equalization of ofdm under time-varying channel conditions · pdf file ·...
TRANSCRIPT
Equalization of OFDM under Time-Varying Channel
Conditions
Michael D. ZoltowskiPurdue UniversityWest Lafayette, [email protected]
David A. HaessigBAE SystemsWayne, [email protected]
14th Annual Workshop on Adaptive Sensor Array Processing (ASAP)
MIT Lincoln Laboratory6 June 2006
2
Outline
• Introduction– OFDM
TransmitterReceiver – delay-only
– Problem Definition Delay-Doppler Channel ModelPrior Work
• Extension of OFDM for large Doppler spread– MMSE-based Receiver – Simulated Performance for complete channel information
• Parameter Estimation– Channel Model– Uncoupled estimation of Doppler, Delay, and Gain
• Conclusions & Future Work
3
Orthogonal Frequency Division Modulation (OFDM) – a waveform, TDMA in nature, designed to accommodate multi-path time dispersion
• Cyclic extensions avoid ISI due to delay spread & aid or enable synchronization• Cyclic Prefix generated by copying end of x(n) to beginning
–Suffix – add beginning to end• Windowing reduces spectral growth
timeOFDM Frame = T + TCE
OFDM Symbol = x(n) = T msecCycle suffix & window
Cycle prefix and window
Windowing
time
… …
time
… …
4
The OFDM Transceiver
Received Power
Delay Spread t
Received Power
Delay Spread t
Tx bit stream
Mod-ulator
S/P IFFTCyclic extension & windowing
P/S
Time dispersive channel
Rx bit stream
De-mod
P/S FFTRemove Cyclic extensions
S/P
noise
Channel Equalization
Extract Pilots
Add Pilots
+
Sync
x(n)x
• A transmit time-domain sequence x(n) is produced from transmit symbol vector s:
where W* is the IFFT matrix, Cp the cyclic prefix operator, N the FFT length and L the number of samples in Cp
sWCx p*=
Cp =
IL 0
IN
s
Estimate Channel
r y(n)
5
The OFDM Receiver -- handles channel time dispersion in a computationally efficient manner
sWCDCr ppr*
τ=
prC =
IN0
Rx bit stream
De-mod
P/S FFTRemove Cyclic extensions
S/P
noise
Channel Equalization
Extract Pilots
+
Sync
• Representing the channel , having K paths, in delay-operator matrix form∑=
−=K
iii nnh
1
)()( τδα
Estimate Channel
( )
W
e
ee
WD
Nj
j
j
i
i
i
i
⎥⎥⎥⎥⎥
⎦
⎤
⎢⎢⎢⎢⎢
⎣
⎡
=
−−
−
−
1
1
0
*
0000
0000
τ
τ
τ
τ ⎟⎟⎠
⎞⎜⎜⎝
⎛= ∑
=
K
ii iDD
1ττ α
r
• The received signal r(n) is then given by:
where Cpr removes cyclic prefix• Channel does not destroy orthogonality of signaling basis functions
where is a diagonal matrix containing the channel frequency response, enabling coherent detection
s~s
rWs =~
H
rWHs *ˆˆ =H
6
• The signal y(n) at channel output given transmit signal x(n) at channel input:
having independent Doppler and Delays on K separate “rays” and assumed to be constant over the OFDM Frame
• The channel is given in delay-Doppler operator form by summing over each of the K paths:
• Converting from serial to parallel vector form:
• An issue, the transfer matrix (in brackets) is no longer diagonal, but diagonally dominant
Consider the general delay-Doppler channel model
⎟⎟⎠
⎞⎜⎜⎝
⎛= ∑
=ΩΩ
K
ii ii
DDD1
ττ α
)()()(1
)( nnxenyK
ii
nji
ii ητα τ +−=∑=
−Ω
( )⎥⎥⎥⎥⎥
⎦
⎤
⎢⎢⎢⎢⎢
⎣
⎡
=
−Ω
Ω
Ω
Ω
1
1
0
0000
0000
Nj
j
j
i
i
i
i
e
ee
D
iΩ iτ
ητ += Ω sWCDy p*
[ ] ητ += Ω sWCDWCs ppr*~
7
The Problem – a time & frequency dispersive channel with a particularly large delay-Doppler product
• In typical wireless applications frequency dispersion is a fraction of a tone spacing.
• In the problem considered here, frequency dispersion greatly exceeds the tone spacing
1/T
time
frequency
tone spacing
OFDM Symbol timeT
Typically Doppler spread << 1/T
Typical Delay spread < T/5
atypically Doppler spread >>1/T
8
Prior work related to the use of OFDM with frequency selective time-varying fading channels:
• Wang, Proakis, Zeidler, Techniques for Suppression of Intercarrier Interference in OFDM Systems, IEEE WCNC, Feb 2005:
– Propose a MMSE-based OFDM receiver structure– Estimate channel by curve fitting over time-frequency using pilots, based on
statistical model of time-varying impulse response
• Linnartz, Gorhkov, Doppler-resistant OFDM receivers for Mobile Multimedia Communications, PIMRC 2000, London, Sept 2000, and
• Mostofi, Cox, ICI Mitigation for Pilot-Aided OFDM Mobile Systems:– MMSE ICI receiver based on piece-wise linear approximation to channel
response at each toneSuitable for limited Doppler spread
• Y. Li, L.J. Cimini, and N.R. Sollenberger, Robust channel estimation for OFDM systems with rapid dispersive fading channels, IEEE, Tr. on Comm. vol. 46, pp. 902-915, July 1998:
• Many others
9
• The received signal vector through the delay-Doppler channel :
• the MMSE solution is given by:
requiring estimate of delay-Doppler channel
• Comparing delay-Doppler MMSE receiver to delay-only OFDM receiver:– OFDM - NxN FFT– MMSE - NsxNs inverse – where Ns – number of symbols; N – FFT size
An MMSE-based OFDM receiver structure is developed for the general, deterministically modeled delay-Doppler channel
ητ += Ω sWCDCr ppr*
rFIFFs n ′+′= − ˆ)ˆˆ(ˆ 12σ
r
*WCDCF ppr Ω= τ
η+= Fsr
*ˆˆ WCDCF ppr Ω= τ
ΩτD
rWHs *ˆˆ =rFIFFs n ′+′= − ˆ)ˆˆ(ˆ 12σ
Simulation results
Comparison of OFDM performance in delay-only case to that of MMSE in delay-Doppler case
11
Simulations comparing the OFDM receiver in the delay-only case to that of the MMSE receiver through the same channel but with arandom Doppler added to each channel tap
Tx packet Add 20% pilots & zero fill
1K IFFT
Inter-leave over 5 packets
BPSK
Rate 1/5 Turbo Encoder
Cyclic extension & windowing
Mod-ulator
S/PExtract pilots
& remove zero fill
1K FFT
De-Inter-leaver
Rate 1/5 Turbo Decoder
Remove Cyclic extensions
BPSK soft Demod
pilots
Channel Equalization
)(ˆ kH
Rx packet
Delay-Doppler Channel
Delay-only Channel
True Channel
1K MMSE
F Matrix Computation
α τ Ω F
Channel Response h(n)
S/PRemove Cyclic extensions
Compute Freq Resp
BPSK soft Demod
De-Inter-leaver
Rate 1/5 Turbo Decoder
Rx packet
Channel Gain, tap location and Doppler 1024FFT Size = NFFT
99
6 samples
160 samples
.0061
Value
No. of bits per packet
Cyclic Suffix
Cyclic Prefix
Tone Spacing -- (rad) -- 2π/NFFT
Parameter Name
12
Channels evaluated -- 11 tap channel with Doppler spread of a few tones, and an several hundred tap channel spanning the entire 160 sample cyclic extension with a Doppler spread of over 500 tones
0 50 100 150 200 250-0.01
-0.005
0
0.005
0.01Tap weights
50 100 150 200-0.4
-0.2
0
0.2
Tap Dopplers (rad/sam ple)
x 104
0.5 1 1.5 2
x 104
-60
-50
-40
-30
-20
Frequency (Hz)
Mag
nitu
de (d
B)
0 0.5 1 1.5 2
x 104
-60
-40
-20
0
Frequency (Hz)
Mag
nitu
de (d
B)
0 50 100 150 200 250
-0.2
-0.1
0
0.1
0.2
Tap weights
0 50 100 150 200 250
-0.2
-0.1
0
0.1
0.2
Tap Dopplers (rad/sample)
• Note: Frequency responses computed for the case in which Doppler = 0
11 tap channel many tap channel
13
-5 0 5 1010
-6
10-5
10-4
10-3
10-2
10-1
100
E b/No (dB )
Pe
B it E rror Rate vs E b/No - Des ign 3 in A W G N, Delay -only , and Delay -Doppler Channels
Performance of OFDM and MMSE Receivers with perfect channel knowledge
• Performance of MMSE receiver in delay-Doppler channel closely matches that of OFDM in delay-only channel
• Performance lost to Doppler recovered with MMSE receiver, but at a significant increase to computational expense
AWGN
Many-tap delay-only & delay-
Doppler channels
11 tap delay-only & delay-Doppler
channels
Parameter Estimation
An estimation approach for a sensor array channel structure
15
We consider the time-varying multi-path channel involving K rays from one transmitter to a P-node array
[ ]∑=
− −M
mm
Mnjm Mnye mm
111
)(1 )(11ωβ
njeNnxns 1)()( 111θα −=
Transmitter
[ ]∑=
− −M
mmPP
MnjmP Mnye mPmP
1
)( )(ωβ
njeNnxns 2)()( 222θα −=
njeNnxns 3)()( 333θα −=
njKKK
KeNnxns θα )()( −=
+)(* nxsWC p =
)(nr
• Each ray involves an unknown gain , fractional delay , and Doppler shift to be estimated
• Array signal processing parameters, , are known
• For this channel we construct the F matrix relating to
kα kNkθ
mpmpmp M ωβ ,,
s r
)(1 nz
)(nzK
)(1 ny
)(nyP
16
Where the channel Dopper & delay matrix operators are:
• Note that F is a function of the known array processing parameters, and the unknown ray parameters
• Requiring estimation of 4 real parameters per ray
The F matrix for the case of K rays
mpmpmp M ωβ ,,
( )⎥⎥⎥⎥
⎦
⎤
⎢⎢⎢⎢
⎣
⎡
=Ω
−Ω
Ω
Ω
1
)1(
0000
00001
)(
Nj
j
e
eD
( )
W
e
eWD
Nj
j
⎥⎥⎥⎥
⎦
⎤
⎢⎢⎢⎢
⎣
⎡
=
−−
−
1
1*
0000
00001
)(
τ
τ
τ τ
( ) ( ) ( )[ ] *
1
),,,,( WCDMNDeCMNF p
M
mmkkmkk
Mjmkprmkmkmkkkk
mkk∑=
ΩΩ− +Ω+=Ω ωβωβ τ
∑=
=K
kkkkk FFF
1
),( αα
kα kNkθ
η+= sFr
17
• Training symbol vectors made entirely of pilots are transmitted– Series of sample-domain vectors used for parameter estimation
• Channel Doppler, delay, and gain estimates generated sequentially, not jointly– Estimates of Doppler generated via ESPRIT, independent and without knowledge of channel
delay and gain estimates– Delay estimates generated via series of 1-D searches, also pairing Dopplers to particular rays– matrices computed, and channel gains estimated as final step
K – coding rate
De-Inter-leaver
Rate 1/K Turbo Decoder
DeMod-ulator
Remove cyclic ext
MMSES/P
ωβ ,, m
r
MMSE receiver and parameter estimation architecture showing separate estimation of Doppler, delay, and channel gain
Rx packetof B bits
F
ESPRIT Doppler Estimation -
Residual Delay Estimation -
Compute FkMatrices
Total F Matrix Computation
Channel Gain Estimation
Array Processing
kF,α
kF
Ω,NΩ
r
18
Use of ESPRIT with Multiple Successive Training OFDM Symbol Blocks to Achieve Separable Estimation of Residual Dopplers and Delays Associated with Multiple MBAJ Channels input to MBAJ
• Estimation Doppler via ESPRIT: – train with K+1 repeated OFDM training symbols – estimates are the eigenvalues of a matrix
avoids a multi-dimensional search and is invariant to unknown gains and time delays
• Subsequent estimation of residual time delays – using Doppler shift estimates– uses eigenvector information from ESPRIT to decouple superimposed rays
allowing for separable estimation of residual time delays (pairing problem addressed)
• Unknown channel gains subsequently estimated via Least Square Error solution
• Total F matrix compute given estimates and used over subsequent OFDM Frames for symbol data detection
19
Send the same OFDM training symbol for 4 successive blocks to generate the following output vectors:
y1 = e j (N + L )ω1 z1 + e j (N + L )ω2 z2 + e j (N + L )ω3 z3
y2 = e j 2(N + L )ω1 z1 + e j 2(N + L )ω2 z2 + e j 2(N + L )ω 3z3
y0 = z1 + z2 + z3
z1 = α1 Cpr β1i Dτ ,1i Dω ,1i1,i∑⎛
⎝⎜⎞
⎠⎟⎡
⎣⎢
⎤
⎦⎥Dτ1
residual Dω1
residual CpW*s( )
z2 = α2 Cpr β2i Dτ ,2i Dω ,2i2,i∑⎛
⎝⎜⎞
⎠⎟⎡
⎣⎢
⎤
⎦⎥Dτ2
residual Dω2
residual CpW*s( )
z3 = α 3 Cpr β3iDτ ,3iDω ,3i3,i∑⎛
⎝⎜⎞
⎠⎟⎡
⎣⎢
⎤
⎦⎥Dτ 3
residual Dω3
residual CpW*s( )
y3 = e j 3(N + L )ω1 z1 + e j 3(N + L )ω2 z2 + e j 3(N + L )ω3 z3
ESPRIT Based Estimation of Delay – consider this example involving 3 rays:
20
Cast these equations as an eigenvalue problem, thereby avoiding a multi-dimensional search to find Doppler values
φ1 = e j (N + L )ω1
φ2 = e j (N + L )ω2 Δ =1 φ1 φ1
2
1 φ2 φ22
1 φ3 φ32
⎛
⎝
⎜⎜⎜
⎞
⎠
⎟⎟⎟
G1Ψ = G2
Φ =φ1 0 00 φ2 00 0 φ3
⎛
⎝
⎜⎜
⎞
⎠
⎟⎟
• With quantities defined as on previous slide, it follows that:
where
G1 = y0 | y1 | y2[ ]= z1 | z2 | z3[ ]Δ G2 = y1 | y2 | y3[ ]= z1 | z2 | z3[ ]ΦΔ
φ3 = e j (N + L )ω3
& are eigenvalues of
Ψ = Δ−1ΦΔ
Thus we can write: allowing to be determined numerically
The solution is implying that
φ1 = e j (N + L )ω1 φ2 = e j (N + L )ω2 φ3 = e j (N + L )ω3 Ψ
Ψ
21
Here is an example involving a 3 ray channel with unknown delays, Doppler, and gains -- the 3 Doppler values are estimated correctly
φ3 = e j (N + L )ω3
φ1 = e j (N + L )ω1
φ2 = e j (N + L )ω2
22
Use eigenvector information from ESPRIT to separate the three channel contributions summed at the array output:
φ1 = e j (N + L )ω1
φ2 = e j (N + L )ω2
G1 = y0 | y1 | y3[ ]= z1 | z2 | z3[ ]Δ
G2 = y1 | y2 | y3[ ]= z1 | z2 | z3[ ]ΦΔ
Ψ = ΕΛΕ−1
THUS: eigenvectors of decouple superposition of sensor array output contributions from channel 1, channel 2 & channel 3:
Ψ
G1E = z1 | z2 | z3[ ]DP
G1Ψ = G2are eigenvalues of solution to φi = e j ( N + L )ω i , i = 1,2, 3
φ3 = e j (N + L )ω3
Φ =φ1 0 00 φ2 00 0 φ3
⎛
⎝
⎜⎜
⎞
⎠
⎟⎟
where is a permutation matrix and is a diagonalmatrix representing unknown scalings, since eigenvectors are only unique to within a multiplicative scalar
P D
Λ = Φ Ε = Δ−1DPwhere and
23
Pairing problem and residual delay estimation:
There is an unknown gain on each channel -- absorb into unknown diagonal scaling matrix
z1 = Cpr β1iDτ ,1i Dω ,1i
1,i∑⎛
⎝⎜⎞
⎠⎟⎡
⎣⎢
⎤
⎦⎥Dτ1
residual Dω1
residual CpW*s( )
z2 = Cpr β2iDτ ,2i Dω ,2i
2,i∑⎛
⎝⎜⎞
⎠⎟⎡
⎣⎢
⎤
⎦⎥Dτ2
residual Dω2
residual CpW*s( )
Combinatoric pairing problem: six possibilities to assess in caseof 3 channels feeding into sensor array
z3 = Cpr β3iDτ ,3i Dω ,3i
3,i∑⎛
⎝⎜⎞
⎠⎟⎡
⎣⎢
⎤
⎦⎥Dτ 3
residual Dω3
residual CpW*s( )
G1E = ζ1 |ζ2 |ζ3[ ]= z1 | z2 | z3[ ]DP
D
24
For each of the permuted pairings, search for the delays producing a match (to within a scalar) of the permuted eigenvector and the vectors
G1E = ζ (1) |ζ (2) |ζ (3)⎡⎣ ⎤⎦
z1(ω
(δ ),τ1) = Cpr β1i Dτ ,1i Dω ,1i1,i∑⎛
⎝⎜⎞
⎠⎟⎡
⎣⎢
⎤
⎦⎥Dτ1
residual Dω (δ )residual CpW
*s( )
z2 (ω (ε ),τ 2 ) = Cpr β2iDτ ,2i Dω ,2i
2,i∑⎛
⎝⎜⎞
⎠⎟⎡
⎣⎢
⎤
⎦⎥Dτ2
residual Dω (ε )residual CpW
*s( )
z3(ω (γ ),τ 3 ) = Cpr β3i Dτ ,3i Dω ,3i
3,i∑⎛
⎝⎜⎞
⎠⎟⎡
⎣⎢
⎤
⎦⎥Dτ 3
residual Dω (δ )residual CpW
*s( )
{ }minimum0 < τ1 < LTs
ζ (δ ) | %z1(ω (δ ),τ1)⎡⎣ ⎤⎦smallest singular value of
+Γ(δ ,ε,γ ) = smallest singular value of{ }minimum
0 < τ 2 < LTs +smallest singular value of{ }minimum
0 < τ 3 < LTs
ζ (ε ) | %z2 (ω (ε ),τ 2 )⎡⎣ ⎤⎦
ζ (γ ) | %z3(ω (γ ),τ 3)⎡⎣ ⎤⎦
(δ ,ε,γ ) ∈{(1,2, 3),(1, 3,2),(2,1, 3),(2, 3,1),(3,1,2),(3,2,1)}
)ˆ,(~1
)()( τως δδ z
)ˆ,(~2
)()( τως εε z
)ˆ,(~3
)()( τως γγ z
z~
25
ESPRIT Based Method for Estimating Residual Dopplers for 3 Channels Feeding into array with Delay-Doppler Taps: Step 2. Pairing number 1.
26
ESPRIT Based Method for Estimating Residual Dopplers for 3 Channels Feeding into array with Delay-Doppler Taps: Step 2. Pairing number 2.
27
ESPRIT Based Method for Estimating Residual Dopplers for 3 Channels Feeding into aray with Delay-Doppler Taps: Step 2. Pairing no. 3 (correct)
28
ESPRIT Based Method for Estimating Residual Dopplers for 3 Channels Feeding into array with Delay-Doppler Taps: Step 2. Pairing number 4.
29
ESPRIT Based Method for Estimating Residual Dopplers for 3 Channels Feeding into array with Delay-Doppler Taps: Step 2. Pairing number 5.
30
ESPRIT Based Method for Estimating Residual Dopplers for 3 Channels Feeding into array with Delay-Doppler Taps: Step 2. Pairing number 6.
31
Channel gain estimation is accomplished as a final step after estimation of individual Fk:
[ ] ααηα SPKPPP
K
kkk FsFsFsFsFr ==+⎥⎦
⎤⎢⎣
⎡= ∑
=
ˆˆˆˆ21
1
• Recall that array processor output:
• Transmission of a train symbol vector allows this equation to be written:
• The least-square solution for alpha:
( )[ ]rFFF SSS ′′= −1α
Ps
η+= sFr
32
Conclusions
• An MMSE-based OFDM receiver robust to large Doppler spread was developed
– No restriction on magnitude of Doppler– BER Performance approaches that of OFDM receiver for same channel
(less Doppler)
• A method for separately estimating channel gains, time delays and Doppler developed and demonstrated
– dimensionality and complexity lower than a joint estimation technique
• Current Activity:– Evaluating sensitivity of MMSE-based OFDM detection performance to
estimation errors– Evaluating sensitivity of channel estimates to noise– Investigating estimation and receiver performance with non-constant
channel coefficients associated with transmitter motion