Wednesday, 06 November 2024

RADAR Signal Processing (19ECE452)

Presented By:
Dr. Harshavardhan Singh
Assistant Professor

Department of Electronics and Communication Engineering

அம் ரிதா விஷ்வா வித்யாபீதம்
अमृता विश्व विद्यापीठम्
Amrita Vishwa Vidyapeetham
Amrita nagar, Ettimadai, Tamil Nadu 641112
INDIA
 Syllabus

Wednesday, 06 November 2024

 Course Plan
Online AUMS Quiz 1- 10 Marks

Online AUMS Quiz 2-10 Marks 20 Marks

Online AUMS Quiz 3-10 Marks

Midterm- 30 Marks

Online AUMS submission Assignment 1-5 Marks

Online AUMS submission Assignment 2-5 Marks

End Semester- 40 Marks

Wednesday, 06 November 2024

Lecture Plan
Lecture Plan
 Introduction to Beamforming and Space-Time
Adaptive Processing
• Beamforming refers to the coherent combination
of data from multiple phase centers to provide
selectivity in the angle of arrival (AOA), i.e., to
form and steer an antenna beam.
• The phase center axis enables analysis and
processing of signals within a range bin based on
their spatial frequency content, which is
equivalent to the AOA.
• STAP combines both spatial and temporal filtering
on a moving radar platform to discriminate targets
from both clutter and jamming.
 Beamforming
• Consider a monochromatic plane wave with
temporal variation Aexp(jΩt) impinging on a
uniform linear one-dimensional array as shown in
Figure.
• If the wave’s AOA relative to the array normal is θ
radians, the signal observed at the nth array
element is

where the phase offset ϕ0 accounts for the absolute

phase at the n = 0 element
 Beamforming
• Now consider the set of samples y[n] formed from the individual array
signals, sampled at a common time t0:

• Assembling the N element samples into column vector form gives a

snapshot of the array at a fixed time:

; where kθ ≡ 2πd sin θ/λ is the normalized spatial frequency in

radians per sample as projected into the plane of the array face and as(θ) is
the spatial steering vector.
• Thus, there is a one-to-one relationship between the AOA of a plane
wave and the spatial frequency across the array face. The range of θ is
±π, so the range of kθ is ±2πd/λ. It is also useful to define the
normalized spatial frequency in cycles per sample, fθ ≡ kθ/2π.
• Conventional nonadaptive beamforming is implemented as a weighted
sum of the element signals, z = [h]T.y ,1 where h is a vector of
complex weights

• A special case of interest occurs when h takes the form

• The symbol represents the Hadamard (element-by-element) product
of two vectors. If a and b are two N-element column vectors, then

• So, in h represents the Hadamard product of two factors: a

shape vector w that provides for any weighting for sidelobe control
and the conjugate of a steering vector as(θ) that provides maximum
coherent integration for signals arriving from angle θ.
• Suppose the weights are matched to an AOA of θ0; the array is said to
be “steered” to θ0. The response of a beamformer steered to θ0 to an
incoming wavefront at angle θ is
• The scalar output z(θ) is just the discrete Fourier transform of the
weight sequence {hn}, shifted to a center spatial frequency of and
multiplied by Â. When all of the weight amplitudes {wn} = 1, this is a
standard asinc pattern in the variable sinθ,

• More commonly, the weights are chosen to reduce antenna sidelobes

at the cost of degraded resolution and gain in the form of a wider
mainbeam.
• Suppose it is desired to maximize the output z(θ) when the input is a
snapshot of a plane wave having AOA θ, plus white noise.
• It need to use for vector matched filtering with (h is a vector of complex
weights)

where SI = covariance matrix of the interference at the N phase center outputs

t = model of the desired target signal vector (i.e., steering vector)
k = arbitrary constant
• When the interference is independent identically distributed (i.i.d.) white
noise with variance at each element, .Choosing gives

• The filter output is then

 Advantage of Beamforming
• Digital beamforming enables several advantageous capabilities for
phased array antennas.
• One is the ability to form multiple simultaneous beams on receive. The
same snapshot data y can be weighted with several different weight
vectors h, effectively forming multiple receive beams steered in
different directions.
• Another major capability is the opportunity to compute the weights
adaptively based on the interference environment characteristics.
• This makes possible adaptive nulling of jammers, discussed next.
 Adaptive Beamforming
The beamformer weights were designed using
 Adaptive Beamforming
• The vector matched filter approach leads directly to a means for
designing array weight vectors h that can steer zeros of the antenna
pattern, often called nulls, in specific directions to cancel interference
sources.
• The antenna pattern that maximizes the output SIR in the presence of
both white noise and jamming is still given by
• However, the model for the interference covariance matrix SI must be
extended to incorporate a model for the noise jammer.
 Adaptive Beamforming
• The temporal variation of the jammer signal Jn(t) received from AOA
θ at each array phase center can be expressed as

• Where is the jammer power and w(t) is a unit variance zero mean
white random process.
• A snapshot of the array response to the jammer gives
 Adaptive Beamforming

• In vector form, this is

• The covariance matrix of the jammer signals is then

• SJ is a Hermitian matrix, . It is also positive semidefinite, but has a
rank of only one because the columns are all linearly dependent; in
fact, they are all simple multiples of the first column.
• The total interference covariance matrix for the sum of receiver noise
and some number P of mutually uncorrelated jammers is

• The beam former weights were designed

using a covariance matrix
• The SIR achieved by the optimum beam former can be computed
using the result of Eq. (5.7), namely

• For the case where the interference is noise only, t = Â.as(θ),

• Coherent integration gain of Gsp = N.

 An Example to discuss
• As an example, consider a case with N = 16 antenna phase
centers. Assume that two jammers are present, one at an AOA of
+18° with a jammer-to-signal ratio (JSR) of +50 dB, and another
at –33° and a JSR of +30 dB. The SNR = 0 dB, resulting in a
total SIR at the element level (before any beamforming) of –
50.04 dB.
• Without jammers the interference is just the noise, so the pre-
beamforming SIR is the single-element SNR.
The post-beamforming SIR is 16 = 12.04 dB and the SIR gain is
therefore also Gsp = 12.04 dB. With jammers present the
prebeamforming SIR is -50.04 dB and the post-beamforming SIR
is 15.64 (11.94 dB), so the gain is 61.98 dB.
The slight reduction in post-beamforming SIR (0.36 dB) when the
jammers are present occurs because some of the degrees of
freedom (DOF) are consumed canceling the jammers instead of
providing coherent integration gain for the signal.
 Distortionless beamformer
• This variation in peak gain in the direction of the desired target signal t
can be compensated by requiring that ℎ𝑇 t = 1, resulting in what is
often called a distortionless beamformer.
 Subbanding
• The elements of the steering vector and thus of the optimum adaptive
weights depend on the RF through the wavelength λ or the spatial
frequency κθ.
• The weights that steer the beam to a particular AOA for one RF steer it
to a different AOA for a different RF.
• Because of this dependence on λ, a single set of phase weights is
incapable of cancelling a jammer over a wide bandwidth. A very
common way to deal with this issue is to use subbanding.
 Subbanding
• This technique, diagrammed in Fig., divides the array spatial data as a
function of fast time into K sub bands, each relatively narrowband.

• Another approach to wideband cancellation uses fast-time filtering in

each spatial channel
 Adaptive Beamforming with Pre-processing
• The discussion so far has applied adaptive interference calculation directly to the
individual array element signals, an approach called element space processing.
• Some systems perform fixed conventional beamforming on the element signals
first, and then apply adaptive processing to the individual beam outputs. This
technique is called beamspace processing.
• Earlier section, Beams are formed as linear combinations of the element signals as
shown in the example of

• A set of P beams formed from an N-element array can thus be represented as a P ×

N transformation matrix T acting on the N × 1 spatial snapshot y to form a new P ×
1 vector of beam outputs ỹ
• For instance, a set of conventional “DFT beams” at spatial frequencies
can be formed with the transformation matrix:

• To see how to apply adaptive processing to the preprocessed

beamspace data, note that the snapshot y is now replaced with ỹ.

• The adaptive weight vector and the filtered output are

 Point to Remember
• Preprocessing can involve filtering, signal enhancement, and noise suppression
techniques before the beamforming step. This further improves SNR by cleaning
up the signal.
• Preprocessing can reduce the complexity of the input signal by filtering out
irrelevant components or compressing the data, thereby reducing the
computational load on the adaptive beamforming algorithms. This is particularly
useful in real-time applications where computational resources are limited.
• In environments with significant multipath propagation (e.g., urban areas, indoor
scenarios), preprocessing can help mitigate the impact of multipath signals.
• Adaptive algorithms often rely on iterative methods to converge to an optimal
solution. Preprocessing can accelerate convergence by providing a cleaner or
more structured input signal, reducing the number of iterations needed to achieve
optimal performance.
 Space-Time Signal Environment
• In a multi-phase center radar, filtering is possible in both Doppler shift and
angle of arrival. It is therefore important to characterize the data in a given range
bin in terms of Doppler and AOA.

• Broadband noise jammers are localized in AOA but

spread across the entire Doppler spectrum.
• This is reflected in Fig. as a ridge of energy localized
in AOA but spread across all Doppler shifts.
• Because the jamming energy is the same at all values
of Doppler. Discrimination against jamming must be
based primarily on spatial filtering as described in
the previous section.
• Clutter is more complicated. Assume a platform with
velocity υ and a side looking radar as shown in Figure
• Assume a platform with velocity υ and a side looking radar
as shown in Figure.
• Scatterers at an angle of θ radians with respect to the antenna
boresight will have a Doppler shift of (2υ/λ) sinθ Hz.
• In normalized unit:

where T is the PRI and the notation ((·))1.0 indicates arithmetic

modulo 1.0 due to the use of discrete-time Fourier analysis. It
can be represented in:

where β ≡ 2υT/d. β is the slope of the clutter ridge when

plotted in (kθ, ωD) coordinates.
Problem
Find the STAP clutter ridge in (kθ, ωD) coordinates assuming the
following system parameters: υ = 150 m/s, T = 1 ms, d = 0.3 m, and F0
= 1 GHz. What is the value of the clutter ridge slope β in this case?
Repeat for d = 0.15 m and T = 0.25 ms.

Solution:
Use simply the formulas given in previous slide.
 Space-Time Signal Modeling
• Space-time adaptive processing (STAP) applies vector matched filtering to the
combined slow-time/phase center data set in each range bin.
• The two-dimensional slice of the datacube in range bin l0, y[l0, m, n], is called a
space-time snapshot (or just snapshot) of the data. To proceed, the N × M two-
dimensional snapshot is converted to an NM × 1 = P × 1 column vector by
stacking the columns:
Two-dimensional spacetime
snapshot and then to a one-dimensional vector.
• Next, the filter weight vector h must be designed using Eq.
• The target model vector t = t(fD, θ) must represent the expected signal
from a target at some specific Doppler shift fDt and AOA θt of interest.
Define a temporal steering vector

• This is simply the model for the slow-time data sequence corresponding
to a target at normalized Doppler frequency fDt.
• The two-dimensional snapshot of the data from a target at Doppler fDt
and AOA θt would have the temporal variation of at(fDt) in each row
and the spatial variation of as(θt) in each column. The snapshot
therefore has the form of the outer product of as with at

• When this matrix is vectorized by stacking the columns, the result is

the Kronecker product of the desired spatial and temporal steering
vectors
• Next, a model for the P × P covariance matrix SI of the interference is needed.
• It is assumed that these are all uncorrelated with one another, with the result that
the total interference covariance is the sum of the covariances of the three
components, SI = Sn + SJ + Sc.
• Now consider a single noise jammer signal. The spatial variation was
given in

• The temporal variation can be modeled as

where each terms are uncorrelated i.i.d. random variables with equal
power .
• The covariance of the jammer temporal variation is
• If the AOA of the jammer is θJ0, the space-time data vector for the jammer
component is

• and its covariance matrix can be shown to be

which is a block-diagonal matrix. If R multiple uncorrelated jammers are present, SJ

is the sum of R terms of the form in Eq. (9.34), each with its own AOA θJR
• Similarly for clutter

• where sigma square term represents the power of the θth clutter patch
as determined by the radar range equation.
• The total clutter vector is

• The covariance matrix of the clutter is

Processing the Space-Time Signal
1. Optimum Matched Filtering
2. STAP Metrics
3. Relation to Displaced Phase Center Antenna Processing
4. Adaptive Matched Filtering
1. Optimum Matched Filtering
Optimal space-time processing of the two-dimensional data snapshot consists of
the following steps. For a range bin of interest,
• Form the interference covariance matrix SI. In practice, SI must be estimated
from the radar data
• Select a Doppler shift and angle of arrival at which to test for the presence of
a target signal, and form the appropriate target space-time steering vector t
using
• Compute the optimal filter weight h using depending on the normalization
desired.
• Form the space-time data vector y
• Apply the weight vector to the data to obtain the test statistic z = z(fD, θ) =
h𝑇 y.
While the optimum weight vector is given by , in practice it is often
desirable to include windowing of the data to reduce sidelobes.

where the windowed target steering vector is

and wf and wθ are the temporal and spatial weight vectors.

#Problem
Consider a small STAP radar system operating at S band (3 GHz) with
M = 2 slow-time samples in the CPI and N = 2 phase centers. The goal
is to compute the optimum matched filter coefficients for detecting a
target directly on boresight of the antenna (θt = 0°) and having a
normalized Doppler shift of fDt = 0.25 cycles per sample.
• a. What is the temporal steering vector for the target, at? Simplify the
resulting expression (e.g., e^(jπ) = –1).
• b. What is the spatial steering vector for the target, as? Simplify.
• c. What is the complete target model vector t?
#Problem
Consider a small STAP radar system operating at C band (6 GHz) with M = 4
slow-time samples in the CPI and N = 2 phase centers. The goal is to compute
the optimum matched filter coefficients for detecting a target directly on
boresight of the antenna (θt = 0°) and having a normalized Doppler shift of fDt
= 0.25 cycles per sample.
• a. What is the temporal steering vector for the target, at? Simplify the
resulting expression (e.g., e^(jπ) = –1).
• b. What is the spatial steering vector for the target, as? Simplify.
• c. What is the complete target model vector t?
• d. What will be optimum beamforming filter coefficient vector h, if
interference is only white noise with power σ2 ω= 2W.
STAP Metrics
A common metric used to visualize STAP filtering performance is the
adapted pattern. This is simply a two-dimensional plot of |z(fD, θ)|2
=|h𝑇 t|2 = h𝐻 t*t𝑇 ℎ as fD and θ are stepped over a regular grid.
• Another important class of metrics involves SIR when the interference
is noise only and when it is noise plus clutter and/or jamming. Metrics
in this class include SIR and SIR loss, which is the reduction in SIR
for the clutter-plus jamming case compared to the noise-only case.
• Each of these is a function of target angle and Doppler; typically they
are plotted as a function of Doppler for a fixed AOA.
• Consider a target that produces an SNR of χ𝑡 for a single pulse and
phase center, i.e., for a single sample of the two-dimensional snapshot.
The optimum SNR is then

• The factor MN represents the coherent integration gain achievable by

combining the MN samples of the snapshot. The SIR is,

when using the optimum weight vector this becomes just

The SIR loss is then defined as

• LSIR is a function of Doppler and angle of arrival.

 • Minimum detectable velocity (MDV)
is the velocity closest to the clutter
notch at which an acceptable SIR loss
is achieved.
• Minimum detectable Doppler(MDD)
is the corresponding Doppler shift in
either absolute or normalized units.

where L0 is the maximum acceptable loss threshold

Another metric related to MDD is the usable Doppler space fraction

(UDSF).
• MDV is λ/2 times the MDD in hertz or λ/2T times the MDD in normalized
frequency units of cycles/sample
Relation to Displaced Phase Center Antenna
Processing
• DPCA processing combines
data from multiple pulses and
multiple antenna phase
centers to form a clutter-
cancelled output; as such, it
appears to be related to STAP
processing.
• Consider non-adaptive DPCA
processing using two phase
centers separated by a
distance dpc so that N = 2.
• A non-adaptive DPCA processor is constrained to be of the form

where yf [m] and ya[m] are the fore and aft phase center outputs,
respectively. It is sufficient to work with a single “sub-CPI”, i.e., an interval
of Ms pulses, that spans the two pulses combined in the DPCA processor.
The complete CPI snapshot is then constructed of a series of M – Ms
overlapping sub-CPIs as shown in the bottom portion of Fig. (previous fig.)
for the case Ms = 2. The weight vector h1 that implements above eq. for a
single sub-CPI and the corresponding weight vector h for the entire CPI are
shown in below Fig in two-dimensional form.
• Viewed as a two-dimensional discrete function, the sub-CPI weight vector
h1 can be written as

• The adapted pattern is the two-dimensional DTFT of this function:

Note that this function has a zero at ωD = kθ/Ms = βkθ, which is exactly the
expression for the clutter ridge in Eq. (9.28). Figure 9.19 shows |H1(fD, kθ)|
for the case Ms = 2. Clearly, the DPCA processor implements a null along the
clutter ridge.
• On the other hand, the constrained form of the weight vector Equation
prevents any combining of samples from different phase centers on the
same pulse so that the DPCA processor cannot provide any spatial
beamforming capability. As a result, DPCA processing can cancel the
clutter but not jammers.
Adaptive Matched Filtering
• Consequently, they can be used to compute a sample mean estimate
of SI, denoted ŜI. Specifically, the sample covariance of the data from
a single reference cell is

• If LS reference cells are used, the estimate of SI is simply

• The reduction in SIR due to the use of ŜI instead of SI is denoted by

LCFAR.It has been shown to be a beta-distributed random variable with
expected value
• This is known as the “Reed-Mallet-Brennan” or RMB rule.

• The Ls in the RMB can be approximated by Rule of thumb

approximation as
Ls ≥ 5P.
Problem
According to the RMB rule, how many range bins should be used in the STAP
system having M=2 and N=4 to estimate the interference covariance matrix
so that the mismatch loss is less than 1 dB?
How does the exact result compare to the rule-of-thumb estimate?

Solution: first get P=M.N,

Then use formulas given in previous slides.
Loss should be taken as 10log X=-1
So X will be ≈ 0.8
Then continue in formulas.