0% found this document useful (0 votes)
186 views15 pages

Matched Filter

This document discusses the matched filter, which is a linear filter designed to detect a known waveform buried in additive noise. It describes the goal of optimizing the signal-to-noise ratio at the output of the filter. The derivation shows that the matched filter is the time reversal of the complex conjugate of the known signal divided by the noise power spectral density. This maximizes the output signal power while minimizing the output noise power, allowing detection of weak signals. Examples of applications in radar and communications are provided.

Uploaded by

jMcLean99
Copyright
© © All Rights Reserved
We take content rights seriously. If you suspect this is your content, claim it here.
Available Formats
Download as PDF, TXT or read online on Scribd
0% found this document useful (0 votes)
186 views15 pages

Matched Filter

This document discusses the matched filter, which is a linear filter designed to detect a known waveform buried in additive noise. It describes the goal of optimizing the signal-to-noise ratio at the output of the filter. The derivation shows that the matched filter is the time reversal of the complex conjugate of the known signal divided by the noise power spectral density. This maximizes the output signal power while minimizing the output noise power, allowing detection of weak signals. Examples of applications in radar and communications are provided.

Uploaded by

jMcLean99
Copyright
© © All Rights Reserved
We take content rights seriously. If you suspect this is your content, claim it here.
Available Formats
Download as PDF, TXT or read online on Scribd
You are on page 1/ 15

EEE3086F

Signals and Systems II


2012

Andrew Wilkinson
[email protected]
http://www.ee.uct.ac.za
Department of Electrical Engineering
University of Cape Town
A.J.Wilkinson, UCT

Correlation Functions

EEE3086FSignalsandSystemsII
212Page1March19,2012

2.7 Matched Filter


Contents:
1. Goal of the Matched Filter
2. Derivation of the Matched Filter
3. Examples

Reference: F.G. Stremler, Introduction to Communication Systems,


A.J.Wilkinson, UCT

Correlation Functions

3rd Edition
EEE3086FSignalsandSystemsII
212Page2March19,2012

Goal of the Matched Filter

A.J.Wilkinson, UCT

Correlation Functions

EEE3086FSignalsandSystemsII
212Page3March19,2012

Aim and Applications of the Matched Filter


The matched filter is a linear filter that is designed to detect the presence of a waveform
(or pulse) x(t) of known structure buried in additive noise.
i.e. within a signal

v (t )= A x (tt 0 )+n(t)

where A and t0 are unknown constants, and n(t) is noise.


The output of a matched filter will typically exhibit a sharp peak in response to the
presence of the desired waveform (or pulse) at its input. From this peak may be
determined time location t0, and amplitude A of the pulse within the received waveform.
E.g. 1. Radars transmit electromagnetic pulses which reflect off distance objects (or
targets). The echoes are captured by a receiving antenna. The echoes may be very
weak signals buried in additive noise (picked up by the antenna, and also created by the
receiver itself). Matched filters are used to process the received waveform, and hence
optimize the chance of detecting a target.
E.g. 2. In telecommunications systems using pulse code modulation, a bank of N filters
may be constructed, with each filter specifically designed to respond to the presence of a
particular pulse code. Comparison of the responses at the outputs identifies the received
code.
EEE3086FSignalsandSystemsII
A.J.Wilkinson, UCT
Correlation Functions

212Page4March19,2012

Model of Transmitted and Received Signal with Additive


Noise
Transmitted pulse x (t)
0
Received signal
v (t )= A x (tt 0 )+n(t)

t0

where t0 is the propagation time delay,


A is an amplitude scaling factor, and
n(t) is additive noise, described by its PSD Sn().
The SNR of the received pulse is defined as the ratio of the the
peak signal power to average noise power, i.e.
2

max {A x (t t 0 ) } A x (0)
peak signal power
SNR=
=
=
2
2
average noise power
n(t)
n(t )
A.J.Wilkinson, UCT

Correlation Functions

EEE3086FSignalsandSystemsII
212Page5March19,2012

Definition of the Matched Filter


Given an input signal v (t )= A x (t t 0 )+n(t)
where x(t) is a pulse of known structure,
t0 is an unknown delay, and A is an unknown scaling factor.
A matched filter is a linear filter h(t) or H() that is
designed to optimize the peak SNR at its output at some
specified instant relative to t0.
The output is v o (t )=v (t ) h(t )= y (t )+no (t )
The aim is to find h(t) or equivalently H() that maximizes
2
(at some instant tx, usually tx >t0):
y (t x )
peak SNR=
2
n o (t )
A.J.Wilkinson, UCT

Correlation Functions

EEE3086FSignalsandSystemsII
212Page6March19,2012

Derivation of the Matched Filter

A.J.Wilkinson, UCT

EEE3086FSignalsandSystemsII
212Page7March19,2012

Correlation Functions

Simplified Signal Model


Let the input be:
Filter output:

Let

plus stationary noise n(t) )

v o (t )=v (t) h(t )= x (t ) h(t )+n(t ) h(t )

y (t )= x (t ) h(t )

(output signal component)

n o (t )=n(t ) h(t )

(output noise component)

Output written as

A.J.Wilkinson, UCT

v (t)= x (t)+n(t ) (where x(t) is a known waveform

v o (t )= y (t )+n o (t )

Correlation Functions

EEE3086FSignalsandSystemsII
212Page8March19,2012

Signal Model Illustrated


t

Input
signal

Filter

x (t)

v (t)

Output
v o (t)= y (t )+n o (t )

H()

peak
t

n(t ) noise

Maximize at t = td

td

y (t d ) peak signal power


output SNR=
=
2
n o (t) average noise power

A.J.Wilkinson, UCT

EEE3086FSignalsandSystemsII
212Page9March19,2012

Correlation Functions

Derivation of the Matched Filter


2

y (t d ) peak signal power at t =t d


output SNR=
=
2
average noise power
n o (t )
+

Output (signal y (t )= x (t ) h(t )= 1 X () H ()e j t d


2
component)

(from inverse FT)

Peak output
signal power

1
y (t d ) = X () H ()e j t d
2
2

+
+
Average noise
2
1
1
2
n
(t
)
=
S
()d
=
S
()

H
()

o
no
n
power
2
2

A.J.Wilkinson, UCT

Correlation Functions

EEE3086FSignalsandSystemsII
212Page10March19,2012

Optimization Problem
The objective is to find a/the filter H() that optimizes the SNR at
the time instant t = td:
2
1
j t
2
X () H ()e d
y (t d ) 2
=
+
2
1
2
no (t )
S n ()H () d

To do this, we make use of a (provable) mathematical relationship called


the Schwartz (or Cauchy-Schwartz) inequality.

A.J.Wilkinson, UCT

EEE3086FSignalsandSystemsII
212Page11March19,2012

Correlation Functions

Schwartz Inequality for Integrals


The Schwartz inequality states that for two complex
functions, f(x) and g(x), integrable over [a,b]
b

f ( x) g ( x)dx f ( x) dx
a

g ( x) dx
a

The equality holds (i.e. LHS = RHS) if g ( x)=k f ( x)


where k is a real constant.

If g ( x)k f ( x) then the left side is less than the right side.
Proof? Google it!
A.J.Wilkinson, UCT

Correlation Functions

EEE3086FSignalsandSystemsII
212Page12March19,2012

To maximize SNR, apply Schwartz Inequality


+

1
j t
2
X () H ()e
d

y (t d ) 2
=
+
2
1
2
no (t)
S n ()H () d

SNR:

[
+

y (t d )
=
2
no (t )

Rewrite expression as:

y (t d )

2
n
(t)
o

X ()
S n ()

1
2

] d

1
2

j t d

EEE3086FSignalsandSystemsII
212Page13March19,2012

X ()

1
2
d
S n (w)H () d

2
S n ()
+
1
2
S n ()H () d

y (t d )
1

2
2
n0 (t )

g ( x ) S n () H () e

The term on the right


of the numerator
cancels with the
denominator.

and
2

Correlation Functions

S n () H ()e

+
X ()
1
j t 2
d
S n () H ()e d

2
S n ()
+
1
2
S n ()H () d

A.J.Wilkinson, UCT

y (t d )

2
n o (t )

j t d

][

1
2
S n ()H () d

Apply Schwartz Inequality with: f ( x)


Notice that the term
on the right of the
numerator cancels
with the denominator.

X ()
S n ()

X ()

S n ()

Swartz inequality becomes an equality i.e. LHS = RHS if g ( x)=k f ( x)


or in our case, if:

S n () H () e
A.J.Wilkinson, UCT

j t d

X ()
=k
S n ()

Correlation Functions

X () j t
=> H ()=k
e
S n ()

EEE3086FSignalsandSystemsII
212Page14March19,2012

Matched Filter

X () j t
H ()=k
e
S n ()

The filter that maximizes


the peak SNR at time td is:

k is a real constant, usually positive, and is often set to 1.


For the special case of white noise, S n ()=constant

H ()=k X ()e
Inverse transforming gives:

j t d

h(t )=k x (t +t d )

Apply Fourier: X () x (t)


j t
and Y ()e
y (tt d )
d

A.J.Wilkinson, UCT

EEE3086FSignalsandSystemsII
212Page15March19,2012

Correlation Functions

Matched Filter (additive white noise)


x (t)

X ()

2 B

2 B

x (t)
t
j t d

H ()= X () e

h(t )= x (t +t d )
td

Flip h(t), slide, mult. and integrate

Y ()=X () e j t

y (t)=x (t ) h(t )c 2 2B Sa (2 B (t t d ))

A.J.Wilkinson, UCT

td

c 2 rect (/ 4 B)

t
Correlation Functions

2 B

2 B

EEE3086FSignalsandSystemsII
212Page16March19,2012

X () j t
Interpretation in Frequency Domain: H ()=k S () e
n

In the frequency domain, the matched filter magnitude boosts the


frequency components for which the signal is large compared to
the noise and suppresses those that are weak compared to the
noise i.e.
X ()
H ()
S n ()
The phase of the filter cancels the phase of X(), H () X ()
and hence the inverse Fourier transform integral gives a strong
peak at t = td. In the white noise case H ()= X ()e j t
d

1
1
2 j t
j t
j t
y (t )= X () H ()e d= X () e
e d
2
2
y (t d )=E x (by Parseval)
d

A.J.Wilkinson, UCT

Correlation Functions

EEE3086FSignalsandSystemsII
212Page17March19,2012

Interpretation in Time Domain


In the time domain, the matched filter for white noise is seen as
the conjugate of a time-reversed (flipped) version of x(t).
It is easiest to consider the case of k=1 and td = 0, which is a

h(t )= x (t )
matched filter with zero-delay.
Performing a graphical convolution, flips h(t), which re-flips

x (t ) such that it will now match (or correlate well) with the
x(t) contained in the received signal v(t) as one slides from left
to right and integrates the product.

v (t) h(t )= x (t ) h(t )+n(t ) h(t )


The width of the peak in the time domain depends on the
bandwidth B of X(). To get y(t), we inverse transform:
2
EEE3086FSignalsandSystemsII
A.J.Wilkinson, UCT
Functions
212Page18March19,2012
X ()Correlation
H ()
X ()

Interpretation in Time Domain:


The width of the peak in the time domain depends on the
bandwidth B of X(), since we are inverse transforming X ()2
If X ()2 has an approximately rectangular bandwidth,
then the output will have a Sa( ) shape.
1
The 3dB width of the lobe will be t
[s].
2B
Bandlimited
input signal x (t )
t
A.J.Wilkinson, UCT

Matched Filter

Output
peak

H()
noise
Correlation Functions

1
2B

td

EEE3086FSignalsandSystemsII
212Page19March19,2012

Physically Realizable Filters


The impulse response of the matched filter for waveform x(t) with

additive white noise is

h(t )= x (t +t d )

If this filter is to be constructed as a filter operating in real-time,


then the output cannot respond before the input arrives.
Thus for physical realizability, the impulse response h(t) must
begin after t = 0. i.e. h(t) = 0 for t <= 0.
In a real matched filter (built either with passive LRC components or
using A/D, D/A and DSP technology), the MF peaks after the entire
waveform has been fed into the input.
In designing a filter, one can always make it physically realizable by
shifting h(t) to the right (by increasing td). This will simply delay the
output response by the corresponding
amount.
Correlation Functions

A.J.Wilkinson, UCT

EEE3086FSignalsandSystemsII
212Page20March19,2012

Relationship to Correlation
Feeding a signal v (t) into a matched filter h(t )= x (t+t d )
matched to waveform x (t) (with white noise), is equivalent
to a correlation operation between x(t) and v(t).

Recall correlation for energy signals: R vx ( )= v * (t ) x (t+ )dt

The matched filter output is:

v o (t ) =

v (t ' )h(tt ' )dt ' = v (t ' ) x (t ' +t d t )dt '

t '=

v o (t ) =
A.J.Wilkinson, UCT

u=

v o (t)=v (t) h(t )

put u=t ' +t d t

x (u)v (u+[ tt d ]) du = R xv (t t d )
Correlation Functions

EEE3086FSignalsandSystemsII
212Page21March19,2012

Relationship to Correlation cont...


Thus if one correlates x(t) and v(t), and shifts the result to the
right by an amount td, one gets the same output as from the
matched filter operation.
The matched filter output is

v o (t )=v (t) h(t )


v o (t)= R vx (tt d )

Note: if v (t)= x (t) (the case where the input matches


perfectly) then the MF output is v o (t )= R x (t t d )
which is just the autocorrelation function of x(t), shifted by td.
A.J.Wilkinson, UCT

Correlation Functions

EEE3086FSignalsandSystemsII
212Page22March19,2012

Output SNR
The output SNR of the matched filter is
2

2
+
y (t d )
X ()
1
=
S () d
2
2

n
n0 (t )

For white noise case, S n ()=/2


2

y (t d ) =

2
1
n o (t ) =

j t d

H ()= X ()e

1
j t
2
X () H ()e
d =E x

2
+

(previously derived
general case)

2
H () d=

2 2

2
X () d= E x

2 2
2

2
Ex
Ex
y (t d )

=
=

2
/ 2
E
n0 (t )
2 x

Thus the output SNR is proportional to the energy of the


received signal (and hence the energy of the transmitted
signal).
A.J.Wilkinson, UCT

EEE3086FSignalsandSystemsII
212Page23March19,2012

Correlation Functions

Example of Matched Filter: detection of a received pulse


Transmitted
pulse.

t T / 2
x (t )= A rect
T

0
A

Matched Filter.

h(t )= x (t )

Delayed, scaled pulse.


Noisy received
waveform.

kA

k x (tt d )

v (t )=k x (t t d )+n(t )

Output of matched filter; h(t ) v (t )


SNR at peak is optimised.

kA2T

td-T td
Note: output peak is proportional to energy in pulse (A2T)
A.J.Wilkinson, UCT

Correlation Functions

td

t
Peak SNR

td+T

EEE3086FSignalsandSystemsII
212Page24March19,2012

Example of Matched Filter: detection of a received pulse


A

x (t )

Transmitted
pulse.

A
Matched Filter.

h(t )= x (t )

Delayed, scaled pulse.


Noisy received
waveform.

kA

k x (tt d )

v (t )=k x (t t d )+n(t )

Output of matched filter; h(t ) v (t )


SNR at peak is optimised.

A.J.Wilkinson, UCT

t
Peak SNR

k Ex

Note: output peak is proportional to energy in pulse

td

td-T

td+T

td

EEE3086FSignalsandSystemsII
212Page25March19,2012

Correlation Functions

Applications of Matched Filter


The matched filter is very useful for detecting known waveforms in the
presence of additive noise .
E.g. 1. Radars use correlation receivers to detect weak echoes from distant
targets. A replica of the transmitted pulse is correlated with the received
waveform. The radar resolution (ability to separate close targets) depends
on the bandwidth of the transmitted pulse.
E.g. 2. In telecommunications systems using pulse amplitude modulation, a
matched filter, matched to the shape of the pulse, can be used to estimate
optimally the pulse amplitude, since the height of the peak at the output of
the matched filter is proportional to the amplitude of the received pulse.
Received v (t)
Output of MF
A.J.Wilkinson, UCT

h(t ) v (t )

t
Correlation Functions

EEE3086FSignalsandSystemsII
212Page26March19,2012

Drill Problem 2.14 (2012)


Design the impulse response of a physically realizable filter
matched to the waveform below:
2V
1V
t

1s 2s

The input noise is bandlimited white noise with a power


spectral density of S n ()=0.5 rect ( /20 )W/Hz into 1 ohm
Sketch the impulse response h(t) of the matched filter.
Plot the output y(t) of the matched filter (excluding noise).
Calculate the peak signal to noise ratio at the input and at the
output.

A.J.Wilkinson, UCT

EEE3086FSignalsandSystemsII
212Page27March19,2012

Correlation Functions

Drill Problem 2.15 (2012)


By application of properties of the inverse Fourier transform,
determine the inverse transform of:
X ()e

j t d

in terms of x (t )=F1 { X ()}


Show each step carefully.

A.J.Wilkinson, UCT

Correlation Functions

EEE3086FSignalsandSystemsII
212Page28March19,2012

EEE3086F
Signals and Systems II

End of handout

A.J.Wilkinson, UCT

Correlation Functions

EEE3086FSignalsandSystemsII
212Page29March19,2012

You might also like