Anda di halaman 1dari 5

2009 International Conference on Advanced Information Networking and Applications Workshops

Improved Channel Estimation Using Wavelet Denoising for OFDM and


OFDMA Systems
Wang Xue, Zhao Lin-jing and Li Jian-dong
Broadband Wireless Communications Laboratory, Information Science Institute
State Key Laboratory of Integrated Service Networks, Xidian University, Xian 710071, China

Abstract
Least Square (LS) channel estimation has been widely used in
OFDM (Orthogonal Frequency Division Multiplexing) and
OFDMA (Orthogonal Frequency Division Multiplexing
Access) systems. However, it's rather sensitive to Guassian
white noise. In this paper, we present a new algorithm which
deals with the LS estimation results through wavelet shrinkage
denoising based on Steins unbiased risk estimation (SURE)
criterion. This algorithm can effectively remove the influence
of noise in the channels and minimize the estimation risk.
Consequently, the sensitivity to noise of LS estimation is
diminished. Simulation in the scenario of IEEE802.16
downlink transmission shows that the proposed algorithm has
significant advantage over LS and modified LS estimators.
Index TermsChannel estimation, OFDM, OFDMA, SURE,
Wavelet denoising

1. Introduction
In Orthogonal Frequency Division Multiplexing (OFDM) and
Orthogonal Frequency Division Multiplexing Access
(OFDMA) systems, channel estimation is one of the most
important techniques. There are two classical pilot-based
channel estimation algorithms namely LS (Least Square) and
MMSE (Minimum Mean-square error) estimation. Since LS
estimation is simpler to implement as it doesnt need any
information about channel statistics, LS estimation has been
widely used[1].However, LS estimation is rather sensitive to
additive white Gaussian noise(AWGN), especially, when the
signal-to-noise ratio (SNR) is low, the performance will
degrade significantly. MMSE estimation is more robust against
noise and performs better than LS, but it is relatively complex
and the MMSE results would be unavailable without the prior
information about the channel.
978-0-7695-3639-2/09 $25.00 2009 IEEE
DOI 10.1109/WAINA.2009.51

129

Many modified algorithms have been proposed to overcome


the shortcomings of LS and MMSE estimations. In order to
reduce the complexity of the MMSE estimation, a low-rank
approximation estimator which uses the singular value
decomposition techniques is proposed. While many other
modified algorithms[2]-[5] are presented to improve the
performance of LS estimator. The basic principle of these
estimators is to suppress the noise existing in LS estimation
results. The Most Significant Tap(MST) algorithm in [2]
chooses the most significant taps of the channel impulse
response in time domain and those taps which contain little
information about channel will be eliminated. MST estimator
can suppress part of the noise, but it is sensitive to timing error.
Another channel estimator proposed by Kang[4]suppresses
noise by choosing a special threshold and those taps which are
lower than the threshold will be ignored. However, the
performance of Kangs estimator is determined by the choice of
threshold, when the threshold is too high, some useful
information would be lost and the performance will degrade
considerably. For MST and Kangs estimators, in order to
suppress the noise effectively, the precise maximal time delay,
as the prior information of the channel should be available.
In this paper, we proposed a channel estimation algorithm
which uses wavelet shrinkage denoising based on Steins
unbiased risk estimation (SURE), namely WDLS. In this
algorithm, LS estimator is used to obtain the channel impulse
response and then we apply wavelet shrinkage to remove the
noise contained in the LS estimation results. Since the wavelet
decomposition results of noise express themselves as
high-frequency ones while that of channel impulse information
express themselves as low-frequency ones, wavelet denoising
can be used to remove noise effectively and the performance of
LS estimator would be improved significantly in this way.
Moreover, no prior information about the channel is required in
WDLS estimation.
The remainder of this paper is organized as follows. In
section 2, system model will be introduced. Section 3 explains

the WDLS algorithm. Simulation and results are presented in


section 4.Conclusions are given in section 5.

Where L is the total number of the propagation paths, hl is


the complex impulse response of the l th path, f Di is the l th
path Doppler frequency shift, T is the sample period and l is
the l th path delay normalized by the sampling time.
At the receiver, CP is removed from ycp ( n) , and then

2. System model
OFDM is used in this system. The binary information bits are
first grouped to signal constellations according to the selected
modulation. After inserting the pilots into the information data
sequence, we obtain the sequence X (k ) , k = 0,1," , N 1 ( N is
the number of subcarrier).In the frequency domain, each
OFDM symbol is created by mapping the sequence of symbols
on the subcarriers. In the IEEE802.16, there are three classes of
subcarriers, namely data subcarriers which are used for
carrying data symbols, pilot subcarriers which are used for
carrying pilot symbols and null subcarriers which have no
power allocated to them, including DC (Direct
Carrier)subcarriers and guard subcarriers towards the edge[6].
The typical frequency domain description of an IEEE802.16
OFDM symbol is shown as Fig 1.

y (n) is sent to the Discrete Fourier Transform (DFT) block.


We obtain the signal sequence which is
Y (k ) = DFT { y (n)}
(4)
k = 0,1," N 1
= X (k ) H (k ) + W (k ),
DFT
results
of
Where H (k ) and W (k ) are
h(n) and w(n) respectively.

3.Channel estimation
According to (4), the LS estimation of the channel frequency
response can be calculated as
H LS (k ) = Y (k ) / X (k )
(5)
And the time domain channel impulse response can be obtained
from H LS (k )
h (n) = IDFT ( H (k ))
(6)
LS

LS

Since LS estimator is much sensitive to noise, we proposed


WDLS estimator, which can remove the noise contained in the
LS estimation results. The principle of WDLS estimator would
be explained in details below.

3.1.The basic principle of WDLS

Fig 1 Frequency domain representation of an IEEE802.16


OFDM symbol
Inverse Discrete Fourier Transform(IDFT) is used to
transform X (k ) into time domain signal x(n) ,
n = 0,1," , N 1 .Following IDFT block, CP(Cyclic Prefix),
which is chosen to be larger than the expected delay spread is
inserted to prevent inter-carrier interference and the resultant
samples xcp (n) can be represented by

The principle of WDLS is as follows. First, the LS estimation


result is transformed into time domain by IDFT, so we get the
time domain channel impulse response h LS .Second, the
empirical wavelet decomposition coefficient matrix d of h is
LS

obtained by applying Discrete Wavelet Transform (DWT) to


h LS .And then, according to Stein's unbiased risk estimator[7]
and Gaos non-negative garrote shrinkage function[8], the
thresholds which will be used to shrink the empirical
wavelet coefficients can be obtained. Then, we use Gaos
shrinkage function to process the empirical wavelet coefficients
d and d wd is obtained as the denoised results. Finally, d wd is
transformed back into time domain by Inverse Discrete
Wavelet Transform (IDWT).So we get the denoised time
domain channel impulse response h wd . The detailed procedure
is shown as Fig 2.

x( N + n), n = N cp , N cp + 1," , 1
xcp (n) =
(1)
n = 0,1," , N 1
x(n),
At last, the transmitted signal is sent to a frequency selective
and time varying fading channel with additive noise. The
received signal is given by
ycp (n) = xcp (n) h(n) + w(n)
(2)
Where w(n) is AWGN, h(n) is the time domain channel
impulse response which can be represented by
L 1

h(n) = hl e

j (2 / N ) f Dl Tn

(n l ),

n = 0,1," N 1

(3)

l =0

130

Kj

h LS

H
LS

k =1

h wd

d wd

Fig 2 The application of wavelet denoising in OFDM


channel estimation

NP

E (hwd (n) h(n))2

n = 0,1," , N 1

defined as j , k , k = 1, 2," , K j

j , k = ( j , d j )

(7)

Kj

i =1

i = k +1

(11)

j , which will minimize ( j ,d j )

j = j arg min
{ j ,k }

dj

(12)

( )

where j = median d j / 0.6745 [9], it is the noise variance


of the j th level.
(8)

)(

3.3. Wavelet shrink

After getting the optimal thresholds at all the levels, we


should choose a proper shrinkage function to shrink the
coefficients d j , k .There are two kinds of traditional shrinkage
functions, the hard and the soft shrinkage functions[10]. The
hard shrinkage function is unstable due to its discontinuity
while the soft shrinkage function is more stable but tend to have
bigger bias. Gaos non-negative function proposed in [8] can
overcome the drawback of the hard and soft shrinkage. The
function is represented by
0,
d j , k j

dwd ,( j , k ) =
(13)
d j , k > j
d j , k j 2 / d j , k ,

the number of the maximal decomposition level, and K j is the


total number of the coefficient of the j th level. j is the
threshold of the j th level, = [1 , 2 ," , J ] . I () is the
indicator function.
Our object is to find out the optimal j at each level, which
can minimize the l2 risk R . In order to guarantee the l2 risk R
is minimal, we should minimize R
. For simplicity, we
unbias

define equation

R unbias = J +
NP
2
2

= d j ,k

At the j th level, we choose the d j , k that minimizes j , k as

4j + 2 2j
I d j , k > j
1 + d 2j , k 2 I d j , k j +
2

d j , k

where d j , k is the j th levels k th wavelet coefficient, and J is

= K j 2k + d 2j ,i + (d 4j , k + d 2j , k ) d j ,2i

Kj



R unbias = J +
N j =1 k =1
2
J

{ }

Where hwd (n) is the expected results of WDLS, Since h(n) is


unavailable, SURE risk defined as R unbias is an unbiased
estimation of the l2 risk R when the true function is unknown.
According to SURE criterion and Gaos shrinkage
function[8], R unbias is represented as
2

j th level. When j is chosen as d j , k , the value of ( j ,d j ) is

n =1

4j + 2 2j

the optimal threshold j within a finite threshold set d j at the

According to the SURE theory, we define the minimum


mean-square error ( l2 risk) as follow
1
N

minimal. In order to get the minimal ( j ,d j ) , we search for

3.2 Calculation of thresholds based on SURE

R=

)(

(10)

>
I
d
j ,k
j

d 2j , k

According to (10), we know that if we can minimize


( j ,d j ) at each level, the unbiased estimation R unbias will be
+

H
wd

( (

( j ,d j ) = 1 + d 2j , k 2 I d j , k j

,d j )

(9)

The denoised results dwd ,( j , k ) can be obtained by shrinking all

j =1

where

the empirical wavelet coefficients d j , k using shrinkage


function (13).

131

In the simulation of WDLS, db1wavelet is chosen and the


number of the maximal decomposition level is 5.
The estimation performance is evaluated in terms of Mean
Square Error (MSE) and Bit Error Rate (BER).The comparison
of LS, MST, Kang and WDLS estimators is shown in Fig 3 and
Fig 4 respectively.
Fig 3 shows the MSE versus SNR for the LS, MST, Kang
and WDLS estimators in the SUI-1 channel which is the
standard channel for IEEE802.16d. We can see that MSE of the
WDLS estimator is 4-6dB higher than LS estimator and 2-4dB
higher than MST and Kang estimators. The MSE of MST and
Kangs estimators is roughly the same. According to SURE
criteria, the performance of WDLS estimator has been
improved without any additional prior information about
channel.

3.4.The procedure of WDLS


Step1, Transform h LS into wavelet domain and get all the
empirical wavelet coefficients d . We order all the empirical
j ,k

wavelet coefficients d j = [d j ,1 , d j ,2 ,..., d j , K j ] at the j th level by


this way d j ,1 d j ,2 " d j , K j .
Step2, According to (11) and (12), we get the optimal
thresholds which minimize the estimation risk.
Step3, After obtaining and the shrinkage function
represented by (13), we shrink all the empirical wavelet
coefficients d j , k , then we can get d wd = {dwd ,( j , k ) } which have
been shrunk.
Step4, Transform the shrunken coefficients d wd back to the
time domain by IDWT.

10

LS estimator
MST estimator
Kang estimator
WDLS estimator

10

MSE

4. Simulation result
In the simulation, the parameters are set as in Table 1.We
assume that the channel is constant over an OFDMA symbol,
but time-varying within an OFDMA frame. In this paper, we
are going to use a pilot scheme called Downlink Fully Used
Subchannelization (DL-FUSC)[6]. In DL-FUSC, the symbol is
allocated with the appropriate pilots and zeros subcarriers, and
then all the remaining subcarriers are used as data subcarriers
which will be divided into subchannels.
Table 1 Simulation parameters
Parameter

Value

FFT size

2048

Number of DC subcarriers

Guard subcarriers: Left,


Right
Number of pilot subcarriers
Pilot subcarrier index

-2

10

-3

10

m=[symbol index]mod 3

16-QAM

Cyclic prefix ratio

1/8

Fading channel

SUI-1

15

20

25

30

Fig 4 shows the BER of LS, MST, Kangs, WDLS and ideal
estimators in the SUI-1 channel. Compared to LS , MST and
Kangs estimators have better BER performance. However, the
BER performance of WDLS estimator is the best among that of
the LS, MST and Kang estimators. When the SNR is as large as
25dB, the BER performance of WDLS estimator is
approaching the performance of ideal estimator.

9k+3m+l,for k=0,,191

Modulation

10

Fig 3 Mean Square Error (MSE) performances


comparison for different estimators

192

1536

SNR

159,160

Number of data subcarriers

-1

10

132

[1]Michele Morelli,Umberto Mengli, A comparison of Pilot-Adided


channel estimation for OFDM systems,IEEE Trans on Signal
Processing,vol.49,no.12,Dec 2001,pp.3065-3073.
[2]Minn H, Bhargava V K. An investigation into time-domain
approach for OFDM channel estimation, IEEE Trans on
Broadcasting, vol.64,no.4, Dec.2000, pp.240-248
[3]Zamiri-Jafarian H, Pasupathy S. Robust and Improved Channel
Estimation Algorithm for MIMO-OFDM Systems, IEEE Trans on
Wireless Communications,vol.6,no.6,2007, pp.2106-2113.
[4]Yeon-S K, Do-Seob A, Ho-Jin L. OFDM channel estimation with
timing offset for satellite plus terrestrial multipath channels,IEEE
VTC 2006,vol.6,2006,pp.2592-2596.
[5] M. R. Raghavendra and K. Giridhar Improving channel
estimation in OFDM systems for sparse multipath channels, IEEE
Trans on Signal Processing Letters, vol.12,no.1,2005, pp.52 55.
[6] IEEE 802.16-2004,Part 16: Air interface for fixed broadband
wireless access systems, Oct. 2004.
[7] Blu T, Luisier F.The SURE-LET Approach to Image Denoising.
IEEE Trans on Image Processing, vol.16,no.11,pp.2778 2786,2007
[8]Gao H Y. Wavelet shrinkage denoising using the non-negative
garrote .J. Comput. Graph. Statist,vol.7,no.4,1998, pp.469-488.
[9]Chen Y, Han C. Adaptive wavelet threshold for image denoising.
Electronics Letters,vol.41,no.10, pp. 586-587.
[10]Donoho D L. De-noising by soft-thresholding. IEEE Trans on
Information Theory,vol.41,no.3,1995, pp. 613-627.

10

LS estimator
MST estimator
Kang estimator
WDLS estimator
Ideal estimator

-1

BER

10

-2

10

-3

10

-4

10

10

15

20

25

30

SNR

Fig 4 Bit Error Rate (BER) performances comparison for


different estimators

5. Conclusion
In this paper, we present that LS estimator is sensitive to
noise while MST and Kangs estimators can suppress part of
noise, but the maximal time delay should be known as the prior
information. In order to solve these problems, we propose
WDLS estimator. It can effectively eliminate the noise in LS
estimation results and no prior information is required in this
estimator. The simulation results show that WDLS has a better
BER and MSE performance than the LS, MST and Kangs
estimators. Compared to LS and the two modified estimators,
there is an increase in computation complexity in WDLS
estimator, however, due to the performance improvement, the
complexity increase is worthwhile.

6. Acknowledgement
This research was supported by National Science Fund for
Distinguished Young Scholars(No.60725105), National Basic
Research Program of China (No.2009CB320404), PSCIRT,
the state 863 projects(No.2007AA01Z288), ISN02080001, the
National Nature Science Foundation of China (No.60572146),
the Research Fund for the Doctoral Program of Higher
Education (No.20050701007), the Teaching and Research
Award Program for Outstanding Young Teachers in Higher
Education Institutions of MOE, P. R. China, the Key Project
of Chinese Ministry of Education (No.107103), the 111
Project(No.B08038).

7. References

133

Anda mungkin juga menyukai