Sie sind auf Seite 1von 5

OFDM Channel Estimation in the Presence of Frequency Offset and Phase

Noise1

Songping Wu, Yeheskel Bar-Ness


Center of Communications and Signal Processing Research
ECE Department, New Jersey Institute of Technology
Newark, NJ 07102, USA
Email: {sw4, yeheskel.barness}@njit.edu

Abstract – In this paper, we consider OFDM channel estimation estimate random phase noise, frequency offset estimation can
in the presence of frequency offset and phase noise. In the be achieved by using pilot signals [5][6]. As these methods
literatures, most channel estimation methods assume perfect cause loss of bandwidth efficiency, non-pilot-aided frequency
frequency synchronization and the knowledge of channel statistics. offset estimation has be used [7]-[10]. The cyclic prefix (CP)
Phase noise and residual frequency offset cause inter-carrier
based method, initially proposed in [9], is quite attractive
interference (ICI), which consequently impairs the accuracy of
channel estimation. The lack of knowledge of channel statistics can among non-pilot-aided approaches due to its simplicity.
make channel estimation much harder. To resolve these problems, Nevertheless, the accuracy of the CP-based method could not
we propose with the aid of cyclic prefix (CP) based frequency be guaranteed for multipath fading channels. Later, as proposed
offset estimation statistics-independent channel estimation. We in [10], the method of [9] was improved by considering the
iteratively search for the most likely channel impulse response channel impulse response (CIR) length. The proposed method
(CIR) length, and use it not only for the optimum compensation of in [10], however, is not feasible in cases when the CIR length is
frequency offset, but also for finding the optimum window to filter unknown.
the least square (LS) channel estimate which further suppress the Furthermore, channel estimation is a very important issue for
effects of ICI and noise. The proposed scheme is compared with
OFDM systems. Blind channel estimation is a desirable
conventional methods for both non-interpolation and interpolation
cases2. Numerical results are presented to illustrate the approach as it does not require pilot signals. It does require,
effectiveness of the proposed scheme. however, a large amount of data and thus higher computational
complexity. With perfect frequency synchronization (without
residual frequency offset), different pilot-symbol-aided channel
I. INTRODUCTION estimation methods can be applied in OFDM [11]-[14]. The
maximum likelihood/least square (ML/LS) estimators of [11]
Orthogonal frequency division multiplexing (OFDM) is a and [12] can readily be implemented without knowing channel
bandwidth efficient transmission technique which provides statistics. The minimum mean square error (MMSE) estimators
high bandwidth efficiency and is quite effective in handling in [12]-[14], however, are more robust against noise and
time dispersion of multipath fading channels. It has been perform better than the ML/LS estimators. Nevertheless, its
chosen as the transmission method of many standards in wire dependence on channel statistics and the operating signal to
and wireless communications, such as Digital Subscribe Line noise ratio (SNR) makes it disadvantageous. Despite its
(DSL), European Digital Audio and Video Broadcasting robustness against mismatch [13][14], when there is no a priori
(DAB/DVB), IEEE 802.11a and European HIPERLAN/2 for knowledge of channel statistics and the operating SNR, the
wireless local area network (WLAN) etc.. performance inevitably degrades. Without the assumption of
Based on multi-carrier modulation [1], OFDM has symbol perfect frequency synchronization, the performance may further
period long enough to eliminate inter-symbol interference (ISI) degrade due to frequency offset and phase noise.
caused by time dispersive channels. Nevertheless, the multi- In this paper, we consider statistics-independent channel
carrier modulation is also sensitive to frequency offset and estimation in the presence of frequency offset and phase noise.
phase noise. Frequency offset and phase noise cause loss of As a function of the CIR length, the LS channel estimate results,
orthogonality among subcarriers and consequently introduce which is based on the CP-based frequency offset estimation and
inter-carrier interference (ICI). The effect of phase noise has compensation, are used to search for the CIR length iteratively.
been analyzed in many papers [2]-[4]. Many approaches have The minimization of channel estimation errors leads to the most
also been proposed to analyze, estimate and compensate likely CIR length, which is then used to optimize frequency
frequency offset [2][5]-[10]. Though it is impossible to offset estimate, and filter the LS channel estimate reducing its
---------------------------------------- - sensitivity to noise and ICI. Thus better performance is
1
This work was partially supported by NSF under Grant No. CCR- achieved.
9903381 and Grant No. CCR-0085846 and by the NJ Communications The paper is organized as follows. The OFDM system model is
Center for Wireless Telecommunications. introduced in Section II. Section III presents and analyzes the
2
In non-interpolation case, samples within some symbol are all pilots; proposed frequency offset and channel estimation scheme.
in interpolation case, pilot signals are inserted into data stream.
Section IV provides the numerical results to illustrate the
effectiveness of the proposed scheme. The paper is concluded
in Section V.

0-7803-7802-4/03/$17.00 © 2003 IEEE


3366
Standard notations are used in this paper. Bold upper/lower where r = [rm (0) rm (1) L rm ( N − 1)]
T
,
case letters are used for matrices/vectors. Other notations are as
z = [ zm (0) zm (1) L zm ( N − 1) ]
T
follows: , and
( )H --- Hermitian
h = [hm (0) hm (1) L hm ( N − 1)] are all N × 1 vectors.
T

( )T --- Transpose
⊗ --- convolution (
X = diag [ xm (0) xm (1) ... xm ( N − 1)] .
T
)
--- 2-norm  I m (0) I m (1) L I m ( N − 1) 
 I (−1) I m ( N − 2) 
F{ } --- Discrete Fourier Transform (DFT) I m (0) L
and P= m are
angle( ) --- The angle of the complex variable  L L O M 
 
diag (a) --- The diagonal matrix using vector a  I m (1 − N ) L L I m (0) 
as its diagonal entries N × N matrices.
The frequency offset and phase noise in P affects the accuracy
II. OFDM SYSTEM MODEL of channel estimation. We cannot measure phase noise, but
frequency offset can be estimated and compensated to reduce
The basic principle of OFDM is to divide each data symbol into its effects on channel estimation. The effects of phase noise and
N samples (subcarriers). The length N discrete Fourier residual frequency offset (due to estimation errors) can possibly
transform (DFT) is applied to those samples and a cyclic prefix be suppressed by filtering channel estimate. For perfect
(CP) is added to eliminate ISI. Data is recovered at the receiver frequency and phase synchronization, P reduces to identity
in reverse order. We define the length of CP as N g and the matrix and therefore the performance of channel estimation can
be guaranteed.
length of CIR as L , and further assume that CIR is finite and
its length is less than that of CP, i.e., L< N g . D
F Difference
At OFDM receiver, following the DFT and due to the presence T LS Channel
Computation
Estimator
of frequency offset and phase noise, the received kth sample of Frequency (eq. (7))
(eq. (11))

the mth symbol in frequency domain can be expressed by Offset Frequency


m( N + N g )+ n Estimation Offset
 j 2π ( ε +φm ( n ))  (eq. (5))
+ ξ m (n)  (1)
compensation
rm (k ) = F  sm (n) ⊗ g m (n) ⋅ e N

 
No yes
where sm (n) , g m (n) and φm (n) denote the transmitted signal, the
CIR and phase noise, respectively. ξ m (n ) indicates the AWGN Output
CIR length
Time Criteria
CIR Length domain Satisfied?
noise. ε is the normalized frequency offset3. We assume Updating
and
channel
window (ineq. (12) &
(eq. (9)) (13))
ε ≤ 0.5 and the 3 dB linewidth of phase noise is much less response

than frequency offset. Equivalently, (1) can be given by Fig.1. Proposed scheme for frequency offset and channel estimation
rm (k ) = xm (k )hm (k ) I m (0)
N −1
(2) III. FREQUENCY OFFSET AND CHANNEL ESTIMATORS
+ ∑ xm (l )hm (l ) I m (l − k ) + zm (k )
l =0
l≠k In the presence of frequency offset and phase noise, both offset
where xm (k ) , hm (k ) and zm (k ) are the corresponding and channel response should be estimated to guarantee good
receiver performance. Phase noise variance is assumed to be
frequency domain expressions of sm (n) , g m (n) and ξ m (n)
much less than unity.
respectively. I m (i) is a function of ε and φm (n) , given by: We propose a new scheme with which, by iteratively searching
j 2π  m ( N + N g )ε / N  N −1 for the most likely CIR length and using it for both frequency
e j 2π [n ( i + ε ) / N +φm ( n )]
I m (i ) =
N
∑ e (3) offset and channel estimation, performance is greatly improved
n=0 in comparison with conventional approaches. The proposed
where i = 0,..., N − 1 . From (2) together with (3), frequency scheme is shown in Fig.1.
offset and phase noise cause the common phase error (CPE)
and introduce inter-carrier interference (ICI) as well. For the A. CP-based Frequency Offset Estimator
mth symbol, Representing (2) by matrix yields CP-based frequency offset estimator in [9] is quite simple and
r = PXh + z (4) bandwidth efficient, but it does not consider the effects of
multipath fading and estimation results may not be accurate as
it is based on the CP of one symbol only. The method proposed
in [10] improves that of [9] by considering CIR length and
3 taking more symbols into consideration. However, when
The normalization of frequency offset means the ratio of the
frequency offset to the OFDM subcarrier spacing.
averaging frequency offset estimates obtained separately from

3367
each symbol, accumulated errors may be larger than expected.
Moreover, its dependence on the CIR length is quite a problem (
B p = diag (b p ) = diag bp (0) bp (1) ... bp ( N − 1) 
T
) is
when channel statistics is not available. A different method is an N × N diagonal matrix defined by the window function
thus proposed in this paper to solve these problems. Like [10],

several symbols are used to estimate frequency offset, but it doe 1 i ∈ [0, p]
not accumulate errors by using the following expression for 
estimation   (i − p )π   2(i − p)π 
bm (i ) = 0.42 + 0.5cos   + 0.08 cos   (9)
1  1 M −1 −1  p  p
εˆ( p) =

angle  ∑ ∑
 N g ⋅ M m = 0 k =− N + p ( rm* (k )rm (k + N ) )  (5)


 i ∈ [ p, 2 p ]
  
 g  
where p is the CIR length which is unknown, M is the number 0 i ∈ [2 p + 1, N − 1]
of symbols used for averaging. The unknown parameter p (as Note that rectangular window is not used here as it introduces
we will show later) can be set initially to one and be found by more high frequency components than is tolerated which causes
iteration. Therefore, we can still get the accurate estimate of (5) a distortion of channel frequency response. Instead, due to its
even without channel statistics. excellent descending properties, Blackman function is used in
(9), as the intermediate part of the designed window.
B. Channel Estimator
Channel estimation is quite crucial for OFDM systems. Also as C. The Most Likely CIR Length and Final Solution
stated earlier, LS method is advantageous over MMSE method The most likely CIR length can be found by minimizing the
due to its simplicity and independence of channel statistics. cost function
Hence assuming in this paper unknown channel statistics, we 2
will focus on LS method. h − hlsp (10)
The estimate of (5) can be used to compensate for the To simplify the process, the window function is not used during
frequency offset, after which, we will get from (4) that the search, and we only have to find the proper p that produces
)
rp = Pp′ Xh + z (6) the frequency offset estimate minimizing (10). Unfortunately,
where Pp′ takes the same form of P except that I (i ) is replaced there are two unknown parameters, h and ε in (10), which
makes such direct minimization difficult.
by However, we notice that, in the absence of AWGN noise, as p
1 N −1 j 2π ( m ( N + N g ) + n )(ε −εˆ ( p )) / N + 2π ni / N +φm ( n ) increases, frequency offset estimate of (5) becomes more
I ′p (i ) = ∑ e  (6a)
N n=0 accurate and the difference of channel estimates of (7) for
As shown in Fig. 1, channel estimate is easily obtained by using adjacent p’s values becomes smaller, and minimum when p is
the LS method, which can be expressed by greater than or equal to CIR length. Therefore, the
) minimization of (10) can be obtained by the first occurrence of
h lsp = X −1rp (7)
the minimum of
The LS method is quite sensitive to interference and noise. 2
Therefore without perfect frequency and phase synchronization, hlsp − h lsp −1 (11)
the effects of frequency offset and phase noise become worse. In the presence of AWGN noise, we have to assert that the
Furthermore, there would still be residual frequency offset even value p that minimizes (10) is the same as that of (11) before
after compensation, which, together with phase noise, we can use (11). Statistically the minimum of (11) would occur
introduces CPE and ICI. Though CPE might partially be when p is close to the CIR length when noise is not so high. (11)
compensated by channel estimation itself, ICI will definitely decreases when p increases from 1 to the CIR length since the
affect the accuracy of such estimation. Therefore, some method CIR effects decreases. For increasing p, which is equivalent to
must be introduced to reduce the sensitivity of channel using fewer samples (see (5)), the frequency offset estimation
estimation to interference and noise. becomes less accurate and so does the channel estimate. Thus
As CIR has a finite length in time domain, the response beyond when p becomes greater than CIR length but less than CP, the
this CIR length is thus due to ICI and noise. Hence, a window difference of (11) is statistically higher when p is greater than
function may be used to filter out these effects of ICI and noise the CIR length than when p is close to the CIR length. Hence,
on channel estimates. In time-domain, using a window function the minimum of (11) occurs with high probability at the point
on (7) yields where p is equal to the CIR length. Hence, the most likely CIR
)
hlsp = WB p W H h lsp (8) length can be found by varying p between 1 and N g , and
1 1 L 1  choosing the value which satisfies the following criteria
1 − j 2π / N − j 2π ( N −1) / N  2 2
1  e L e  h lsp − hlsp −1 ≤ h lsp −1 − h lsp − 2 (12)
where W= is
N L L O M  and
 − j 2π ( N −1) / N − j 2π ( N −1)( N −1) / N  2 2
 1 e L e  h lsp − hlsp −1 ≤ h lsp +1 − h lsp (13)
an N × N Fourier transform matrix, and

3368
To examine the effectiveness of the criteria, we resort to percent of total possibilities, which indicates the effectiveness
computer simulation. The final channel estimate is expressed of CIR length searching method. Note that, since the AWGN
by noise affects the searching process and we use the exponential
) )
hlsp) = WB p) W H X −1rp) (14) power delay profile with the maximum delay spread of 6Ts , the
) search result in between 5-7 is quite reasonable.
where p represents the estimated value of p. Note that CIR
With the estimated CIR length obtained, the receiver
length can found with only a single search as in most cases it
performance is shown in Fig. 4-5, where conventional method
does not change even in a time variant channel and the result
is designated by LMMSE+FOE and the perfect case indicates
might be used for quite a few OFDM symbols.
OFDM signal reception with perfect frequency synchronization
and the LMMSE channel estimator. Note that, in order to meet
D. Interpolated Pilot Symbols
The aforementioned channel estimator is for non-interpolation the Nyquist sampling theorem, D f must be less than N /(2 L )
case. However, interpolation case is often used where pilot to guarantee the estimation accuracy for the interpolation case.
signals are multiplexed into the transmitted data stream, i.e., The proposed scheme outperforms the conventional method
pilot signals are inserted into data stream every D f samples. and approaches the perfect situation for both non-interpolation
and interpolation cases. It is shown that the proposed scheme
Without loss of generality, we assume that K = N / D f is integer,
successfully eliminates the effect of phase noise and frequency
i.e., there are K pilot samples per symbol. In this case, the offset by approaching the perfect case as SNR increases, while
principle of the proposed scheme remains correct, except that conventional method exhibits an error floor.
the size of DFT matrix W and the window diagonal
matrix B p become K × K diagonal matrices. The searching V. CONCLUSION
process for p remains the same, but the interpolation must be
applied to the result of (8) to get the complete channel estimate. In this paper, we proposed a new statistics-independent channel
estimation scheme in the presence of frequency offset and
IV. NUMERICAL RESULTS phase noise. By searching for the most likely CIR length, we
optimize the frequency offset estimation result. Based on the
The proposed scheme was evaluated by simulation. Part of searched CIR length, a time domain window is designed to
simulation parameters is based on IEEE 802.11a standard e.g., suppress noise as well as ICI caused by phase noise and
DFT length, CP length and sample period Ts are 64, 16 and residual frequency offset. By this means, the excellent
performance is achieved. This scheme approaches the
0.05µ s , respectively. 3dB linewidth of phase noise equals
performance of the LMMSE channel estimator in the absence
0.1% of subcarrier spacing. The actual frequency offset ε is of frequency offset and phase noise while outperforms the
set to 0.1382. Number of symbols used to estimate frequency conventional methods for frequency offset estimation and
offset M equals 8. Exponential Rayleigh fading channel is used channel estimation.
with the exponential power delay profile specified by
eτ / τ rms REFERENCES
, where τ rms , Ts and L are the mean delay
(
τ rms 1 − e )
LTs / τ rms
[1] J.A.C Bingham, “Multicarrier modulation for data transmission:
spread, sample period and CIR length respectively. τ rms is set an idea for whose time has come,” IEEE Commun., Mag., no.28,
pp.5-14, 1990.
to 0.05µ s , which equals Ts . L is set to 6. There are 16 symbols [2] T. Pollet, M.V. Bladel and M. Moeneclaey, “BER sensitivity of
per packet. The total energy of CIR has been normalized to one. OFDM systems to carrier frequency offset and Wiener phase
Channel changes independently from symbol to symbol, but noise,” IEEE Trans. Commun., vol. 43, no. 2, pp. 191-193, Feb.
remains static within a symbol. 1995.
16QAM is used to examine our scheme. The proposed scheme [3] S. Wu and Y. Bar-Ness, “Performance analysis on the effect of
is compared with the frequency offset estimator of [10] plus the phase noise in OFDM systems,”, to be presented in ISSSTA’02,
Prague, Czech, Sep. 2002.
LMMSE channel estimator of [13] (which is termed
[4] P. Robertson and S. Kaiser, “Analysis of the effects of phase
conventional method) for both non-interpolation and noise in orthogonal frequency division multiplexing (OFDM)
interpolation cases. For fair comparison, the frequency offset systems,” ICC'95, Seattle, vol. 3, pp. 1652-1657, 1995.
estimator uses the first M symbols of each packet with [5] P.H. Moose, “A technique for orthogonal frequency division
unknown CIR length. Simulation results are shown in Fig. 2-5. multiplexing frequency offset correction,” IEEE Trans.
As can be seen from Fig. 2, the proposed scheme performs Commun., vol. 42, pp. 2908-2914, Oct. 1994.
quite well in estimating frequency offset. For SNR ≥ 5dB , the [6] T.M. Schmidl and D. C. Cox, “Robust frequency and timing
synchronization for OFDM,” IEEE Trans. on Commun., vol. 45,
mean square error of the estimation is of the order of 10 −3 or no.12, pp. 1613-1621, Dec, 1997.
less. The accurate frequency offset estimate also reflects the [7] M. A. Visser and Y. Bar-Ness, “OFDM frequency offset
fact that the proposed scheme is quite successful in searching correction using an adaptive decorrelator,” Proceedings of the
for the CIR length. From Fig. 3, the most likely CIR length is 5 32nd Annual Conference on Information Science and Systems,
and an estimated length between 5 and 7 accounts for over 80 Princeton, NJ, pp. 483-488, Mar. 1998.

3369
[8] M. A. Visser, P. Zong and Y. Bar-Ness, “A novel method for
0
blind frequency offset correction in OFDM Systems,” PIMRC'98, 10
Perfect
Boston, MA, pp. 816-820, Sep. 1998. LMMSE+FOE
[9] F. Daffara and O. Adami, “A new frequency detector for Proposed
orthogonal multicarrier transmission techniques,” IEEE 45th -1
10
Vehicular Technology Conf., Chicago, IL, pp. 804-809, Jul.1995.
[10] X. Ma, G. B. Giannakis and S. Barbarossa, “Non-data-aided

Word Error Rate


frequency-offset and channel estimation in OFDM and related
block transmissions,” ICC’ 2001, pp. 1866-1870, Helsinki, -2
10
Finland, Jun. 2001.
[11] R. Negi and J. Cioffi, “Pilot tone selection for channel
estimation in a mobile OFDM system,” IEEE Trans. Consum.
Electron., vol. 44, pp. 1122-1128, Aug. 1998. -3
10
[12] M. Morelli and U. Mengali, “A comparison of pilot-aided
channel estimation methods for OFDM systems,” IEEE Trans.
on Signal Processing, vol. 49, pp. 3065-3073, Dec. 2001. -4
[13] O. Edfors, M. Sandell, J. V. D. Beek, S. Kate and P. O. 10
5 10 15 20 25 30 35 40
Borjesson, “OFDM channel estimation by singular value SNR, dB
decomposition,” IEEE Trans. Commun., vol. 46, no. 7, pp.931- Fig.4. Proposed scheme vs. conventional methods for non-
939, Jul. 1998.
interpolation case, D f = 1
[14] Y. Li, L. J. Cimini Jr., and N. R. Sollenberger, “Robust channel
estimation for OFDM systems with rapid dispersive fading 0
10
channels,” IEEE Trans. Commun., vol. 46, pp. 902-915, Jul. Perfect
1998. LMMSE+FOE
Proposed
0.1386
Df=1,Frequency offset estimation
Df=2,Frequency offset estimation -1
10
0.1384 Df=4,Frequency offset estimation
Word Error Rate

Real Frequency offset


Normalized frequency offset

0.1382

o o
0.138 -2
Df=2
10

Df=4
0.1378

0.1376
-3
10
5 10 15 20 25 30 35 40
0.1374 SNR, dB

Fig.5. Proposed scheme vs. conventional methods for interpolation


0.1372
5 10 15 20 25 30 35 40 case
SNR, dB

Fig.2. Proposed frequency offset estimation results

0.4

Df=1
0.35
Df=2
Df=4
0.3

0.25
Percentage

0.2

0.15

0.1

0.05

0
0 2 4 6 8 10 12 14 16
CIR Length

Fig.3. Estimated CIR length distribution over 10000 Monte Carlo trials,
SNR=30dB

3370

Das könnte Ihnen auch gefallen