Sei sulla pagina 1di 6

J Shanghai Univ (Engl Ed), 2009, 13(6): 438–443

Digital Object Identifier(DOI): 10.1007/s11741-009-0604-3

Particle filter for joint frequency offset and channel estimation in

ty
MIMO-OFDM systems

i
ZHANG Jing (张 静)1,2 , LUO Han-wen (罗汉文)1 , JIN Rong-hong (金荣洪)1

rs
1. Department of Electronic Engineering, Shanghai Jiao Tong University, Shanghai 200240, P. R. China
2. Department of Communication Engineering, Shanghai Normal University, Shanghai 200234, P. R. China

e
Abstract A particle filter is proposed to perform joint estimation of the carrier frequency offset (CFO) and the channel in
multiple-input multiple-output orthogonal frequency division multiplexing (MIMO-OFDM) wireless communication systems.

iv
It marginalizes out the channel parameters from the sampling space in sequential importance sampling (SIS), and propagates
them with the Kalman filter. Then the importance weights of the CFO particles are evaluated according to the imaginary
part of the error between measurement and estimation. The varieties of particles are maintained by sequential importance

Un
resampling (SIR). Simulation results demonstrate this algorithm can estimate the CFO and the channel parameters with high
accuracy. At the same time, some robustness is kept when the channel model has small variations.

Keywords multiple-input multiple-output (MIMO), orthogonal frequency division multiplexing (OFDM), frequency offset,
channel estimation, sequential Monte Carlo, particle filter
ai
Introduction to track time-variant parameters[14−15].
gh

Recently, a sequential Monte Carlo approach in the


The combination of orthogonal frequency division Bayesian framework has been widely investigated in the
wireless communication systems[16−22] . These particle
an

multiplexing (OFDM) and multiple antennas at both


the transmitter and the receiver, referred to as multiple- filtering algorithms estimate the state information with
input multiple-output (MIMO), has been one of the a posterior probability[23]. Moreover, they can achieve
Sh

theoretical optima in the presence of nonlinear and non-


promising schemes in B3G and 4G wireless communi-
cation systems. In the MIMO-OFDM transmission sys- Gaussian models. In comparison with other nonlinear
tem, the carrier frequency offset (CFO) induced by the estimation methods such as EKF and unscented Kalman
local oscillator and the Doppler shift results in the inter- filter (UKF), a particle filter will yield more accurate
results along with the flexibility of the algorithm de-
of

ference of sub-carriers and further incurs an error floor


effect. It is required to estimate and compensate for sign. However, in terms of this high-dimensional pa-
the CFO at the receiver side. Moreover, the unknown rameter estimation, it is very hard to get appropriate
channel parameters mixed in the transmission will de- particles with small distribution variance on each di-
mension by the general sequential importance sampling
teriorate the acquisition condition in practice.
l

In contrast to frequency synchronization schemes of (SIS) method.


a

single-input single-output (SISO)-OFDM systems[1−7] , When the transmit/receive antennas in a MIMO sys-
tem share one oscillator, the difference of CFOs among
rn

this problem is extended to multi-dimensional param-


eter acquisition in MIMO systems. Hence, it incurs different antenna pairs can be negligible[24] . Under such
much difficulty for the coherent detection. Various a scenario, the particle filter is investigated to marginal-
ize out the unknown time-variant Rayleigh channel from
u

methods have been investigated to estimate multiple


time-invariant frequency offset and channel parame- the sampling space. The channel is estimated by the
Jo

Kalman filter on the assumption that its response model


ters, such as maximum-likelihood estimation (MLE)[8] ,
correlation-based method[9] , expectation maximization is known. The particle weights of CFO are evaluated
(EM)[10−11] and iterative methods[10−13] . They all use according to the estimation error. Simulation results
demonstrate the algorithm can estimate the CFO and
the Cramér-Rao bound as the precision reference. In
addition, an extended Kalman filter (EKF) is designed the channel jointly with high accuracy, as well as its

Received Apr.8, 2009; Revised Jul.20, 2009


Project supported by the National Natural Science Foundation of China (Grant No.60572157), and the International Cooper-
ation Foundation (Grant No.2008DFA11950)
Corresponding author ZHANG Jing, PhD, Assoc Prof, E-mail: jannety@shnu.edu.cn
J Shanghai Univ (Engl Ed), 2009, 13(6): 438–443 439

robustness to small variations of the channel model. The model of the frequency offset is
The rest of the paper is organized as follows. In
Section 1, the transmission model is briefly introduced. γ(n) = bγ(n − 1) + vγ (n), (5)
In Section 2, the particle filter is presented. In Section
3, the detailed performance evaluation results are pre- where b is the coefficient related to the mobile speed,
and vγ (n) ∼ N (0, σγ2 ).

ty
sented. Finally, in Section 4, the conclusions are given.

1 Model description 2 Particle filter for joint estimation

i
When the frequency offsets among antenna pairs are The received signal model (1) does not have an ana-

rs
the same, the received signal can be given by lytical solution for unknown parameters. In a Bayesian
framework, the optimum decision on {hn , γn } at the n-
Nt
 th symbol instant can be obtained by maximizing a pos-
r
y (n) = W (n) diag{dt (n)}F1 hr,t (n) + F η r (n), (1)

e
terior criterion, which is expressed as (The time index
t=1
n is rewritten as subscript for clarity)

iv
r
where y (n) denotes the receive signal on the r-
th antenna; W (n) = diag{1, e−j2πγ(n)/K , · · · , {h n , γn } = arg max f (hn , γn |y1:N ) , (6)
hn ,γn
e−j2πγ(n)(K−1)/K } represents the K × K diagonal fre-

Un
quency offset matrix induced by the frequency offset where f (hn , γn |y1:N ) is a posterior distribution, and
γ(n) and normalized by subcarrier spacing; dt (n) de- the subscript 1:N denotes a collection of the measure-
notes the transmit signal from the t-th transmit an- ment y from 1 to N .
tenna at the n-th symbol; diag{dt (n)} denotes the di-
ai
The posterior distribution can not be explicitly ob-
agonal matrix with the entries of K×1 transmit sig- tained but be estimated by a set of random weighted
nal dt (n); F1 means the first column of the unitary particles δ{h(i) , γ (i) } drawn from the posterior distribu-
DFT matrix F , F = [Ω(0), · · · , Ω(K − 1)], where
gh

√ i0 i1 i(K−1) T ik
tion,
Ω(i) = 1/ K[ωK , ωK , · · · , ωK ] , ωK = e−j2πik/K ,
r,t
k = 0, · · · , K − 1; h (n) is the Rayleigh sub-channel Np
  
an

between the t-th transmit and the r-th receive antenna; f ( hn , γn | y1:N ) = wn(i) δ h(i) , γ (i) , (7)
and η r (n) is the complex zero-mean noise with the inde- i=1
pendent real part and imaginary part of power spectral
Sh

(i)
2
density σ2 . where δ(·) denotes the random particles, wn is the im-
Then the received signal model can be simply ex- portance weight of the i-th particle, and Np is the num-
pressed as a function C of the unknown parameters ber of particles. However, the direct sampling from the
hr,t (n) and γ(n): posterior density function is still impossible. It may
of

  be necessary to sample from a proposal distribution


y r (n) = C hr,t (n), γ(n) + E(n). (2) q(hn , γn |y1:N ).
2.1 Sequential importance sampling (SIS)
The Rayleigh sub-channel hr,t (n) can be further repre-
In the SIS procedure, the selection of the proposal
sented by an autoregressive model:
distribution q( hn , γn | h0:n−1 , γ0:n−1 , y1:n ) is very criti-
l

hr,t (n) = −a1 hr,t (n − 1) − a2 hr,t (n − 2) + vhr,t (n), (3) cal. Its analytical solution is not achievable. Therefore,
a

we approximately choose the distribution


where vhr,t (n) ∼ N (0, σh2 ), a1 and a2 are the coefficients
rn

related to the Doppler shift and the power delay profile, q( hn , γn | hn−1 , γn−1 , y1:n ) ≈ q( hn , γn | hn−1 , γn−1 )
and we have ≈ f ( hn , γn | hn−1 , γn−1 ). (8)
u

a1 = −2rd cos (2πfd T ), a2 = rd2 , (4)


Hence, the recursive expression of wn can be
Jo

where fd is the maximum Doppler shift, T the period


of one symbol, and rd the fading velocity of the power wn = wn−1 f (yn |hn , γn ) . (9)
delay profile. (i)
The synchronization can be performed at each re- At each step, the normalized importance weights w
n
ceive antenna, respectively. Therefore, the superscript r are given by
can be omitted. Thus the unknown channel parameters (i)
for the sub-channels can be written as wn
w
n(i) = N
. (10)
Nt
 (m)
p

h(n) = [h (n), · · · , h
1
(n)] .
T wn
m=1
440 J Shanghai Univ (Engl Ed), 2009, 13(6): 438–443
 
(i)
In the particle filter, the weights are proportional to the where Im yn − ŷn denotes the absolute sum of the
following importance ratio imaginary part of the error between the measurement
f (h0:n , γ0:n |y1:n ) and the estimation.
wn ∝ . (11) Based on the above algorithm design, the proposed
q (h0:n , γ0:n |y1:n )
particle filer algorithm can be shown in Fig.1.

ty
It can be shown that the variance of the importance
ratio increases over time, thus resulting in the algorithm i) For i = 1,· · · , N
np , o “ ”
(i) (i)
degeneracy. The effective sample size N̂eff is defined as draw the state γ0 from the prior f γ0 , set

i
(i)
1 w0 = N1p and h0 ∼ N (ĥ0 , Ph,0 ), Ph,0 = σ 2 I.

rs
N̂eff = N  2 , (12)
p (i) ii) For n =1, · · · , loop,
wn Sequential importance sampling step
i=1
For i = 1, · · · , Np ,

e
where small N̂eff indicates that a large computational do the Kalman filtering with the model and the ini-

iv
effort will occur with updated particles, of which the tials h0 and Ph,0
(
contribution to the approximation to f (h0:n , γ0:n |y1:n ) hn = −a1 hn−1 − a2 hn−2 + vh,n
is almost zero. To avoid the degeneracy problem, the

Un
Re yn = Re C(hn , γn ) + Re En
sequential importance resampling (SIR) can be used.
2.2 Sequential importance resampling obtain ĥn , Ph,n .
(SIR) Calculate
To maintain the varieties of particles, in the resam- ˛ ”
ai
˛ 1
f (yn ˛hn , γn(i) ≈ ˛˛ “ ”˛
pling step, only particles with high importance weights (i) ˛
˛Im yn − ŷn ˛
are selected and split them into other Ns particles with “ ˛ ”
(i) (i) ˛
wn(i) = wn−1 f yn ˛hn , γn(i) .
gh

weights wn = N1s , and those particles with lower


weights are discarded.
2.3 Algorithm design For i=1, · · · , Np ,
(i) (i)
an

The direct algorithm with the SIS and resampling normalize wn to obtain w̃n .
Decision
needs to sample in high-dimensional channel spaces. Np
P (i) (i)
Thus it will produce large particle variances and incur a Choose γ̂n = wn γn .
Sh

i=1
large amount of computation. Marginalizing out some Sequential importance resampling step
of the variables can reduce the size of the sample space. If N̂eff < Nthres , multiply/discard particles,
n oNp
Because the channel experiences fast fading compared use γn
(i) (i)
related to high w̃n to obtain Ns un-
to the CFO as well as the channel and the CFO are un- i=1
n oNs
of

(i)
correlated, the channel particles are marginalized out of weighted particles γn , discard particles with
i=1
the estimation. lower weights.
(i)
With the particles
 of  the CFO γ0 which are drawn Fig.1 Proposed algorithm
(i)
from the prior f γ0 , hn can be estimated by the

(i) 3 Performance evaluation


Kalman filter in sequence with initials h0 , γ0 and
a

the real part of the measurement. That is, the Kalman 3.1 System parameters
rn

filter is performed with the model: In the simulation studies, a MIMO-OFDM system
with four-transmit and two-receive antennas is consid-
hn = −a1 hn−1 − a2 hn−2 + vh,n , ered. The Rayleigh sub-channel between each trans-
u

(13)
Re yn = Re C(hn , γn ) + Re En . mit and receive antenna is an independent propagation
Jo

path. The coefficients of one sub-channel are as fol-


The estimation variance can be obtained in this pro- lows. a1 = −1.994, a2 =0.996, fd =80 Hz, T =5 ms
cess. It represents the estimation accuracy of the chan- and rd =0.998. The parameter of the CFO is set as
nel. Moreover, the imaginary part of the error between b=0.99. The maximum normalized frequency offset is
the measurement and the estimation means the particle less than 0.25. The mean square error (MSE) is defined
 2

fitness of the CFO. Hence, the probability of the parti-  


as hMSE = E h − ĥ and γMSE = E |γ − γ̂|2 ,
cles can be approximated to
  respectively. The bit error rate (BER, Pb ) performance
1
f yn hn , γn(i) ≈   , (14) of the MIMO-OFDM system is based on the maximum
(i)
Im yn − ŷn ratio combination and the orthogonal space-time block
J Shanghai Univ (Engl Ed), 2009, 13(6): 438–443 441

decoding. 0.30
NP = 20
3.2 Result analysis 0.25 NP = 50

Estimation error of γ
The estimation results of the sub-channel h1,1 and 0.20 NP = 100
the frequency offset γ with Np =50 particles at signal
0.15
to noise ration (SNR, σs )=0 dB, 6 dB and 12 dB are

ty
shown in Fig.2. The BER performance with the estima- 0.10
tion results under stable state is shown in Fig.3. The 0.05
estimation results with particle numbers of Np =20, 50

i
0
and 100 at σs =6 dB are shown in Fig.4. From Figs.2–4,
it can be shown that with the increment of σs , the es-

rs
−0.05
0 10 20 30 40 50
timation accuracy of the channel tends to be higher. It Calculation steps
can also be observed that the number of the particles is (a) CFO estimation result

e
another direct factor of the estimation accuracy of γ. 0.35
0.30 NP = 20

iv
0.16 NP = 50

Estimation error of h1,1


σs = 0 dB 0.25 NP = 100
0.14
σs = 6 dB 0.20
0.12 σs = 12 dB
Estimation error of γ

Un
0.15
0.10
0.10
0.08 0.05
0.06 0
0.04 − 0.05
ai
0.02 − 0.10
0 0 10 20 30 40 50
Calculation steps
−0.02
gh

0 10 20 30 40 50 (b) Channel estimation result


Calculation steps
Fig.4 Results of joint channel and CFO estimation with
(a) CFO estimation result
0.3 different number of particles
an

σs = 0 dB
0.2 σs = 6 dB
Estimation error of h1,1

σs = 12 dB When the particles can capture the real distribution,


0.1
Sh

larger number of the particles means the approximation


0
distribution becomes more accurate. In addition, the
− 0.1 particles with larger weights have larger number than
− 0.2 that of the particles with lower weights. Therefore, it
of

− 0.3 can provide more accurate results. It can also be seen


that the computational step is about 10 when the esti-
− 0.4
0 10 20 30 40 50 mation process reaches a stable state.
Calculation steps
(b) Channel estimation result Moreover, the MSE performance of the particle fil-
l

Fig.2 Results of joint channel and CFO estimation at dif- ter with Np =100 in comparison with EKF and UKF is
a

ferent SNRs shown in Fig.5. The expansion point of the EKF is at


γ=0 and a random channel. The real part and imagi-
rn

10−2 nary part of the measurement are used to estimate pa-


rameters in the EKF and the UKF. We can see that
u

the particle filter yields more accurate results in both


10−3
channel and CFO.
Jo
Pb

Real
3.3 Robustness to the channel model vari-
10−4 σs = 0 dB ation
σs = 6 dB
σs = 12 dB
The proposal distribution is based on the model of
the CFO and the channel. It needs a priori knowledge of
10−5 the transmission process. When the mismatch exists be-
−3 0 3 6 9 12 15
σs/dB tween the model and the real process, the particles can
Fig.3 BER versus SNR with respect of the joint estimation still catch the real one, since the particles are various and
results in the region of the real distribution. It can be found
442 J Shanghai Univ (Engl Ed), 2009, 13(6): 438–443

that when the coefficients of the channel have a small OFDM system [J]. IEEE Transaction on Broadcasting,
variation produced by the Doppler shift of fd =70 Hz, 2007, 53(2): 560–566.
the MSE performances are still with the same order as [6] Biao C, Hao W. Blind estimation of OFDM carrier
shown in Fig.5. Hence, the algorithm has some robust- frequency offset via oversampling [J]. IEEE Transaction
ness. on Signal Processing, 2004, 52(7): 2047–2057.

ty
10−1
[7] Francois S, Yves D, Atika M, Christelle G. Non-
EKF-γ
pilot-aided sequential Monte Carlo method to joint sig-
UKF-γ

i
PF-γ nal, phase noise, and frequency offset estimation in mul-
10−2
ticarrier systems [J]. EURASIP Journal on Advances in

rs
rMSE and hMSE

EKF-h Signal Processing, 2008: 1–14.


10−3 UKF-h
PF-h [8] Besson O, Stoica P. On parameter estimation of

e
MIMO flat fading channels with frequency offsets [J].
10−4 IEEE Transaction on Signal Processing, 2003, 51(3):

iv
602–613.
10−5 [9] Yao Y, Ng T S. Correlation-based frequency offset
0 3 6 9 12 15

Un
σs/dB estimation in MIMO system [C]// IEEE Vehicle Tech-
nology, Florida, USA. 2003: 438–442.
Fig.5 MSE of EKF, UKF and the particle filter
[10] Pham T H, Nallanathan A, Liang Y C. Joint
channel and frequency offset estimation in distributed
ai
4 Conclusions MIMO flat-fading channels [J]. IEEE Transaction on
Wireless Communication, 2008, 7(2): 648–656.
The paper has presented a particle filtering algo-
gh

rithm for frequency synchronization in MIMO-OFDM [11] Sun Y, Xiong Z X, Wang X D. EM-based iterative
systems. It is based on the sequential Monte Carlo ap- receiver design with carrier-frequency offset estimation
proaches in the Bayesian framework, and simplified to for MIMO OFDM systems [J]. IEEE Transaction on
an

Communication, 2005, 53(4): 581–586.


a one-dimensional particle sampling in order to improve
the computation efficiency. Numerical simulation shows [12] Lu Z, Li J D, Zhao L J, Pang J Y. Iterative param-
eter estimation in MIMO flat-fading channels with fre-
Sh

that the algorithm can obtain high accuracy of CFO and


channel estimation performance. Additionally, it can si- quency offsets [C]// IEEE Advanced Information Net-
working Applications, Vienna, Austria. 2006: 802–805.
multaneously estimate time-varying CFO and channel.
It also has some robustness to the variation of the chan- [13] Yu Q, Lambotharan S. Iterative (Turbo) estima-
tion and detection techniques for frequency selective
of

nel model.
channels with multiple frequency offsets in MIMO sys-
References tem [J]. IEEE Signal Processing Letter, 2007, 14(4):
236–239.
[1] Zhao Y, Haggman S G. Intercarrier interference self-
l

cancellation scheme for OFDM mobile communica- [14] Roman T, Enescu M, Koivunen V. Joint time-
tion systems [J]. IEEE Transaction on Communication, domain tracking of channel and frequency offsets for
a

2001, 49(7): 1185–1191. MIMO-OFDM systems [J]. Wireless Personal Commu-


rn

nication, 2004, 31(3): 181–200.


[2] Schmidl T M, Cox D C. Robust frequency and tim-
ing synchronization for OFDM [J]. IEEE Transaction [15] Liang Y M, Luo H W, Huang J G. Extended Kalman
on Communication, 1997, 45(12): 1613–1621. filtering-based channel estimation for space-time coded
u

[3] Moose P H. A technique for orthogonal frequency divi- MIMO-OFDM systems [J]. Journal of Shanghai Uni-
Jo

sion multiplexing frequency offset correlation [J]. IEEE versity (English Edition), 2007, 11(5): 469–473.
Transaction on Communication, 1994, 42(10): 2908– [16] Punskaya E, Andrieu C, Doucet A, Fitzgerald
2914. W J. Particle filtering for demodulation in fading chan-
[4] van de Beek J, Magnus S, Boriesson P O. ML nels with non-Gaussian additive noise [J]. IEEE Trans-
estimation of time and frequency offset in OFDM sys- action on Communication, 2001, 49(4): 579–582.
tems [J]. IEEE Transaction on Signal Processing, 1997, [17] Ghirmai T, Bugallo M F, Miguez J, Djuric P M.
45(7): 1800–1805. A sequential Monte Carlo method for adaptive blind
[5] Hao M J. Decision feedback frequency offset estima- timing estimation and data detection [J]. IEEE Trans-
tion and tracking for general ICI self-cancellation based action on Signal Processing, 2005, 53(8): 2855–2865.
J Shanghai Univ (Engl Ed), 2009, 13(6): 438–443 443

[18] Joaquin M, Tadesse G, Monica F B. A sequen- [21] Huang Y F, Djuric P M. A blind particle filtering
tial Monte Carlo technique for blind synchronization detector of signals transmitted over flat fading chan-
and detection in frequency-flat Rayleigh fading wire- nels [J]. IEEE Transaction on Signal Processing, 2004,
less channels [J]. Signal Processing, 2004, 84(11): 2081– 52(7): 1891–1900.
2096. [22] Qian Y, Guaoan B, Chunru W. SMC-based blind

ty
[19] Elena P, Arnaud D, William J F. Particle fil- detection for DS-CDMA systems over multipath fad-
tering for joint symbol and code delay estimation in ing channels [J]. IEEE Transaction on Communication,
DS spread spectrum systems in multipath environment 2006, 54(6): 971–974.

i
[J]. EURASIP Journal on Applied Signal Processing, [23] Arnaud D, Simon G, Christophe A. On sequen-

rs
2004(15): 2306–2314. tial Monte Carlo sampling methods for Bayesian filter-
[20] Huang Y F, Zhang J Q, Luna I T, Djuric P M, ing [J]. Statistics and Computing, 2000, 10(3): 197–208.

e
Padiuo D P R. Adaptive blind multiuser detection [24] Magnus S, Darren M, Steve P. Analysis of
over flat fast fading channels using particle filtering [J]. frequency-offset tracking in MIMO OFDM systems [J].

iv
EURASIP Journal on Wireless Communication and IEEE Transaction on Communication, 2006, 54(8):
Networking, 2005(2): 130–140. 1481–1491.

Un
ai
gh
an
Sh
of
a l
u rn
Jo

Potrebbero piacerti anche