ut austin 1 biao lu 1 wireline channel estimation and equalization ph.d. defense biao lu embedded...
TRANSCRIPT
UT Austin1
Biao Lu 1
WIRELINE CHANNEL ESTIMATION ANDEQUALIZATION
Ph.D. Defense
Biao Lu
Embedded Signal Processing Laboratory
The University of Texas at Austin
Committee Members
Prof. Brian L. Evans
Prof. Alan C. Bovik
Prof. Joydeep Ghosh
Prof. Risto Miikkulainen
Dr. Lloyd D. Clark
UT Austin2
Biao Lu 2
OUTLINE
Wireline channel equalization Wireline channel estimation
Channel modelingMatrix pencil methodsContribution #1: modified matrix pencil
methods for channel estimation Discrete multitone modulation
Minimum mean squared error equalizer Contribution #2: matrix pencil equalizer Maximum shortening SNR equalizer Contribution #3: fast implementation
» Divide-and-conquer methods
» Heuristic search
Summary and future research
UT Austin3
Biao Lu 3
WIRELINE CHANNEL EQUALIZATION
fjc efAfH
transmitter channel equalizer detector
noise
nm ns nw ny ns nm
+hc(n)
Wireline digital communication system
Ideal channel frequency response Amplitude response A( f ) is constantPhase response ( f ) is linear in f
Channel distortions Intersymbol interference (ISI)
Additive noise
0 1 1.00.75
1.0 0.75 0.5
1 1
UT Austin4
Biao Lu 4
COMBATTING ISI IN WIRELINE CHANNELS
Channel equalizer response Heq( f ) compensates for channel distortion
Equalizers may compensate forFrequency distortion: e.g. ripplesNonlinear phaseLong impulse response
Channels may haveSpectral nullsNonlinear distortion, e.g. harmonic
distortion Goal: Design time-domain equalizers
Shorten channel impulse responseReduce intersymbol interference
1 ceq fHfH
UT Austin5
Biao Lu 5
OUTLINE
Wireline channel equalization Wireline channel estimation
Channel modelingMatrix pencil methodsContribution #1: modified matrix pencil
methods for channel estimation Discrete multitone modulation
Minimum mean squared error equalizer Contribution #2: matrix pencil equalizer Maximum shortening SNR equalizer Contribution #3: fast implementation
» Divide-and-conquer methods
» Heuristic search
Summary and future research
UT Austin6
Biao Lu 6
WIRELINE CHANNEL ESTIMATION
Problem: Given N samples of the received signal, estimate channel impulse response Training-based: transmitted signal knownBlind: transmitted signal unknown
Time-domain channel estimation methodsLeast-squares [Crozier, Falconer & Mahmoud, 1996]
Singular value decomposition (SVD) [Barton & Tufts, 1989; Lindskog & Tidestav, 1999]
Frequency-domain channel estimationDiscrete Fourier transform [Tellambura, Parker & Barton, 1998; Chen & Mitra, 2000]
Discrete cosine transform [Sang & Yeh 1993; Merched & Sayed, 2000]
UT Austin7
Biao Lu 7
WIRELINE CHANNEL ESTIMATION
Broadband channel impulse responses have long tails
Model channel as infinite impulse response (IIR) filterTransfer function with K poles
iii fjdp 2
K
i
p ze
zB
zA
zBzH
i
1
11
UT Austin8
Biao Lu 8
WIRELINE CHANNEL ESTIMATION
All-pole portion of an IIR filter
Problem: given a noisy observation of channel impulse response h(n)
Estimate Least-squares method to compute {ai}
from
1,,1,0 ),()()( Nnnwnhny
nueanhK
i
npi
i
1
ap
K
i
p zezH
i
1
1ap
1
1
ai: complex amplitude
Kie ip ,,2 ,1,
Kie ip ,,2 ,1,
Assuming no duplicate poles
UT Austin9
Biao Lu 9
MATRIX PENCIL METHOD [Hua & Sarkar, 1990]
Matrix pencil of matrices A and B is the set of all matrices AB,
Noise-free case: N samples of h(n)
L is the pencil parameter (K L N K)
H, H0 and H1 are Hankel and low rank, where rank is K.
)1()(1211
21432
1321
1210
LLNNNLNLNLN
LL
LL
LL
hhhhh
hhhhhhhhhh
hhhhh
H
1H
0H
UT Austin10
Biao Lu 10
MATRIX PENCIL METHOD [Hua & Sarkar, 1990]
Noise-free data
1. Form matrices H, H0 and H1
2. Calculate C = H0†H1 († is pseudoinverse)
3. K non-zero eigenvalues of C are
Noisy data
1. Form matrices Y, Y0 and Y1
2. Calculate
: rank-K SVD truncated pseudoinverse
: rank-K SVD truncated approximation
» vi and ui are left and right singular vectors
i is ith largest singular value
3. Calculate
4. K non-zero eigenvalues of C are
K
i
Hiii
Hii
K
i i 11
† ˆ and 1ˆ vuYuvY 10
†ˆ0Y
1Y
Kie ip ,,2 ,1 ,
10 YYC ˆˆ †
Kie ip ,,2 ,1 ,
UT Austin11
Biao Lu 11
LOW-RANK HANKEL APPROXIMATION
Problem in noisy data caseNoise destroys rank deficiencySVD truncation restores rank deficiency,
but destroys Hankel structure Low-rank Hankel approximation (LRHA)
[Cadzow, Sun & Xu, 1988]
Replaces each matrix cross-diagonal with average of cross-diagonal elements
Restores low rank after SVD truncation Iteratively apply SVD truncation and LRHA [Cadzow, Sun & Xu, 1988]
Modified Kumaresan-Tufts method (MKT) uses LRHA instead of SVD truncation
[Razavilar, Yi & Liu, 1996]
Hankel
low-rank
Hankel
low-rank
SVD truncation LRHA
Hankel
approximately low-rank
AAA
UT Austin12
Biao Lu 12
CONTRIBUTION #1: PROPOSED MATRIX PENCIL METHODS
Modified MP methods 1 and 2 in dissertation Modified MP method 3 (MMP3)
Maintain relationship between partitioned matrices
Y
SVD truncation
0Y 1Y
steps 3-4 in MP method
YLRHA
Ypartition
UT Austin13
Biao Lu 13
COMPUTER SIMULATION
321
21
1275.02757.03501.11
5056.05354.01
zzz
zz
zA
zBzH
Channel [Al-Dhahir, Sayed & Cioffi, 1997]
Zeros at 1.0275 and 0.4921Poles at 0.8464, 0.7146, and 0.2108
Parameters for matrix pencil methodsK = 3, N = 25, L = 17
Additive Gaussian noise with variance
SNR varied from 0 to 30 dB at 2 dB steps500 runs for each SNR value
Performance measure
210 2
1log 10SNR
2
10 original estimated MSE ,MSE
1log10 E
UT Austin14
Biao Lu 14
COMPUTER SIMULATION
Pole 1 at 0.8464
Pole 2 at 0.7146 Pole 3 at 0.2108
UT Austin15
Biao Lu 15
OUTLINE
Wireline channel equalization Wireline channel estimation
Channel modelingMatrix pencil methodsContribution #1: modified matrix pencil
methods for channel estimation Discrete multitone modulation
Minimum mean squared error equalizer Contribution #2: matrix pencil equalizer Maximum shortening SNR equalizer Contribution #3: fast implementation
» Divide-and-conquer methods
» Heuristic search
Summary and future research
UT Austin16
Biao Lu 16
MULTICARRIER MODULATION
Divide frequency band into subchannels
Each subchannel is ideally ISI free
Based on fast Fourier transform (FFT)
Orthogonal frequency division multiplexing
Discrete multitone (DMT) modulation
ADSL standards use DMT: ANSI 1.413, G.DMT and G.lite
etc.
Mag
nitu
de
Frequency
channel frequency response
subchannel
UT Austin17
Biao Lu 17
COMBAT ISI IN DMT SYSTEMS
Add cyclic prefix (CP) to eliminate ISI
Problem: Reduces throughput by factor of ADSL standards use time-domain equalizer
(TEQ) to shorten effective channel to (+1) samples
Goal: TEQ design during ADSL initializationLow implementation complexity“Acceptable” performance
vN
N
CP CP
samples i th symbolN samples
(i+1) th symbolN samples samples
UT Austin18
Biao Lu 18
MINIMUM MSE METHOD
MMSE method [Falconer & Magee, 1973][Chow & Cioffi, 1992][Al-Dhahir & Cioffi, 1996]
Constraints to avoid trivial solutionUnit tap constraint:Unit norm constraint: ADSL parameters: Lh = 512, Nw = 21,
= 32, Lh + Nw - - 2 Computational cost for a candidate delay
Inversion of Nw Nw matrix Eigenvalue decomposition of Nw Nw
matrix (or power method)
h w
z - b
kxkn
ky ke
2 MSE
)()1()0(
)()1(
kT
kT
T
Tw
E
bbb
Nww
xbyw
b
w
, ,1 ,0 ,1)( iib
1or 1 wb
bRRRRb
RwRb
yxyyxyxx
yyxy
MSE 1
T
TT
UT Austin19
Biao Lu 19
CONTRIBUTION #2:MATRIX PENCIL TEQ
From MMSE TEQ
MMSE TEQ cancels poles Matrix pencil (MP) TEQ
Estimate pole locations using a matrix pencil method on» Channel impulse response
» Received signal — blind channel shortening
Set TEQ zeros at pole locations
zW
zBzH
UT Austin20
Biao Lu 20
MAXIMUM SHORTENING SNR METHOD
Maximum shortening SNR (SSNR) method: minimize energy outside a window of (+1) samples [Melsa, Younce & Rohrs, 1996]
Simplify solution by constraining Computational cost at each candidate delay
Inversion of Nw Nw matrix Cholesky decomposition of Nw Nw matrix Eigenvalue decomposition of Nw Nw
matrix (or power method)
h wkx
knky
Aww
Bww
BwwwHHwhh
AwwwHHwhh
T
T
TTTT
TTTT
10
winwinwinwin
wallwallwallwall
log10SSNR
1BwwT
wallh
winh
UT Austin21
Biao Lu 21
MOTIVATION
MMSE method minimizes MSE both inside and outside window of (+1) samples
For each , maximum SSNR method requiresMultiplications:
Additions:
Divisions: Delay search
32
3
25
2
5
6
7wwwh NNNL
32
3
25
2
3
6
5wwwh NNNL
2wN
499020 wh NL
T9080.0,4189.0
dB 27.13SSNR
1067.6MSE 9wall
w T6913.0,7226.0
dB 90.29SSNR
1083.6MSE 11wall
w
MSE = 0.0019 with 1w
UT Austin22
Biao Lu 22
CONTRIBUTION #3:DIVIDE-AND-CONQUER TEQ
Divide Nw TEQ taps into (Nw - 1) two-tap filters in cascade
The ith two-tap filter is initialized as Unit tap constraint (UTC)
Unit norm constraint (UNC)
Calculate gi or i using a greedy approachMinimize : Divide-and-conquer TEQ
minimizationMinimize energy in hwall: Divide-and
conquer TEQ cancellation Convolve two-tap filters to obtain TEQ
ii g
1w
SSNR
1
i
ii
cos
sinw
UT Austin23
Biao Lu 23
CONTRIBUTION #3:DC-TEQ-MINIMIZATION (UTC)
Objective function
At ith iteration, minimize Ji over gi
Closed-form solution
Bww
AwwT
T
J SSNR
1
2
,3,2,1
2,3,2,1
,3,2
,2,1
,3,2
,2,1
2
2
11
11
iiiii
iiiii
iii
iii
iii
iii
iTi
iTi
i gbgbb
gagaa
gbb
bbg
gaa
aag
J
Bww
Aww
iiiiiiiiiiii
iiiiiiii
iiiii
babababababaD
baba
D
baba
babag
,2,1,1,2,3,2,2,32
,3,1,1,3
,3,2,2,3,3,2,2,3
,3,1,1,32,1
4
22
UT Austin24
Biao Lu 24
CONTRIBUTION #3:DC-TEQ-CANCELLATION (UTC)
Objective function to cancel energy in hwall
At ith iteration, minimize Ji over gi
Closed-form solution
wallwallhhTJ
1~
2
11wallwall
,,2 ,,,2 ,1
,1~~~~
ih
Skiii
Ti
LS
khgkhJ
hh
Ski
Skii
ikh
khkhg
)1(~
)(~
)1(~
21
11
UT Austin25
Biao Lu 25
CONTRIBUTION #3:DC-TEQ-MINIMIZATION (UNC)
Each two-tap filter
At ith iteration, minimize Ji over i
Calculate i in the same way as gi for DC-TEQ-minimization (UTC)
ii
iii
i
ii
1
sinsincos
1sin
cos
sinw
iii
iii
iii
iii
ii
ii
iiii
ii
ii
iiii
iTi
iTi
i
bb
bb
aa
aa
bb
bb
aa
aa
J
11
11
1sin1sin
1sin1sin
,3,2
,2,1
,3,2
,2,1
,3,2
,2,1
,3,2
,2,1
Bww
Aww
UT Austin26
Biao Lu 26
CONTRIBUTION #3:DC-TEQ-CANCELLATION (UNC)
Each two-tap filter
At ith iteration, minimize Ji over i
Closed-form solution
i
ii
cos
sinw
1~
2
11wallwall
,,2 ,,,2 ,1
,cos1~
sin~~~
ih
Skiiii
Ti
LS
khkhJ
hh
Skii
Skii
ii
khkhbkhkha
ba
a
ba
a
)(~
)1(~
,1~~
415.0cos,
415.0sin
1121
21
22
2
22
2
UT Austin27
Biao Lu 27
COMPUTATIONAL COMPLEXITY
Computational complexity for each candidate for G.DMT ADSL
Lh = 512, = 32, Nw = 21
Divide-and-conquer TEQ design methods vs. maximum SSNR methodReduce multiplications and additions by a
factor of 2 or 3Reduce divisions by a factor of 7 or 22Reduce memory by a factor of 3Avoids matrix inversion, and eigenvalue
and Cholesky decompositions
Method Memory(words)
MaximumSSNR
120379 118552 441 1899
DC-TEQ-mini-mization(UTC)
53240 52980 60 563
DC-TEQ-can-cellation(UNC)
42280 42160 20 555
DC-TEQ-can-cellation(UTC)
41000 40880 20 554
UT Austin28
Biao Lu 28
KNOWN CHANNEL
Dedicated data channel
Carrier-Serving-Area (CSA) ADSL channel 1
UT Austin29
Biao Lu 29
UNKNOWN CHANNEL
Dedicated data channel
Carrier-Serving-Area (CSA) ADSL channel 1
UT Austin30
Biao Lu 30
HEURISTIC SEARCH DELAY
Estimate optimal delay before computing TEQ taps
Computational cost for each Multiplications:Additions:Divisions: 1
Reduce computational complexity of TEQ design for ADSL by a factor of 500 over exhaustive search
h
h
original of windowa outsideenergy
original of windowa insideenergy maxargratio
2hLhL
UT Austin31
Biao Lu 31
HEURISTIC SEARCH
Maximum SSNR method for CSA DSL channel 1
DC-TEQ-cancellation (UTC) for CSA DSL channel 1
UT Austin32
Biao Lu 32
SUMMARY
Channel estimation by matrix pencil methodsNew methods to estimate channel poles by
applying low-rank Hankel approximation to multiple matrices [Lu, Wei, Evans & Bovik, 1998]
Time-domain equalizer channel shorteningMatrix pencil TEQ [Lu, Clark, Arslan & Evans, 2000]
» From known channel impulse response
» From received signal: blind channel shortening
Reduce computational cost [Lu, Clark, Arslan & Evans, 2000]
» Divide-and-conquer TEQ minimization method
» Divide-and-conquer TEQ cancellation method
» Heuristic search for delay
Other contributions: cascade two neural networks to form a channel equalizer
[Lu & Evans, 1999]
Multilayer perceptron to suppress noise Radial basis function network to equalize
the channel
UT Austin33
Biao Lu 33
FUTURE RESEARCH
Discrete multitone systemsMaximize channel capacity
» Optimize channel capacity at TEQ output
» Jointly optimize a TEQ with other blocks
Frequency–domain equalizersTEQ to shorten time-varying channels
» Fast and accurate channel estimation
» Convert time-varying channels to additive white Gaussian noise channel
Reduce computational complexityFast training for neural networksParallelize matrix pencil method
UT Austin34
Biao Lu 34
ABBREVIATIONS
ADSL: Asymmetrical Digital Subscriber Line CP: Cyclic Prefix CSA: Carrier-Serving Area DC: Divide-and-Conquer DMT: Discrete Multitone DSL Digital Subscriber Line FFT: Fast Fourier Transform IIR: Infinite Impulse Response ISI: Intersymbol Interference LRHA: Low-Rank Hankel Approximation MKT: Modified Kumaresan-Tufts MLP: Multilayer Perceptron MMP: Modified Matrix Pencil MMSE: Minimum Mean Squared Error MP: Matrix Pencil RBF: Radial Basis Function SNR: Signal-to-Noise Ratio SSNR: Shortening Signal-to-Noise Ratio SVD: Singular Value Decomposition TEQ: Time-domain Equalizer UNC: Unit Norm Constraint UTC: Unit Tap Constraint
UT Austin35
Biao Lu 35
NEURAL NETWORK EQUALIZERS
Equalization is a classification problem Feedforward neural network equalizers
Multilayer perceptron (MLP) equalizer» Has to be trained several times
» Reduces additive uncorrelated noise
Radial basis function (RBF) equalizer» The number of hidden units increases
exponentially with the number of inputs
» Adapts to local patterns in data
Cascade MLP and RBF networksUse MLP to suppress noiseUse RBF to perform equalization
UT Austin36
Biao Lu 36
PROBLEMS FROM NN EQUALIZER
Computational cost: training NN takes time Number of symbols used in training [Mulgrew, 1996]
where
M : number of constellations
Lh : length of channel impulse response
Nin: number of neurons in the input layer
e.g., M = 4, Lh = 8, Nin = 3 means that
number of symbols = 1,048,576 Channel length is unknown
GoalsEstimate channel impulse response —
Lh can be knownShorten channel impulse response to be
less than Lh
1 inh NLM
UT Austin37
Biao Lu 37
BACKUP INFORMATION
Derivation from Hap(z) to hap(n)
K
i
npi
K
i p
iK
i
p
i
ii
eanh
ze
a
zezH
1ap
1 1
1
1ap
11
1
UT Austin38
Biao Lu 38
KUMARESAN-TUFTS (KT) AND MODIFIED KT METHOD
KT-method: noisy data
1. Form matrix
2. Solve
3. Form
4. Calculate zeros of B(z)
5. All the zeros outside unit circle gives
Modified KT (MKT) method: apply LRHA
to matrix A before step 2
1
1
0
2
1
11
132
21
LNy
y
y
Lb
b
b
NyLNyLNy
Lyyy
Lyyy
A
Liib ,,2 ,1 ,
LzLbzbzbzB 21 211
Lie ip , 2, ,1 ,*
UT Austin39
Biao Lu 39
COMPARISON BETWEEN MMP3 AND MKT
Common proceduresIterative LRHASVD-truncated pseudoinverse
MMP3 onlyMatrix partitionEigenvalue decomposition
MKT onlySolve equation
UT Austin40
Biao Lu 40
CONTRIBUTION #1:PROPOSED MP METHODS
Modified MP method 1 (MMP1)
Noise may corrupt and to lose the connection
0Y 1Y
partition
Y
0Y
0Y 1Y
0Y 1Y
Steps 3-4 in MP method
1Y
LRHA
SVD truncation
LRHA
SVD truncation
UT Austin41
Biao Lu 41
CONTRIBUTION #1:PROPOSED MP METHODS
Modified MP method 2 (MMP2)
SVD truncation may destroy the connection between Y0 and Y1
SVD truncation
Y
0Y 1Y
0Y
0Y
Joint LRHA
Y
partition
SVD truncation
1Y
partition
Y
Step 3-4 in MP method
1Y
UT Austin42
Biao Lu 42
COMPUTER SIMULATION
Data model
whereK=2, N=25, L=17, A1= A2= 1 pi = -di+ j2 fi , i = 1, 2
where d1= 0.2 and d2= 0.1,
f1= 0.42 and f2= 0.52 w(n) is complex zero-mean white Gaussian
noise with variance 2
Signal-to-noise ratio (SNR)
SNR varied from 5 to 25 dB at 2 dB step500 runs for each SNR value
Performance measure
1,,1 ,0
),()()()(1
Nn
nweAnwnhny npK
ii
i
2102
1log 10
SNR
210 original - estimated MSE ,
MSE
1log10 E
UT Austin43
Biao Lu 43
ESTIMATION OF DAMPING FACTORS
d1 = 0.2
d2 = 0.1
UT Austin44
Biao Lu 44
ESTIMATION OF FREQUENCIES
f1 = 0.42
f2 = 0.52
UT Austin45
Biao Lu 45
PREVIOUS WORK
Maximum channel capacityBased on geometric SNR
» Nonlinear optimization techniques [Al-Dhahir & Cioffi, 1996, 1997]
» Projection onto convex sets [Lashkarian & Kiaei, 1999]
Based on model of signal, noise, ISI paths [Arslan, Evans & Kiaei, 2000]
» Equivalent to maximum SSNR when input signal power distribution is constant over frequency
lbits/symbo SNR
1log geom2DMT
Nb
UT Austin46
Biao Lu 46
COMPUTER SIMULATION
Simulation parameters
Parameters DedicatedChannel
CSA DSLchannels 1-8
Sampling rate 300 kHz 2.208 MHz
Samples per symbol 512 512
Power of AWG noise 200 dBm/Hz 113 dBm/Hz
Transmitter power 1 W 1 W
Transmittertermination resistance
50 50
FFT size 512 512
Cyclic prefix 16 32
Bit error rate 10-7 10-7
Field margin 6 dB 6 dB
Coding gain 0 dB 0 dB
Number of runs 50 50
UT Austin47
Biao Lu 47
FREQUENCY RESPONSE OF A TRANSMISSION LINE
Model as a RC circuit
Characteristic impedance of the line
R L
C
Z0
fLjZL 2
fCjZC 2
1
fL
Rj
C
LZ
410
UT Austin48
Biao Lu 48
SSNR VS. DATA RATE
CSA DSL channel 1
SSNR = 40 dB