Applications of Adaptive Filters

CHAPTER 5:
APPLICATIONS of ADAPTIVE FILTERS
1. Introduction to adaptive filters 2. Applications of adaptive filters Duration: 3 hrs
Introduction to adaptive filters
Outline: 1. Preliminaries 2. Problem formulation
Adaptive filtering
When x(n) and y(n) dynamically change their characteristics, unlike the standard LPF/HPF/BPF/BSF, then h(n) should be adaptively updated! 2 types of adaptive filter: FIR and IIR s+w Example: noise cancellation.
w
Adaptive filter
Adaptive filter structure

Typical FIR adaptive filter: most practical and widely used
(no linear phase requirement)
Introduction to adaptive filters
Outline: 1. Preliminaries 2. Problem formulation
Problem formulation
Goal: to optimize the adjustable filter parameters Criterion: - Meaningful measure of filter performance - Practically realizable algorithm Methods: - Least Mean Square (LSE): to minimize E[e2(n)], mean squared error, then y(n) will closely approximate d(n).
LMS algorithm (cont.)

Adjustable coefficients: {h(k)}, 0k -1
y(n) = h(k ) x(n k ),
k =0 1
n = 0,1,2,...., M n = 0,1,2,...., M
M
e(n) = d (n) y(n),

M M 2 2
= e ( n ) = ( d ( n ) y ( n )) = ( d ( n ) h ( k ) x ( n k ) ) 2
n =0 n =0 1 1 2 n =0 1 k =0
d
n =0
( n ) 2 h ( k ) rdx ( k ) + h ( k ) h ( l ) rxx ( k l )
k =0 M k =0 l =0
where rdx ( k ) = and rxx ( k ) =
d ( n ) x ( n k ),
n =0
0k 0k
1 crosscorrelation 1
autocorrelation
x ( n ) x ( n + k ),
n =0
LMS algorithm (cont.)

Goal: to determine {h(k)} so that
can be minimized
= 0, 0 m h ( m )
1 Yule - Walker Eq. 1
h(k )r
k =0
xx
( k m ) = rdx (m ), 0 m
Since the data characteristics keep on changing, we have to also update rxx ( k ) & rdx ( k ) so that {h(k)} can be updated!
Gradient descent search

Goal: to dynamically minimizes the instantaneous squared error without explicitly computing {rxx(k)} and {rdx(k)}.
e 2 (n ) hn (k ) = hn 1 ( k ) , 0k 2 hn 1 (k ) [d (n ) y (n )] hn (k ) = hn 1 ( k ) e( n ) hn 1 ( k ) hn (k ) = hn 1 ( k ) e( n )
[d (n ) h( m) x (n m )]
m =0
hn 1 (k )
hn (k ) = hn 1 ( k ) + e(n ) x ( n k )
where is adaptation rate (or step size)
Gradient descent search

e2(n)
Move this way Move this way optimum point
e2min(n) hn-1
e 2 ( n ) hn ( k ) = hn 1 ( k ) , 0k 2 hn 1 ( k )
hn-1
1 Move to the reverse direction of derivative sign
The step size of LMS

It controls the rate of convergence toward the optimal solution. Larger step size more rapid convergence of oscillation (unstable). Smaller step size Rule of thumb: slower convergence. the potential
1 0<< , where N is the filter length 10 P x

M 1 Px x 2 ( n ) , average signal power M + 1 n=0
Some variations of LMS

Sign-Error LMS: h n ( k ) = h n 1 ( k ) + sgn[ e ( n )] x ( n k ),
where 1, sgn[ u ] = 1, u 0 u <0
Sign-Data LMS:
hn ( k ) = hn 1 ( k ) + e ( n ) sgn[ x ( n k )]
Sign-Sign LMS:
hn ( k ) = hn 1 ( k ) + sgn[ e ( n )] sgn[ x ( n k )]
Variable step size: is large at the beginning and then is progressively reduced to a smaller size.
Applications of adaptive filters

Outline: 1. System identification 2. Signal prediction 3. Noise cancellation 4. Channel equalization 5. Inference suppression
System identification (system modeling)

To identify (model) an unknown system (all-zero or all-pole) by an FIR filter of length N. Both systems are excited by the same signal.
e(n) = d (n) y(n)

x(n) d(n)
y(n) e(n)
Example of System Identification

Special features offered in many Hi-Fis that consists of several scenarios, e.g., church.

Signal prediction The signal x(n) is highly correlated among samples The current sample are linearly predicted x(n) d(n) by immediate past samples.
y(n) e(n)

Noise cancellation
Noise: signal carries no correlation among samples not be removed through signal prediction. can
d ( n ) = x ( n ) + v ( n ) received in the primary microphone , the reference noise r(n) received by secondary microphone is correlated to v(n) .
d(n) x(n) e(n)
r(n) v(n)
y(n)
Noise cancellation in cockpit

Adaptive Noise Canceller for speech in adverse environments (e.g., inside airplane) Adaptive system filters the noise reference signal to make it more similar to that of the oise reference Main microphone main channel. microphone
Noise is not completely removed, but its level is reduced.
Acoustic echo cancellation

Getting important in video conferencing.
y ( n ) = z ( n ) + T [ x ( n )] where T[x(n)] is the reflected (summation of multiple delayed) speaker output received in the microphone .

Channel equalization
To remove (or to minimize) the adverse effects caused by the communication channels.
d (n) = x(n M ) delay " M" is due to the black box or the filter processing
x(n)
d(n)
y(n) e(n)
Channel equalization
In high speed modem (>2400 bps), all the processing introduces amplitude/phase distortion throughout the channel need to compensate that distortion.

Narrowband interference suppression

Signal source {x(n)} consists of a desired wideband signal {w(n)} (e.g., spread spectrum), corrupted by an additive narrowband interference {s(n)} (e.g., jamming signal). Goal: to suppress {s(n)}
Narrowband interference suppression

Using {x(n-D)} to perform adaptive filtering to approximate {s(n)}, so as to suppress {s(n)}.
s (n) = h(k ) x(n k D),
k =0 1
e(n) = x(n) s (n), hn (k ) = hn 1 (k ) + e(n) x(n k D), k = 0,1,...,
1, n = 1,2,...
Interference cancellation in ECG

Lots of interference in Electrocardiography, but 50 Hz from the main power distribution is the strongest. Notch filter (very narrowband bandstop filter) requires high order, therefore adaptive filtering is a good alternative.
Performance of ECG interference cancellation

Applications of Adaptive Filters

Uploaded by

Applications of Adaptive Filters

Uploaded by

CHAPTER 5:

APPLICATIONS of ADAPTIVE FILTERS

1. Introduction to adaptive filters 2. Applications of adaptive filters Duration: 3 hrs

Introduction to adaptive filters

Outline: 1. Preliminaries 2. Problem formulation

Adaptive filter structure

(no linear phase requirement)

Introduction to adaptive filters

Outline: 1. Preliminaries 2. Problem formulation

LMS algorithm (cont.)

e(n) = d (n) y(n),

where rdx ( k ) = and rxx ( k ) =

LMS algorithm (cont.)

1 Yule - Walker Eq. 1

Gradient descent search

Gradient descent search

Move this way Move this way optimum point

The step size of LMS

1 0<< , where N is the filter length 10 P x

Some variations of LMS

Applications of adaptive filters

System identification (system modeling)

e(n) = d (n) y(n)

Example of System Identification

Applications of adaptive filters

Applications of adaptive filters

Noise cancellation in cockpit

Acoustic echo cancellation

Applications of adaptive filters

Applications of adaptive filters

Narrowband interference suppression

Narrowband interference suppression

e(n) = x(n) s (n), hn (k ) = hn 1 (k ) + e(n) x(n k D), k = 0,1,...,

Interference cancellation in ECG

Performance of ECG interference cancellation

You might also like