International Journal of Information and Mathematical Sciences 1:4 2005

Neuro-Fuzzy System for Equalization

Channel Distortion
Rahib H. Abiyev

and numerical information the fuzzy adaptive filters are

Abstract— In this paper the application of neuro-fuzzy system for developed. The use of such approach improves the adaptation
equalization of channel distortion is considered. The structure and speed.
operation algorithm of neuro-fuzzy equalizer are described. The use In some case the construction of fuzzy rules for equalizers is
of neuro-fuzzy equalizer in digital signal transmission allows to difficult. In this case one of the effective technologies for
decrease training time of parameters and decrease the complexity of construction of equalizer’s knowledge base is the use of neural
the network. The simulation of neuro-fuzzy equalizer is performed.
The obtained result satisfies the efficiency of application of neuro-
network. In this paper the channel equalization by using
fuzzy technology in channel equalization. neuro-fuzzy network is considered. The use of neuro-fuzzy
technology allows to use small number of parameters, fast and
Keywords— neuro-fuzzy system, noise equalization, neuro-fuzzy easy train equalizer. The equalizer based on neural networks
equalizer, neural system doesn’t need appropriate knowledge about channel dynamics.
These equalizers give better results in bit error rate (BNR), at
I. INTRODUCTION the cost of computational strength.

S IGNAL transmitted through a channel suffers from linear,

nonlinear and additive distortion. The conventional
method for compensation of channel distortion is based on
II. NEURO-FUZZY INFERENCE SYSTEM
The structure of neuro-fuzzy inference system is given in
introducing of the linear equalizer (linear inverse filter to the figure 1. The input signals applied to the network at time k are
channel frequency response) to the output of the channel. This xi(k) (i=1..N) and output signal of the network are u(k) . N is
number of neurons in the input layer.
design methodology is appropriate when the channel model is
In first layer the number of nodes is equal to the number of
precisely known and characteristics of the channel are not
external inputs. In second layer each node corresponds to one
changed with time. When channel has time-varying
linguistic term. For each input signal entering the system the
characteristics the adaptive equalizers [1] are used. A various membership degree to which input value belongs to a fuzzy
approaches have been used for nonlinear channel equalization. set is calculated. To describe linguistic terms the Gaussian
Classical approaches based on knowledge of the parametric membership function is used.
channel model [2]. Next type is decision feedback equalizer
that improves the performance of equalizer. Nowadays neural ( xi  c1ij ) 2

networks are widely used for channel equalization [3-9]. One V 1ij2
of the class nonlinear adaptive equalizer is based on P1 j ( x i ) e i=1..n, j=1..J (1)
multiplayer perceptions (MLP) and radial basis functions
(RBF). The MLP equalizers require long training and are c1ij and V 1ij are the center and width of the Gaussian
sensitive to the initial choice of network parameters. RBF
equalizers are simple and require less time for training, but membership function of the jth term of ith input variable,
usually require a large number of centers, which increase the respectively. P1j(xi) is membership function of i-th input
variable for j-th term. n is number of external input signals. J
complexity of computation.
is number of linguistic terms assigned for external input
One of effective way for development of adaptive
signals xi, P is number of linguistic terms assigned for one-,
equalizers for nonlinear channels is the use of fuzzy
two-,…, D- delayed output signal of network.
technology in their development. In [10] for adaptive In the third layer the number of nodes corresponds to the
equalization of channel distortion the fuzzy equalizer is number of rules. Each node represents one fuzzy logic rule.
proposed. Human experts determine the fuzzy rules using Here to calculate the values of output signals of the layer AND
input-output data pairs of the channel. These rules are used to (min) operation is used.
construct the filter for nonlinear channel. The RLS and LMS
algorithms are applied to change parameters of the member-
ship functions of rules and develop equalizers. Using linguistic
Pl – P1
j
j ( xi ) , l=1..L, j=1..J+P, i=1,..,n (2)

Manuscript received July, 2003. This work was supported by the Near East Here 3 is min operation. Pl signals are input signals for the
University, Lefkosha, TRNC, Turkey. next layer. This layer is a consequent layer. In this fourth
Rahib H. Abiyev is with the Department of Computer Engineering, Near
East University, Mersin-10, TRNC, Turkey (e-mail: [email protected]).

229
 International Journal of Information and Mathematical Sciences 1:4 2005

c111
x1 R1
c 21
c112
u1
R2 c 22
:
u

:
x2 ‘ u2
c2 L

c12l Rl

Figure 1. Structure of neuro-fuzzy inference system

layer the output signals of previous layer are multiplied to the proper values of parameters of neuro-fuzzy system are
weight coefficients of network and in the next layer sum of determined.
these products is calculated. Weight coefficients of neuro- The unknown parameters of neuro-fuzzy system are c2 l
fuzzy system are represented by fuzzy set of output variables.
parameters of last layer and membership functions of the first
They are described by Gaussian function. If as a
layer. Using gradient method the values of unknown
defuzzification operation we use “center of average” then only
parameters of last layer are corrected as
the center of Gaussian function is used in the next layer for
defuzzification. In this case during development of control
system the width of Gaussian function is not used. In formula wE
c 2 l (t  1) c 2 l (t )  J (6)
(3) the parameters c2 l will represent the center of fuzzy wc 2 l
coefficients. Outputs of fifth layer are calculated as
wE
L here J is learning rate. is determined as
wc 2 l
u1 ¦P
l 1
l * c 2 l , l=1..L (3)

L wE wE wu
and u 2 ¦P l
(u (t )  u d (t ))
wc 2 wu wc 2 (7)
l 1

Using the values of calculated variables the output of the The adjusting of the membership functions of input layer is
fuzzy system is determined. carried out by correction unknown coefficients c1ij and V 1ij .
L The following formulas can be used for learning these

u1
¦ P l * c2 l
l 1
coefficients.
u (4)
u2 L wE
¦P c1ij (t ) c1ij (t )  J ,
l 1
l
wc1ij
(8)
wE
After calculating output signal of neuro-fuzzy network the V 1ij (t ) V 1ij (t )  J
value of error is determined. Using value of error the learning wV 1ij
of the parameters of network start.
where
1 m
E ¦
2i1
(u i (t )  u id (t )) 2 (5)
wE wE wu wP l
¦ wu wP ,
wc1ij j l wc1ij
where u i (t ) and u id are current and desired outputs of the (9)
wE wE wu wP
system, m is number of outputs.
wV 1ij
¦j wu wP wV 1l
For learning network parameters the gradient and least l ij
square methods are iteratively used. As a result of learning the

230
 International Journal of Information and Mathematical Sciences 1:4 2005

n(k)

x (k ) +
s(k) Channel
6 z-1 z-2 ... z-m
+
x(k) x(k-1) x(k-2) x(k-m)

Equalizer

s (k )
Figure.1. Signal transmitting scheme

Here transmitted signals are given to the channel input. The additive
L noise is added to the transmitted signal. In the equalizer-using
wE wu
¦ c2 l u target transmitted signal the deviation from the current
d
u (t )  u (t ), l 1
(10) network output is determined. This error is used to adjust
L
wu wP l equalizer parameter. Training is continued until the value of
¦P
l 1
l error for all training sequence of signals would be minimum.
During simulation we use the following channel model.

x(k ) 0.3482y(k) 0.8704y(k- 1)  0.3482y(k- 2)  n(k)

­ 2( x j  c ji )
°P l ( x j ) if j node (12)
° V 2ji A channel input, the transmitted data sequence s(k) is
wP l ( x j ) ° assumed to be independent sequence taking values from {-
® is connected to rule node l
wc ji °0, 1,1} with equal probability. In the channel we add additive
otherwise distortion as white Gaussian noise.
°
° During simulation the four inputs of neural equalizer are
¯
x(k), x(k-1), x(k-2), x(k-3). The three layer neuro-fuzzy
­ 2( x j  c ji ) 2 network structure is taken. In the hidden layer number of
P
° l j( x ) if j node neurons are 27.The training of neuro-fuzzy network equalizer
° V 3ji
wP l ( x j ) ° (NFNE) have been carried out. In figure 2 the error plot of
® is connected to rule node l training result of NFNE is given. The figure 3 illustrates the
wV ji °0, otherwise curve that describes the performance (Bit error rate versus
° (11)
signal-noise ratio) analysis of neuro-fuzzy network equalizer
°
¯ (NFNE) for channel (12). The obtained result satisfies the
efficiency of application of neuro-fuzzy technology in channel
equalization.
Using (6) - (11) the corrections of the parameters of neuro-
fuzzy system are carried out. IV. CONCLUSION
III. SIMULATION The development of neural-fuzzy equalizer for
equalization channel distortion is carried out. The operation
Structure of digital signal transmission system is shown in principle and learning algorithm of neuro-fuzzy network are
figure 1. The random binary input signals s(k) are transmitted represented. The construction of neuro-fuzzy equalizer is
through communication channel. s(k) are input known carried out for the channel in presence of additive distortion.
samples with an equal probability of being –1 and 1. These The obtained results of simulation of neuro-fuzzy equalizer
signals are corrupted by additive noise n(k). The corrupted are given.
signals are given to the equalizer. The problem of the channel
equalization is to estimate the input signals using the sequence REFERENCES
of {x(k), x(k-1), x(k-2),…,x(k-m)}.
[1] Qureshi, S.U.H. “Adaptive equalization”. Proc.IEEE, 1985, 73, (9),
The neuro-fuzzy network structure and its training algorithm pp.1349-1387.
is used to design equalizer for equalization of noise. The [2] D.D.Falconer, “Adaptive Equalization of Channel Nonlinearites in QAM
inputs of neuro-fuzzy equalizer are channel output signals. Data Transmission Systems”, Bell System Technical Journal, 1978,
The output of neural network is the recovered transmitted vol.27, no.7.
sequence of signal. During equalizer design the sequence of

231
 International Journal of Information and Mathematical Sciences 1:4 2005

0
0 10 20 30 40
-0.2
-0.4

log10(BER)
-0.6
-0.8
-1
-1.2
-1.4
SNR

Figure 2. Error plot Figure 3. Performance of NFNE equalizer

[3] Chen, S., Gibson, G.J., Cowan, C.F.N., and Grant, P.M. “Adaptive
equalization of finite non-linear channels using multiplayer
perceptrons”. Signal Process, 1990, 20, (2), pp.107-119.
[4] Chen, S., Gibson, G.J., Cowan, C.F.N., and Grant, P.M. Reconstruction of
binary signals using an adaptive radial-basis function equalizer”, Signal
Processing, 1991,22, (1), pp.77-93
[5] Chen, S., Mclaughlin, S. and Mulgrew, B. Complex valued radial based
function network, Part II: application to digital communications channel
equalization”, Signal Processing, 1994,36, pp.175-188.
[6] M.Peng, C.L.Nikias, J.G.Proakis, “Adaptive Equalization for PAM and
QAM Signals with Neural Networks”, in Proc. Of 25th Asilomar Conf.
On Signals, Systems & Computers, 1991,vol.1, pp.496-500.
[7] M.Peng, C.L.Nikias and J.Proakis, “Adaptive equalization with neural
networks: new multiplayer perceptron structure and their evaluation”
Proc.IEEE Int. Conf.Acoust., Speech, Signal Proc., 1992, vol.II,(San
Francisco,CA), pp.301-304.
[8] J.S.Lee, C.D.Beach and N.Tepedelenlioglu, “Channel equalization using
radial basis function neural network” Proc.IEEE Int. Conf.Acoust.,
Speech, Signal Proc., 1996, vol.III, (Atlanta, GA),pp.1719-1722.
[9] Deniz Erdogmus, Deniz Rende, Jose C. Principe, Tan F. Wong.
Nonlinear channel equalization using multiplayer perceptrons with
information-theoretic criterion. In Proc. Of 2001 IEEE Signal
Processing Society Workshop, 2001, pp.443-451.
[10] Li-Xin Wang, Jerry M. Mendel. Fuzzy adaptive filters, with application
to Nonlinear Channel Equalization. IEEE Transaction on Fuzzy
Systems, 1993, vol.1, No.3.
[11] Rahib Abiyev, Fakhreddin Mamedov. Channel equalization using
recurrent neural network// First International Conference on Soft
Computing and Computing with Words in System Analysis, Decision
and Control (Dedicated to Professor Lotfi Zadeh), ICSCCW-2001,
Antalya, Turkey, June 6-8,2001,pp.327-330.

Rahib Hidayat Abiyev was born in Azerbaijan, in 1966. He received Ph.D

degree in Electrical and Electronic Engineering from Azerbaijan State Oil
Academy (old USSR). He worked research assistant at the research laboratory
“Industrial intellectual control systems” of Computer-aided control system
department at Oil Academy for six years. From 1999-present he is working as
associate professor at the department of Computer Engineering of Near East
University, TRNC, Turkey. He is vice chairman of Computer Engineering
Department. He has published more than sixty papers in related fields. He is
member of IJCI, IJSP and IJIT from 2003 and member of IEEE. His research
interests are Intellectual Control Systems, Fuzzy Systems, Neural Networks,
Genetic Algorithms, Chaos Theory, Signal Processing, Pattern Recognition,
Optimization.

232