Modelling and Control of Dynamic

Systems

Controllability and Observability

Sven Laur
University of Tartu
Closed-loop Controllers:
Basic Structure
 Closed-loop control with state estimation

Controller System
r[k] u[k] y[k]
State Ĝ[z]
Estimator

The instability of a open loop controller is caused by gradual accumulation of

disturbances. The control signal becomes unsynchronised with the system.

⊲ If we can estimate the system state from the output, we can circumvent
such synchronisation errors. The system must be observable for that.
⊲ Even if we know the system state, it might be impossible to track the
reference signal. In brief, the system must be controllable.
1
 Controllability
The state equation is controllable if for any input state x0 and for any final
state x1 there exists an input u that transfers x0 to x1 in a finite time.
⊲ T6. The
 n dimensional pair (A, B) is controllable iff the controllability
matrix B AB A2B An−1B has the maximal rank n.

The desired input can be computed from the system of linear equations

x1 = Anx0 + An−1Bu[0] + An−2Bu[1] + · · · + Bu[n − 1] .

Although the theorem T6 gives an explicit method for controlling the

system, it is an off-line algorithm with a time lag n.
A good controller design should yield a faster and more robust method.

2
 Observability

The state equation is observable if for any input state x0 and for any input
signal u, finite the output y sequence determines uniquely x0.

⊲ T7. The pair (A, C) is observable iff the observability matrix

 
C
 CA 
 
 ... 
CAn−1

has the maximal rank n.

3
 Offline state estimation algorithm

Again, the input state x0 can be computed from a system of linear equations


 y[0] = Cx0 + Du[0]



 y[1] = CAx0 + CBu[0] + Du[1]


 ···


 y[n − 1] = CAn−1 x + · · · + CBu[n − 2] + Du[n − 1]
0

Although this equation allows us to find out the state of the system, it is
offline algorithm, which provides a state estimation with a big time lag.

A good state space estimator must be fast and robust.

4
 General structure of state estimators

u[k] System y[k]

x[k] = Ax[k − 1] + Bu[k − 1]
y[k] = Cx[k] + Du[k]

x̂[k]
State Estimator

A state estimate x̂[k] is updated according to u[k] and y[k]:

⊲ Update rules are based on linear operations.
⊲ The state estimator must converge quickly to the true value x[k]
⊲ The state estimator must tolerate noise in the inputs u[k] and y[k].
5
 Example

0.5z+0.5
Consider a canonical realisation of the transfer function ĝ[z] = z 2 −0.25

   
0 0.25 1
A= B=
1 0 0
C = [0.5 0.5] D=0

Then a possible stable state estimator is following

x̂[k + 1] = Ax̂[k] + Bu[k] + 1(y[k] − C x̂[k]) .

| {z }
ŷ[k]

In general, the feedback vector 1 can be replaced with any other vector to
increase the stability of a state estimator.
6
Kalman Decomposition
 Equivalent state equations

Two different state descriptions of linear systems

( (
x[k + 1] = A1x[k] + B 1u[k] x[k + 1] = A2x[k] + B 2u[k]
y[k] = C 1x[k] + D 1u[k] y[k] = C 2x[k] + D 2u[k]

are equivalent if for any initial state x0 there exists an initial state x0 such
that for any input u both systems yield the same output and vice versa.
⊲ T8. Two state descriptions are (algebraically) equivalent if there exists
an invertible matrix P such that

A2 = P A1P −1 B2 = P B1
C 2 = C 1P −1 D2 = D1 .

7
 Linear state space transformations

The basis e1 = (0, 1) and e2 = (1, 0) is a canonical base in R2. However a

basis a1 = (1, 1) and a2 = (1, −1) is also a basis.
Now any state x = x1 e1 + x2e2 can be represented as x = x1a1 + x2a2
and vice versa. The latter is known as a basis transformation:
 x1 + x2 (
 x1 =

x1 = x1 + x2
2
 x2 = x1 − x2
 x2 = x1 − x2
2

For obvious reasons, we can do all computations wrt the basis {a1, a2} so
that the underlying behaviour does not change. The same equivalence of
state descriptions hold for other bases and larger state spaces, as well.

8
 Kalman decomposition
⊲ T9. Every state space equation can be transformed into an equivalent
description to a canonical form
      
xco[k + 1] Aco 0 A13 0 xco[k] B co
xco[k + 1] A21 Aco A23 A24 xco[k] B co
xco[k + 1] =  0
    +  u[k]
0 Aco 0   xco[k]   0 
xco[k + 1] 0 0 A43 Aco xco[k] 0
y[k] = [C co 0 C co 0] x[k] + Du[k]

where
⋄ xco is controllable and observable
⋄ xco is controllable but not observable
⋄ xco is observable but not controllable
⋄ xco is neither controllable nor observable
9
 Minimal realisation

⊲ T10. All minimal realisations are controllable and observable. A

realisation of a proper transfer function ĝ[z] = N (z)/D(z) is minimal iff
it state space dimension dim(x0) = deg D(z).

10
Closed-loop Controllers:
Design Principles
 General setting

According to Kalman decomposition theorem, the state variables can be

divided into four classes depending on controllability and observability.
⊲ We cannot do anything with non-controllable state variables.
⊲ Non-observable variables can be controlled only if they are marginally
stable. We can do it with an open-loop controller.
⊲ For controllable and observable state variables, we can build effective
closed-loop controllers.

Simplifying assumptions
⊲ From now on, we assume that we always want to control state variables
that are both controllable and directly observable: y[k] = x[k].
⊲ If this is not the case, then we must use state estimators to get an
estimate of x. The latter just complicates the analysis.
11
 Unity-feedback configuration

r[k] Controller u[k] System y[k]

p +
Ĉ[z] ĝ[z]

-1

Design tasks
⊲ Find a compensator ĉ[z] such that system becomes stable.
⊲ Find a proper value of p such that system starts to track reference signal.

It is sometimes impossible to find p such that y[k] ≈ r[k].

⊲ The latter is impossible if ĝ[1] = 0, then the output y[k] just dies out.
12
 Overall transfer function
Now note that
pĝ[z]ĉ[z]
ŷ = ĝ[z]ĉ[z](pr̂ − ŷ) ⇐⇒ ŷ = r̂
1 + ĝ[z]ĉ[z]

and thus the configuration is stable when the new transfer function

pĝ[z]ĉ[z]
ĝ◦[z] =
1 + ĝ[z]ĉ[z]

has poles lying in the unit circle. Now observe

N (z) B(z) pB(z)N (z)

ĝ[z] = , ĉ[z] = =⇒ ĝ◦[z] =
D(z) A(z) A(z)D(z) + B(z)N (z)

13
 Pole placement

We can control the denominator of the new transfer function ĝ◦[z]. Let

F (z) = A(z)D(z) + B(z)N (z)

be a desired new denominator. Then there exists polynomials A(z) and

B(z) for every polynomial F (z) provided that D(z) and N (z) are coprime.
The degrees of the polynomials satisfy deg B(z) ≥ deg F (z) − deg D(z).
ℑ(s) ℑ(z)

BIBO
stable
BIBO
ℜ(s) ℜ(z)
stable

14
 Signal tracking properties

Let ĝ◦[z] be the overall transfer function.

⊲ The system stabilises if ĝ◦[z] is BIBO stable.
⊲ The system can track a constant signal r[k] ≡ a if ĝ◦[1] = 1.
⊲ The system can track a ramp signal r[k] ≡ ak if ĝ◦[1] = 1 and ĝ◦′ [1] = 0.
The latter follows from the asymptotic convergence

y[k] → aĝ◦′ [1] + kaĝ◦[1]

for an input signal r[k] = ak. Moreover, let ĝ◦ = N (z)/D(z). Then

ĝ◦[1] = 1 ∧ ĝ◦′ [1] = 0 ⇐⇒ N (1) = D(1) ∧ N ′(1) = D ′(1) .

15
 Trade-offs in pole placement

A pole placement is a trade-off between three design criteria:

⊲ response time
⊲ overshoot ratio
⊲ maximal strength of the input signal

There is no general recipe for pole placement. Rules of thumb are given in
⊲ C.-T. Chen. Linear System Theory and Design. page 238.

16
 Illustrative example

1
Consider a feedback loop with the transfer function g[z] = z−2 . Then we
can try many different pole placements
 2 2

2
F (z) ∈ z − 0.5, z + 0.5, z − 0.25, z + z + 0.25, z − z + 0.25

The corresponding compensators are

 
3 5 15 24 9
ĉ[z] ∈ , , , , ,
2 2 4z + 8 4z + 12 4z + 4
 
1 3 1 9 1
p∈ , , , , , .
3 5 5 25 9

They can be found systematically by solving a system of linear equations.

17
 Robust signal tracking

Sometimes the system changes during the operation. The latter can be
modelled as an additional additive term w[k] in the input signal.

If we know the poles of reference signal r[k] and w[k] ahead, then we can
design a compensator that filters out the error signal w[k].
1
For instance, if w[k] is a constant bias, then adding an extra pole z−1
cancels out the effect of bias. See pages 277–283 for further examples.

In our example, the robust compensator for F [z] = z 2 − 0.25 is

12z − 9
ĉ[z] = p=1 .
4z − 4

18
 Model matching

Find a feedback configuration such that ĝ◦ is BIBO stable and ĝ◦[1] = 1.

r[k] Controller u[k] System y[k]

+
ĉ1 [z] ĝ[z]

Feedback
ĉ2 [z]

⊲ The two-parameter configuration described above provides more flexibility

and it is possible to cancel out zeroes that prohibit tracking.
⊲ There are many alternative configurations. A controller design is
acceptable if every closed-loop configuration is BIBO stable.
19