Stability
X(s) Y(s)
Y(s) =
8X(s)
G0 +
2
(s + 2) (s 2) +
System has 2 poles: points where Y(s) -> G1
at s = +2 and s = -2 - x3
+
G2
If all poles are in region where s < 0, system is stable
x7
in Fourier language s = j
can only have positive frequencies, ie s > 0 Im(s)
so this system is unstable
will see why from solution Re(s)
Pole location s could have imaginary part stable unstable
=> oscillatory solution
[Link]@[Link] [Link]/Instrumentation/ 9 13 December, 2001
Response to step
x(t) = u(t) = 1, for t > 0 so X(s) = 1/s
8X(s) 8 A B C D
Y(s) = = = + + +
(s + 2)2 (s 2) s(s + 2)2 (s 2) s (s + 2) (s + 2)2 (s 2)
Solve by expressing as partial fractions
Find A, C, D by taking limit s -> a of (s+a)NY(s) N is highest power term
Find A by multiplying by s
Bs Cs Ds
RHS limit
23 KsY(s) = A + + + =A
1
s >0
(s + 2) (s + 2) 2
(s 2) A = 1
8 8
LHS limit
1 23 KsY(s) = = = 1
2
s >0 (s + 2) (s 2) 4(2)
Find C by multiplying by (s+2)2
D(s + 2)2
C=1
2 2
RHS limit
1 23 K(s + 2) Y(s) = A(s + 2) + B(s + 2) + C + =C
s > 2
(s 2)
2 8 8
LHS limit K(s + 2) Y(s) = = =1
1 23
s > 2
s(s 2) (2)(4) similarly D = 1/4
[Link]@[Link] [Link]/Instrumentation/ 10 13 December, 2001
Step response... continued
8X(s) 8 A B C D
Y(s) = = = + + +
(s + 2)2 (s 2) s(s + 2)2 (s 2) s (s + 2) (s + 2)2 (s 2)
Find B by multiplying by (s+2)2, differentiate, then take limit
d d 8 1 1
2
RHS (s + 2) Y(s) = [ ] = 8 +
ds ds s(s 2) s (s 2) s(s 2)
2 2
1 1 1 1 3
1 23(8 2
limit + ) = 8
2
+
4(4) (2)(4 ) 4
2
=
s > 2 s (s 2) s(s 2)
d
(s + 2)2 Y(s) =
d 3
LHS limit
1 23 K
ds ds
B(s+ 2) = B B=
s > 2
4
now have the solution in s
1 4 3 4 1
Y(s) = + + +
4 s (s + 2) (s + 2)2
(s 2)
[Link]@[Link] [Link]/Instrumentation/ 11 13 December, 2001
Finally solution
1 4 3 4 1
Y(s) = + + +
4 s (s + 2) (s + 2)2
(s 2)
n!
Recall F(s) = is LT of f(t) = tne-at
n+1
(s + a)
1
and F(s) = is LT of u(t) = unit step
s
x(t)=u(t) + y(t)
24te-2t
+
y(t) =
1
4
[ 4u(t) + 3e2 t + 4te 2 t + e2 t ] d/dt
3
t t 1 2 t
- x3
y(t) = u(t) + e + te + e
2 2
4 4 t>0 +
..dt
Can now see the reason for instability x7
term with e2t
By the way: this problem could equally well be solved with Fourier
[Link]@[Link] [Link]/Instrumentation/ 12 13 December, 2001
z transforms
Laplace transform applies to continuous signals in time domain
Extend idea to discrete, sampled signals
from Laplace Transform definition
F(s) = 0
f(t).[Link],
sample waveform f(t) at intervals t
sampled signal
f(t) = f(0), f(t), f(2t), f(3t), f(4t),, f(nt),
We will assume functions for which f = 0 for t < 0
transform f(t)
F(s) = n=0 f(nt).e -snt
Define z = est
F(z) = n=0 f(nt).z -n = n=0 fn.z-n ZT[f] = F(z)
each term in z-1 represents a delay of t, ie z -n => delay of nt
[Link]@[Link] [Link]/Instrumentation/ 13 13 December, 2001
Examples
(1) f n = 0 = 1 0 0 0 0 ...
F(z) = 1
(2) fn = 1 represents a step function, since f(t) = 0 for all t < 0
F(z) = 1 + z-1 + z-2 + z-3 + z-4 + + z-n +
Should recognise geometric series, or binomial expansion of (1-x)-1
1
F(z) =
(1 z 1 )
(3) fn = e-na a = t/ = time constant t = sampling interval
F(z) = 1 + e-az-1 + e-2az-2 + e-3az-3 + e-4az-4 + e-naz-n +
1
F(z) =
(1 e az 1 )
(4) fn = 1 - e-na
1 1 z 1 (1 ea )
F(z) = =
(1 z ) 1 a 1
(1 e z ) (1 z1 )(1 e az 1 )
[Link]@[Link] [Link]/Instrumentation/ 14 13 December, 2001
Digital filters
What is the output if every previous input sample is summed with weight e-na?
ie compute gm = nme-nafn
Convolution in time, so becomes z-transform multiplication G(z) = H(z)F(z)
1 F(z)
H(z) = Z T [ ena ] = G(z) =
(1 e a z1 ) (1 ea z1 )
F(z) = (1 e az 1 )G(z) = G(z) G(z)e a z1
fn = gn e a gn1 or gn = fn + ea gn1
ie - Latest value of output sampled waveform
= current input sample + previous output sample x e-a
Impulse response corresponding to H(z)?
h(t) = e-nt/ which is impulse response of Low Pass Filter (Problems 2, No 8)
Conclusion
Low pass digital filter can be made using just two samples gn = fn + ea gn1
well suited for simple digital processor operation
[Link]@[Link] [Link]/Instrumentation/ 15 13 December, 2001
Step response of previous digital filter
To be more exact
Impulse response of Low Pass filter
output
R
vin vout input
C
1
h(t) = e t / 0 100 200 300 400 500
fn
gn = + ea gn1
closeup view
50 60 70 80 90 100
[Link]@[Link] [Link]/Instrumentation/ 16 13 December, 2001
Deconvolution
Suppose a signal has been filtered by a system with a known response
How to recover the input signal from the samples?
In t: input = f output = g, filter impulse response = h
In z: F(z) G(z) and H(z)
Since g(t) = f(t)*h(t), then G(z) = F(z)H(z)
so to recover input F(z) = H-1(z)G(z)
Low pass filter again
1
H(z) = Inverse filter H1 (z) = (1 e az 1 )
(1 e az 1 )
fn = gn e a gn1
terms in z-1 identify which delayed samples to use
This time gn are the measured samples, fn the result of digital processing
[Link]@[Link] [Link]/Instrumentation/ 17 13 December, 2001
An example of a deconvolution filter
Integrator + CR-RC bandpass filter waveform
form weighted sum of pulse samples
gn = [Link]+1 + [Link] + [Link]-1 CRRC waveform
Weighted sum
for correct choice of wi
(Problems 6)
Note gn needs fn+1
doesn't violate causality if data
are digital, in storage - 0 5 10 15 20 25
or could simply delay output
in applications such as image processing, causality does not apply
[Link]@[Link] [Link]/Instrumentation/ 18 13 December, 2001
CMS experiment at Large Hadron Collider
uses this deconvolution filter CRRC waveform
implemented in CMOS IC
Weighted sum
beam crossings at 40MHz (t = 25ns)
many events per crossing
small number of weights 0 5 10 15 20 25
implemented as analogue calculation
process only data which are to be read out
400
1.0 Ideal deconvolution output [mV]
Real
0.8 300
late early
0.6
200
w(t)
0.4
100
0.2
0.0 0
-2 -1 0 1 2
t/ t -75 -50 -25 0 25 50 75
test signal injection time [nsec]
[Link]@[Link] [Link]/Instrumentation/ 19 13 December, 2001