Sei sulla pagina 1di 9

1

Neural Fuzzy Systems for Power Peaks reduction in


Multi-Carrier Modulations
Younes Jabrane

, Vctor P. Gil Jim enez

, Member, IEEE, Aomar Tarki

,
Ana Garca Armada

, Senior Member, IEEE, Brahim Ait Es Said

and Abdellah Ait Ouahman

Team of Telecommunications and Informatics Networks


University of Cadi Ayyad, Marrakech, Morocco

Department of Signal Theory and Communications


University Carlos III of Madrid, Spain
Email: y.jabrane,a.tarki,aitessaid,ouahman@ucam.ac.ma; vgil,agarcia@tsc.uc3m.es
AbstractIn this paper, a novel scheme for reducing the power
peaks in OFDM (Orthogonal Frequency Division Multiplexing)
and OFDMA (Orthogonal Frequency Division Multiple Access)
signals applying ANFIS (Adaptive Neural Fuzzy Inference Sys-
tems) is proposed and analyzed. Once trained, the ANFIS
synthesizes the ACE-AGP (Active Constellation Expansion -
Approximate Gradient Project) for OFDM(A) signals without
the complexity and the convergence problems of ACE-AGP; and
with similar performance. Besides, the proposed scheme obtains
signals with lower mean transmit power than ACE-AGP, which
impacts in a lower BER (Bit Error Rate). The PAPR (Peak to
Average Power Ratio) and the CM (Cubic Metric) are used for
evaluating the peak reductions.
Index TermsOFDM, OFDMA, PAPR, Cubic Metric, ANFIS,
ACE-AGP.
I. INTRODUCTION
Multi-carrier modulation such as OFDMA (Orthogonal Fre-
quency Division Multiple Access) has been advocated for LTE
(Long Term Evolution) for wireless personal communications
[1] due to its good performance in multipath environments,
exibility and the multi-user diversity [2]. While in OFDM
(Orthogonal Frequency Division Multiplexing) all the sub-
carriers are assigned to one user at a time, in OFDMA, each
sub-carrier (or group of sub-carriers) can be allocated to a
different user, i.e., in OFDM the link is time shared whereas
in OFDMA the link is frequency-and-time shared. However, as
in OFDM, the OFDMA signal exhibits a high power peaks,
which cause a loss in energy efciency due the needed of
back-off at the HPA (High Power Ampliers).
Since this problem is one of the most important drawbacks
in OFDM, there is a large number of proposals in the literature
to reduce it. Some of these proposals are valid for both OFDM
and OFDMA, whereas other are only useful for OFDM. These
proposals in the literature can be classied into two groups,
namely, distortion-less and distortion-based techniques.
In the rst group, the signal is not distorted but only encoded
or re-organized (maybe interleaved) in such a way the obtained
peaks power are minimized. Examples of these techniques are
the PTS (Partial Transmit Sequences) [3], [4], SLM (Selective
Mapping) [5], [6], CDA (Circular Data Allocation) [7], CEM
(Cross-Entropy Method) [8], BSPTM (Blind Selected Pilot
Tone Modulation) [9], TR (Tone Reservation) [10]; coding
such as Golay Sequences [11], [12] or Reed-Solomon codes
[13], and interleaving [14], [15]. The reduction in these
techniques relays on the idea that, different OFDM symbols
exhibit different power peak. While in PTS [3] the input data
block is partitioned into disjoint sub-blocks and then weighted
by a phase factor that should be optimized to obtain a nal
OFDM symbol with low power peaks, in SLM [5], the blocks
are re-ordered in such as way the obtained OFDM symbol
presents low power peaks. In order to reduce the complexity
of exhaustive search, in CDA [7], phases for SLM are selected
in a random circular fashion and thus, no side information
is needed, whereas BSPTM [9] links the index of the phase
rotation sequence in SLM to the location of the pilot tones
that are used to estimate the channel, and thus, the exibility
and reductions depend on the transmit structure. In the case
of PTS, in CEM [8], rst, a score or tness function based
on the corresponding peak power reduction performance is
dened and next, the score function is then translated into
a stochastic approximation problem which can be solved
effectively and with similar results as exhaustive search. In TR
[10], some dummy sub-carriers - do not carry information - are
optimized to obtain a low power peaks OFDM symbol. Those
techniques need send side information and thus there are data
rate loss, although there exist proposals without this constrain
such as [4] or [6], at expenses of increase in complexity
and lower reductions. Another solution is to encode input
data into a sequence that exhibit low power peaks such as
complementary Golay codes [11], [12] or Reed-Solomon [13],
with the advantage that some error correction capabilities can
be expected. However, the usefulness of these techniques is
limited since they can only be applied to M-PSK (Phase Shift
Keying) modulations with small number of sub-carriers.
If some distortion is allowed, the signal peak power can
be reduced signicantly. In this group of distortion-based
techniques can be found the AC (Amplitude Clipping) [16],
ACF (Amplitude Clipping and Filtering) [17], TI (Tone In-
jection) [10], ACE (Active Constellation Expansion) [18] or
similar approaches to ACE such as APS (Adaptive Projected
Subgradient) [19], MMSE-Threshold (Minimum Mean Square
Error) [20], SDPA (Semi-Denite Programming Algorithm)
[20] and MBSP (Metric Based Symbol Predistortion) [21],
[22]. Besides, there are also companding methods such as in
2
[23] for peak power reduction, or even applying TI, SLM or
PTS on an hexagonal modulation [24], [25].
In APS [19] the reduction is obtained by slightly disturbing
the symbols in carriers used to transmit information and by
sending dummy symbols in unused sub-carriers. The optimal
choice of the data and dummy symbols is determined by the
solution of a convex optimization problem. In order to reduce
the complexity, a modied version of the adaptive projected
subgradient method is applied. In [20] the constellation points
with lower average energy are selected from a larger set
of points, similar to TI, however, the selection algorithm is
formulated in terms of a zero-one quadratic problem and
optimized by the SDPA. Simulation results show that the
power peaks of SDPA is noticeably better than MMSE-
Threshold, while the complexity of the second is smaller.
In [21] authors proposed a simple amplitude predistortion
of a subset of the input symbol blocks which are selected
according to a metric which measures their contribution to the
output signal samples of large magnitude. This metric-based
algorithm is simple, highly exible, and can be implemented
as a one-shot process, although its performance can be further
improved by iterating the process one or more times. Next,
authors extent this algorithm with different metrics in [22].
As in voice or other signals, companding methods modify
the values of the signal without modifying its nature. It is well-
known that companding the signal using -Law or similar
yields a reduced power peaks signal without restrictions on
system parameters such as number of subcarriers, frame format
and constellation type [23].
Similar as in ACE, the use of more dense constellations
such as hexagonal modulation can be exploited for power
peaks reduction due to more degrees of freedom. So in [24]
and [25], these hexagonal modulation is used with the already
mentioned PTS, SLM and TI techniques.
For an overview of some of these solutions, reader is
referred to [26]. These techniques above mentioned achieve
the reduction at the expense of increase the transmit power,
the BER (Bit Error Rate) or the complexity, and a data rate
loss due to need of side information or coding.
Among them, the ACE is one of the most interesting
techniques since it is able to achieve large reductions, it does
not need side information and it only needs a small increase
in transmit power. Indeed, it is the base for several of the
distortion-based techniques described above. In this paper,
ACE will be used to obtain the training set for the neural
fuzzy system.
Adaptive Neural Fuzzy Inference System (ANFIS) is a well-
known tool to solve problems when the physical description of
the behavior is unknown and it is able to provide some fuzzy
(heuristics) rules that synthesize or approximate that behavior.
Examples of neural fuzzy systems are [27], [28] [29].
When dealing with the power peaks, we are only able to
calculate it after the OFDM symbol has been generated and
then, try to modify it to obtain another one with lower peaks.
By using the ANFIS, the original signal can be introduced
and then make learn the system on which output signals have
good power peak characteristics. Since ACE is able to obtain
very low power peak signals, this technique is used to obtain
the training samples.
In most of the literature, the PAPR (Peak-to-Average Power
Ratio) - also commonly denoted as PAR - is used as a measure-
ment for the power peaks, however, recently, a new metric has
been addressed for this purpose with better properties because
it does not only take into account the maximum peak, as PAPR
does, but also the others peaks into the contribution. This is
important for the evaluation of the power amplier de-rating
factor. This new measurement is the CM (Cubic Metric) [30],
[31]. Most of the above described techniques can be adapted
to use the new measurement. In this paper, both metrics have
been evaluated for the proposed ANFIS-based power peaks
reduction method, although CM has been used for the training
part of the system.
The balance of this paper is organized as follows. Section
II introduces system model and the two utilized metrics. In
Section III, the neuro fuzzy systems adapted to solve the power
peaks reduction problem are described and analyzed. Then,
the obtained results are presented and discussed in Section
IV. Finally, conclusions are drawn in Section V.
Note: Through the paper, the following notation will be
used. Bold-face symbols will be used for vectors while normal-
face for single-values. Time-domain will be denoted with non-
capitalized, whereas frequency-domain will use capitalized
ones.
II. SYSTEM MODEL
In a multi-carrier system such as OFDM, the time-domain
complex base-band transmitted signal x

for -th symbol can


be written as
x

[] =
1

=0

,
= 0 1
= 0
(1)
where is the number of sub-carriers and

is the frequency-
domain complex base-band modulated symbol on -th sub-
carrier at symbol .
The classical metric to evaluate the peak power is the PAPR,
dened as
(x) = 10 log
10
(
x
2

(x
2
2
)
)
(2)
where () denotes the expectation, and x

and x
2
represent the -norm and the 2-norm of x, respectively.
However, this metric only accounts for the maximum peak
and therefore, it mights do not have appropriately into account
the distortion effect due to the non linear response of the High
Power Amplier. As an example, it can be found that, although
signals in Figure 1 would exhibit similar PAPR (9.61 and 9.62
dB respectively), if the amplier clips the signal at certain
threshold shown in the gure, 1 will experiment larger
degradation than 2. For this reason, other metrics have
been recently proposed such as the Cubic Metric [30], [31]
or [32]. The CM uses higher statistics to evaluate the power
de-rating factor in an HPA, and it is dened by the Third
Generation Partnership Project (3GPP) as
=
()

()

(3)
3
0 20 40 60 80 100
0
50
100
150
200
250
300
Time Index
A
m
p
l
i
t
u
d
e


Signal 1
Signal 2
Clipping Threshold
Figure 1. Time-domain signals example
where is the Raw Cubic Metric, and for a signal x is
dened as
= 20 log
10

{(x)}
)
3

, (4)

is the reference RCM which for multi-carrier mod-


ulation is 1.52 dB, and is an empirical slope factor that is
used to complete the estimate of the power de-rating required
to accomplish a target ACLR (Adjacent Channel Leakage
Ratio). This value for multi-carrier is 1.56. The CM for signals
in Figure 1 are 6.75 and 4.9 dB, respectively. Here in this
example, it can be observed the large difference between both
signal metrics, which correspond with the large difference in
behavior at the HPA. Since the global performance will be
obtained after going through the HPA, the CM accounts better
for this behavior and give more insight on what is the expected
performance. In this paper, the two metrics have been used for
comparison.
As anticipated, the ACE method offers some advantages
such as it does not need side information, obtains large reduc-
tions and can be used for different metrics. In the following
the ACE - AGP (Approximate Gradient-Project) will be briey
described.
The ACE method modies and expands the constellation
points within an allowable region which does not affect the
demodulation slicer (see Figure 2 for an example of allowed
expansion regions (shaded) for QPSK), and thus, it does not
need side information. By using this new degree of freedom,
multi-carrier signals with arbitrarily low power peaks can
be obtained. In [18] different algorithms to achieve PAPR
reduction are provided, namely, POCS (Project onto Convex
Set), AGP and SGP (Smart Gradient-Project). In this paper,
the AGP method will be used since the utility of ACE is to
obtain the training samples, and therefore the convergence is
more important than speed or maximum reduction.
Lets [] be the frequency-domain version from [], i.e.,
its FFT, and the clipping amplitude value. The algorithm
1.5 1 0.5 0 0.5 1 1.5
1.5
1
0.5
0
0.5
1
1.5
In Phase (I)
I
n

Q
u
a
d
r
a
t
u
r
e

(
Q
)
1

2
(1 + j) 1

2
(1 + j)
1

2
(1 j)
1

2
(1 j)
Figure 2. Active Constellation Expansion for QPSK
proceeds as follows [18]:
1) Initialization: x
0
= x. = 0. accounts for the iteration
index.
2) Clip any

[] and form
[] =
{

[] ,

[]

[]
,

[] >
, (5)
where

[] =

[]
[]
.
3) Compute the clipped signal:

[] = []

[].
4) Calculate the FFT of c
clip
to obtain C
clip
.
5) Maintain the values of C
clip
when they are valid point
extensions on the constellation and set to 0 when not.
Apply an IFFT to obtain c.
6) Determine the step size according to some criterion
and compute new version of the time-domain signal
x
i+1
= x
i
+ c.
7) Calculate the PAPR or the CM for the new signal. If
acceptable, stop the algorithm and return x
i+1
as output,
otherwise, increase and go to step 2. Iterate until target
is accomplished or a maximum number of iterations is
reached.
The complexity of this algorithm is ( log ) due to the
FFT and IFFT operations, however, it is able to obtain rather
large reductions with enough iterations. These signals will be
used for training our neural fuzzy system with the CM as the
measurement for the stop criteria. The idea is to make learn
the neural fuzzy system which signals exhibit low power peaks
and how to obtain them.
III. NEURO-FUZZY SYSTEMS
The fusion of Adaptive Neural Networks (ANN) and Fuzzy
Inference system (FIS) has attracted a growing interest of
researchers in various scientic and engineering areas due to
the growing need for adaptative intelligent systems to solve
real world problems. In this paper we use the typical fuzzy
rule in a Sugeno fuzzy model [33], which has the format:
= (, )
4

1 2 3 4


Figure 3. Corresponding ANFIS architecture
where and are fuzzy sets in the antecedent and =
(, ) is a crisp function in the consequent. We consider the
rst-order Sugeno fuzzy inference system which contain two
rules:

1

1

1
=
1
+
1
+
1

2

2

2
=
2
+
2
+
2
The ring strengths
1
and
2
are usually obtained as the
product of the membership grades in the premise part, and the
output is the weighted average of each rules output.
{

1
=
1
+
1
+
1
=

1

1
+
2

1
+
2

2
=
2
+
2
+
2
=
1

1
+
2

2
(6)
To facilitate the learning of the Sugeno fuzzy model, it is
convenient to put the fuzzy model into framework of adapta-
tive networks that can compute gradient vectors systematically.
The resultant network architecture, called ANFIS is shown in
Figure 3, where node within the same layer performs functions
of the same type, as detailled bellow (

denotes the output


of the -th node in -th layer) [33]:
Layer 1: Each node in this layer generates a mem-
bership grades of linguistic label. For instance, the
node function of the -th node may be generalized
gauss membership function

() = exp
(

)
2
2
2

)
(7)
where is the input to node ,

is the linguistic
label (small, large, etc.) associated with this node,

and
2

are, respectively, the center and variance


corresponding to gauss membership function.
Layer 2: Each node in this layer calculates the ring
strength of a rule via multiplication:

()

(), = 1, 2 (8)
Layer 3: Node in this layer calculates the ratio of
the rules ring strength to total of all ring
stregths:

1
+
2
, = 1, 2 (9)
Layer 4: Node in this layer compute the contribu-
tion of rule toward the overall output, with the
following node function:

) (10)

[]
/ X x

x
x

Figure 4. Integrated Time-domain Fuzzy Models

and

Where

is the output of layer 3, and

is
the parameter set. Parameters in this layer will be
referred to as the consequent parameters.
Layer 5: The single node in this layer computes the
overall output as the summation of contribution from
each rule:

= =

(11)
Using a given input/output data set, ANFIS constructs a
FIS (Fuzzy Inference System), whose membership function
parameters are tuned using either a backpropagation algorithm
alone, or in combination with a least squares type of method.
This allows the fuzzy systems to learn from the data they are
modeling. These techniques provide a method for the fuzzy
modeling procedure to learn information about a data set, in
order to compute the membership function parameters that best
allow the associated fuzzy inference system to track the given
input/output data. From the ANFIS architecture in Figure 3,
it is observed that given the values of premise parameters, the
overall output can be expressed as linear combinations of
the consequent parameters
=
1

1
+
2

2
=
1
(
1
+
1
+
1
) +
2
(
2
+
2
+
2
) . (12)
Since, the reduction of power peaks using the ACE-AGP al-
gorithm undergoes a high complexity of parameters included,
and an important time to get the reduced value of PAPR [18],
in this paper, we use the ANFISs proposed in Figures 4 and
9, which are trained by data obtained from the ACE-AGP
algorithm. The training is made using 70% of the data while
30% is used for testing and validation.
5
Figure 5. Constellation after ACE-AGP and proposed Time-domain Fuzzy
and Time-Frequency-domain fuzzy
A. Time-domain Fuzzy
This model is based on ACE-AGP algorithm by using
the time-domain signals as a training data. The process is
enumerated in the following:
1) Split into two groups the time-domain original data x,
namely, the training set and the test set.
2) Introduce the original time-domain data x in the training
set into the ACE-AGP algorithm to obtain x

.
3) Decompose into real and imaginary part the original
data x (x

, x

), and ACE-AGP output x

(x

,
x

).
4) Obtain x

and x

by training the two models

and

with the pairs [x

, x

] and
[x

, x

] in the training set, respectively.


5) Test with the values of x in the test set to validate the
models

and

.
After training and testing the two neural fuzzy models (one for
the real part and the other for the imaginary part), the models
are ready to be used. In Figure 4, the proposed time-domain
neural fuzzy scheme for power peak reduction is shown.
As it will be seen in section IV, the power peaks reduction
metrics of the proposed time-domain neural fuzzy system
are similar to ACE-AGP, with the advantage of reduction in
complexity (it does not need several FFT and iterations to
obtain the nal signal).
Next, in Figure 5, the constellation obtained by ACE-AGP
and the proposed time-domain fuzzy are presented. It can
be seen that, although similar performance, constellations are
different. In fact, some constellations points in Figure 5 are
in not allowable regions, which will decrease the system
BER performance, as it can be observed in Figure 6. In this
gure, all the schemes (original, ACE-AGP and proposed
time-domain and time-frequency-domain fuzzy) have been
obtained for the same SNR. It can be seen that proposed time-
domain neural fuzzy outperforms ACE-AGP for low SNR but
not for high SNR. And moreover, both methods (ACE-AGP
and proposed time-domain fuzzy) obtain worse performance
than original. The reasons are the following. Since all the
systems have been simulated at the same SNR, the noise power
6 7 8 9 10 11 12 13 14
10
7
10
6
10
5
10
4
10
3
10
2
10
1
BER comparison. N = 1024
SNR (dB)
B
E
R


Original
AGP
T Fuzzy
TF Fuzzy
Figure 6. BER performance for ACE-AGP and proposed Time-domain Fuzzy
and Time-Frequency-domain Fuzzy
0.5 1 1.5 2 2.5 3 3.5 4
0.5
1
1.5
2
2.5
3
3.5
4
Real Axis
I
m
a
g

A
x
i
s
Time Domain Fuzzy


Probability (dB)
60
55
50
45
40
35
30
25
20
15
10
Original
Figure 7. 2D histogram for proposed TF
will be dened by the transmitted power. If transmitted power
increases, so the noise power. Thus, for non-extended points in
the constellation (70% of total in ACE-AGP), the noise power
will be larger than in the original scenario and so, the BER
increases. The mean constellation energy for original signal is
1, whereas for ACE-AGP is 1.29 and proposed TF is 1.2. For
low SNR regime, since the proposed TF concentrates more
the energy (as it can be seen in the 2D-histogram in Figure 7)
than ACE-AGP (Figure 8), the performance is better in terms
of BER. However, in high SNR regime, the points into the
not-allowable region (20% of total) degrade the performance
for proposed TF.
In order to overcome these problems, a second scheme is
proposed in the following.
B. Time-Frequency-domain Fuzzy
The main problem with the time-domain training scheme is
that the neural fuzzy is not able to learn which regions in the
constellation are allowed and which ones are forbidden. Thus,
a second ANFIS working on frequency-domain is proposed.
The procedure is as follows:
6
0.5 1 1.5 2 2.5 3 3.5 4
0.5
1
1.5
2
2.5
3
3.5
4
Real Axis
I
m
a
g

A
x
i
s
AGP


Probability (dB)
60
50
40
30
20
10
Original
Figure 8. 2D histogram for ACE-AGP
1) Apply on output of

and

to obtain
X

.
2) Separate the four constellation regions in order to train
eight ANFISs (two sets): 1st set concerning real parts
and the 2nd set concerning the imaginary parts, as it can
be seen in Figure 9.
3) Train the second set of ANFIS by (X

) to generate

,1
,

,2
,

,3
and

,4
for
each quadrant.
4) Train the rst set of ANFIS by (X

) to generate

,1
,

,2
,

,3
and

,4
for
each quadrant.
5) Test with the values of X in the test set to validate the
models.
Once the neural fuzzy models have been trained, the signal
is easily obtained by taking the original time-domain data x
and introduce them into the two time-domain models

and

. Next, apply an FFT to obtain the frequency-


domain signal, separate into the four quadrants and intro-
duce into the frequency-domain models
,1
,
,2
,

,3
,
,4
and
,1
,
,2
,
,3
,

,4
, to obtain the nal frequency-domain signal X

.
This way, the obtained signal ts the AGP constraints, as it
can be seen in Figure 5, where the constellations for the ACE-
AGP and the proposed time-frequency-domain fuzzy system
are displayed.
In Figure 10, the 2D-histogram for the proposed time-
frequency-domain fuzzy system constellation is shown. It can
be seen that this proposal concentrates more than ACE-AGP
energy, indeed, the average energy constellation is 1.21, and
moreover, it is constrained to the allowable regions.
C. Analysis of Mean Square Error
Since we use the gradient-based learning algorithm, we
consider the following learning cost function for the error
=
1
2

=1
(

)
2
(13)

x
x

{
X

{
X

}

X

,1
X

,2
X

,3
X

,4
X

,1
X

,2
X

,3
X

,4
Figure 9. Integrated frequential Models

1
,

2
,

3
and

4
for real and imaginary part
where is the number of the output (in our case = 1),

and

are the desired and the current output values of the


network, respectively.
Since the time-domain AGP signal (desired signal) is a
truncated Gaussian (to amplitude threshold ) and so the
obtained by neural fuzzy models, the Probability Density
Function (PDF) of the mean square error will be

() =
1
2
1

exp
(

)
2
(
2

+
2

)
)

2
(

+
2

)
(14)
where

are the means of AGP and (time-


domain or time-frequency-domain) neural fuzzy systems, re-
spectively, and

and

are the variances for AGP


and (time-domain or time-frequency-domain) neural fuzzy sys-
tems, respectively. And thus, the cumulative density function
7
0.5 1 1.5 2 2.5 3 3.5 4
0.5
1
1.5
2
2.5
3
3.5
4
Real Axis
I
m
a
g

A
x
i
s
TimeFrequency Domain Fuzzy


Probability (dB)
60
55
50
45
40
35
30
25
20
15
10
Original
Figure 10. 2D histogram for proposed Time-Frequency-Domain Fuzzy
for MSE can be analytically evaluated by
( ) = erf

+
2

erf

+
2

,
where the error function (erf) is dened as
erf() =
2

2
. (15)
On the other hand, the linear parameters

and

, with
= 1, 2, ...., ( = 2 and = 8 for time-domain system and
time-frequency-domain system, respectively), and the parame-
ters

of the membership functions ( = 1, 2, ..., , where


is the number of fuzzy rules) of the neuro-fuzzy structure
are adjusted by using the following formulas

( + 1) =

()

( + 1) =

()

(16)

( + 1) =

()

( + 1) =

()

(17)
where is the learning rate. Since = 1, we have removed
the sub-index in the following for clarity purposes.
Evaluating the partial derivatives in eq. (16),

(18)
where
=

=1

=1

=1

=
, (19)
yields

= (

=1

. (20)
And

(21)
where

=1

= 1
, (22)
yields

= (

=1

. (23)
By applying similar procedure to eq. (17) we obtain that

(24)
where

=1

(25)
yields

= (

=1

. (26)
And

(27)
where

=1

()
2

(28)
yields

= (

=1

)
2

. (29)
By using equations (20), (23), (26) and (29), eqs. (16) and
(17) are obtained.
IV. RESULTS
The results have been obtained by using Monte Carlo
simulations with 2000 randomly generated OFDM symbols for
= 256, 512 and 1024 sub-carriers, with QPSK modulation.
For the ACE-AGP, the maximum number of iterations was
xed to be 2000, whereas for the PTS and SLM, the number
of phases and blocks was 2(1, 1) and 8, respectively. Both
metrics, PAPR (gure 11) and CM (gure 12) have been
evaluated for all the systems.
In gures 11 and 12, several aspects can be observed. The
rst one is that the loss in performance of the proposed time-
frequency-domain neural fuzzy scheme with respect the time-
domain neural fuzzy system is negligible in terms of PAPR
reduction and below 0.3 dB in terms of CM. The second is
that both proposed neural fuzzy schemes outperform the PTS
8
3 4 5 6 7 8 9 10 11 12
10
3
10
2
10
1
10
0
PAPR comparison
(dB)
P
r
o
b

(
P
A
P
R

>

)


Original
AGP
T Fuzzy
TF Fuzzy
PTS
SLM
N = 256
N = 512
N = 1024
Figure 11. Comparison of PAPR reduction results using QPSK with N=256,
512 and 1024
1 2 3 4 5 6 7 8
10
4
10
3
10
2
10
1
10
0
Cubic Metric comparison
(dB)
P
r
o
b

(
C
M

>

)


Original
AGP
T Fuzzy
TF Fuzzy
PTS
SLM
N = 256
N = 512
N = 1024
Figure 12. Comparison of CM reduction results using QPSK with N=256,
512 and 1024
and SLM algorithm, and, moreover, without the complexity
and convergence time of PTS or SLM. The third one is that
the performance loss, in terms of PAPR, of proposed neural
fuzzy models with respect the ACE-AGP is less than 0.7 dB
for = 256 sub-carriers and 1.1 dB for = 1024, although
without the complexity and convergence problems of ACE-
AGP.
In gure 6, the BER for the proposed neural fuzzy systems,
the original signal and the ACE-AGP is presented. It can
be seen that the time-frequency-domain ANFIS outperforms
the other and is close to the original system. The reason
is similar to the time-domain system. As it can be seen in
gure 10, the proposed time-frequency-domain neural fuzzy
model concentrates even more than ACE-AGP the energy and
moreover, most of the constellation points are moved from the
original position, and thus, only a few of points will experience
effective lower SNR.
Finally, in Figure 13, the analytical MSE for values of

= 2.72 10
5
,

= 2.98 10
5
,

=
1.3 10
3
,

= 1.2 10
3
and = 7.3 10
2
is shown.
It can be observed in this gure that the MSE is very small.
10
9
10
8
10
7
10
6
10
5
10
4
10
3
10
3
10
2
10
1
10
0

P
r
o
b
(
E

)
Figure 13. CCDF for the MSE for values

= 2.7210
5
,

=
2.98 10
5
,

= 1.3 10
3
,

= 1.2 10
3
and = 7.3 10
2
V. CONCLUSION
In this paper, a two novel methods for power peaks reduc-
tion in OFDM and OFDMA signals based on ACE-AGP by
applying neural fuzzy models, have been proposed, described
and analyzed. It has been shown that the proposed time-
domain neural fuzzy system, although simpler than time-
frequency-domain one, offered good performance in terms of
peak power reduction but not in terms of BER. On the other
hand, the proposed time-frequency-domain neural fuzzy model
offered similar reductions as to TF but higher performance
in terms of BER, at expenses of a small increase in system
complexity. Both proposals can be applied to OFDM and
OFDMA systems because they are independent of input signal.
It has been shown that both proposals outperform other already
proposed schemes for peak reduction.
ACKNOWLEDGMENT
This work has been partly funded by projects MULTI-
ADAPTIVE (TEC2008-06327-C03-02) and AECI Program of
Research Cooperation with Morocco.
REFERENCES
[1] 3rd Generation Partnership Project (3GPP), UTRA-UTRAN Long
Term Evolution (LTE) and 3GPP System Architecture Evolution
(SAE), http://www.3gpp.org/Highlights/LTE/LTE.htm, Accessed on 2
July 2009.
[2] E. G. Larsson, On the combination of spatial diversity and multiuser
diversity, IEEE Communications Letters, vol. 8, no. 8, pp. 517519,
Aug. 2004.
[3] L. Jr and N. Sollenberger, Peak-to-average power ratio reduction of an
OFDM signal using partial transmit sequences, IEEE Communications
Letters, vol. 4, no. 3, pp. 86 88, Mar. 2000.
[4] A. Jayalath and C. Tellambura, SLM and PTS peak-power reduction
of OFDM signals without side information, IEEE trans. on Wireless
Communications, vol. 4, no. 5, pp. 2006 2013, Sep. 2005.
[5] R. W. Bauml, R. F. H. Fisher, and J. B. Huber, Reducing the peak-to-
average power ratio of multicarrier modulation by selected mapping,
IEE Electronics Letters, vol. 32, no. 22, pp. 2056 2057, Oct. 1996.
[6] M. Breiling, S. H. Muller-Weinfurtner, and J. B. Huber, Slm peak-
power reduction without explicit side information, IEEE Communica-
tions Letters, vol. 5, no. 6, pp. 239241, Jun. 2001.
[7] S. Abedi, Novel PAPR Reduction Technique based on Circular Data
Allocation, in International Conference on Digital Telecommunications,
vol. 0. Los Alamitos, CA, USA: IEEE Computer Society, 2006, p. 35.
9
[8] J.-C. Chen, Partial transmit sequences for peak-to-average power ratio
reduction of ofdm signals with the cross-entropy method, IEEE trans.
on Signal Processing, vol. 16, no. 6, pp. 545 548, Jun. 2009.
[9] N. Chen and G. T. Zhou, Peak-to-average power ratio reduction in
ofdm with blind selected pilot tone modulation, IEEE trans. on Wireless
Communications, vol. 5, no. 8, pp. 2210 2216, Aug. 2006.
[10] J. Tellado, Peak-to-average power reduction, Ph.D. dissertation, Stand-
ford University, Sep. 1999.
[11] J. A. Davis and J. Jedwab, Peak-to-mean power control in OFDM,
golay complementary sequences and reed-muller codes, IEEE trans.
on Information Theory, vol. 45, no. 7, pp. 2397 2417, Nov. 1999.
[12] V. P. G. Jim enez, M. J. F.-G. Garca, M. P. S. Fern andez, and A. G.
Armada, Efcient Implementation of Complementary Golay Sequences
for PAR Reduction and Forward Error Correction in OFDM-based
WLAN systems, AE

U - International Journal on Electronics and
Communications, vol. 62, no. 9, pp. 683 694, Oct. 2008.
[13] R. F. H. Fischer and C. Siegl, Reed-solomon and simplex codes
for peak-to-average power ratio reduction in ofdm, IEEE trans. on
Information Theory, vol. 55, no. 4, pp. 1519 1528, Apr. 2009.
[14] P. V. Eetvelt, G. Wade, and M. Tomlinson, Peak to average power
reduction for ofdm schemes by selective scrambling, IEE Electronics
Letters, vol. 32, no. 21, pp. 1963 1964, Oct. 1996.
[15] A. D. S. Jayalath and C. Tellambura, Reducing the peak-to-average
power ratio of orthogonal frequency division multiplexing signal through
bit or symbol interleaving, IEE Electronics Letters, vol. 36, no. 13, pp.
1161 1163, Jun. 2000.
[16] D. Wilich, N. Dinur, and A. Glinowiecki, Level clipped high-order
OFDM, IEEE trans. on Communications, vol. 48, no. 6, pp. 928930,
Jun. 2000.
[17] X. Li and L. J. Cimini, Effect of clipping and ltering on the
performance of ofdm, IEEE Communications Letters, vol. 2, no. 5,
pp. 131 133, May 1998.
[18] B. S. Krongold and D. L. Jones, Par reduction in ofdm via active
constellation extention, IEEE trans. on Broadcasting, vol. 49, no. 3,
pp. 258 268, Sep. 2003.
[19] R. L. G. Cavalcante and I. Yamada, A exible peak-to-average power
ratio reduction scheme for ofdm systems by the adaptive projected
subgradient method, IEEE trans. on Signal Processing, vol. 57, no. 4,
pp. 1456 1468, Apr. 2009.
[20] F. Kohandani and A. K. Khandani, A new algorithm for peak/average
power reduction in OFDM systems, IEEE trans. on Broadcasting,
vol. 54, no. 1, pp. 159 165, Mar. 2008.
[21] S. Sezginer and H. Sari, OFDM peak power reduction with simple
amplitude predistortion, IEEE Communications Letters, vol. 10, no. 2,
pp. 65 67, Feb. 2006.
[22] , Metric-based symbol predistortion techniques for peak power
reduction in OFDM systems, IEEE trans. on Wireless Communications,
vol. 6, no. 7, pp. 2622 2629, Jul. 2007.
[23] S. A. Aburakhia, E. F. Badran, and D. A. E. Mohamed, Linear
companding transform for the reduction of peak-to-average power ratio
of OFDM signals, IEEE trans. on Broadcasting, vol. 55, no. 1, pp. 155
160, Mar. 2009.
[24] S. H. Han, J. M. Ciof, and J. H. Lee, On the use of hexagonal
constellation for peak-to-average power ratio reduction of an OFDM
signal, IEEE trans. on Wireless Communications, vol. 7, no. 3, pp. 781
786, Mar. 2008.
[25] , Tone injection with hexagonal constellation for peak-to-average
power ratio reduction in ofdm, IEEE Communications Letters, vol. 10,
no. 9, pp. 646 648, Sep. 2006.
[26] S. H. Han and J. H. Lee, An overview of peak-to-average power
ratio reduction techniques for multicarrier transmission, IEEE Wireless
Communications, vol. 12, no. 2, pp. 56 65, Apr. 2005.
[27] K. R. Lo and C. J. Chang, A neural fuzzy resource manager for
hierarchical cellular systems supporting multimedia services, IEEE
trans. on Vehicular Technology, vol. 52, no. 5, pp. 11961206, Sep.
2003.
[28] J. Bas and A. P. Neira, Fuzzy adaptive signal predistorter for ofdm
systems, in International Conference on Digital Telecommunications,
vol. 4. Tampere, Finlandia: EUSIPCO, Septiembre, 2000.
[29] J. S. R. Jang, Adaptive-network-based fuzzy inference systems, IEEE
Trans. on Systems, Man and Cybernetics, vol. 23, no. 3, pp. 665685,
1993.
[30] TDocs R4-040367, R1-040522, and R1-040642, Comparison of PAR
and Cubic Metric for power de-rating, 3GPP TSG RAN WG1 and
3GPP TSG RAN WG4, Tech. Rep., May 2004.
[31] TDoc R1-060023, Cubic Metric in 3GPP-LTE, 3GPP TSG RAN WG1,
Tech. Rep., Jan. 2006.
[32] A. R. S. Bahai, M. Singh, A. J. Goldsmith, and B. R. Saltzberg, A
new approach for evaluating clipping distortion in multicarrier systems,
IEEE Journal on Selected Areas in Communications (JSAC), vol. 20,
no. 5, pp. 1037 1046, Jun. 2002.
[33] L. E. Aik, S. Yogan, and O. Jayakumar, A study of neuro-fuzzy system
in approximation-based problems, MATEMATIKA, vol. 23, no. 2, pp.
113130, 2008.

Potrebbero piacerti anche