Cognitive radar - introduction
 Cognitive MIMO radar for beamforming and detection
 Conclusion
MIMO Radar at a Glance
Data model: X  H(Θ)S 
 X - Rx signal
W at sensor m and time index
n
S i,nm,n - Tx signal by element i and time index
n

H (Θ) - Tx-Rx channel matrix
Θ - Unknown targets'
parameters
1 H Tx elements
RS SS
 N
MIMO - Rs  cI
Phased array - rank(Rs )
=1
Rx elements
MIMO Radar at a Glance
Orthogonal Tx signals can be decomposed at the
receiver, allowing adaptive beamforming of the Tx
signals → Virtual receiving elements:
Virtual
Tx elements Rx elements

Single
Tx element

Rx element Rx element
MIMO Radar at a Glance
Virtual receiving elements: Orthogonal Tx signals can
be decomposed at the receiver, allowing adaptive
beamforming of the Tx signals.
Virtual
Tx elements Rx elements

Single
Tx element

Rx elements Rx elements

Colocated (mono-static) MIMO radar: Bekkerman-Tabrikian 2004

Distributed (multi-static) MIMO radar: Fishler et al. 2004
MIMO Radar Properties
Array aperture extension:
Receive elements
Transmit elements Virtual elements

Receive elements

Transmit elements

Virtual elements
MIMO Radar Properties
Array aperture extension:
Transmit/Receive elements

Virtual elements
MIMO Radar Advantages
 More degrees of freedom due to the virtual sensors:
 Higher angular resolution.
 Higher number of targets/clutter in a given range-Doppler
cell, which can be detected and localized.
 Lower sidelobes by virtual spatial windowing.
 Digital beamforming of the Tx beams in addition to the
Rx beams, and therefore avoid beam shape loss in cases
that the target is not in the center of the beam.
 Decrease the spatial power density of the Tx signal – spatial
spread spectrum (SSS) which is critical for low probability
of intercept radars (LPIR).
MIMO Radar Disadvantage
 Implementation
 Gain loss (omni-directional transmission)
 Not a real problem in search mode: omni-
directional coverage allows large time-on-target
(requires quasi-stationarity or track-before-detect).
 A real problem in track/acquisition modes:
If the target direction is known with a given degree
of accuracy, then MIMO radar “wastes” its energy
towards undesired directions.
Solution: Cognitive MIMO Radar
Cognitive Radar
Proposed by Simon Haykin 2006.

A cognitive radar employs

adaptive Tx-Rx based on
history observation and
environmental information.
Cognitive Radar
Why the term cognitive is used?
NIH definition:
“Cognition: conscious mental activity that informs a person
about his or her environment. Cognitive actions include
perceiving, thinking, reasoning, judging, problem solving and
remembering.”
Biological Cognitive Properties versus Cognitive Radars
Cognitive Property Cognitive Radar Equivalent
Perceiving Sensing
Thinking, Reasoning, Expert Systems, Adaptive
Judging, Problem Solving Algorithms, and Computation
Remembering Memory, Environmental Database
[Guerci 2011]
 Cognitive MIMO Radar
Data model at the kth step: Xk  Hk (Θ)Sk  Wk
Hk (Θ) - MIMO channel Θ - Target parameters
matrix,
Detection/
Side Information Estimation/
Tracking/
Optimal Adaptive Sk Xk Optimal Classification
Waveform H k (Θ)
Receiver/Processor
Design
noise

Optimal processor: Detect/localize/track/classify the target(s) based on available

measurements, X(k )  [X , 1 , X ].
k
Optimal adaptive waveform design: Design the transmit signal at the kth pulse, Sk ,
based on the measurements during the previous pulses, X (k 1)
 [X ,
1 , k ] to
optimize a given criterion. 1 X
Cognitive MIMO Radar
Data model at the kth step: Xk  Hk (Θ)Sk  Wk
Hk (Θ) - MIMO channel Θ - Target parameters
matrix,
Detection/
Side Information Estimation/
Tracking/
Optimal Adaptive Sk Xk Optimal Classification
Waveform H k (Θ)
Receiver/Processor
Design
noise

S(opt
k
)
 arg max C(S , X (k

S
1)k
) k

s.t. Tx power
constraint
(k 1)
Cognitive Beamforming
Criterion for estimation accuracy: performance bound on the mean-
squared-error (MSE):
 Bayesian Cramér-Rao bound (BCRB): Simple, but not tight.
 Bobrovski-Zakai, Reuven-Messer, or Weiss-Weinstein
bounds: High computational complexity, but tighter.

It can be shown that with Gaussian noise, the bounds depend on

1
the Tx auto-correlation matrix: R  N S kSH
Sk
k
2
Power constraint: k F
S  tr R
S k

or  S  P / NT , nP 1, ,
R  n,n
k
NT
Cognitive Beamforming
For single unknown parameter, θ, with total Tx
power constraint, and zero-mean Gaussian noise with
cov. R v:
R(opt )  Pu uH
Sk k k

uk - Γ
eigenvector
k
(k
1) 
) corresponding
E HH  )R1Hto(the
k  ) maximum
(k
1)
v
eigenvalue of
(X ( k X

Vector parameter case, Θ  Q - weighted BCRB:
Convex optimization problems, and thus can be solved
efficiently (Boyd and Vandenberghe (2004)).
Example – Cognitive Beamforming

Scenario :
 Uniform linear array of transceivers N R  NT  7 elements
with  /2 inter-element spacing.
 AWGN with covariance R   2I N .
R

 ASNR   2 NPN /  2
6dB.
R
Example – Cognitive Beamforming
Posterior pdf’s versus transmit beampatterns Pk ()  a*T ()RS a
T
k
.
() T

Auto-focusing effect: Automatic beamforming before detection/estimation.

Example – Cognitive Beamforming
Single target – direction estimation accuracy, 7 transceivers
Example – Cognitive Beamforming
Probability of resolution compared to space-reversal method.
Two targets – SNR=-2 dB, k=10, 7 transceivers.
Cognitive Detection
Sequential Hypothesis Testing (SHT):

 H1 : x k ,l  H k ,l (Θ)sk ,l  w k ,l k  1, 2,, l  1, ,
 ,
H0 : x k ,l  w k ,l L,

Goal: Minimize Average

Sample Number (ASN) to
achieve given error
probabilities: 1 P , P . D FA

f (X ( k ) | H1 ) PD
Decide H1 if: log X( k ) 
f ( k ) (X ( k ) | H0 ) 1 D
X
P
f ( k ) (X ( k ) | H1 )
Decide H 0 if: log X 1 PFA
f (X (k )
| H0)   PFA
X( k )
Cognitive Detection
Two hypotheses:
log(1 PD) log PFA 
 , .
ASN  max  
KLD k (H 1 || H 0 ) KLDk (H 0 || H1) 

C S k , X ( k 1) 

Optimal signal design:

log(1 PD ) log PFA 
Sk  arg min max , .
Sk
 ,opt  KLDk (H 1 || H 0 ) KLDk (H 0 || H1 )


KLD1)(H
. k ||m H ) - conditional
n
Kullback-Leibler Divergence given X (k
Cognitive Detection
L

Sk
,opt
 arg min  skH, E Hk ,
Sk
l l
 H
l
v1
k ,l 
(Θ) X (k 1) s l
k,

(Θ)RL H 1
2
P
s.t.  s k ,l
l 1

2
 jDTk  j T l
H k ,l (Θ)  a R ( )a ( )
T
T , k  1, 2,, l  1, L,
e e
R s ,opt  Pumax u Hmax
k

umax - Eigenvector corresponding to the maximal eigenvalue of the matrix


E a T*( )a T1)
(T ) X (k
.
Example - Cognitive Detection
4Tx, 16 Rx
 /2 inter-element
spacing NF  7dB
RCS  1m2
Range=50m

Azimuth=30
Example - Cognitive Detection
4Tx, 16 Rx, NF  7dB, RCS 1m 2 , range=50m, azimuth=30
Conclusions and Future Research
 MIMO radar offers great advantages but needs to be used with care.
 In cognitive MIMO radar, Tx signal auto-correlation matrix is adaptively
optimized. The optimized signal is not necessarily orthogonal (MIMO) or
fully correlated (phased array).
 Two new cognitive Tx beamforming approaches were presented to
optimize: localization accuracy and detection performance
This approach provides an automatic focusing array: beamforming before
detectionqestimation.
 Future research:
Considering other criteria, such as probability of resolution, or target
classification performance.
Thank you!