Upload
others
View
8
Download
0
Embed Size (px)
Citation preview
LRA
DSP
LRA
DSP MultiMulti--Rate DSP Rate DSP
Applications:
Oversampling, Undersampling,
Quadrature Mirror Filters
Professor L R Arnaut © 1
LRA
DSP
LRA
DSP MultiMulti--Rate DSP Rate DSP
Oversampling
Professor L R Arnaut © 2
LRA
DSP
LRA
DSP
Professor L R Arnaut © 3
Optimal Sampling vs. Optimal Sampling vs.
OversamplingOversampling
Sampling at Nyquist rate Fs =2FB
Allows perfect reconstruction in principle, but…
Pre-sampling anti-aliasing filter must have very steep roll-off:
High-order analogue filter: expensive, difficult, imprecise, large phase distortion, …
Sampling at Fs >>2FB & decimation to 2FB
Larger separation between images easierfiltering of aliases (lower-order filter)
Cheaper analogue component; easier digital than analogue VLSI filters; greater digital complexity
LRA
DSP
LRA
DSP Oversampling Noise ReductionOversampling Noise Reduction
Quantisation step (b-bit ADC, range R):
Noise power density (per unit sampling bandwidth):
Total in-band noise power:
Low Pin for high Fs:
b
RQ
2
]W/Hz[62/
12/
2/
222
sss
NN
F
Q
F
Q
Fp
FsFB FB’ Fs
’
pN
N2
pN’
Pin’
Pin
2'2' NinBs PFF
2/12
)2/(
6d)(
22
0s
Bb
s
BFNin
F
FR
F
FQffpP B
Professor L R Arnaut © 4
LRA
DSP
LRA
DSP
Oversampling:Oversampling:
Effective ResolutionEffective Resolution
Equivalent -bit ADC operating at Fs=2FB
giving same noise power as b-bit ADC operating at Fs>>2FB over same range R:
i.e.,
M = oversampling ratio, -b = resolution increase
2/212
)2/(
2/12
)2/( 22
B
B
s
Bb
F
FR
F
FR
2
)(log
2
)2/(log 22 Mb
FFb Bs
Professor L R Arnaut © 5
LRA
DSP
LRA
DSP Oversampling RatioOversampling Ratio
Example:
M=106: =b+10, i.e., standard 16-bit ADC @ Fs=2FB has equivalent resolution w.r.t. noise power as an oversampling6-bit ADC @ Fs=2 106FB
or as oversampling 11-bit ADC @ Fs=2000FB
Thus, 0.5 bit length reduction of ADC per doubling of MProfessor L R Arnaut © 6
LRA
DSP
LRA
DSP (Sigma-Delta) Converters
(digital)Analogue
integrator
1-bit
ADC
Mth band
Digital LPF
M
1-bit
DAC
-
++(analogue)
(2MFB) (2MFB) (2MFB)
1 b
2FB
quantiser M downsampler
(b-bit multipliers)
-High-rate oversampling allows for differential encoding: only 1 bit needed to quantify change between input and delayed output of 1-bit ADC, for closely spaced consecutive samples in output of quantiser (bitmap of sample increments)
-Mth order alias digital LPF eliminates out-of-band quantisation noise (sharp cut-off); heavy comp.; remedy: perform multiplications at later stage (down-rate 2FB) instead
-Long word length of digital o/p determines overall oversampling rate
Analogue
integrator
z -1
+x(n)-y(n-1)w(n)
Professor L R Arnaut © 7
LRA
DSP
LRA
DSP
Professor L R Arnaut © 8
ConvertersConverters
Accumulator (integrator) & quantiser outputs:
noise transfer function:
Output psdf (one-sided spectrum):
z -1
+w(n)
z -1
-
++x(n)
quantiser
y(n)+
eADC(n)
eDAC(n)
)1()1()1( nenwny
)1()1()()( nwnynxnw
)]1()([)()()()( nenenxnenwny
11)( zzH
)(2
sin4)()( 22
fpT
fpeHfp Ns
NTj
ys
LRA
DSP
LRA
DSP OversamplingOversampling
Precise psdf of output noise depends on pdf & spectral characteristics of {x(n)}
Assume: {e(n)} is random,
uncorrelated, white:
psdf of output noise:
Professor L R Arnaut © 9
z -1
+w(n)
z -1
-
++x(n)
quantiser
y(n)+
eADC(n)
eDAC(n)
2/
12/
2/
22
ss
NN
F
Q
Fp
s
sy
F
QTffp
22
3
)(sin2)(
LRA
DSP
LRA
DSP
Oversampling:Oversampling:
Performance ComparisonPerformance Comparison
For efficient oversampling :
Output noise power for modulator:
Improvement over standard oversampling:
Professor L R Arnaut © 10
),1( sFfM
3
22222
3
2
3
)/(2)(
ss
sy
F
fQ
F
QFffp
3
322
09
2d)(
s
BFyy
F
FQffpP B
dB)](log2017.5[3
log10log10 102
2
1010 MM
P
P
y
in
LRA
DSP
LRA
DSP MultiMulti--Rate DSP Rate DSP
Undersampling
Professor L R Arnaut © 15
LRA
DSP
LRA
DSP NyquistNyquist ConditionCondition
Alias-free (sub)sampling of (discrete) function
Spectrum (Fourier transformation): convolution
Thus,
No spectral overlap (aliasing, stroboscopy) iffProfessor L R Arnaut © 16
Bs FF 2
kss
ksTn kTtkTxkTttxttxtx
s
)()()()()()()(
ks
s
ks
s
n
dkFfXT
dkFT
fXfX
)()(1
1)()(
ks
s
n kFfXT
fX )(1
)(
LRA
DSP
LRA
DSP
Professor L R Arnaut © 17
UndersamplingUndersampling: Aliasing: Aliasing
LRA
DSP
LRA
DSP UndersamplingUndersampling: Baseband: Baseband
Shannon sampling theorem:
Applies to baseband signals (DC-coupled): FB islargest frequency component in signal
Motivation: avoid spectral overlap of baseband frequency responses that are periodically continued due to sampling operation
For bandpass (narrowband modulated) signals (e.g., radio- and optical communications, IF filters, etc.): condition is too conservative: large spectral gaps occur because Fc >> |FB - Fc |
Professor L R Arnaut © 18
Bs FF 2
LRA
DSP
LRA
DSP UndersamplingUndersampling:: BandpassBandpass
Aliasing of bandpass signal is avoided if baseband can be folded periodically around carrier frequency without causing overlap
Range of permissible sample frequencies:
i.e.,
Yields additional permissible lower samplingrates for narrowband signals without aliasing
Practically useful (slower computations)Professor L R Arnaut © 19
1
22
n
FF
n
F ls
u
lu
u
FF
Fn1
LRA
DSP
LRA
DSP UndersamplingUndersampling: Applications: Applications
58.7
(n=4)(n=3)
Fu
=108f (MHz)Fl
=88
20 54 2 Fu
=216
2 Fl
=176
…(n=1)(n=2)
72
Professor L R Arnaut © 20
Example 1: digitization of analogue FM audio
Fl = 88 MHz, Fu = 108 MHz (nonzero gap)
•n=1: classical ( Nyquist rate)
•n=2: between modulated (Iu=1) signal and doubled (Il=2) signal
•n=3: Du=2/3, Il=1
•n=5:
51 n
MHz44MHz2.43 sF MHz1105.2...2MHz4.86 sF
LRA
DSP
LRA
DSP MultiMulti--Rate DSP Rate DSP
Quadrature Mirror Filters
for Subband Coding
Professor L R Arnaut © 21
LRA
DSP
LRA
DSP Subband CodingSubband Coding
Problem statement:
Efficient transmission of realistic speech or video signals
Contain most energy at relative low frequencies (time/space)
Coding scheme to be tailored to assign more bits to LF band
Solution:
Subband coding:divide total frequency band in unequal subbands;
narrowest subband for interval with highest energy (equalization ofpower across band)
each subband is encoded separately
is alternative to companding (pre-distortion) Professor L R Arnaut © 22
LRA
DSP
LRA
DSP
Example:
(approximate equalisation)
- Multi-rate conversion by factor I/D after each frequency subdivision (LPF/HPF)
- Reduced bitrate of digitized signal (bandwidth compression) due to nonuniform coding (variable number of bits per sample)
Professor L R Arnaut © 23
Subband CodingSubband Coding
S(f)
fmfm/29fm/329fm/960
1 bit/sample5 3 2
(D=2)(I/D=9/16)(I/D=1/3)
LRA
DSP
LRA
DSP
Implementation:
Brickwall Filter: Quadrature Mirror Filter (QMF):Physically aliasing for decimated subbands can be removed
unrealizable by judicious choice of H0( ) and H1( )
Professor L R Arnaut © 24
Subband CodingSubband Coding
LRA
DSP
LRA
DSP
Implementation (analyzer / synthesizer):
Professor L R Arnaut © 25
TwoTwo--Channel QMFChannel QMF
(for I/D=1/2)
LRA
DSP
LRA
DSP
QMF Analyzer:
TwoTwo--Channel QMF: AnalysisChannel QMF: Analysis
2D,2
exp2
exp1 1
0
11
00
D
k
DD zkD
jXzk
D
jH
DzX
22222
1000 XHXHX
22222
1111 XHXHX
Professor L R Arnaut © 26
LRA
DSP
LRA
DSP
QMF Synthesizer:
Cascaded QMF analyzer-synthesizer:
Professor L R Arnaut © 27
TwoTwo--Channel QMF: SynthesisChannel QMF: Synthesis
2I,)( 00
IzYzV
22 1100 YGYGY
1100 , XYXY
XGHGHXGHGHY 110011002
1
2
1
Aliasing (k=1)
LRA
DSP
LRA
DSP
Elimination of aliasing for any input signal:
results in time-invariant filter
example: alias-free symmetric subband coding
Professor L R Arnaut © 28
QMF AntiQMF Anti--AliasingAliasing
e.g.
01100 GHGH
0110 , HGHG
0100 , HGHG
LRA
DSP
LRA
DSP
Distortion-free & alias-free reconstruction:
Example: symmetric subband
i.e., independent of (all-passfilter), but may exhibit phase distortion!
It can be shown: linear-phase FIR QMF causes amplitude distortion
Professor L R Arnaut © 29
QMF Perfect ReconstructionQMF Perfect Reconstruction
2),exp(1100 DjkDGHGH
)exp(0110 jkDHHHH
)exp(20
20 jkDHH
2
0
2
0 HH
LRA
DSP
LRA
DSP
M branches; M in analyzer, M in synthesizer
Output kth analyzer branch (BPF+D):
Output synthesizer (I+BPF):
Professor L R Arnaut © 30
MM--Channel QMF BankChannel QMF Bank
1
0
M
k
M
kk zYzGzY
)(,2
exp2
exp1 1
0
/1/1 DMM
mjzX
M
mjzH
MzX
M
m
MMkk
1
0
1
0
2exp
2exp
1 M
m
k
M
k
kM
mjzX
M
mjzH
MzGzY
,2
exp1
0
M
m
mM
mjzXzL
1
0
)(2
exp1 M
k
kkm zGM
mjzH
MzL
LRA
DSP
LRA
DSP
Alias-free QMF:
Distortion-free & alias-free QMF:
independent of (all-pass filters)
Professor L R Arnaut © 31
MM--Channel QMFChannel QMF
XLY )()( 0)(,0
2exp
1
1
zXM
mjzXzL
M
m
miff
i.e. 11,0 MmzLm
0L