Upload
stephan
View
236
Download
0
Tags:
Embed Size (px)
DESCRIPTION
multirate digital signal processing
Citation preview
Multi rate Digital Signal
ProcessingProcessing
1
Syllabus
– Basics of Multirate systems and its application
– Up-sampling and Down-Sampling
– Fractional Sampling rate converter
– Poly-phase decomposition
– Efficient realization of Multirate systems– Efficient realization of Multirate systems
– Uniform filter banks
• Its implementation using poly-phase decomposition
– Two channel Quadrature Mirror Filter Banks
– Perfect Reconstruction
– M-channel PR QMFB
2
Basics of Multirate systems and its application
3
Introduction
– Sampling rate conversion � Changing the
sampling rate of a signal
• Required in many practical applications of DSP
• Change in sampling rate can be an increase or decrease
– Example on need for Sampling rate conversion
• Communication systems may transmit and receive
different types of signals � facsimile, speech, video etc
– Each of these signal is to be processed at a different rate
corresponding to its band width
4
Introduction
– Multi-rate Digital signal processing systems
• Systems that employ multiple sampling rates in the
processing of digital signals
• Example: Playing CD music in a studio
– CD music has a data rate of 44.1 kHz
– To play this in a studio that handles data at 48 KHz rate, CD
data rate must be increased to 48 KHz
5
Advantages of Sampling rate conversion
• Computational efficiency with sampling rate reduction
– Number of computations is less at lower sampling rate
– Relevant only when sampling rate reduction does not cause
loss of any desired information
• Achieve compatibility with another system
– Playing CD music in a studio
6
Operations in Multi-rate DSP
• Decimation that reduces sampling rate � Compresses
data to retain only desired information
• Interpolation that increases sampling rate
7
Concept of sampling rate conversion
– Resampling after reconstruction
• Pass the signal through a DAC and LPF
• Sample the analog signal at the desired rate
• Advantage of this method
– New sampling rate can be arbitrarily selected
– Do not need any special relationship to the old sampling rate
• Disadvantage is added Signal distortion
– In DAC (signal reconstruction) and
– In ADC (quantization)
8
Concept of sampling rate conversion
– Reason for the disadvantage
• In the conceptual method, signal is reconstructed
• Noise gets added in signal reconstruction
– The disadvantage is overcome if sampling rate – The disadvantage is overcome if sampling rate
conversion is performed in the digital domain
• Anti aliasing can be done in digital domain with linear
phase and sharper magnitude response
• Convert the signal to analog form only when it is
essential
9
Applications in Industry
1. Reduce cost of analog anti aliasing filters
• We can sample the analog signal at a rate much higher
than that specified by sampling theorem
– This calls for a much simpler anti aliasing filter
– Once in digital form, the signal can be reduced to the desired – Once in digital form, the signal can be reduced to the desired
rate using multi rate approach
2. Efficient implementation of DSP functions
• Example: Narrow band digital FIR filters
– Direct implementation requires a large number of coefficients
to meet their frequency response specifications
– Multi-rate techniques can reduce the Number of coefficients
10
Applications in Industry
3. Dispense with expensive high resolution ADCs
• Advantage of High resolution ADCs is lesser
quantization noise
• With multi-rate techniques, we can achieve similar
performance using cheap ADCsperformance using cheap ADCs
– Oversampling spread the Quantization noise over a wider
frequency range
– Signal bandwidth is reduced with Sampling rate reduction,
simultaneously reducing the in-band noise too
11
Theory of sampling rate conversion
– Mathematical formulation of the theory:
• Developed around the concept of resampling after
reconstruction
– An aid to understand the sampling rate conversion process in
the digital domainthe digital domain
2)( ofBandwidth
1);( signal timediscrete Generates
)(
x
xxx
x
Ftx
FTnTx
FrateNyquistatsampledistxsignaltimeContinuous
=
=⇒
12
Theory of sampling rate conversion
{)(
sin
)(
,'
2,
,0
FFT
OtherwiseFx FG
t
T
t
tg
isfunctionionInterpolatIdealtheoremsShannonFrom
xx
=→←=
=
≤
π
π
summation infinite needsit since practical,not istion reconstruc Ideal
)()()(*)()(
)()()( )( samples theingInterpolat
xn
xx
x
x
nTtgnTxtgnTxty
txtytsreconstructgwithnTx
T
t
−==
=
∑∞
−∞=
π
13
Theory of sampling rate conversion
sequence desired theof samples )(
)()()(
instants at time )( evaluating is conversion rate Sampling
⇒
−=
=
∑∞
−∞=
y
xyn
xy
y
mTy
nTmTgnTxmTy
mTtty
samplesinput andoutput ofinstant Time &
andfunction tion Reconstruc )(
sequenceinput theof Samples )(
need we,)( compute To
sequence desired theof samples )(
→→→
⇒
xy
x
y
y
nTmT
tg
nTx
mTy
mTy
14
Theory of sampling rate conversion
≥==
>
⇒<
yxyy
y
xy
FFFXifonlymTxmTy
Ftx
FF
2
,2
minFfor 0)(),()(
2/)( of components
frequency out theFilter if occurserror Aliasing
( ){ }
−=∆
=
−∆+=
−=−=
∑
∑∑
∞
−∞=
∞
−∞=
∞
−∞=
x
y
x
ym
x
ym
mmxn
x
x
yx
nxxy
nxy
T
mT
T
mT
T
mTkwhere
nkTgnTx
nT
mTTgnTxnTmTgnTxmTy
&
)(
)()()()(
15
Theory of sampling rate conversion
( ){ } ( ){ }
{ } )()()()(
)()(
fromsummation ofindex Changing
kxmxmxxmmy
m
nTgnTxTngnTx
TkkxTkTgTkymTy
nkkton
∗=∆+∗=
−∆+=∆+=
−=
∑∞
−∞=
• This is the fundamental equation for the discrete time
implementation of sampling rate conversion
• Process in this Discrete time sampling rate conversion
is demonstrated in the figure below
{ } )()()()( xmxxmx nTgnTxTngnTx ∗=∆+∗=
16
Theory of sampling rate conversion
17
Theory of sampling rate conversion
• mTy = (km + ∆m)Tx
�For each m, the fractional interval ∆mTx determines the
Impulse response coefficients
�km specifies the input samples to compute y(mTy)
• This sampling rate conversion needs different impulse
responses gm(nTx) for each output sample y(mTy)
– Thus we have a Discrete time linear and time varying system
• The conversion process becomes inefficient when the
number of required values of gm(nTx) is large
18
Theory of sampling rate conversion
19
Theory of sampling rate conversion
number rational a ,
when simplified gets process conversion rate sampling The
mDmDmTmTmThen
I
D
T
T
yy
x
y
−=
−=∆
=
( )
( )
valuesof setsdistinct only can take )(,
,1
,.....2
,1
,0 valuesonly the can take
11
InTgSoI
I
IIm
mDI
II
mDmD
I
I
mD
I
mD
T
mT
T
mTmThen
xm
I
x
y
x
y
−∆
=
−=
−=
−=∆
20
Theory of sampling rate conversion
system DT varyingly timeperiodical andlinear a have We
,...2,1,0),()(
period with periodic is)(
±±== + rnTgnTg
InTg
xrImxm
xm
• This system is much simpler than the continuously
time varying DT system
• The two cases of sampling rate conversion are
– Decimation or Down-sampling
– Interpolation or Up-sampling
21
Downsampling by an integer factor D
nTmTgnTxmTy
DTTI
I
D
T
T
xyxy
xy
x
y
)()()(
& 1Then
D integer,an by ngDownsampliFor
−=
==
==
∑∞
mallfornTgI
DTT
Ttgandtx
TnmDgnTxmDTyT
nTmTgnTxmTy
xmx
xy
x
xn
xxy
xyn
xy
)( oneonly have weand0 1, Since
of incrementsat shifted is response Impulse
period with sampled are)()(
))(()()(,for ngSubstituti
)()()(
=∆=
=
−=
−=
∑
∑
∞
−∞=
−∞=
22
Downsampling by an integer factor
23
Process of Downsampling
Upsampling by an integer factor I
)()(get Then we
So 1,D integer,an by lingFor Upsamp
−=
=
==
∑∞
−∞=nT
I
TmgnTx
I
TmymTy
I
TT
xx
nx
xy
xy
1 .., 1, 0, mfor )( responses Impulse of sets
)( of resampling requires shifting fractionalEach
of incrementsat shifted is response Impulse
period with sampled are)()(
−=⇒
=
InTgI
tg
I
TT
Ttgandtx
xm
xy
x
24
Upsampling by an Integer factor I
x(t)
Two impulse responses
required – for odd
numbered outputs and
even numbered outputsy(t)
(a)
25
x(t)
y(t)
Only one impulse
response required
(b)
(c)
Upsampling by an Integer factor I
– Upsampling can also be achieved in the method
shown in (c) above
• Determine an impulse response g(nTy) as shown
• Create a new sequence v(nTy) by inserting (I-1) zero
valued samples between successive samples of x(nT )valued samples between successive samples of x(nTx)
• Compute y(nTy) as the convolution of the sequences
g(nTy) and v(nTy)
26
Decimation by an integer factor D
Decimation
FFinzerononbeXLet
sequencesignalthebeXnxLet
x
F
πωω
ω
≤≤≤−
→←
reduction rate sampling is2
;0)(
)()(
DD
FF
nxD
Fnx
nx
x
x
πω ==
⇒
maxmax
th
or 2
to)( ofBW thereducemust wealiasing, avoid To2
at frequency foldingwith ),( of version aliasedan get We
)( of sample Devery Select
27
Decimation by an integer factor D
• Decimation process consists of a digital anti aliasing
filter h(k) and a Down sampler
– Down sampling is a data compression operation, where (D-1)
samples are discarded for every D samples of signal v(n)
28
Decimation by an integer factor D
{
range in the )( of spectrum theeliminatesIt
)()(
sticscharacteri ideal thehas LPF
,1
,0
DX
Hnh DOtherwiseD
F
<<
=→←≤
πωπω
ωπω
)()()()(
r,Downsample ofOutput
)()()( LPF ofOutput
range in the )( of spectrum theeliminatesIt
0
0
mykmDxkhmDv
knxkhnv
DX
k
k
=−=
−=⇒
<<
∑
∑
∞
=
∞
=
πωω
29
Decimation by an integer factor D
Overall Decimation operation is not Time invariant
• Low pass Filtering is a Linear Time Invariant operation
• The Down sampler is not time invariant
– x(n) produces y(m) does not imply that x(n-n0) � y(m-n0),
except when n is a multiple of Dexcept when n0 is a multiple of D
• Thus overall Decimation operation is not Time invariant
30
Decimation by an integer factor D
)()()(
)(
)( sequenceoutput of sticscharacteridomain Frequency
,...2,,0),(
,0
npnvnv
nvsquencetheLet
my
DDnnv
Otherwise
== ±±=
)()()()(
1 )(DFS, of In terms
D period with impulses of train periodic a is )(
)()()(
1
0
2
mDpmDvmDvmy
eD
np
np
npnvnv
D
k
D
knj
==
=
=
∑−
=
π
31
∑
∑∑
∞
−∞=
−
∞
−∞=
−∞
−∞=
−
=
==
m
D
m
m
m
m
m
zmpmv
zmDpmDvzmyzY
)()(
)()()()(
32
=
∞
−∞=
−−
=∑ ∑
m
D
mD
k
D
kmj
zeD
mvzY
mp
1
0
21
)()(
),(for ngSubstituti
π
=
=
=
−−−
=
−−
=
−−−
=
∞
−∞=
∑
∑∑ ∑
DD
kj
DD
kjD
kD
DD
kjD
k
m
DD
kjD
k m
zeXzeHD
zeVD
zemvD
12121
0
121
0
121
0
1
1)(
1
ππ
ππ
33
( )( ) πωπωω
ω
−
−=
⇒=
∑−
=
yD
k
yDy
y
D
kX
D
kH
D
zYmy
aliasing eliminatecan filter suitableA
221Y
circleunit on the )( Y )( of Spectrum
1
0
( )
πωπω
ωππω
ωωωω
≤≤≤≤
===
=
=
yx
xxy
y
yyyDy
toD
ORDF
FD
F
F
DX
DDX
DH
D
00 stretches ngDownsampli
22
11YThen
aliasing eliminatecan filter suitableA
34
Decimation in detail
35
Decimation by factor D = 3
36
Decimation by factor D
37
Decimation by factor D
38