Professional Documents
Culture Documents
Solutions Manual: Computer-Controlled Systems
Solutions Manual: Computer-Controlled Systems
Third Edition
Solutions Manual
Preface
This Solutions Manual contains solutions to most of the problems in the fourth edition of str m, K. J. and B. Wittenmark 1997: Computer controlled Systems Theory and Applications, Prentice Hall Inc., Englewood Cli s, N. J. Many of the problems are intentionally made such that the students have to use a simulation program to verify the analytical solutions. This is important since it gives a feeling for the relation between the pulse transfer function and the time domain. In the book and the solutions we have used Matlab Simulink. Information about macros used to generate the illustrations can be obtained by writing to us. It is also important that a course in digital control includes laboratory exercises. The contents in the laboratory experiments are of course dependent on the available equipment. Examples of experiments are ? Illustration of aliasing ? Comparison between continuous time and discrete time controllers ? State feedback control. Redesign of continuous time controllers as well as controllers based on discrete time synthesis ? Controllers based on input-output design ? Control of systems subject to stochastic disturbances. Finally we would like to thank collegues and students who have helped us to test the book and the solutions. Karl J. str m Bj rn Wittenmark
Department of Automatic Control Lund Institute of Technology Box 118 S-220 00 Lund, Sweden
Solutions to Chapter 2
Problem 2.1
,
b xkh + h = e,ah xkh + 1 , e,ah a ukh y kh = cxkh The pole of the sampled system is exp,ah. The pole is real. For small values of h the pole is close to 1. If a 0 then the pole moves towards the origin when h increases. If a 0 then the pole moves along the positive real axis.
Problem 2.2
a. Using the Laplace transform method we nd that
,,sI , A,1 = L,1 1 8 s = eAh = L,1 s2 + 1 : ,1 8 cos h sin h 9 = : , sin h cos h ; 8 1 , cos h 9 Zh Zh 8 sin s 9 As B ds = : ,= e cos s ; ds = : sin h ;
0 0
1 = s;
9!
b. The system has the transfer function + 2 1 s Gs = s2 + +s3+ 2 = s +s1s3+ 2 = s + 1 , s + 2 3 Using the Table 2.1 gives
,2h e,h H q = 2 1 , e,h , ,1 , e ,2h q, 2 q,e
c. One state space realization of the system is 80 0 09 819 x = :1 0 0;x + :0;u _ 0 80 1 09 y = :0 0 1;x 1
Now
then
80 0 A2 = : 0 0 1 0
09 0 ; A3 = 0 0
,= and
Zh
0
8 1 0 09 eAh = I + Ah + A2 h2 =2 + : : : = : h 1 0 ; h2=2 h 1
eAs B ds =
0
Problem 2.3
= eAh a.
1 A = h ln
yk , 0:5yk , 1 = 6uk , 1 y , 0:5q ,1 y = 6q ,1 u qy , 0:5y = 6u xkh + h = 0:5xkh + 6ukh xt = axt + but _ ykh = xkh y t = xt discrete-time system continuous time system
ah = ln 0:5 a = , ln 2 h
b h b
eas b ds = a eas = a eah , 1 = 6
0
ln b = eah6a 1 = 12h 2 ,
b.
1 = ,0:5 2 = ,0:3
c.
Both eigenvalues of on the negative real axis No corresponding continuous system exists.
yk + 0:5y k , 1 = 6uk , 1 , yk + 1 = ,0:5y k + 6ut H q = q +60:5 one pole on the negative real axis.
Problem 2.4
h sin xk + 1 = xk + ,uk h = : , cos h cos h ; sin y k = Cxk 8 1 , cos h 9 h ,h = : sin h ;
a. 0 h = = : ,1 1 ; , = : 1 ; 2 0 1
8 9
y = :1 0;x
Pulse transfer operator
z 1 Y z = H zU z = zz2 + 11 z , 1 = z2 1 1 + z , 1 + + k , 1k , 1 + k , 1 =
1 , cos k k , 1 yk = sin 2 2
3
where k , 1 is a step at k = 1.
Y s = ss21+ 1 = 1 , s2 s 1 s +
b. The same way as a. y t = 1 , cos t, and y kh = 1 , cos k Notice 4 that the step responses of the continuous time and the zero-order hold sampled systems are the same in the sampling points.
Problem 2.5
Do a partial fraction decomposition and sample each term using Table 2.1: 1 5 1 1 Gs = s2 s + 2s + 3 = 1 s12 , 36 1 + 1 s + 2 , 1 s + 3 6 s 4 9 5 1 , ,2 1 , ,3 1 +1 H q = 12 qq, 12 , 36 q , 1 + 1 1 , e,2 , 27 1 , e,3 8q e q e
Problem 2.6
x = Ax + Bu _
gives
kh+h Z
kh Ah xkh + Bukh =e
eAs B s , khukhds
Problem 2.7
where
81 = :0 8 C = :1
h; 1 9 0;
8 h2=2 9 8 9 ~ = : 2 ,1 ; ,=: h ; 1 0 8 2 2 9 ~ C = : h =2 h =2 ;
~ = T T ,1 ~ , = T, ~ C = CT ,1 or or ~ T = T ~ CT = C
~ , = :1; 0
8 9
t11 = 2t11 , t21 t21 = t11 ht11 + t12 = 2t12 , t22 ht21 + t22 = t12 h2 t + ht = 1 12 2 11 h2 t + ht = 0 22 2 21 h2 t + h2 t = 1 2 11 2 21 h2 t + h2 t = 0 2 12 2 22
Equations 1 4 now give
1 2
3 4
or
1 h T = 2h2 : 2 ,h ; 2
Problem 2.8
Problem 2.9
8 ,a b 9 8 f 9 x = : c ,d ; x + : g ; u = Ax + Bu _
5
The eigenvalues of A is obtained from + a + d , bc = 2 + a + d + ad , bc = 0 which gives r a + d a , d2 + 4bc =, 2 4 The condition that a; b; c, and d are nonnegative implies that the eigenvalues, 1 and 2, are real. There are multiple eigenvalues if both a = d and bc = 0. Using the result in Appendix B we nd that eAh = 0I + 1Ah and e1h = 0 + 11h e2h = 0 + 12h This gives 2 h , 1 h 8 , ah 9 = 1e , 2e 0 1 2 0 1 1 bh =: ; 1 ch 0 , 1 dh e1h , e2h 1 = , h To compute , we notice that
0= 1= 1 2
Zh
0
and
0 sds =
, e2h , 1 , 2 e1h , 1 1 , 2 2 1
1 1
1 ,
depend on h. Introduce
Zh
0
s 1 sds = , 1 2 1
1,
e1 h , 1 ,
1 ,e2 h , 1 2
then
, = 0 B + 1AB
Problem 2.10
a. Using the result from Problem 2.9 gives 1 = ,0:0197 e1 h = 0:7895 2 = ,0:0129 e2 h = 0:8566 Further 0 = 0:9839 1 = 0:8223 and 8 0:790 0 9 = 0I + 12 1 A = : 0:176 0:857 ;
0 = 11:9412 1 = 63:3824
which agrees with the pulse transfer operator given in the problem formulation.
Problem 2.11
The motor has the transfer function Gs = 1=ss + 1. A state space representation is given in Example A.2, where
A = : ,1 0 ; 1 0
= expAh = L,1
B = :1; 0
sI , A,1
8 9
C = :0 1;
,=
8 e,h =: 1 , e,h Zh
0
0; 1
0
= L,1
s 0 1 ss + 1 : 1 s + 1 ; =
9!
eAs Bds =
This gives the sampled representation given in Example A.2. a. The pulse transfer function is
H z = C zI , ,1, =
8 9 8 z , e,h 0 9,1 8 1 , e,h 9 = :0 1;: ,1 + e,h z , 1 ; : h + e,h , 1 ; 8 9 8 9 8 1 , e,h 9 :0 1; z,1 0 = z , e,h z , 1 : 1 , e,h z , e,h ; : h + e,h , 1 ;
,
1 0
Figure 2.1
The pole solid and zero dashed in Problem 2.11.d as function of the sampling period.
10 Sampling period h
20
This gives
C k,1 , = : 0
An alternative way to nd hk is to take the inverse z -transform of H z . c. A di erence equation is obtained from H q
ykh + 2h , 1 + e,h ykh + h + e,h y kh = = h + e,h , 1ukh + h + 1 , e,h , he,h ukh
d. The poles are in z = 1 and z = exp,h. The second pole will move from 1 to the origin when h goes from zero to in nity. The zero is in e,h he,h z = , 1 , + e,, , 1 = f h h h The function f h goes from ,1 to 0 when h goes from zero to in nity. See Fig. 2.1.
Problem 2.12
Gs = 1 e,s s
h = 1; = 0:5 h
in nite order
x = ut , = 0 x + 1 ut , _
8
h
eAs ds B =
Z0:5
0
1 ds 1 = 0:5
= eAh,
Z
0
eAs ds B =
Z0:5
0
ds = 0:5
98
9 8 9
8 9 ,1 : H z =C zI , ,1, = : 1 0 ; : z , 1 ,0:5 ; : 015 ; 0 z 8 9 8 8 9 1 8 z 0:5 9 0:5 9 1 8 :1 0; : z 0:5 9 0:5 ; ; = : :1 z z , 1: 0 z , 1 ; 1 ;= zz , 1 : + 1 = 0z5z, 1 z
: + 1 z z H z = 0z5z, 1 = 0:5 z ,1 z , 1 + z ,2 z , 1 z | z | z k=0 hkh = 0:5 k = 1 1 k 1
9 8 9
0
=1; k1
=1; k2
Problem 2.13
a. This is the same system as in Example 2.6 but with = 1:5. In this case we have d = 2 and 0 = 0:5 and we get
= e,1 = 0:37 ,0 = 1 , e,0:5 = 0:39 ,1 = e,0:5 , e,1 = 0:24 A state representation is obtained from 2.12 8 xk + 1 9 8 ,1 ,0 9 8 xk 9 8 0 9 k : uu,1 ; = : 0 0 1 ; : uk , 2 ; + : 0 ; uk k 0 0 0 uk , 1 1 8 xk 9 8 9 yk = : 1 0 0 ; : uk , 2 ; uk , 1 b. The pulse transfer function is 9 8 9 8 8 9 z , ,,1 ,,0 ,1 0 z ,1 ; : 0 ; = H z = : 1 0 0 ; : 0 1 0 0 z + :24 , = z 20z + ,1 = 0:239z, 0037 z , z z : Some calculations give 80 k1 hk = : ,0 k=2 ,k,2 + ,1 e,k,3 k 3 ,0 e Using Theorem 2.4 and the de nition of the z -transform 2.27 we can also get the pulse response by expanding H z in z ,1 , i.e., H z = 0:39z,2 + 0:38z ,3 + 0:14z,4 + 0:05z,5 + 0:02z,6 + : : : The pulse response is now given by the coe cients in the series expansion. c. There are three poles p1 = p2 = 0 and p3 = 0:37 and one zero z1 = ,0:62.
where
Problem 2.14
Sampling the given di erential equation using the same procedure as in Problem 2.13 gives a = e, h Using the de nition of 0 and with h = 1 it follows that d = 4 and = 3+ 0 Further h, 0 Z
b = e, sb ds = b 1 , e, h, 0
3 0
b4 = e, h,
= 10
0
Z0
0
e, s b ds
b
, h, 0 e
, e,
ab3 + b4 = e, h, b3 + b4
Thus and
0
0 = 1 ln ab3 + b4 + h b3 + b4
Problem 2.15
yk , 0:5y k , 1 = uk , 9 + 0:2uk , 10 , yk + 10 , 0:5yk + 9 = uk + 1 + 0:2uk q 10 , 0:5q 9y k = q + 0:2uk Aq = q10 , 0:5q9 System order= deg Aq = 10 B q = q + 0:2
Also
Remark
10 B q = q + 0:2 = q,, q + 0:2 = q,9 + 0:2q ,10 = Aq q10 , 0:5q9 q,10 q 10 , 0:5q 1 1 , 0:5q ,1 , ,1 , ,1 ,9 ,1 + 0:2q = q ,d B , q =q 1 , 0:5q ,1 A q,1
11
Problem 2.16
FIR lter:
, H q ,1 = b0 + b1q,1 + : : : + bnq ,n
yk = b0 + b1q,1 + : : : + bnq,n uk = = b0uk + b1 uk , 1 + : : : + bn uk , n yk + n = b0uk + n + b1uk + n , 1 + : : : + bnuk
qnyk = b0qn + b1qn,1 + : : : + bn uk n,1 n n,1 H q = b0q + b1q qn + : : : + bn = b0 + b1q +n: : : + bn q
q H q = D + Bq A
8 ,a1 1 0 : : : 0 9 8 0 1 0 : : : 0 9
0 ,= . . 1; : . ; bn 0
8 b1 9
Problem 2.17
y k + 2 , 1:5y k + 1 + 0:5yk = uk + 1 , q2yk , 1:5qy k + 0:5yk = quk
Use the z -transform to nd the output sequence when
uk = 0 k 0 1 k 0
Table 2.2 page 57:
Z qnf = zn F , F1 ; F1z =
12
,
n,1 X j =0
f jhz,j
Compute
z = e,k=T ; e,1=T = 0:5 T = 1 ln 2 z , e,1=T Z ,1 z,1 z ,z0:5 = e,k,1ln2 1 z Z ,1 z , 12 = Z ,1 z,1 z , 12 = k , 1 h = 1 yk = 0:5e,kln2 + 2k , 1 + e,k,1 ln2
Z ,1
8 y1
Problem 2.18:
Verify that
Z 2
kh2
h2zz + 1 =
d
z,k = X ,kz ,k,1 = d
z = z , 1 , z = ,1 dz dz z , 1 z , 12 z , 12 X ,k z multiply by z kz = z , 12
z z ,k = z , 1 ;
di erentiate twice
1 ,1 = X k2z ,k f z
0 h2
Remark A necessary condition for convergence of f z,1 is that f is analytic for jz j 1. In that case there exists a power series expansion and
termwise di erentiation is allowed. Double integrator: The rst step is to translate Gs to the corresponding pulse transfer operator H z . Use the method of page 58. The sampled system. z uk = step U z = z , 1 U s = 1 s 1= 1 Y s = Gs s s3
2 1 Y z = h zzz,+13 Table 3:3 2 Y z = H z U z 2 + 1 z 2 1 H z = Y z = h zzz,+13 z , 1 = h zz, 12 U 2 z 2 cf. example A1: H z = C zI , ,1,.
14
Problem 2.19
a. The transfer function of the continuous time system is Gs = s bc a + Equation 2.30 gives
sh H z = z , 1 ,ah Ress=,a e s, 1 s bc a e + ,ah , = z , 1 ,ah e ,a 1 bc = e ,ah = bc 1 , e,ah a z,e This is the same result as obtained by using Table 2.1. b. The normalized motor has the transfer function Gs = ss 1 1 + and we get
H z =
s=si
To compute the residue for s = 0 we can use the series expansion of esh , 1=s and get
,h , h H z = z ,1e,h e,121 + z , 1 =
s+ is lead if s+ Consider the discrete time system z+b z+a ei!h + b
cos !h + b + i sin !h
arg ei!h + a = arg cos !h + a + i sin !h = sin !h
sin !h
= arctan b + cos !h , arctan a + cos !h
15
Problem 2.21
10
Output 0 0
Figure 2.2
b = ,0:9 upper solid, ,0:75 upper dashed, ,0:50 dash-dotted, 0 dotted, 0:5
10 Time
20
Phase lead if
0 !h
A state space representation of the system is 8 1:1 ,0:4 9 8 1 9 xk + 1 = : 1 0 ; x + : 0 ; u 9 1 8 yk = b + 1 : 1 b ; xk Simulation of the system for b = ,0:9; ,0:75; ,0:5; 0; 0:5 and 1 is shown in Fig. 2.2.
Problem 2.22
Problem 2.23
x = ,ax + b , au _ y =x+u The assumption that the system is stable implies that a 0. Sampling this system gives xkh + h = e,ah xkh + 1 , e,ah b , a ukh a y kh = xkh + ukh
16
where
b 2a b 2a
Problem 2.28
y k = yk , 1 + y k , 2 y 0 = y 1 = 1 The equation has the characteristic equation z2 , z , 1 = 0
which has the solution The solution has the form
z12 = 1 2 5
y k = c1
Using the initial values give
1 + p5
k
2
+ c2
1 , p5
k
2
1 p c1 = p 5 + 1 2 5 1 p c2 = p 5 , 1 2 5 17
The system is given by ,1 , 0:5q,1 + q,2 yk = ,2q,10 + q,11 uk Multiply by q 11. This gives , q9 q 2 , 0:5q + 1 yk = 2q + 1 uk The system is of order 11 = deg Aq and has the poles p 1 i 15 p12 = 4 p3:::11 = 0 multiplicity 9 and the zero z1 = ,0:5
Problem 2.29
Problem 2.30
The system H1 has a pole on the positive real axis and can be obtained by sampling a system with the transfer function Gs = s K a + H2 has a single pole on the negative real axis and has no continuous time equivalent. The only way to get two poles on the negative real axis as in H3 is when the continuous time system have complex conjugate poles given by 2 s2 + 2
!0s + !0 = 0 Further we must sample such that h = =! where p ! = !0 1 ,
2 We have two possible cases !2 i G1s = s2 + 2
!0 s + ! 2 0 0 s ii G2s = s2 + 2
! s + ! 2
0 0
Problem 2.31
Problem 2.32 8
9,1 ,1 s , 1 ; = 9 8 1 8 eh 0 9 0 s,1 ,1 =: =L ; 1 1 : s , 12 s , 1 ; heh eh h, 8 s 9 8 es 9h, h, 8 0 9 Z Z e ,0 = : ses ; ds = : ses ; , : es ; ds = 0 0 8 9 80 ,1 + eh, 9 h, , 1 e =: = h , eh, , eh, + 1 ; : 1 + h , , 1eh, ; 98 9 Z 8 ,s 9 8 eh, 0 ; : ,1 + e ; = Ah, : e ; ds = : ,1 = e se,s 1 + , 1e h , eh, eh, 0 8 9 eh, ,1 + e = : h, ; e 1 , h + + h , 1eh
= L,1 : ,sI , A,1 ; = L,1 : s , 1
x = : 1 0 ; x + : 1 ; ut , _ 1 1 0
8 9
= 0:2 h = 0:3 0
8 0:105 9 ,0 = : 0:005 ;
8 0:245 9 ,1 = : 0:050 ;
19
Solutions to Chapter 3
Problem 3.1
Jury's criterion is used to test if the roots are inside the unit circle. a. 1 ,1:5 0:9 0:9 ,1:5 1 2 = 0:9 0:19 ,0:15 ,0:15 0:19 1 = ,0:15=0:19 = ,0:79 0:07 The roots are inside the unit circle since the underlined elements are positive. The roots are 0:75 0:58i. b. 1 ,3 2 ,0:5 ,0:5 2 ,3 1 3 = ,0:5 0:75 ,2 0:5 0:5 ,2 0:75 2 = 2=3 5=12 ,2=3 ,2=3 5=12 1 = ,8=5 ,13=20 One of the underlined elements is negative and there is thus one root outside the unit circle. The roots are 2:19 and 0:40 0:25i. c. 1 ,2 2 ,0:5 ,0:5 2 ,2 1 3 = ,0:5 0:75 ,1 1 1 ,1 0:75 2 = 1:33 ,0:58 0:33 0:33 ,0:58 1 = ,0:57 ,0:39 There are two roots outside the unit circle. The roots are 0:35 and 0:82 0:86i.
20
d.
One root is outside the unit circle. The roots are ,5; ,0:5, and 0:5. e. 1 ,1:7 1:7 ,0:7 ,0:7 1:7 ,1:7 1 3 = ,0:7 0:51 ,0:51 0:51 0:51 ,0:51 0:51 2=1 0 0 0 0 The table breaks down since we can not compute 1 . There is one of the underlined elements that is zero which indicates that there is at least one root on the stability boundary. The roots are 0:7 and 0:5 0:866i. The complex conjugate roots are on the unit circle.
1 5 ,0:25 ,1:25 ,1:25 ,0:25 5 1 ,0:56 4:69 6 6 4:69 ,0:56 63:70 54:92 54:92 63:70 16:47
Problem 3.2
K 0
The stability can be determined using root locus. The starting points are z = 0; 0:2 and 0:4. The asymptotes have the directions =3 and ,. The crossing point of the asymptotes is 0:2. To nd where the roots will cross the unit circle let z = a + ib, where a2 + b2 = 1. Then a + iba + ib , 0:2a + ib , 0:4 = ,K Multiply with a , ib and use a2 + b2 = 1.
Im
1
Figure 3.1
0 Re
The solution is
This gives K = 2a , 0:6 = 0:70 and ,1:30. The root locus may also cross the unit circle for b = 0, i.e. a = 1. A root at z = ,1 is obtained when
K = 1:68
Problem 3.3
ukh = Kekh and the characteristic equation of the closed loop system becomes Kh + z , 1 = 0 The system is thus stable if K 2=h When there is a delay of one sample = h then the regulator is ukh = K ekh q and the characteristic equation is Kh + z z , 1 = z2 , z + Kh = 0 The constant term is the product of the roots and it will be unity when the poles are on the unit circle. The system is thus stable if K 1=h
b. Consider the continuous system Gs in series with a time delay of seconds. The transfer function for the open loop system is Go s = K e,s s The phase function is arg Go i! = , , w 2 and the gain is jGoi!j = K ! The system is stable if the phase lag is less than at the cross over frequency jGoi!cj = 1 !c = K The system is thus stable if
The continuous time system will be stable for all values of K if = 0 and for K =2h when = h. This value is about 50 larger than the value obtained for the sampled system in b. 23
,=2 , !c , =0 K =2 = 1 =h 2h
Im
1
Figure 3.2
0 Re
Problem 3.4
The Nyquist curve is Ho ei! for ! = 0; . In this case 1 Hoei! = ei! , 0:5 = = cos ! , 0:1 + i sin ! 5
cos ! , 0:5 , i sin ! cos2 ! + sin2 ! + 0:25 , cos ! : , sin = cos ! :, 0,5cosi! ! 1 25 For ! = 0 then Ho 1 = 2 and for ! = then Ho ,1 = ,2=3. The argument is ! arg Ho = ,arctg cossin, 0:5 ! The argument is ,=2 for ! = =3cos ! , 0:5. The following table gives some values for the real and imaginary parts. ! ReHo ImHo 0 2 0 =6 0:97 ,1:32 =3 0 ,1:16 =2 ,0:40 ,0:80 2=3 ,0:57 ,0:50 ,0:67 0 The Nyquist curve is shown in Fig. 3.2. 24
Problem 3.5
81 xk + 1 = : 1 8 yk = : 0
8 9
We have
x11 = 1
81 09819 819 829 x2 = x1 + ,u1 = : 1 1 ; : 0 ; + : 0 ; 1 = : 1 ; 81 09829 819 819 x3 = : 1 1 ; : 1 ; + : 0 ; ,1 = : 3 ;
The possibility to determine x1 from y 1, y 2 and u1 is called observability.
Problem 3.6
a. The observability matrix is
8 C 9 8 2 ,4 9 W o = : C ; = : 1 ,2 ;
The system is not observable since det Wo = 0. b. The controllability matrix is 8 9 8 9 : , , ; = : 6 1 ; Wc = 4 1 det Wc = 2 and the system is reachable.
Problem 3.7
1 0; For instance, the rst two colums are linearly independent. This implies that Wc has rank 2 and the system is thus reachable. From the input u0 we get the system 81 0 9 809 xk + 1 = : 0 0:5 ; xk + : 1 ; u0k In this case rankWc = 1 and the system is not reachable from u0 . 25
8 9 81 1 1 Wc = : , , ; = : 1 0 0:5
80 0 9 Wc = : 1 0:5 ;
Problem 3.8
a.
80 1 xk + 1 = : 0 0 0 0 80 1
x1 = : 0 0 80 x2 = : 0 0
0 0 1 0 0
809 29 3 ; xk + : 1 ; uk 0 0 9819 8 0 9 8 3 9 2 3 ; : 1 ; + : u0 ; = : 3 + u0 ; 0 1 0 0 2 9 8 3 9 8 0 9 8 3 + u0 9 3 ; : 3 + u0 ; + : u1 ; = : u1 ; 0 0 0 0 8 9 : 0 0 0 ;T x2 =
u0 = ,3 u1 = 0
b. Two steps, in general it would take 3 steps since it is a third order system. c. 8 9 8 9 0 1 0 Wc = : , , 2, ; = : 1 0 0 ; 0 0 0 8 not full rank not reachable, but may be controlled. 9 : 1 1 1 ;T is not in the column space of Wc and can therefore not be reached from the origin. It is easily seen from the state space description that x3 will be 0 for all k 0.
Problem 3.11
and
0:5 + K ,1 + 1:5
The z -transform for a ramp is given in Table 2.3, also compare Example 3.13, and we get z Uc z = z , 12 Using the pulse transfer functions from Problem 3.11 and the nal value theorem gives , lim ek = klim uc k , y k k!1 !1 , = zlim z , 1 1 , Hc z Uc z !1 z if K is chosen such that the closed loop system is stable. a. To use the nal value theorem in Table 2.3 it is required that 1 , z ,1F z does not have any roots on or outside the unit circle. For the proportional controller we must then look at the derivative of the error, i.e. 2, + , z lim e0 k = zlim z , 1 z z 2 0:50z:5z K K K z , 1 = K 0:50:5 !1 z , + + k!1 The derivative is positive in steady-state and the reference signal and the output are thus diverging. b. For the integral controller we get 2, + , z lim ek = z!1 z , 1 z z z 2 1:51z:5z 0:50+ K K K z , 12 = 0:5 lim z k!1 z , + :5 + K
Problem 3.12
Problem 3.13
Output
The continuous time solid and sampled step dots responses for the system in Problem 3.13.
Figure 3.3
25 Time
50
Problem 3.14
a. When Hr = K then the pulse transfer function for the closed loop system is 030q 0: Hc q = q2 , 1:65qK 0::68 + + 0026q + 0:026 +0 K :030 The characteristic equation is
,0:54 K 12:31
The steady state gain is 0:056K Hc 1 = 0:03 + 0:056K The steady state error when the input is a unit step is 0 03 e1 = 1 , Hc 1 = 0:03 +:0:056K The minimum error is about 4. 28
Im
2 3
Figure 3.4
The root locus when the tank system in Problem 3.14 is controlled with an integral controller.
1 Re
Output 0 0
Step response of the system in Problem 3.14 when the proportional controller is used with K = 0:5 solid, 1 dashed, and 2 dash-dotted.
Figure 3.5
250 Time
500
b. If the closed loop system is stable then the steady state error is zero if the integral controller is used. The characteristic equation is then z 2 , 1:65z + 0:68z , 1 + Kz 0:03z + 0:026 = 0 A scetch of the root locus is shown in Fig. 3.4. c. Fig. 3.5 and 3.6 show the step responses when the proportional controller with K = 0:5, 1, and 2 and the integral controller with K = 0:01, 0:025, and 0:05 are used on the tank system.
29
1 Output 0 0
Step response of the system in Problem 3.14 when the integral controller is used with K = 0:01 solid, 0:025 dashed, and 0:05 dash-dotted.
Figure 3.6
1000 Time
2000
Problem 3.16
H Hcz = 1 + o zz H
o
30
Problem 3.17
Using 3.17 we nd that the transformation is given by ~ Tc = Wc Wc,1 ~ where Wc is the controllability matrix of the original system and Wc is the controllability of the transformed system.
9 8 9 8 : , , ; = : 3 11 ; Wc = 4 11 8
H z = C zI , ,1, = : 5
4 = 39z + z z2 , 3 ~ = :3 0; 1 0 and
9 8 z , 1 ,2 9,1 8 3 9 6 ; : ,1 z , 2 ; : 4 ;
~ , = :1; 0
8 9 8
~ C = : 39 4 ;
,1 1 3 ; : 3 11 ; = 11 : 1 ,2 ; 4 11 4 3 1
9 9
9 8
9 8
Probem 3.18
a i b c d e f g i iii iii ii ii iii
Poles are mapped as z = esh . This mapping maps the left half plane on the unit circle. see a When a system with relative degree 1, new" zeros appear as described on p. 63 CCS. These zeros may very well be outside the unit circle. See Example 2.18 p. 65 CCC Sample the harmonic oscillator, Example A.3 p. 530 CCS. 81 09 For h = 2=!; = : 0 1 ; not controllable See e See p. 63, CCS 31
Problem 3.19
a. The controllability matrix is
8 9 8 9 0 1 0 Wc = : , , 2, ; = : 1 2 0 ; 2 3 0
Since one column is zero we nd that the system is not reachable det Wc= 0, see Theorem 3.7. b. The system may be controllable if the matrix is such that n x0 = 0. In this case 80 0 09 3 = : 0 0 0 ; 0 0 0 and the origin can be reached from any initial condition x0 by the control sequence u0 = u1 = u2 = 0.
Probem 3.20
Ho = q2 +10:4q Hr = K
a.
K 1 K ,1 + 0:4 K ,1 , 0:4
b.
,0:6 K 1
ek = uc , y , E z = 1 , Htotz Ucz If K is chosen such that the closed loop system is stable e.g. K = 0:5 the nal-value theorem can be used.
z lim ek = zlim z , 1 E z = zlim z , 1 1 , z 2 + 0Kz + K z , 1 = !1 z !1 z k!1 :4
2+ :4 = zlim z 2 z 0:40z:4z K = 1:41+ K = 0:74; K = 0:5 !1 + +
Problem 3.21
32
Im
2 Re
Figure 3.7
a.
:8K H Htot = 1 + r Ho = q2 , 1:2q0:4q5+ 0K 0:4q + 0:8 = Hr Ho + 0: + 0 = q 2 + q 0:40:4q +:2:8K:5 + 0:8K K,1 +0 The system is stable if, see p. 82 CCS,
0:5 + 0:8K 1 0:5 + 0:8K ,1 + 0:4K , 1:2 0:5 + 0:8K ,1 , 0:4K + 1:2
,0:25 K 0:625
b.
0: q + H Htot = 1 + rHo = qq2 , 1:2q +40:5 0:8K :4q + 0:8 = Hr Ho + K 0 4q = q 3 , 1:2q 2 0:q 0+ 0:8:KK + 0:8K + :5 + 0 4 Using root locus, Fig. 3.7, we can determine that the closed loop system is stable if p ,17 + 489 ,0:25 K 16
Hr = K q
33
Solutions to Chapter 4
Problem 4.1
The characteristic equation of the closed loop system is detzI , , ,L = z 2 , a11 + a22 , b2`2 , b1`1 z + a11 a22 , a12a21 + a12b2 , a22b1`1 + a21b1 , a11 b2`2 Identifying with the desired characteristic equation gives
8 b 9 8 ` 9 8 p + tr 9 b2 1 : a12b2 , a22b1 a21b1 , a11b2 ; : `1 ; = : p21, det ; 2 8 ` 9 1 8 a b , a b ,b 9 8 p + tr 9 21 1 : `1 ; = : ,a12b2 + a11b2 b12 ; : p21, det ; 22 1 2
where tr = a11 + a22 and det = a11a22 , a12 a21. The solution is
where
= a21b2 , a12b2 + b1b2 a22 , a11 1 2 To check when 6= 0 we form the controllability matrix of the system
and we nd that
= det Wc There exists a solution to the system of equations above if the system is controllable. For the double integrator with h = 1 and dead beat response we have a11 = a12 = a22 = b2 = 1, a21 = 0, b1 = 0:5, and p1 = p2 = 0. This gives
Problem 4.2
In this case the desired characteristic equation is z , 0:1z , 0:25 = z 2 , 0:35z + 0:025: Using the result from Problem 4.1 we nd that = 0:5 34
and L is obtained from 8 9 8 98 9 8 9 T = : `1 ; = 1 : 0:5 0 ; : 0:75 ; = : 0:75 ; L `2 0:1 1 ,0:025 0:1 To check the result we form 8 1 0:1 9 8 0:75 0:1 9 8 0:25 0 9 , ,L = : 0:5 0:1 ; , : 0 0 ; = : 0:5 0:1 ; The matrix , ,L thus has the desired eigenvalues 0.25 and 0.1.
Problem 4.3
8 = :1,
0 ; 1;
8 1, 9 , = :h, 1+ ;
p1 = p2 = 0:
From Problem 4.1 it follows that 8 2 T = 1 : 1 , , h , 1 + 1 , h , L 1, 1, 8 1 , 21 + + 21 , h , 9 1 = : ; 1, 8 9 1 1, ,h 2 = : 1, ; where If x0 = : 1 1 ; then
981+ 9 ;: , ;
= 1 , 3 + 1 , 2 h , 1 + = h1 , 2
9T
h = f h
35
Output u(0) 2 0
1
Figure 4.1
u0 s function
2 3 Sampling period h
of h in Problem 4.3.
hk+1 = f hk
Starting with h0 = 2 gives
k 0 1 2 3 4
with h = 2:21 we get L = : 0:49 0:51 ;. Fig. 4.2 shows the response of the motor controlled with a deadbeat controller when h = 2:21. We see that the system settles after two samples and that the constraint on the control signal is ful lled.
Problem 4.4
a. Using the results in Problem 4.1 gives
LT
:1600 = ,0:1 : ,0:0880 ,00100 ; : ,0:5900 ; = : 9:22 ; 0029 ,0:0125 0: 0:1595 3:11
98
9 8
36
1 Output
1 0 1 5 10
Input
1 0
Figure 4.2
5 Time
10
p1 = ,2e,
!h cos!h 1 ,
2 = ,0:63 p2 = e,2
!h = 0:21 This has the solution
0:7 and ! 5:6, so the characteristic polynomial becomes s2 + 7:8s + 31:7. In Matlab you can do
rd=roots 1-0.63 0.21 rc=Logrd h Ac=polyrc
The chosen sampling interval is higher than recommended by the rule of thumb, since !h 1:1 0:6 8 9 c. The closed loop system when using L = : 9:22 3:11 ; is shown in 8 9T Fig. 4.3 when x0 = : 1 0 ; .
Problem 4.5
a. In this case 8 C 9 8 0 19 Wo = : C ; = : 0:22 1 ; ; 809 8 0 9 = : C , ; = : 0:03 ; ;
Wo,1 = : ,41:55
4:55 0 ;
1 Output 0 0 5 0 Input 5 10 15 2 4 0
The 9 8 response and the control signal of the system in Problem 4.4 8 9 T when x0 = : 1 0 ; and L = : 9:22 3:11 ;.
Figure 4.3
2 Time
We get
8 1 ,k 9 8 0:78 I , KC = : 0 0 1 ; : 0:22
9 8
38
k1 = 0:78=0:22 = 3:55:
The observer is then 8 0 ,3:55 9 8 0:114 9 8 3:55 9 xkjk = : 0 0 ; xk , 1jk ,1+ : 0 ; uk , 1+: 1 ; y k ^ ^ Since x2 kjk = y k we get ^
Problem 4.6
8 0:790 0 9 8 0:281 9 xkh + h = : 0:176 0:857 ; xkh + : 0:0296 ; ukh 8 9 ykh = : 0 1 ; xkh
The continuous time poles of the system are ,0:0197 and ,0:0129. The observer should be twice as fast as the fastest mode of the open loop system. We thus choose the poles of the observer in
z = e,0:039412 = 0:62
The desired characteristic equation of , KC is thus
z2 , 1:24z + 0:38 = 0
Using the results from Problem 4.1 gives
K = : 0::139 ; 0 407
9 8 9T
Fig. 4.4 shows the states and the estimated states when x0 = : 1 1 ; and when ukh is zero up to t = 250 and one thereafter.
Problem 4.7
x1 and xe1
1 0.5 0 0.5
250
500
x2 and xe2
1 0.5 0 0.5
Figure 4.4
Problem 4.6
The states solid and their estimates dots for the tank system in
250 Time
500
then
k = I , KC k , 1 + Kyk , 1 + I , KC ,uk , 1 = I , KC k , 1 + I , KC Ky k , 1 , I , KC ,L k , 1 + Kyk , 1 = I , KC , ,L k , 1 + I , KC , ,LKy k , 1 = o k , 1 + ,o y k , 1
The output of the regulator can be written
Problem 4.8
819 xw = : 0 ;
a. If the state and v can be measured then we can use the controller
xk + 1 = xk + xw wk , ,Lxk , ,Lw wk = , ,Lxk + xw , ,Lw wk yk = Cxk
In general it is not possible to totally eliminate the in uence of wk. This is only possible if xw , ,Lw is the zero matrix. We will therefore only consider the situation at the output in steady state
Hw 1 = 0
This will be the case if
22 Lw = 1 ,1 , 'c+ ' 'c22 2 c12 1
where 'cij is the i; j element of , ,L and i is the i:th element of ,. Assume that L is determined to give a dead beat regulator then
8 9 : 3:21 5:57 ; L=
and and
Lw = 5:356 b. In this case is the state but not the disturbance measurable. The disturbance can now be calculated from the state equation
xw wk , 1 = xk , xk , 1 , ,uk , 1: The rst element in this vector equation gives
0.2
Output
0 0.2 0 2 5 10 15
Output
0 2 0 5 10 15
Estimate ve
0 5 Time 10 15
Output
0 2
0 0 5 Time 10 15
The output of the system in Problem 4.8 for the regulators in a upper left, b upper right and c lower left and right. The estimate of v is also shown for case c. Notice the di erence in scale in the upper left curve.
Figure 4.5
c. If only the output is measurable then the state and the disturbance can be estimated using an observer of the form 4.41 8 xk + 1 9 8 9 8 xk 9 8 , 9 8K9 ^ ^ xw : wk + 1 ; = : 0 1 ; : wk ; + : 0 ; uk + : Kw ; "k ^ ^ "k = y k , C xk ^ The gain vector can now be determined such that the error goes to zero provided the augmented system is observable. The error equation is 8 xk + 1 9 8 , KC 9 8 xk 9 ~ ~ xw : wk + 1 ; = : ,Kw C 1 ; : wk ; ~ ~ The characteristic equation of the system matrix for the error is z 3 +k1 , 2:2z 2 +1:05 , 1:7k1 + k2 + kw z +0:7k1 +0:15 , 0:7kw , k2 = 0: The eigenvalues are in the origin if k1 = 2:2 k2 = ,0:6433 kw = 3:3333: The controller now has to be uk = ,Lxk , Lw wk ^ ^ where L and Lv are the same as in a. The solutions above have the drawback that there may be an error in the output due to the disturbance if there are small errors in the model. Fig. 4.5 show that the output when the controllers in a, b and c are used.
42
Problem 4.9
a. The state equation for the tank system when h = 12 was given in the solution to Problem 4.6. The desired characteristic equation is
z2 , 1:55z + 0:64 = 0
Using the result in Problem 9.1 give
L = : 0:251 0:8962 ;
b. An integrator can be incorporated as shown in Section 4.5 by augmenting the system with
,0:281`2 ,0:281`3 9 8 xk 9 : 0:176 ,00:0296`1 0:857 ,,01:0296`2 ,0:0296`3 ; : x3k ; 1 8 0:281`c 9 + : 0:0296`c ; uc k
1 The characteristic equation is
z 3 + ,2:647 + 0:281`1 + 0:0296`2z 2 + 2:3240 , 0:5218`1 , 0:0035`2 , 0:0296`3z + + ,0:6770 + 0:2408`1 , 0:0261`2 , 0:0261`3 = 0
Assume that two poles are placed in the desired location and that the third pole is in p. We get the following system of equations to determine the state feedback vector. 8 0:281 0:0296 0 9 8 `1 9 8 2:647 , 1:55 , p 9 0 5218 0 : ,0::2408 ,0::0035 ,0::0296 ; : `2 ; = : ,2:3240 + 1:55:p +p 0:64 ; , 0261 ,0 0261 `3 0:6770 , 0 64 This gives
Output 0 0
Figure 4.6 The stepresponse for the tank process when p = 0 solid and 0.5 dashed, and when `c = 0.
250 Time
500
Output 0 0
The step response for the tank process when p = 0 and when `c = 0 solid, 3 dashed and 6 dash-dotted.
Figure 4.7
250 Time
500
c. The parameter `c will not in uence the characteristic equation, but it is a feedforward term from the reference signal, see Fig. 4.11 CCS. Fig. 4.6 shows the step response for some values of p. Fig. 4.7 shows the in uence of `c when p = 0.
44
Problem 4.10
The process is
h ; xkh + : h2=2 ; uk + : h2=2 ; v k 1 h h where v k is a sinusoidal. I.e. it can be described by 8 cos ! h , sin ! h 9 wkh + h = : sin !oh cos ! o ; wkh oh 8 o9 v k = : 1 0 ; wkh
The augmented system 9.33 is now 8 9 8 2 9 2 8 xkh + h 9 1 h h h=2 0 8 xkh 9 h h=2 0 : wkh + h ; = 0 1 0 0 , ; : wkh ; + : 0 ; ukh :0 0 0 where = cos !o h and = sin !o h. Assume rst that the control law is ukh = ,Lxkh , Lw wkh where 9 8 L = : h12 23h ;
81 xkh + h = : 0
i.e., a dead beat controller for the states. Further if Lw = : 1 0 ; we would totally eliminate v since v and u are in uencing the system in the same way. Compare the discussion in the solution of Problem 4.8. Since xkh and kh cannot be measured we use the observer of the structure 9.35 where 8 h2=2 0 9 8 , 9 xw = : h 0 ; and w = : ; The error equation is then 8 xk + 1 9 8 , KC 9 8 xk 9 ~ ~ xw : wk + 1 ; = : ,Kw C 1 ; : wk ; ~ 8 1 , k1 h h2=2 ~0 9 8 xk 9 ,k2 1 h 0 ~ = ,k : wk ; , ; ~ : ,kw1 0 w2 0 Let the desired characteristic equation of the error be z , 4 = 0 If h = 1 and !o = 0:1 then for = 0:5 we get the following system of equations 8 1 0 0 0 9 8 k1 9 8 ,4 + 3:9022 9 ,2:9022 1 0:5 0 k2 = 6 2 , 5:8044 3 :9022 9022 : 2,1 ,1:1 ,0::0245 ,0::1545 ; : kw1 ; : ,4 4,,3:19022 ; 0 4756 ,0 1545 kw2 45
x1 and xe1
10
20
30
x2 and xe2
10 Time
20
30
The states of the double integrator solid and their estimates dots 8 9 when vt = sin!o t. The controller is de ned by L = : 1 1:5 ; and Lw = 8 9 : 1 0 ;.
Figure 4.8
k2 = 1:1018 :kw1 = 0::2288 kw2 = 0 1877 Fig. 9.8 shows the states of the double integrator and their estimates when vt = sin!o t. It is seen that the controller is able to eliminate the disturbance.
We have to determine the feedback vector L such that a. 8 0:9 , ` ,` 9 , ,L = : 1 1 0:72 ;
8 k = 1:9022 1
Problem 4.11
has all eigenvalues in the origin. This gives the condition detI , + ,L = 2 + ,1:6 + `1 + 0:63 , 0:7`1 + `2 = 2 I.e., or
46
l2
1 Deadbeat
1 1
l1
Figure 4.9
Problem 4.12
The stationary gain of the closed loop system is given by stationary gain of the m C I , + ,L,1 , = m To get unit steady state gain we choose m = 1 b. The closed loop characteristic equation is stable if See Example 3.2 0:63 , 0:7`1 + `2 1 0:63 , 0:7`1 + `2 ,1 + ,1:6 + `1 0:63 , 0:67`1 + `2 ,1 , ,1:6 + `1 This gives ,0:7`1 + `2 0:37 ,1:7`1 + `2 ,3:23 0:37`1 + `2 ,0:03 The stability area is shown in Fig. 4.9. Assume that the deadbeat control in a. is used. The closed loop system will be unstable if the feedback from x1 is disconnected i.e., if `1 = 0, but the system will remain stable if x2 is disconnected if `2 = 0.
a. The deadbeat requirement implies that the characteristic equation of the closed loop system is 8 9 `2 , 0:5 ; 2 = detI , + ,L = det : , 0:25 + `1 4`1 , 1 , 2 + 4 `2 2 + ` + 4` , 2:25 2 = 1 2 There are in nitely many solutions, one is
L = : 2:25 0 ;
47
9 8 9 8 : , , ; = : 1 2:25 ; Wc = 4 9
det Wc = 0 implies that the system is not reachable and arbitrary states 8 9T cannot be reached from the origin. However, xk = : 2 8 ; is in the column space of Wc and the point can thus be reached.
x2 = 2x0 + ,u0 + ,u1 8 0:5625 1:125 9 8 2:25 9 819 = : 2:25 4:5 ; x0 + : 9 ; u0 + : 4 ; u1
With x2 = : 2 8 ; and x0 = : 0 0 ; we get 2 = 2:25u0 + u1 8 = 9u0 + 4u1
9T
9T
One solution is
8 1:85 9 K = : 7:48 ;
48
Solutions to Chapter 5
Problem 5.1
Euclid's algorithm de nes a sequence of polynomials A0; : : :; An where A0 = A, A1 = B and Ai = Ai+1 Qi+1 + Ai+2 If the algorithm terminates with An+1 then the greatest common factor is An . For the polynomials in the problem we get A0 = z4 , 2:6z3 + 2:25z 2 , 0:8z + 0:1 A1 = z3 , 2z2 + 1:45z , 0:35: This gives Q1 = z , 0:6 A2 = ,0:4z2 + 0:42z , 0:11 = ,0:4z2 , 1:05z + 0:275 The next step of the algorithm gives Q2 = z , 0:95=,0:4 A3 = 0:1775z , 0:08875 = 0:1775z , 0:5 Finally Q3 = z , 0:55=0:1775 A4 = 0 The greatest common factor of A and B is thus z , 0:5.
Problem 5.2
a. To use Algorithm 5.1 we must know the pulse-transfer function B z =Az and the desired closed-loop characteristic polynomial Acl z . Since the desired pulse-transfer function from uc to y is Hm z = 1 + =z + , we know at least that z + must be a factor in Acl z . Step 1. You easily see that, with A and Acl being rst order polynomials and B a scalar, you can solve the equation for the closed loop system using scalars Rz = r0 and S z = s0 : AzRz + BzS z = Acl z z + a r0 + 1 s0 = z + Identi cation of coe cients gives the following equation system:
r0 = 1 ar0 + s0 =
49
Ao = 1 and Ac = z + . Choose
1 T z = t0 Ao z = Ac1 Aoz = 1 + B With this choice of T , the static gain from uc to y is set to 1 Hm 1 = 1, and the observer dynamics are cancelled in the pulse-transfer function from uc to y . In this case, there are no observer dynamics, though, since deg Ao = 0. The resulting control law becomes
Solution with higher order observer: The solution above is not the
we must increase the order of R by one. This gives and the equation system becomes
only one solving the original problem. We can, for example, decide to have another closed loop pole in z = , , say.
Step 1. To solve the equation for the closed loop characteristic polynomial
z + ar0z + r1 + 1 s0 = z + z +
r0 = 1 ar + r1 = + :ar0 + s0 = 1
8r = 1 0 r1 = + , a :s0 = , a + , a
1 T z = t0 Aoz = Ac1 Ao z = 1 + z + B The resulting control law becomes
uk = , + , a uk , 1 + 1 +
u k + u k , 1 c c , , a + , ayk , 1 ,
The controller thus is a dynamical system. In this case there is a delay of one sample from the measurements y to the control signal u. This could have been avoided by choosing deg S = 1. 50
Output
1 0
Figure 5.1
The output of the = 1, and uc k is 0 for k 25 and ,1 for k = 25. The dots corresponds to the zero order controller, and the crosses to the rst order controller.
50
b. The closed loop characteristic polynomial is given by AR + BS , i.e. z + in the rst solution, and z + z + in the second one. In both cases we get the same closed loop pulse-transfer function from uc to y since the observer polynomial is cancelled by T z : 0 Hmz = AzRBzzTBzzS z = tABzz = 1 + + z+
c
Fig. 5.1 shows the response when the two di erent controllers are used. It is assumed in the simulations that a = ,0:99, and that the design parameters are = ,0:7 and = ,0:5. You can see the e ect of the observer polynomial when regulating a nonzero initial state, but not in the response to a set point change.
Problem 5.3
a. The desired closed loop pulse-transfer function is Hmz = z2 ,Bm zz+ 0:7 1:5 In this case, the process zero is cancelled by the controller, so z + 0:7 is not a factor of Bm . By choosing Bm = 0:2z we make the steady state gain Hm 1 = 1, and the pole excess in Hm equals the pole excess in H . Cancellation of process poles and zeros is handled by Algorithm 5.3 or through the following discussion. First, the process numerator is factored as B = B + B , , where B + is the part of the numerator which should be cancelled by the controller, i.e., B + = z + 0:7 and B, = 1. B + must be a part of the R polynomial as 51
well as Acl . This gives the Diophantine equation AzB+ z Rz + B+ z B, z S z = B + z Acl z z 2 , 1:8z + 0:81Rz + S z = Acl z If Rz is a constant r0, the left hand side is of second order, and so must cl be. With this choice of R, the causality condition deg S deg R = 1 A leads us to set S z = s0 z + s1 . Now, we can solve the Diophantine equation above, since we have 3 indeterminates r0, s0 and s1 and 3 coe cients to set: z 2 , 1:8z + 0:81 r0 + s0 z + s1 = z 2 , 1:5z + 0:7 8 8 r0 = 1 r0 = 1 s0 = 0:3 ,1:8r0 + s0 = ,1:5 :s1 = ,0:11 : 0:81r0 + s1 = 0:7 Thus, we have Rz = B + z Rz = z + 0:7 and S z = 0:3z , 0:11. To obtain the desired + ,T 02 Hm z = ARBTBS = B + z2B, B:5z + 0:7 = z 2 , 1::5zz + 0:7 + 1 we must select T z = 0:2z The controller is now uk = ,0:7uk , 1 + 0:2uck , 0:3y k + 0:11yk , 1: b. In this case we do not want to cancel the process zero, so 0:2 z + :7 Hm z = z 2 ,Bm zz+ 0:7 = z21,1:5z 0:7:7 1:5 +0 in order to get Hm 1 = 1. The closed loop characteristic polynomial is now given by the identity z 2 , 1:8z + 0:81Rz + z + 0:7S z = Acl z The simplest choice, a zero order controller, will not su ce in this case since it would only give 2 parameters r0 and s0 to select the 3 parameters in the second order polynomial z 2 , 1:5z + 0:7. Thus, we must increase the order of the controller by one and, consequently, add an observer pole which is placed at the origin, i.e. Ao = z and Acl = z 3 , 1:5z 2 + 0:7z . Letting R = r0z + r1 S = s0 z + s1 the identity then gives the system of equations 8r0 = 1 8 r0 = 1 r1 = 0:0875 ,1:8r0 + r1 + s0 = ,1:5 2125 s 1 :s0 = 0:0:1012 :0:81r0 , 1:8r0:+ 0:17+00+7s1 = 0:7 s1 = , 81r : s1 = 0 52
Output
Output
0 20 40
0 0.2
0 0.2
20
40
Control
Control
Figure 5.2 The output and the control signal for the controllers in Case a left and Case b right in Problem 5.3. The ringing in the control signal in Case a is due to the cancellation of the process zero on the negative real axis.
Problem 5.4
a. Using the controller
BS = Bm AR + BS Am
Section 5.10 gives one solution to the problem
R = BAm , Bm S = ABm
With the given system and model we get
R = 1z + , 1 , = z , 1 S = z + a1 +
The controller contains an integrator. Further the pole of the process is cancelled. 53
AR + BS = z + z + a
The closed loop system will contain an unstable mode if jaj controller can be written S = A Bm R B Am , Bm 1. The
From this we can conclude that in order to get a stable closed loop we must ful ll the following constraints. i. Bm must contain the zeros of B that are outside the unit circle. ii. Am , Bm must contain the poles of the process that are outside the unit circle. The rst constraint is the same as for the polynomial design discussed in Chapter 5.
Problem 5.5
a. Equation 5.33 gives the pulse transfer operator from uc and v to y :
R=1 S = ,a
We thus get
BR = 1 AR + BS z + If v k is a step there will thus be a steady state error 1=1 + in the output.
b. By inspection of the transfer function from v to y we see that we must make R1 = 0 in order to remove the steady state error after a load disturbance step. By forcing the factor z , 1 into Rz we thus have obtained integral action in the controller. The design problem is solved by using the general Algorithm 5.3 or through a discussion like the one below. With Rz = z , 1Rz the closed loop characteristic equation becomes Azz , 1Rz + Bz S z = Acl z z + az , 1Rz + 1 S z = Acl z If Rz is a constant r0, the left hand side is of second order, and so must Acl be. With this choice of R, the causality condition deg S deg R = 1 leads us to set S z = s0 z + s1 . Now, we can solve the Diophantine 54
Output
1 0 0
Figure 5.3
The output of the process in Problem 5.5 when the controller does not contain an integrator dots and when an integrator is introduced crosses.
25 Time
50
equation above, since we have 3 indeterminates r0, s0 and s1 and 3 coe cients to set: z + az , 1 r0 + s0 z + s1 = z + z + 8r = 1 r0 = 1 0 s0 = + , a + 1 a , 1r0 + s0 = + : s1 = + a : ,ar0 + s1 =
To obtain the desired T Hmz = AzRBzz+ BzzS z = z + Tzz + = 1 + z+ we must select The controller is now
T z = 1 + z +
y k , 1
Fig. 5.3 shows the controllers in Problem 5.2 and the controller with an integrator. The reference value is zero and there is an initial value of the state in the process. At t = 25 a constant load disturbance is introduced. It is assumed that a = ,0:99, and the design parameters are chosen as = ,0:7 and = ,0:5.
55
It is assumed that the design is based on the model H = B=A while the true model is H 0 = B 0 =A0 . The pulse transfer operator of the closed loop system is T=R B0 Hcl = A0R +TB 0S = A0=B0 + S=R The design gives 0 T = Bm Ao and AR + BS = A0Am B+ or S = B + Am Ao , A R BR B This gives 0 0 Bm Ao Bm Ao R R ! Hcl = 0 + Ao Am A = Ao Am A B 1 1 + BR , B B, R + H 0 , H B0 Ao 0 B, 1 R m ! = Hm ! = BA m RB, 1 1 Ao B , 1 1 1+ R + Am H 0 , H Ao Am H 0 , H
Problem 5.6
Problem 5.7
a. The design in Problem 5.2 gives R=1 S = ,a T =1+ Assume that the true process is 1 z + a0 Equation 5.41 gives
ja0 + , aj 1
With the numerical values in the problem formulation we get
,1:4 a0 0:6
56
10
0
Figure 5.4
1 Frequency
The left hand side of the inequality in Problem 5.7 when z = ei! , 0 ! for a0 = ,0:955 dashed, ,0:9 dash-dotted and 0.6 dotted. The right hand side of the inequality is also shown solid.
b. Equation 5.40 gives the inequality z Hz H z , H 0z H z Hff z = 1 +z + z + a 1 + a Hfb , m = z , 0::5 2:5 z,0 9 or 1 z , 0:5 2:5 1 z , 0:9 , z + a0 z , 0:9 Fig. 5.4 shows for z = ei! the left hand side of the inequality for di erent values of a0 . The right hand side is also shown. Section 5.6 shows that the control signal is given by 5.52 uk = Hmq uck = 1 +q + q + a uc k Hq We may assume that both the process and the model have a continuous time correspondence. This implies that a and are less than zero. Further the desired model is stable, i.e. j j 1. The control signal is now obtained by studying the step response of Hm =H , which is a stable rst order system. The largest value is then either at the rst step or the nal value. The magnitude at the rst step can be determined either through the initial value theorem or by using series expansion and the value is 1 + . The nal value is 1 + a. If j j jaj then the closed loop system is faster than the open loop system and the control signal is largest at the rst step. If the desired response is slower than the open loop system then the nal value is the largest one. 57
Problem 5.8
Problem 5.14
a. The rule of thumb on p. 130 gives
!h = 0:1 , 0:6
Identifying with
s2 + 2
!s + ! 2 gives ! = 0:1. Thus an appropriate sampling interval is h = 1,6
b. Using Example 2.16 we get sampled data characteristic equation
z 2 + a1 z + a 2 = 0 a1 = ,2e,
!h cos 1 ,
2!h = ,1:32 a2 = e,2
!h = 0:5 The poles are in 0:66 0:25i.
This solution demonstrates how to use Algorithm 5.3. Data: The process is given by A = q 2 , 1:6q +0:65 and B = 0:4q +0:3. Acl will at least contain Ac = q 2 , 0:7q + 0:25, other factors may be added later on. Rd = Sd = 1 since no given factors are forced into the controller. The desired response to command signals is assumed to be Hm = Bm =Am = Bm =Ac = 2 , 0:7q + 0:25 cancelled process zero, Hm 1 = 1. 0:55=q
Pole excess condition:
where
Problem 5.15
deg Am , deg Bm deg A , deg B 2,0 2,1 zero in Bm causes the delay from the command signal to be one time unit more than the delay of the process. Bm = B, Bm Bm = 0:55=0:4 = 1:375
Degree condition: Remark: The fact that we cancel one zero and do not introduce any other
deg Acl = 2 deg A + deg Am + deg Rd + deg Sd , 1 = = 22+2+0+0,1 = 5 with Acl = A+ B + Am Acl and Acl = Ac Ao . Step 1. A+ = 1, A, = A = q 2 , 1:6q + 0:65, B + = q + 0:75 and B , = 0:4 achives cancellation of the process zero, but no cancellation of process poles. 58
deg Ao = deg Acl , deg A+ , deg B + , deg Am , deg Ac = = 5,0,1,2,2 = 0 The Diophantine equation to solve thus becomes A, Rd R + B, SdS = Acl 2 , 1:6q + 0:65R + 0:4S = q 2 , 0:7q + 0:25 q Since this is of second order, R must be a constant, r0 , say. In order to solve the identity we must have two more parameters, so we let S = s0 q + s1 : q 2 , 1:6q + 0:65r0 + s0 q + s1 0:4 = q 2 , 0:7q + 0:25 This gives the system of equations
r0 = 1 s0 = 2:25 s 1 = ,1
R = AmB + Rd R = Amq + 0:75 S = AmA+ Sd S = Am 2:25q , 1 T = Bm A+ Acl = 1:375 Ac Since, in this case, Am = Ac , this factor can and should of course be cancelled in all controller polynomials, giving
Problem 5.16
This gives
and
S q = 2:5 a0 + 1:9q2 , 0:7a0 + 2q + 0:25a0 + 0:65 Using a0 = ,0:25, 5.45 gives after cancelling the common factor Am :
,2:6 + 0:4s0 = a0 , 0:7 2:25 + 0:4s1 = ,0:7a0 + 0:25 ,0:65 + 0:4s2 = 0:25a0
Problem 5.17
The minimum degree solution has deg A0 = 1 and gives a unique solution to the Diophantine equation. Let us instead use deg A0 = 2 and deg S = deg R , 1. This gives the equation z + 1z + 2z 2 + r1z + r2 + z s1 z + s0 = z 2 z 2 with the solution
R0 = z 2 , 3z S0 = 7z + 6 The controller is causal. Using Theorem 5.1 we also have the solutions R = R0 + Qz S = S0 , Qz , 1z , 2
where Q is an arbitrary polynomial. Choose for instance Q = ,1. This gives
R = z 2 , 3z , z = z 2 , 4z S = 7z + 6 + z2 , 3z + 2 = z2 + 4z + 8
This is also a causal controller. The closed loop systems when using R0; S0, T0 = S0 and R; S , T = S respectively are
BS0 = z7z + 6 AR0 + BS0 z4 BS = zz2 + 4z + 8 AR + BS z4 The number of zeros are di erent.
60
Solutions to Chapter 6
Problem 6.1
In the rst case it is assumed that we have a control structure as in Fig. 6.1. There are three subsystems each with the transfer function K Gi s = s + iK i and the total transfer function from uc to y is K2 G = s +G1G2 G3 G = ss + K s + KK1s +K3 + K K K K K4G1 G2 3 K3 1 2 1 2 3 4 If either of the gains Ki is increased su ciently much the closed system will become unstable. Fig. 6.2 shows the response when uc is an impulse and when K1 = K2 = K3 = 1 and K4 = 0:1, 0.25, and 0.75. A disturbance in the process will propagate in the direction of the ow. In the case of control in the direction opposite to the ow each of the subprocesses
Raw material flow u c x2
1 s
K1
1 s
K2
x3 1 s K3 1 s
Final product x 4 =y
K4
Figure 6.1
Block diagram for the control in the direction of the ow in Problem 6.1.
Output
0 0
Figure 6.2
K4 = 0:1 solid,
Impulse response for the control in the direction of the ow when 0.25 dashed, and 0.75 dash-dotted.
25 Time
50
61
x 4 =y 1 s
uc x3 1 s K3 x2 K2 1 s
K4
1 s
1 s
K1
1
Figure 6.3
Problem 6.1.
has a transfer function of the type Gi . The system is then represented with the block diagram in Fig. 6.3. Notice the order of the states. The system will remain stable for all positive values of Ki . A disturbance will now propagate in the direction opposite the ow. A disturbance in uc will now only in uence the rst subprocess and will not propagate along with the ow. The reader is strongly recommended to compare with the case where the disturbance appears at the nal product storage instead.
Problem 6.2
Fig. 6.3 in CCS contains several examples of couplings of simple control loops. a. Cascade control loops are found for the cooling media ow and for the output product ow. b. Feedforward is used for the level control loop where the input ow is used as a measurable disturbance. The input ow is also used as feedforward for the cooling of the jacket. c. Nonlinear elements are used in the ow control loops of the product output and the coolant ow. The ow is probably measured using di erential pressure which is proportional to the square of the ow. The square root device is thus used to remove the nonlinearity of the measurement device. An intentional nonlinearity is introduced in the selector. Either the temperature or the pressure is used to control the coolant ow depending on the status of the process.
62
Solutions to Chapter 7
Problem 7.1
!s = 2h
where F ! is the Fourier transform of the time continuous signal. The Fourier transform of sin !0 t has its support i.e., , set , it is 6= 0 in the two , the+ ! where! , ! . Thus, if the points !0 . More precisely, it equals i ! 0 , 0 signal sin !0 t is sampled with the sample interval h its Fourier transform will be 6= 0 in the points
Problem 7.2
We have the following speci cations on the choice of sampling period and presampling lter: 1. All frequencies in the interval ,f1 ; f1 should be possible to reproduce from the samples of the continuous time signal. 2. We want to eliminate the disturbance with the known and xed frequency f2 5f1. 63
10
a) h=2/10
10
20
30
b) h=2/20
0 c) h=2 /50
10
20
30
Figure 7.1
10
25
50
The sampling theorem states that the rst speci cation will be satis ed if and only if the sample frequency fs is chosen such that
fs 2f1
Moreover, for the disturbance f2 not to fold on the data signal fs =2 , f1 f2 , fs =2
fs f2 + f1 = 6f1
Two cases: 1. Filter out the disturbance using an antialiasing lter. Then sample with fs 2f1 . Suppose the disturbance should be attenuated 20 dB without e ecting the datasignal. A n:th order lter gives maximally n 20 2 dB decade. So to achieve 20 dB in log f1 = 0:699 decades takes n 2. f 2. If fs 6f1 , the disturbance does not mix with the data signal. It can instead be removed using digital lters.
The magnitude of the spectrum of the sampled signal can be obtained by folding the spectrum of the time continuous signal around the angular frequency !N = !s=2 = =h. See Fig. 7.1 and Fig. 7.2. 64
100
100
a) s =120
120
100
60
20
20
60
100
120
b) s =240
140
100
100
140
Figure 7.2
s r
r s
4 3
1 2
1 4 2
Figure 7.3
Problem 7.7
The rotation frequency of the wheel !r = 2r. The frequency of the camera shutter !s = 2=h. The picture will not move if !r = n !s ; for integer values n. A correct picture will be seen, if !s 2!r according to the sampling theorem. The eye acts like a low pass lter. The wheel will appear to rotate with a frequency lower than r if !s 2!r . See Fig. 7.3. For instance let !s = 4=3 !r . Aliasing will give a frequency ! = 1=3 !r. The wheel then appears to rotate three times slower and in the wrong direction. If !s = !r the wheel will appear to stand still. Compare the stroboscope.
65
Problem 7.9
The signal is 1 ut = sin4!0t cos2!0t = 2 sin6!0t + sin2!0t Sampling the signal with
!s = 2h = 6!0 gives the Nyquist frequency !N = 3!0. Sampling the signal ut gives the alias of sin6!0 t in ! = 0. We thus get the frequencies f1 = 0 f2 = !0
66
Solutions to Chapter 8
Problem 8.1
The three transformations Euler's method forward di erence 8.4, backward di erence 8.5 and Tustin's approximation 8.6 have di erent stability properties. This can be seen by nding how the left half s-plane is transformed into the z -plane. For Euler's method we have z = sh + 1 This implies that the stability boundary in the sh-plane the imaginary axis is translated one unit to the right, see Fig. 8.1a. When the backward di erence is used then 1 z = 1 , sh For s = i! we get 1 1 , !h This represents a circle with radius 0.5 and going through the points 0 and 1, see Fig. 8.1b. Finally for Tustin's approximation with s = i! z = 1 + i!h=2 1 , i!h=2 Now jzj = 1 arg z = 2 arctan !h The imaginary axis is thus transformed into the unit circle in the z -plane. If a transfer function is stable in the s-plane it will be translated into a stable discrete time system if using the backward di erence or Tustin's approximation.
Problem 8.2
a Gs = s + a ;
a 0
Forward differences
Figure 8.1
Transformation of the left half s-plane when using a. Eulers method, b. Backward di erence and c. Tustin's approximation.
Backward differences
Tustin
67
ah a H z = z , 1=h + a = z , 1 + ah
This corresponds to the di erence equation
jah , 1j 1
0 h 2=a: The approximation may, however, be poor even if the di erence equation is stable. b. For Tustin's approximation we get or
z+1 a a= = 1 + a= H z = z , 1 a= , 1 z+ +a z+1 a= + 1
where Thus
a = tanah=2
tanah=2 H z = 1 + tanah=2
Problem 8.3
should be approximated using di erent methods. At ! = 1:6 rad s it has the argument 19 and the gain 2.95. 68
Gk s = 4 s + 1 s+2
Within two decimals this is the same as in c. e. Zero order hold sampling gives
,2h ,2h ,2h 0 1 HZOH z = 4 , 4 2 1 , e,2h = 4 z , e z , 1,,h e =2 = 4 z , 0::803 z,e ,e 2 z , 607
All ve approximations have all the form H z = K z + a z +b The gain and the phase at ! = 1:6 are obtained from
i!h i!h e,i!h + b H ei!h = K ei!h + a = K e i!h + ae,i!h + b e +b e + b 1 + ab + a + b cos!h + ib , a sin!h =K 1 + b2 + 2b cos!h b , sin arg H ei!h = arctan 1 + ab + aa + b!h !h cos
jH ei!hj = K
1 + a2 + 2a cos!h 1 + b2 + 2b cos!h
10
Gain
10 2 10 30
0
10
10
10
10
Phase (deg)
20 10 0 2 10
Figure 8.2
10
The Bode diagrams for the lter in Example 8.3 when h = 0:25 continuous time lter full; Euler's method dashed; backward di erence dotted; Tustin dash-dotted.
10 Frequency (rad/s)
10
10
Euler Backward Tustin Tustin with prewarping Zero order hold Fig. 8.2 shows the Bode diagrams for the continuous time system and for the Euler, backward and Tustin approximations. Fig. 8.3 is the same as Fig. 8.2 but with h = 0:05. The shorter sampling period gives a better approximation. It is assumed that the sample and hold circuit can be approximated by a delay of h=2 seconds. Further we will allow a decrease of the phase margin of 5 , 15 . This approximately corresponds to a decrease of the damping by 0:05 , 0:15. A time delay of h=2 seconds gives at the crossover frequency a decrease of ! ch ' = !c h=2 rad = 180 c h = 0!035 = 5 , 15 2 : This gives !c h = 0:17 , 0:52 or approximately !c h 0:15 , 0:5: 70
Problem 8.4
10
Gain
10 2 10 30
0
10
10
10
10
Phase (deg)
20 10 0 2 10
10
Figure 8.3
10 Frequency (rad/s)
10
10
Problem 8.5
K GI s = T s
i
HI z = T Kh 1 z,
i
which is the same integral part in 8.23. The derivative part of 8.22 has the transfer function KT s GD s = 1 + T ds=N Using backward di erence gives
d
a. Approximation of the integral part with backward di erence gives Khz HI z = T z , 1 i An error will then directly in uence the computation of the integral part. Euler's approximation gives a delay of one sampling interval before an error will in uence the integral part. The sampling interval is, however, usually short for digital PID-algorithms. b. Euler's approximation for the derivative part gives KN , 1 Hdz = z , 1 +zhN=T d A small value of Td can make Hd unstable. Since the D-part of a PIDcontroller sometimes is not used it is necessary that the regulator remains stable when Td = 0.
Problem 8.6
= K 1 + 2T 1 + 2T + hz , 1 i i This is of the same form as 8.24 with h Kd = K 1 + 2T i Tid = Ti + h=2
1 0 h 1 1 HT z = K B1 + 2 z , 1 C = K 1 + 2h + T z , 1 A @ Ti i Ti h z + 1 h 2h
Problem 8.7
The tank process in Problem 2.10 has the transfer function 0:000468 Gs = s + 0:0197s + 0:0129 a. At the desired cross over frequency we have jGi!cj = 0:525 arg Gi!c = ,115 We will use a PI controller of the form Gr s = K Ts + 1 Ts and we want the gain 1=0:523 and the phase ,15 degrees at !c . This gives K = 1:85 T = 149 72
Output
0 0
Step response of the tank process when controlled with a continuous time solid and a discrete time PI controller. The sampling interval is 6 dashdotted and 12 seconds dashed.
Figure 8.4
250 Time
500
d. Fig. 8.4 shows simulations of the step response of the system controlled with the continuous time and the approximate discrete time PI-controller when h = 6 and 12 seconds. 73
Output
0 0
2 Time
Stepresponses for the motor in Problem 8.9 when a continuous time solid, a discretized dash-dotted and a modi ed discretized state dashed feedback controller is used when h = 0:25.
Figure 8.5
Problem 8.9
a. The continuous time controller is
c. Fig. 8.5 shows the stepresponse of the system when using the continuous controller and the controllers in a and b when h = 0:25. It is possible to calculate backwards to nd out the corresponding damping and natural frequency for the controllers in a and b. A discrete time state space 8 9 representation of the motor is given in A.6. Using L = : `1 `2 ; gives
8 0:336 ,0:885 9 , ,L = : 0:164 0:885 ; 8 9 and for h = 0:25 and L = : 1:75 3 ; we get 8 0:392 ,0:664 9 , ,L = : 0:171 0:914 ;
These two matrices have the characteristic equations
z2 , 1:221z + 0:442 = 0
and
z 2 , 1:305z + 0:471 = 0: From the equations given in Example 2.16 we can calculate the corresponding continuous time systems. For the discretized controller L = 8 9 : 2 4 ; we get
= 0:71 !0 = 2:31
= 0:77 !0 = 1:96
The change in the damping is smaller when the modi ed controller is used and the change in the undamped natural frequency is also smaller.
Problem 8.10
a. We rst want to compute a state feedback such that A , BL has the characteristic equation s2 + 8s + 32 = 0: 8 9 Assume L = : `1 `2 ; then
, A , BL = : ,`3 ,2 1 ` ; , 2 1
The characteristic equation of A , BL is s + 3s + 2 + `2 + `1 = s2 + 5 + `2 s + 6 + 3`2 + `1 s + 8s + 32: This gives
L = : 17 3 ;
75
Problem 8.12
a. Using 8.16 and 8.17 give 1 ~ L = L I + A , BL h = L : ,h=2 1h=2h ; 2 , 9 8 9 8 9 8 1 h=2 9 8 = : 1 2 ; : ,h=2 1 , h ; = : 1 , h 2 , 3 h ; = : 0:8 1:7 ; 2 8 9 ~ M = : I , LBh=2 ; M = 2 , 2h = 1:6 b. Using the backward di erence approximation gives 1 , q ,1 I xk = A , KC ^k + Buk + Ky k x h ^ I , Ah + KCh^k = q ,1 xk + Bhuk + Khy k x ^ Introduce 0 = I , Ah + KCh,1 = This gives 1 h 1 2 : ,h 1 + h ; 1+h+h
xk = 0xk , 1 + 0Bhuk + 0Khy k ^ ^ 8 0:81 0:16 9 8 0:03 9 8 0:19 9 = : ,0:16 0:97 ; xk , 1 + : 0:19 ; uk + : 0:16 ; y k ^
76
0.5
Output
0 0.5 0
The response of the system in Problem 8.10 when the state feedback controller in a is used with h = 0:1 dashed , 0.2 dash-dotted and 0.3 dotted. The response for the continuous-time controller is also shown solid.
Figure 8.6
2 Time
0.5
Output
0 0.5 0
The response of the system in Problem 8.10 when the modi ed state feedback controller in b is used with h = 0:1 dashed, 0.2 dash-dotted and 0.3 dotted. The response for the continuous-time controller is also shown solid.
Figure 8.7
2 Time
77
Solutions to Chapter 10
Problem 10.1
a. Better sensors, for instance a tachometer with less noise. b. Flow control with local feedback. c. Temperature control in houses.
Problem 10.2
a. The time function y t = sin!t has the z-transform z Y z = z 2 , 2zsin !h + 1 cos !h See Table 2.1. Consider a system with
Hdz = Y z
The impulse response of Hd will thus be sinkh! . That this is the correct answer is easily seen by making long division.
Problem 10.3
Using the model of the disturbance gives yk + m = C q wk + m: Aq Introduce the identity
The operator qGq =C q is casual since deg C = deg G + 1. Let Aq = q , 0:5, C q = q and m = 3 then q2 q = q , 0:5q2 + f1q + f2 + g0 This gives the system of equations 0 = ,0:5 + f1 0 = ,0:5f1 + f2 0 = ,0:5f2 + g0 with solution f1 = 0:5 f2 = 0:25 g0 = 0:125 The predictor at k + 3 given data up to and including k is thus y k + 3jk = 0:125q y k = 0:125y k ^ q Let wk be zero except at k = 0 and 5 when it is assumed to be one then k y k = 0:5y k , 1 + wk ykjk , 3 ^ ,1 0 0 0 1 0 1 0:5 0 2 0:25 0 3 0:125 0:125 4 0:063 0:063 5 1:031 0:031 6 0:516 0:016 7 0:258 0:008 8 0:129 0:129 9 0:064 0:064 10 0:032 0:032
where deg F = m , 1 and deg G = n , 1. Then yk + m = F q wk + 1 + qGqq wk = F qwk + 1 + qGqq y k A C If wk + 1; : : :; wk + m are assumed to be zero then the best prediction of yk + m is y k + m = qGqq y k: ^ C
Problem 10.4
Using 10.11 we nd that the stationary variance of x ful ls stable P = P T + R1 The stationary covariance exists since the system is stable. Since R1 is symmetric P is also symmetric. Introduce 8p p 9 P = : p11 p12 ; 12 22 79
then
8 p p 9 8 0:4 0 9 8 p p 9 8 0:4 ,0:6 9 8 1 11 12 : p12 p22 ; = : ,0:6 0:2 ; : p11 p12 ; : 0 0:2 ; + : 0 12 22
p11 = 0:16p11 + 1 p12 = ,0:24p11 + 0:08p12 p22 = 0:36p11 , 0:24p12 + 0:04p22 + 2
0 2;
This gives
The solution is
The stationary covariance function is Exk + xkT = P It remains to compute . The eigenvalues of are 1 = 0:4 and 2 = 0:2. Using the results on matrix functions in Appendix B we get = 0I + 1 where 0 and 1 are obtained from 1 = 0 + 1 1 2 = 0 + 1 2 The solution is ,1 ,1 = ,12 1, , 2 0 1 2 1 , 2 1= , 1 2 and 8 9 0:4 0 = : ,30:4 , 0:2 0:2 ; Finally
8 9 1:19 0:4 ,0:31 0:4 rx = : ,3:57 0:4 + 3:27 0:2 0:93 0:4 + 1:68 0:2 ; 0
From the state space description we get the input-output description of the process yk + a1yk , 1 + + an yk , n = c1v k , 1 + + cnv k , n where ai ; i = 1; : : :; n are the coe cients of the characteristic polynomial of the matrix . Multiply the equation above by y k , and take the mathematical expectation. y k , is independent of all the terms on the right hand if n + 1. Thus ry + a1ry , 1 + + an ry , n = 0: This is called the Yule-Walker equation. 80
Problem 10.5
Problem 10.6
There are two noise sources v1 and v2 that is used to generate y k. Using Theorem 10.2 we nd that the spectral density of y is
y
= H z v H T z ,1
where z = ei!
H z = C zI , ,1, = : 1
and Thus
98z +a 1;:
0
2
0 ,1 1 = 8 1 : z+a z + b;
1 z+b
9 ;
8 2 09 1 v=: 0 2; 8 1 = : z+a y
1 z +b 2
98 ;:
2 1
9 8 ,11 9 z +a 2;: 1 ;
2
z,1 +b
2 1 = z + az ,1 + a + z + bz ,1 + b
Using the spectral factorization theorem Theorem 10.3 we nd that we can generate the same spectral density by sending white noise through H1z = z +za+zc + b this gives the spectral density
1
Problem 10.7
xk + 1 = axk + v k yk = xk + ek This is the same as in Example 10.3 with the exception that Ev kes = r12 k , s. The covariance function for x will thus be the same but the covariance of y will contain an additional term due to the correlation between v and e. From Example 10.3 we know that r rx = aj j 1 ,1a2
81 The process is
The covariance of y is
ry = E fy k + ykg = E f xk + + ek + xk + ek g = = rx + rxe + rex + re = rx + rxe + rxe, + re
where it has been used that rex = rxe , in stationarity.
80
0 = 1 = 0 ,1 r a 12 1
0 1
0 =0 = 0
8 a,
6= 0 =0
r , 2 r r 12 12 = 21 1 ,1a2 , ra ei! , 1 ea i! , a , 1 ,1a2 , ra + 1 ,1a2 + r2 = , a 2 i! ,i! ei! , a e,i! , a = 21 r1a + r121 , a , r12aei! ,e,i! ,a+ r2ae , ae , a a 1 where it has been used that
1 1 X r e,i! = y ! = 2 y =,1
1 X
2
82
r2a , r12 = 2c A more elegant solution The ouput can be written as
8 9 r ,1 = 21 : H z 1 ; : r 1 rr12 ; : H z ; y 1 12 2 1 r H z H z ,1 + r H z + r H z ,1 + r = 2 1 12 12 2 1
r1 = 21 z , az ,1 , a + r12 z , a + z ,11, a + r2 ,1 + 21 + 2 = 21 z r12 , r2a + r1 , 2zar12 zr,1 , aa + z r12 , r2a ,a
r21 + a2 + r1 , 2ar12 = 2 1 + c2 ar2 , r12 = 2 c
Problem 10.8
The process can be written as q,c a,c yk = q , a ek = q , a ek + ek = xk + ek where
aj j ry = 1 , r12 + aj j,1r12 a
83
with a = 0:7 and c = 0:5 we get for 6= 0 02 0 ry = 0:7j j 1 ,:0449 + 0::7 0:7j j = 0:360:7j j 0: For = 0 r a,c 2 ry 0 = 1 ,1a2 + r2 = 1 , a + 1 = 1:08 2 The variance can also be obtained from Theorem 10.4. 2 ry 0 = I1 = 1 +1c,,22ac = 1:08 a Further 10.17 gives z , cz ,1 = 21 z , az ,1 , c = 21 1:1:25 ,:cos ! ! y , a 49 , 1 4 cos where z = ei! .
Problem 10.9
The recursive scheme in Section 10.4 can also be used to get 1 ,1:5 0:7 1 0:2 0 0:7 ,1:5 1 0:7 ,1:5 1 2 = 0:7 0:51 ,0:45 1 0:2 ,0:45 0:51 0:51 1 = ,0:8824 ,0:45 0:1129 1:1765 0:1129 This gives
I2 = 10:4167 1:1765 + 0:3922 0:2 12:33 These calculations have been performed in high precision and thereafter rounded.
84
Problem 10.10
The process is
ry 0 = E ek , 2ek , 1 + 3ek , 2 , 4ek , 32 = E ek2 + 4ek , 12 + 9ek , 22 + 16e6 , 32 + crossterms = 1 + 4 + 9 + 16 = 30
The mean value of the crossterms are zero since
E fek + ekg = 0
In the same way
6= 0
ry 1 = 1 ,2 + ,2 3 + 3 ,4 = ,20 ry 2 = 1 3 + ,2 ,4 = 11 ry 3 = 1 ,4 = ,4 ry 4 = 0 k 4
Problem 10.11
a.
b H z = z , a
b2=2 1 + a2 , 2a cos !
2 2 1 +2a b
1 2a , 2 ,b2 cos !
Identifying with the desired spectral density gives 2 1 + a2 = b2 1:36 2 2a = ,1:2b2 85
H z = z +20:6
Problem 10.12
8p : p11 12
P = P T + R1 9 8 98 98 9 8 9 p12 ; = : 0:3 0:2 ; : p11 p12 ; : 0:3 0 ; + : 1 0 ; p22 0 0:5 p12 p22 0:2 0:5 0 0:5 p11 = 0:09p11 + 0:12p12 + 0:04p22 + 1 p12 = 0:15p12 + 0:1p22 p22 = 0:25p22 + 0:5 8 1:1385 0:0784 9 P = : 0:0784 0:667 ;
Problem 10.13
b H z = z , a
gives the spectral density
r1 b2 ! = 2 1 + a2 , 2a cos !
In this case r1 = 1. Identify with the desired spectral density gives 1 3 1 1 2 5:43 , 5:40 cos ! = 2 1:81 , 1:8 cos ! This gives a = 0:9 and b = 1. 86
Solutions to Chapter 11
Problem 11.1
kh; kh = 1
Zt
kh
b e,at,s b ds = a 1 , e,at,kh
The discrete time loss function is obtained from CCS 11.6-11.8, which gives
Q1 = Q12 = Q2 =
Z kh+h
Zkh +h kh
kh2 b = a2
Zkh +h b2 kh
ds a2
b2
+ h , 2a3 3 , 4e,ah + e,2ah
Problem 11.2
80 x = :0 _ 8 y = :1
= eAh ,=
1 x+ 0 u 0; :1; 9 0;x
8 9
Zh
0
81 = :0
eA
8 h2=2 9 Bd =: h ;
87
h; 1
Sample the loss function Q1c = I Q2c = 1 Q12c = 0. Using 11.6-11.8 we get
9 8 h 9 h2=2 2 + 1 ; d = : h2 =2 h3 =3 + h ; 0 0 h8 9 8 2=2 9 Zh 8 2=2 9 8 h3=6 9 Z Q12 = : 1 0 ; : ; d = : 3=2 + ; d = : h4=8 + h2=2 ; 1 0 0 ! h 8 8 2=2 9 ! Zh Z 9 3 5 4 2+ 2 =2 ; : ; +1 d = 1+ d = h+ h + h Q2 = : 4 3 20 Z 0 1 d = 1;:0 1;
h
0 0
Zh 8 1 Q1 = :
98
81 :
Problem 11.3
The Riccati equation11.17 gives + 12 sk = a2sk + 1 + 1 , a2b2 bs2skk + 1 = 1 k = N , 1; ::; 1 and 11.19 gives which gives the controller
Lk = b,2 ba = a b
xk + 1 = 0
The state is zero after one step and the resulting controller is thus a dead-beat controller.
Problem 11.5
a. The loss function is y 2 + u, i.e. 0 0; The steady state value of the Riccati equation is obtained from
Q1 = C T C
81 = :0
S = T S + Q1 , T S ,Q2 + ,T S ,,1,T S
Let 88
8s s 9 S = : s11 s12 ; 12 22
Pole
0.5
0 2 10
Figure 11.1
10
10
10
22
s12 = s11 , + s
The solution is The feedback vector is where
8 9 L = K : 1 1 ;
s12 = 1 + p1p 4 + K = + s 22 2 + 1 + 1 + 4 b. The dynamics of the closed loop system is 8 1 9 1 , ,L = : ,K ,K ; The poles are obtained from , 1 + K + K = , 1 + K = 0 There is one pole in the origin and one in 1 , K . For = 0 then 1 , K = 0 and as ! 1 then 1 , K ! 1. Fig 11.1 shows how the pole varies as a function of .
89
a) y(k), u(k)
2 0 2 0 5 10
b) y(k), u(k)
2 0 2 0 5 10
c) y(k), u(k)
2 0 2 0 5 Time 10
d) y(k), u(k)
2 0 2 0 5 Time 10
The output solid and the control signal dashed for the system in Problem 11.5 when a = 0, b = 0:5, c = 5 and d = 25.
Figure 11.2
The poles of the closed loop system can also be determined from 11.40. For the inventory system Az = qq , 1 B z = 1 and we get
r = 2 + 1 , 1 + 4 = 2 + 1 +2 1 + 4 p p1 = , 2 + 1 , 1 + 4 2
The poles of the closed loop system are thus one at the origin and one at
2 2p
,p1 = 2 + 1 , 1 + 4 2
It is easily seen that this is the pole in 1 , K .
90
Im
1 1
Figure 11.3
0 Re
The system has the transfer function 0 030 H z = z2 :, 1:z + 0:026 65z + 0:68 Only the output and the control signals are penalized in the loss function. The closed loop system has poles that are determined by the stable roots of
Problem 11.6
+ H zH z ,1 = 0
This gives
,1 :026 0 030 + z2 :, 1:z + 0:026 z ,02:030z65z+10+ 0:68 65z + 0:68 , 1: ,
or
2 :030z 0 0 030 = + z 2 :, 1:z + 0:026 10, 1:65+ +:026zz 2 = 0 65z + 0:68 z 0:68
0:00078z 3 + 0:001576z 2 + 0:00078z + 0:68z 4 , 2:7720z 3 + 4:1849z 2 , 2:7720z + 0:68 = 0 Fig. 11.3 shows the closed loop poles when is varying.
Problem 11.9
Solving LQ-problem for system of higher order than one by hand causes much work. With numerical packages, like Control toolbox for Matlab, the design is signi cantly simpli ed for the control engineer. The following Matlab-macro illustrates the solution of Problem 11.9, i.e. sampling of the system, sampling of the loss function and solving the Riccati equation. 91
Macro for Problem 11.9 CCS alpha=0.001; k=0.0005; rho=0.08; h=5; A= 0 1; 0 -alpha ; B= 0; k ; Q1c= 1 0; 0 0 ; Q12c= 0; 0 ; Q2c=rho; Transform continuous-time LQG problem to the corresponding discrete-time LQG problem Phi,Gam,Q1,Q2,Q12,R1,Je = lqgsampA,B,h,Q1c,Q2c,Q12c,zeros2,2; Linear quadratic regulator design for discrete-time system L,Lv,S = lqrdPhi,Gam,Q1,Q2,Q12; L
L = 3:055 108:7
Problem 11.10
In Problem 11.5 we have determined r, then 2 = r 2 + 1 + p1 + 4 Equation 11.37 may be used to get the exact values of the gain margin. With
A z = z 2 , z P z = z2 + p1z
we get
z 2 , z + z 2 + p1z , z 2 + z = z z + p1 + , 1 = 0
I.e., the system is stable if
p1 + , 1 1 4 min = 0 ,1 + p1 + 4 =
min
,1
max
and
max
92
10
Gain margin
0 0
Figure 11.4
min
dashed and
max
Problem 11.11
xk + 1 = 0:5xk + v k y k = xk + ek Theorem 11.5 gives that the Kalman lter is de ned by 8xk + 1jk = 0:5^kjk , 1 + K kyk , xkjk , 1; x0j , 1 = 0 ^ x ^ ^ 5 K k = r0:+PPkk 2 2 : : P k + 1 = 0:25P k + r1 , 0225PPkk ; P 0 = r0 r + The dynamics of the lter is determined by r , KC = 0:5 , K k = r 0:5P2k 2+ The steady state variance is given from P 2 + 0:75r2 , r1P = r1 r2
Consider three cases r1 r2 , r1 = r2 and r1 r2. In the rst case P r1 and , KC 0. In the second case P = 1:13r1 and , KC = 0:23. Finally if r1 r2 then P 1:33r1 and , KC 0:5. Fig. 11.5 shows , KC for di erent values of r1 when r2 = 1. The system is
Additional problem
Suppose that the system in Problem 11.11 has a control signal uk, i.e. the system is xk + 1 = 0:5xk + v k + uk yk = xk + ek Determine a steady-state LQG-controller when Q1 = 1, Q12 = 0 and Q2 = . 93
0.5
Pole 0 0
The pole of the Kalman lter in Problem 11.11 for di erent values of r1 when r2 = 1.
Figure 11.5
5 State weighting r1
10
Using the Kalman lter from Problem 11.11 and the LQ-controller gives
xk + 1jk = ^kjk , 1 + ,uk + K y k , C xkjk , 1 ^ x ^ = ^kjk , 1 + , ,Lxkjk , 1 + K y k , C xkjk , 1 x ^ ^ = , ,L , KC ^kjk , 1 + Ky k x
and thus or
Problem 11.12
a. Equation 11.47 gives
4 2 0 4 2 0
0 Time
Figure 11.6
with
1 1; 8 9 T = :0 0; R1 = ,v ,v 8 90 1 C = :1 0;
2
81 = :0
we get p11k + 1 = p11k + 2p12k + p22k , p11kp +p12k 11 k 2 = p11kp22kk, p12k p11 p12k + 1 = p12k + p22k , p12kp11k + p12k = p11k + 1 p11k 2 p22k + 1 = p22k + 1 , p p12k = 1 + p11k + 1 11 Further
k 0
The poles of the lter are found from detI , , KC = 2 = 0 The lter has a dead beat response. b. The initial values of the lter are
x0j , 1 = : 1 1 ; ^
9T
and assume that P 0 = 3I . Fig. 11.5 shows the elements of the covariance matrix as a function of time. 95
Problem 11.14
Introduce the notation 81 19 809 8 0:5 9 = : 0 1 ; ; ,1 = : 1 ; ; and ,2 = : 1 ; The state at time k + 3 can now be determined xk + 3 = xk + 2 + ,1 vk + 2 + ,2 = 2 xk + 1 + ,1 v k + 1 + ,2 + ,1 v k + 2 + ,2 = 3 xk + 2,1 v k + ,1 v k + 1 + ,1 v k + 2 + 2,2 + ,2 + ,2 The best estimate of xk given y k is determined from 11.50. Since v k, v k +1 and vk +2 are independent of y k then the best estimate of xk +3 is given by
81 2 = :0
8 9
The variance of the estimation error is 2, P k + 3jk = 3 P kjk3T +8 :019 1 ,T 28T + ,1 ,T T + ,1 ,T 0 1 1 81 39 9 1 1 0 + 0:01 5 3 = : 0 1 ; P kjk : 3 1 ; :3 3; If x0 is known then P 0j0 = 0 and
Problem 11.15
xk + 1 = axk + vk y k = xk + ek We use the exponential smoothing estimator
cov v = 1 cov e =
xkjk = xk , 1jk , 1 + 1 , y k ^ ^ xkjk,xk = 1q, q yk , xk ^ , ! 1 , q 1 v k + ek , 1 v k = q, q,a q ,a q, = , q , aq 1 v k + 1q, q ek , , Using Theorem 10.4 we get
2 2 var xkjk = 1 + a1 + 1 , a + 1 , 1 ~ + 96
This will have the same form as the exponential smoothing estimator only when a = 1. Kalman variance a2 P 2 P = a2 P + 1 , P + This gives
K = PaP +
Numerical values: = 1; a = 0:5. Exp. smoothing: Kalman: = 0:4222; K = 0:2650; var xkjk = 0:6181 ~ var xkjk = 0:5311 ~
Problem 11.16
Let and
8 : x mv me 9T ; X= 819 8a 1 09
819
8 C 9 81 0 Wo = : CA ; = : a 1 a2 a + 1 CA2
19 1; 1
rank Wo = 2 This means that me and mv are not both observable and no Kalman- lter can be designed. It is, however, possible to design a second order observer with reconstruction of a linear combination of me and mv . Rede ning the state vector X as 8x 9 1 X = :m; where x1 = x + me m = a , 1me + mv gives 8a 19 819 819 8 X k + 1 = : 0 1 ; X k + : 0 ; uk + : 0 ; v1k 8 9 : yk = : 1 ; X k + e1k 0 Reconstruction of x1 and m is possible if these states are observable. The observability matrix is
with
8 C 9 81 Wo = : C ; = : a
0 1;
Problem 11.17
The constants a1 and a2 are determined by the following two conditions 1. The correct mean value should be obtained. 2. The variance should be minimized. Condition 1 gives that a1 + a2 = 1 The variance of the estimation error is V = E xk , xk2 = E xk , a1 xk , a1e1k , a2xk , a2 e2 k2 ^ 2 1 + a2 9 = a2 + 1 , a 2 9 = 10a2 + 9 , 18a = a1 1 1 2 1 1 Taking the derivative with respect to a1 gives the condition 9 20a1 , 18 = 0 a1 = 10 The estimator is thus 1 9 xk = 10 y1k + 10 y2 k ^ 98
9 Vmin = 10 Using only y1 gives the variance 1 and only y2 the variance 9. Thus, a combination of the measurements gives a better result than using only the best measurement. Assume that the a priori estimate of x is zero and the variance of x is p, i.e. x0j0 and P 0j0 = p ^ From 11.50 11.54 we get P 1j0 = p and 8 9 , K 1 = P 1j0C T R2 + CP 1j0C T ,1 = 10pp+ 9 : 9 1 ; This gives 9 xkjk = 10p p 9 y1 k + 10pp+ 9 y2 k + 10p9+ 9 xkjk , 1 ^ ^ + If p is large then the weights for y1 and y2 will be those calculated above. In the example R1 = 0 the steady state gain will then be zero since the estimate will approach the true value of x.
Problem 11.20
Q12 = 0 Q2 = 0 The steady state solution is obtained from 11.17. Using Matlab we get the solution 8 0:25 0:19 9 S = : 0:19 0:1444 ; 8 9 L = : 2:21 ,0:45 ; An alternative solution is to use 11.40 AzAz,1 + Bz B z,1 = rP z P z ,1 = 0 P z = b1 zB z = 2z 0:5z + 0:38 = z z + 0:76 1 Now we have to nd L such that 8 ,0:76 0 9 , ,L = : 1 0 ; where controllable canonical form have been used. This gives 9 8 L = : 2:21 ,0:45 ;
99
8 1:45 ,0:45 9 819 xk + 1 = : 1 0 ; xk + : 0 ; uk 8 9 yk = : 0:5 0:38 ; xk 9 8 T C = : 0:25 0:19 ; Q1 = C 0:19 0:1444
Problem 11.21
a. First assume that = 0 and use linear quadratic design with
Q1 = Q12 = 0 Q2 = Q0 = 1 N =2
Theorem 11.1 gives the Riccati equation
S N = S 2 = Q0 = 1
0 52 S 1 = :+ 1
: L1 = 0+51
The control law is uk = ,Lkxk = ,Ly k. For di erent values of we get 1:0 0:1 0 L0 0:056 0:093 0:1 L1 0:250 0:455 0:5 b. In this case = 1 and xk is reconstructed using a Kalman lter
R1 = R12 = 0 R2 = = 1 , R0 = E x2 0 = 1
5 K h = 10:+P k Pk
Problem 11.22
a.
8 2 09 R1 = 0:01; R2 = : 01 2 ; 2 8 1 9
xk + 1jk = 1 , K : 1 ; xkjk , 1 + Ky k ^ ^ 8 9 8 12 0 9 8 1 9 8 9
,1 8 1 9 :1 1; : p:1 1; :1; 2 + 0 2 ; :1; 8 98 2 p 9,1 8 1 9 = 0:01 2:1 1;: 1 + p p ; :1; 2 p 2 +p
p2
b. 11.47
p = p + 0:01 , p2
c.
2 2
101
Solutions to Chapter 12
Problem 12.1
where G is obtained from the identity 12.17 and the variance of the prediction error is given by 12.19. For m = 1 we get q 2 , 1:4q + 0:5 = q2 , 1:2q + 0:4 + g0q + g1 which gives Gq = g0q + g1 = ,0:2q + 0:1 The predictor is then given by q2 0 1 y k + 1jk = q,0:21:4++ :0:q5 y k ^ 2, q and the variance of the prediction error E y2k + 1jk = 2 = 4 ~ For m = 2 qq 2 , 1:4q + 0:5 = q2 , 1:2q + 0:4q + f1 + g0 q + g1 This gives F q = q , 0:2 Gq = ,0:14q + 0:08 and E y 2k + 2jk = 21 + f12 = 4:16 ~ For m = 3 we get q 2q 2 , 1:4q + 0:5 = q2 , 1:2q + 0:4q2 + f1 q + f2 + g0q + g1 which gives F q = q2 , 0:2q , 0:14 Gq = ,0:088q + 0:056 2k + 3jk = 2 1 + f 2 + f 2 = 4:24 Ey ~ 1 2 Using Theorem 10.4 we can compute the variance of y to vary = 2 2 2 : : 0 5 22 1+1:4 +0:5 1+044+2,1::4,1+4,:1:2 1:2+2 00::51:2 :,0:41+0:4 = 1 , 0: 21 + 0 4 + 1:2 4 1 2 = 4:28 This implies that the prediction variance is almost the same as the variance of y when m 3. 102
Problem 12.2
The solution is
Problem 12.3
a. The C -polynomial has a zero outside the unit circle. Example 12.1 shows how the C -polynomial should be changed. It is replaced by
C z = 5z + 1 = 5z + 0:2
The equivalent process is thus
:99 yk + 2jk = q0+ 0q2 y k ^ : E y 2k + 2jk = 251 + 1:12 = 55:25 ~
Problem 12.4
Using the data given up to time k = 7 it is possible to calculate y k and zd k = z k , y k. zd is the deterministic part of z .
k zk y k zd k 1 320 10 310 2 320 0 320 3 325 ,5 330 4 330 ,10 340 5 350 0 350 6 370 10 360 7 375 5 370 The prediction of the demand for August through November is z 8j7 = zd8 + y 8j7 ^ ^ . . . z11j7 = zd11 + y11j7 ^ ^
We thus need the 1, 2, 3 and 4 step ahead predictors of y. Those are given by solving the identity 12.17 and give
m 1 2 3 4
The prediction is
m 1 2 3 4
104
Problem 12.5
A = q3 , q2 + 0:5q B = q + 0:5 C = q3 + 0:8q 2 + 0:25q It is easily seen that C is a stable polynomial, e.g. by the stability triangle. This is a necessary condition for the minimum variance design method. The pole excess is d = deg A , deg B = 2 The Diophantine equation is q d,1 C q = Aq F q + Gq qq3 + 0:8q2 + 0:25q = q 3 , q 2 + 0:5qq + f1 + g0q2 + g1q + g2 Identifying coe cients of powers of q gives 0:8 = ,1 + f1 0:25 = 0:5 , f1 + g0 0 = 0:5f1 + g1 0 = g2 Solving these equations f1 = 1:8 g0 = ,0:25 + f1 = 1:55 g1 = ,0:9 g2 = 0 The minimum variance regulator is obtained from 12.27 ,1 q 2 + 0 9 q uk = , BGFq yk = q +:555q + :1:q y k q 0: 8 The loss function is Ey2 = 0:521 + 1:82 = 1:06
The noise sequence has a non zero mean value. Introduce ek = 2 + k uk = u + uk ~ where k is zero mean white noise. The process is now yk , 0:5y k , 1 = u + uk , 2 + 2 + k , 0:72 + k , 1 ~ = uk , 2 + k , 0:7 k , 1 + u + 0:6 ~ Choose u = ,0:6 and the problem is reduced to the standard problem. The identity gives q u = q 0:10:2 y k ~ , and uk = q 0:10q:2 y k , 0:6 , 105
Problem 12.6
Problem 12.7
a. The identity gives
F q = q + c , a Gq = aa , cq and the minimum variance controller is cq uk = , qa+a,,a yk c
Problem 12.8
b. The expression above gives the optimal controller uk = 0 if a = 0. The process is then a moving average process of rst order. This implies that uk , 2 cannot be used to decrease the variance of y k.
Problem 12.9
and the minimum variance in the two cases are d=1: Ey 2 = 1 d=2: Ey 2 = 1 + 3:22 = 11:24 b. Fig. 12.1 shows the output of the open loop system for one realization of the noise sequence. The output is drifting since the A-polynomial contains an integrator. Fig. 12.2 and Fig. 12.3 shows the output and the control signal for the same noise realization when the minimum variance controller is used with d = 1 and d = 2. 1 H z = z + a Sending white noise ek through this lter gives the spectral density see Theorem 10.2 1 2 ! = 2 1 + a2 + 2a cos ! This implies that = 1 and a = 0:6 gives the desired spectral density. The process is now described by y k = 1 , 01:5q,1 q +10:6 ek + 1 uk q or q 2 + 0:1q , 0:3 y k = q + 0:6 uk + q ek
F q = q + 3:2 Gq = 5:64q2 , 2:24q The minimum variance controller is q uk = , BGFq y k q
a. Assume that
106
200
Output
100
0
Figure 12.1
50
100 Time
150
200
10
Output
10
50
100
150
200
20
Input
20 0
Figure 12.2
50
The output and the control signal when d = 1 and the minimum variance controller is used for the process in Problem 12.8.
100 Time
150
200
q 2 + 0:1q , 0:3 y k = ,q + 0:6 K y k + q ek yk = q2 + 0:1 + K qq + 0:6K , 0:3 ek The system is stable if
,0:5 K 1:5
107
10
Output
10
50
100
150
200
20
Input
20 0 50
Figure 12.3
100 Time
150
200
Theorem 10.4 gives an expression for the variance of a second order process. : + : I2 K = 1:3 , 0:6K 0077+ 0066K 2 , 0:1 + K 2 : : K
I20:004 = 1:12
d. The minimum variance is Ey 2 = 1 since d = 1.
Problem 12.10
a. With the proportional controller
uk = ,Kyk
we get the closed loop system
I2 1 = 4 3
This is minimal variance for the present control law. With minimum variance control we would get E y 2 = 1. b. From Example 3.2 we nd that the closed loop system is stable if 0:5 ,1 + K , 0:25 0:5 ,1 , K + 0:25 or
,1:25 K 1:75 Both K = 3:25 the second root in a. and K = 2:125 give an unstable closed loop system and the calculation of I2 is no longer valid.
Problem 12.11
y k , 1:5yk , 1 + 0:7y k , 2 = uk , 2 , 0:5uk , 3 + v k
gives
Aq = q 3 , 1:5q 2 + 0:7q Bq = q , 0:5 Note that the process zero 0.5 is stable and well damped. This means that the process zero can be cancelled without problem. The degree conditions gives deg Am , deg Bm deg A , deg B = 2 deg Ao 2 deg A , deg Am , deg B + , 1
109
Rq = R1qB + q = q + 1:5q , 0:5 T q = Bm Ao = q 2 Assuming that uc k = 0 gives S :55 2 , : uk = , Rq yk = , q1+ 1q5q1,05:q5 q : 0
b. v k = ek , 0:2ek , 1; Minimum variance The polynomial C is given by C q = q3 , 0:2q2 The minimum variance control law can be written as q uk = , BGFq y k q where the polynomials F and G are obtained from
deg F = d , 1 = 1; deg G = 2
f1 = 1:3 g0 = 1:25 g1 = ,0:91 g2 = 0 The minimum variance controller is thus given by :25 2 , : uk = , q1, 0q5q0+91:q3 yk : 1
c. The output is in the deadbeat case given by CR1 yk = ARCRBS ek = A A ek + m o = q ,4 C q R1q ek = C q ,1 Rq ,1 ek 1 = 1 , 0:2q ,1 1 + 1:5q ,1ek = 1 + 1:3q ,1 , 0:3q ,2ek which is a moving average MA process. The variance of the output is then simply calculated as
d. Fig. 12.5 shows the output and the sum of the square of the output for the regulators in a and b.
Problem 12.12
A1q = AqDq B1q = B qDq C1q = AqC q and the noise e1 = e. The system can then be written as C yk = B1 uk + A1 e1k A1 1
12.1 111
10
Output
0 10 10
50
100
150
200
Output
0 10 500
50
100
150
200
Loss
0 0
50
The output and the sum of the square of the output for the two regulators in Problem 12.11. The deadbeat controller is used in the upper diagram and the minimum variance controller in the middle diagram. The accumulated loss functions are shown in the lower diagram, deeadbeat dashed and minimum variance solid.
Figure 12.4
100 Time
150
200
Since A, C and D are monic, A1 = AD and C1 = AC will also bo monic. Let d1 = deg A1 , deg B1 = deg A , deg B = d. The minimum-variance controller for the rewritten system 12.1 is then calculated as
12.2 12.3
where
qd1 ,1 C1q = A1qF1q + G1q Equation 12.3 is equivalent to qd,1 AqC q = Aq DqF1q + G1 q
which implies that A must divide G1, i.e. G1q = Aq Gq . Putting F = F1 gives the minimum-variance control law
,1 A ,1 q q uk = , BGDAFq yk = , B qG 1q q ,1qF q,1 y k , D q q
Problem 12.13
In this case
g0 = c
Problem 12.15
We have
,1 y1k = 1 , 017q,1 uk , 1 + 1 , 0::5q ,1 ek : 1 , 0 7q ,1 q y2 k = 1 + q,1 y1k
1 = A B uk , 2 + C ek , 1 A 1 A , 0:5 ,1 1 = 1 + q ,1 1 , 0:7q ,1 uk , 2 + 1 + 1,1 1 q 0:7q ,1 ek , 1 q , To normalize the notations it is convenient to introduce a new noise "k = ek , 1. a. Assume that y1 can be measured. The minimum variance controller for y1 is then uk = ,0:2y1k b. The variances of y1 and y2 are
2 Ey1 = 1 1 2 Ey2 = 1 ,
2:78
c. The minimum variance controller for y2 when y2 is measurable is obtained by using the identity 1 , 0:5q ,1 = 1 + q ,1 1 , 0:7q ,11 + f1 q ,1 + q ,2 g0 + g1q ,1 This gives
y2k = 1 + q,1 "k = "k , 1 + ek , 2 y1 k = 1 + q,11 + q ,1 ek + 1 The variances of the two signals are
2 Ey1 = 1 + + 12 + 2 Ey2 = 1 + 1 = 2 2 = 1:68
and
d. In this case both y1 and y2 are measurable and y1 k will contain more recent information about the noise process than y2 k. It is now only necessary to predict one step ahead. Introduce the identity
C = AA + q,1G 1 1
This gives
y2 k + 2 = ACA ek + 2 + ABA uk 1 1 G ek + 1 + B uk 1 = ek + 2 + A A A A
1 1
G A y k , B uk , 1 + B uk 1 y2 k + 2 = "k + 2 + A A C 1 C A A 1 1 G y k + B ,C , G q,1 uk 1 = "k + 2 + A A 1 1 AA C 1 1 1 1 = "k + 2 + C G y1 k + B uk A1
The variance is thus minimized when 1 uk = , AGB y1 k 1 which is an admissible control law. For the speci c system we get
,1 uk = , 11, 00::56qq,1 y1 k , 8
114
2 Ey2 = 1 We can thus conclude that the extra measurement will greatly improve the control of the output of the system.
and
Problem 12.16
The same arguments can be used in this case as when deriving the normal minimum variance controller. Assume that the system has a stable inverse. Also assume that deg A = deg B = deg D i.e. that the process can be written as
115
Problem 12.17
Aqyk = Bquk + C q ek Aq = q , 0:9; Bq = q; C q = q , 0:5 LQG-Control: Minimize E y2 + u2. Let P z be the closed loop system characteristic equation 12.45 rP zP z,1 = AzAz,1 + B z Bz,1 P z contains stable zeros of the right hand expression Lemma 12.1 rz + p1z ,1 + p1 = z , 0:9z,1 , 0:9 + 1 1 r1 + p2 + rp1 z + rp1 z ,1 = 1:81 + 1 , 0:9 z ,1 , 0:9 z 1
This gives the system of equations
p1 = , 1 + :1:81 + 1 + :1:81 2 , 1 18 18 Determine the LQG-regulator by means of pole placement: Am = P z = z + p1 Ao = C z = z , 0:5
Control law:
where S 0 = 0. See computational procedure in Section 12.5. The Diophantine equation to be solved is
or This gives
r1 = 5 p1 9
uk = , q s0 qr y k +
The input u is
For the closed loop system we get B uk + C ek = B , S yk
+ C ek y k = A A A R A or CR q+r yk = ARCRBS ek = PC ek = R ek = q + p1 ek + P 1 Theorem 10.4 gives 2, V ar y = 1 + r1, p2r1p1 1 2
1
In the following table the calculations are summarized for = 0:1; 1 and 10.
s2 V ar u = 1 ,0p2
1
Problem 12.24
P q = q d,1 Bq = h
In LQ , design we use a state-space description of the process see Example 12.16 xkh + h = xkh + hukh + v kh + h , v kh ykh = xkh + kh To obtain the LQ-controller we have to sample the loss function. From 11.6 - 11.8 we get
Q1 = Q12 =
Zh Zh
0
1 ds = h
2
and thus
The Riccati equation is S = T S + Q1 , LT Q2 + ,T S ,L L = Q2 + ,T S ,,1,T S + Q12 and the solution is S = ph 12 p 1 L = h 3+p3 2+ 3 The closed loop system has a pole in ,p1 p 1 3 + p3 = p3 , 2 ,p1 = , ,L = 1 , h h 2+ 3 and P z = z + p1 To get the regulator we solve the Diophantine see p. 481 PC = AR + BS S 0 = 0 q + p1 q + c = q , 1q + r1 + hs0 q which has the solution r0 = ,p1c 1 s0 = h p1 + c + 1 + p1 c
1 h p1 + c + 1 + p1cq y k uk = , q , p1 c
s ds = h 2 Z0 h 3 Q2 = s2 ds = h 3 0
Problem 12.29
AzF z + Gz = z 2 Gz = 1:4z , 0:65 , : u = , 1:4q, 00265 y q : c. Minimum variance control
vary = 1 Dead-beat control vary = 1 + 0:42 2 = 4:64
2=4
which gives
Problem 12.30
a. Assume that C = 0 and use the identity
C = AF + G
This gives and the controller
F =1 G = ,a
b. Assume that C z = z + c. The minimum variance controller for this ^ model is given by F =1 G = c,a ^ The closed loop system is now q c yk = q + c ek +^ which has the variance, see Theorem 6.4 ^ vary = 1 +1c, c, 2cc 2 ^
2
Problem 12.31
The C polynomial has a pole outside the unit circle. Replace C by a new polynomial obtained from spectral factorization
where " has the standard deviation 1:25. a. The 2-step ahead predictor is given by
z m,1 C z = Az F z + Gz zz2 , 0:8z = z2 , 1:1z + 0:3z + f1 + g0z + g1
which gives and
F z = z + 0:3 Gz = 0:03z , 0:09 yk + 2jk = qGqq y k = 0:03q0,83 y k ^ C q, : E y = 1:2521 + 0:32 = 1:70 ~
120
Problem 12.32
a. d = 2. This gives the identity zC z = AzF z + Gz where deg F = 1 and deg G = 2, i.e., z3 z , 0:1 = z 3 , 1:7z2 + 0:8z , 0:1z + f1 + g0z 2 + g1z + g2 The solution is F z = z + 1:6 Gz = 1:92z 2 , 1:18z + 0:16 and the controller is G 92 2 + :16 uk = , BF y k = , 1:2qq,,:1:18q+ 106 y k 0 9q : b. The output variance when using the minimum variance controller in a. is given by vary = 1 + f12 = 1 + 1:62 = 3:56 c. Since B z has a root outside the unit circle we use the procedure in Theorem 12.3. Factor B z as Bz = B + z B,z with B , q monic, so B + z = ,2 B, z = ,0:9q + 1 The Diophantine to solve is qd,1 C q B , q =Aq F q + B , zGq q3q , 0:1q , 0:9 =q3 , 1:7q2 + 0:8q , 0:1f0q 2 + f1 q + f2 + ,0:9q + 1g0q 2 + g1 q + g2 This gives the system of equations 1 = f0 ,1 = ,1:7f0 + f1 0:09 = 0:8f0 , 1:7f1 + f2 + 1:8g0 0 = ,0:1f0 + 0:8f1 , 1:7f2 , 2g0 + 1:8g1 0 = ,0:1f1 + 0:8f2 , 2g1 + 1:8g2 0 = ,0:1f2 , 2g2 which has the solution f0 = 1:000 f1 = 0:700 f2 = 2:721 g0 = 2:490 g1 = ,1:862 g2 = 0:272 121
or
CF yk = AF CF , ek = qd,1 CB , ek = qd,FB, ek 1 + GB 2 + 0:7q + 2:721 = q q q , 0:9 ek Theorem 10.4 gives the output variance V ar y = 94:7
Problem 12.33
q , 1q , 0:7q + r1 + 0:9q + 1s0q + s1 = q 2q , 0:7 This gives the system of equations