Download as pptx, pdf, or txt
Download as pptx, pdf, or txt
You are on page 1of 19

Project Presentation

on
SYSTEM IDENTIFICATION USING ADAPTIVE
EXPONENTIAL FUNCTIONAL LINK NETWORK-BASED

Click to edit
NONLINEAR FILTERS
PRESENTED

Master title style


BY
ABHYANAND TIWARI (1804213)
ANKITA KUMARI (1804223)
SHUBHAM SAURAV(1804268)
CLICK TO EDIT MASTER SUBTITLE STYLE

UNDER THE SUPERVISION OF


ASST. PROF B.N.RAO

SCHOOL OF ELECTRONICS ENGINEERING


KALINGA INSTITUTE OF INDUSTRIAL TECHNOLOGY (KIIT)
UNIVERSITY
BHUBANESWAR-751024
APRIL 2022
Outline

1. Introduction
2. Functional Link Network
3. Nonlinear Filter
4. Non-Linear Adaptive Filter
5. Adaptive Signal Processing
6. Least Mean Square Algorithm
7. System Identification Using FLN
Filter
8. Adaptive Exponential FLN
9. Results and Simulation
10. Conclusion
 Adaptive filters are used as models to
estimate the characteristics of an unknown
system.there is an unknown system called
Introduction ‘plant’, that is to be identified or model it
by finding its transfer function
 The system is modeled by an FIR filter
with M adjustable coefficients. Both the
plant and model are excited by the same
input sequence x(n). If y(n) denotes the
output of the plant and y^(n) denotes the
output of the model then

 The error sequence e(n)=y(n)-y^(n) ;


n=0,1….. And select the coefficients h(k)
to minimize the squared error

 By solving the above equation, we obtain


the filter coefficients for the model.
Functional link network
 Functional link network provide significant
importance as it offers less computational
complexity.functional link is basically a
functional operator that allows to represent a
input pattern.
 Functional link network is mainly used for the
function recognition and for pattern recognition.
 It mainly works on the non linear modeling
capabilities of functional link and the filtering
properties of the given algorithm
 Functional link network has found application in
many areas such as channel equalisation and
echo cancellation.
 Functional link network has flexible
architecture thats why the functional link
network provide the best and the effective way
to model non linearities which affect the speech
signal.
Nonlinear Filter

A nonlinear filter is a filter whose output is not a linear


function of its input,as they are not based on linear
relationships between an input and an output
Non-linear filters have many applications, especially in
the removal of certain types of noise that are not
additive.
Nonlinear filters have been widely used and appreciated
in active noise control (ANC).also they are based on
functional link networks (FLNs) have received significant
attention owing to its single layer structure,which offers
lesser complexity.
For nonlinear filters,the filter output or response of the
filter does not obey the principles outlined
earlier,particular scaling and shift invariance.
Non-Linear Adaptive Filter

Some of the techniques available to model nonlinear


systems using nonlinear adaptive systems.

The nonrecursive polynomial model


The recursive polynomial model
The multilayer perceptron (MLP) neural network
The radial basis function (RBF) neural network
The Volterra LMS algorithm is presented for a second-
order series and Nth-order filter.and reduces the
computational complexity to an acceptable level for some
applications and also simplifies the derivations.
The RLS algorithms are known to achieve fast convergence
. The objective of the RLS algorithm is to choose the
coefficients of the adaptive filter such that the output
signal y(k), during the period of observation, will match
the desired signal as closely as possible in the least-
squares sense.
ADAPTIVE SIGNAL PROCESSING

An adaptive filter is no more than a digital filter, which can adjust its
characteristics. It adapts to changes in its input signals automatically
according to a given algorithm. The algorithm will vary the coefficients
according to a given criteria, typically an error signal to improve it's
performance. In essence an adaptive filter is a digital filter combined with
an adaptive algorithm, which is used to modify the coefficients of the
filter. Adaptive filters are used in many diverse applications in today's
world for example telephone echo canceling, radar signal processing,
equalization of communication channels and biomedical signal
enhancement. Adaptive filters are useful · when the characteristics of a
given filter must be variable · when the spectrum of a given signal
overlaps with the noise spectrum · if the frequency band occupied by noise
is unknown or may vary with time. In most real world scenarios adaptive
filters are realized.using Finite Impulse Response (FIR) filters, since they
are guaranteed to be stable and are simple to use
LEAST MEAN SQUARE ALGORITHM

LMS stands for least mean square which was developed by


Widrow and Hoff in 1960 and it was the first linear adptive
filtering algorithm for the problem such as prediction.there
are many features of LMS algorithm such as it is simple to
code,robust in nature with respect to external disturbance
and it also has linear computational complexity with
respect to adjustable parameters.lms algorithm consist of
the two major processes that is filtering process and
adptation process.
SYSTEM identification deals with the
process of estimating the
mathematical model of a system
based on measured observations.
Nonlinear filters based on functional
link networks (FLNs) have received
SYSTEM significant attention owing to its
single layer structure, which offers
IDENTIFICATIO lesser computational complexity and
N USING FLN due to the simple learning rule . FLN-
based nonlinear system identification
FILTER by using FLN-based filters has been
attempted in and FLN filters have
found applications in other areas of
nonlinear filtering including channel
equalization and echo cancellation.
ADAPTIVE EXPONENTIAL FLN
FLN is a popular nonlinear filter. In an FLN, the functionally
expanded input signal is multiplied with a set of adaptive weights
and then added together to obtain the output signal. Assuming x(n)
as the input signal, in a TFLN,
the functionally expanded signal vector may be written as

f(n)= {1, x(n), sin [πx(n)] , cos [πx(n)] ,. . ., sin [Bπx(n)] ,


cos[Bπx(n)], x(n−1), sin [πx(n−1)],cos[πx(n −1)], . . . ,
x(n−N+1),sin[πx(n−N+1)],cos[πx(n−N+1)] , . . . ,
sin [Bπx(n − N + 1)] , cos [Bπx(n − N + 1)]}T
where B is the order of the functional expansion.
The output of the TFLN may be written as

y(n) = wf0(n) + wf1(n)x(n) + wf2(n)sin[πx(n)] + wf3(n)


cos[πx(n)] + · · ·wf(M−2)(n)sin[Bπx(n)] + wf(M−1)
(n)cos[Bπx(n)]

with wf (n) = [wf0(n), wf1(n), . . . , wf(M−1)(n)]T as the


adaptive weight vector of the TFLN. The weights of the
TFLN are usually updated using a gradient descent approach
as wf (n + 1) = wf (n) + μte(n)f(n)
where e(n) is the error signal and μt is the learning rate.
SIMULATION RESULTS
CONCLUSION

A new nonlinear filter, which is based on an exponentially decaying


trigonometric functional expansion has been designed in this project.
The performance of the filter has been further
enhanced by incorporating an adaptive exponential factor. The
effectiveness of the proposed filter has been tested using an extensive
simulation study. In the simulations, the new filter has been shown to
provide improved system identification accuracy. In a similar manner,
an enhancement in noise cancellation has also been obtained in an
ANC scenario.
MATLAB program
MATLAB program:

clc;

no_of_independent_trials = 10;

zeta = 0.1;

zeta_a = 0.1;

for itr=1:no_of_independent_trials

clc;

disp(['Independent Trial No: ',num2str(itr)])

no_of_inputs = 2e3;

input=rand(1,no_of_inputs) - 0.5;

desired_output = zeros(1,no_of_inputs);

N=10;

AEFLN_order = 1;

input_buffer= zeros(1,N);

M = (2*AEFLN_order+1)*N + 1;

AEFLN_weights= zeros(1,M);

mu_weight= 0.05;

mu_a= 0.1;

noise = awgn(input,30)-input;

a(1)= 0;

x_tap = zeros(1,4);
for i=1:length(input)

input_buffer=[input(i) input_buffer(1:end-1)];

x_tap = [input(i) x_tap(1:end-1)];

desired_output(i) = (0.6 * (sin(pi*x_tap(1)))^3) - (2/((x_tap(1))^3 + 2)) -


(0.1*cos(4*pi*x_tap(4))) + 1.125 + noise(i);

AE_FEB=[];

for k =1:N

for l =1:AEFLN_order

AE_FEB=[AE_FEB, exp(-1*a(i)*abs(input_buffer(k)))*sin(pi*l*input_buffer(k)),
exp(-1*a(i)*abs(input_buffer(k)))*cos(pi*l*input_buffer(k))];

end

end

AEFLN_after_FEB= [1,input_buffer,AE_FEB];

AEFLN_output(i)= AEFLN_weights * AEFLN_after_FEB';

error(i)= desired_output(i) - AEFLN_output(i);

z=[];

for k =1:N

for l =1:AEFLN_order

z=[z, -1*abs(input_buffer(k))*exp(-
1*a(i)*abs(input_buffer(k)))*sin(pi*l*input_buffer(k)), -1*abs(input_buffer(k))*exp(-
1*a(i)*abs(input_buffer(k)))*cos(pi*l*input_buffer(k))];

end

end

z_final= [0,zeros(1,N),z];

normalized_a = (z_final * AEFLN_weights')^2 + zeta_a;

a(i+1) = a(i) + ((mu_a * error(i) * z_final * AEFLN_weights')/normalized_a);

normalized_weight = (AEFLN_after_FEB*AEFLN_after_FEB')+zeta;

AEFLN_weights= AEFLN_weights + ((mu_weight * error(i) *


AEFLN_after_FEB)/normalized_weight);
end

err(itr,:)=error.^2;

end

disp('Please Wait! Smoothing Operation is Going On...')

length_of_smoothing_filter = 200;

smoothing_filter_coeff =
(1/length_of_smoothing_filter)*ones(1,length_of_smoothing_filter);

for i=1:no_of_independent_trials

err_smooth(i,:) = filter(smoothing_filter_coeff,1,err(i,:));

end

figure

plot(10*log10(mean(err_smooth)));

xlabel('Iterations');

ylabel('MSE (dB)');

grid on;

title('Nonlinear System Identification using Normalizedc AEFLN (Case-II)');

AEFLN_mse=10*log10(mean(mean(err(end-1000:end))));

fprintf('Average MSE Value over the last 1000 iterations is %f', AEFLN_mse);

xlabel('Iterations');

ylabel('MSE (dB)');

grid on;

title('Nonlinear System Identification using AEFLN Filter');


SIMULATION_RESULTS
Thank You

You might also like