ELEC-C8201 - Control and automation

Design of State Variable Feedback Systems

Themistoklis Charalambous

Room 2561, TUAS Building

[email protected]
In the previous lecture...

You:

I Learned how to plot the open-loop frequency response of the return ratio L(s),

with the imaginary part Im L(jω) plotted against the real part Re L(jω) on
an Argand diagram.

I Used the Nyquist diagram to ascertain stability of the closed-loop system.

I Inferred more detailed information about the behavior of the closed-loop system:
(a) No need to explicitly compute the poles of the system, so it can be applied to systems
defined by non-rational functions.

(b) Relatively easy to see how changing C(s) affects L(s), but difficult to see how C(s)
affects L(s)/(1 + L(s)) directly.

(c) Gain and phase margins measure how close the closed-loop system is to instability.

Themistoklis Charalambous
ELEC-C8201 - Control and automation
1/26
Learning outcomes...
...the student will:
I Become familiar with the concepts of controllability and observability.

I Be able to design full-state feedback controllers and observers.

I Appreciate pole-placement methods and the application of Ackermann’s formula.

I Understand the separation principle and how to construct state variable controllers.

Themistoklis Charalambous
ELEC-C8201 - Control and automation
2/26
Introduction
State variable design typically consists of three steps.

1. We assume that all the state variables are measurable and utilize them in a
full-state feedback control law
I Not practical because it is not possible (in general) to measure all the states - in
practice, only certain states (or linear combinations of them) are measured.

2. Construct an observer to estimate the states that are not directly sensed and
available as outputs.

3. Design process: appropriately connect the observer to the full-state feedback

control law.
System
u x y
ẋ = Ax + Bu C

Observer +
Controller x̂ ỹ = y C x̂
K x̂˙ = Ax̂ + Bu + Lỹ ⌃

Themistoklis Charalambous
ELEC-C8201 - Control and automation
3/26
Fundamental questions:

1. Is it possible to steer a system from a given initial state to any other state?

2. Is it possible to determine a state from observations?

state given initial state

ẋ(t) = Ax(t) + Bu(t), x(0) = x0

y(t) = Cx(t)

observation

These questions were posed and answered by Kalman

(1930–2016), who also introduced the concepts of
controllability and observability

Rudolph Kalman was a central figure in the development of mathematical systems theory upon
which much of the subject of state variable methods rests. Kalman was well known for his role in the
development of the so-called Kalman filter, which was instrumental in the successful Apollo moon
landings.

Themistoklis Charalambous
ELEC-C8201 - Control and automation
4/26
Controllability and Observability

They can be roughly defined as follows:

Controllability: In order to be able to take the current state to whichever state we

want with the given dynamic system under control input, the system must be
controllable

Observability: In order to see what is going on inside the system under

observation, the system must be observable

In this lecture we:

I show that the concepts of controllability and observability are related to linear
systems of algebraic equations
I a solvable system of linear algebraic equations has a solution if and only if the rank of
the system matrix is full

I Observability and controllability tests will be connected to the rank tests of certain
matrices: the controllability and observability matrices

Themistoklis Charalambous
ELEC-C8201 - Control and automation
5/26
Controllability

Definition (Controllability)

A linear time-invariant (LTI) system is controllable, if for every x∗ (t) and every
finite time T > 0, there exists an input function u(t), 0 < t < T , such that
the system state goes from x(0) = x0 to x(T ) = x∗ .

I The state of the system at any time t is given by

Z t
x(t) = eAt x(0) + eA(t−τ ) Bu(τ )dτ
0

I Without loss of generality, assume that x(0) = 01 . Need only to consider the
forced solution to study controllability
Z t
xf (t) = eA(t−τ ) Bu(τ )dτ
0

1
Starting at 0 is not a special case – if we can get to any state in finite time from the origin, then
we can get from any initial condition to that state in finite time as well.

Themistoklis Charalambous
ELEC-C8201 - Control and automation
6/26
I Change of variables τ2 = t − τ , dτ = −dτ2 gives a form that is a little easier to
work with:
Z t
xf (t) = eAτ2 Bu(t − τ2 )dτ2
0

I Assume system has n states (i.e., x(t) ∈ Rn ) and m inputs (i.e., u(t) ∈ Rm ).

I Note that, regardless of the eigenstructure of A, the Cayley-Hamilton theorem

gives
n−1
X
eAt = Ai αi (t)
i=0

for some computable scalars αi (t), so that

n−1 Z t n−1
X X
xf (t) = Ai B αi (τ2 )u(t − τ2 )dτ2 = Ai Bβi (t)
i=0 0 i=0

for coefficients βi (t) that depend on the input u(τ ), 0 < τ < t.

Themistoklis Charalambous
ELEC-C8201 - Control and automation
7/26
I Result can be interpreted as meaning that the state xf (t) is a linear combination
of the nm vectors Ai B (with m inputs)
I All linear combinations of these nm vectors is the range space of the matrix formed
from the Ai B column vectors, called the controllability matrix:

Pc = B AB A2 B · · · An−1 B

I Definition: Range space of Pc is the controllable subspace of the system

I If a state xc (t) is not in the range space of Pc (i.e., it is not a linear combination of
these columns) ⇒ it is impossible for xf (t) to ever become equal to xc (t) – called
uncontrollable state

Theorem

A linear time-invariant (LTI) system is controllable, if and only if (iff) it has no

uncontrollable states.
I Necessary and sufficient condition for controllability is that

rank(Pc ) , rank B AB A2 B · · · An−1 B = n

Remark: For a single-input, single-output system, the controllability matrix Pc is an

n × n matrix. Hence, if the determinant of Pc is nonzero, the system is controllable.

Themistoklis Charalambous
ELEC-C8201 - Control and automation
8/26
I When the system is not completely controllable, but where the states (or linear
combinations of them) that cannot be controlled are inherently stable, then these
systems are classified as stabilizable
I If a system is completely controllable, it is also stabilizable (but not vice versa!)

Example: Let us consider the system

" # " #
0 1 0 0
ẋ = 0 0 1 x+ 0 u
−a0 −a1 −a2 1

y= 1 0 0 x
Is the system controllable?

Answer: First, we find the controllability matrix:

" #
0 0 1
Pc = B AB A2 B = 0 1 −a2
1 −a2 a22 − a1
Then, we compute the determinant:

0 0 1
0
−a2 =
1
det(Pc ) = 0 1 = −1 6= 0 (system is controllable)

1 −a2
1 −a2 a22 − a1

Themistoklis Charalambous
ELEC-C8201 - Control and automation
9/26
Observability

Definition (Observability)

A linear time-invariant (LTI) system is observable, if the initial state x(0) can
be uniquely deduced from the knowledge of the input u(t) and output y(t) for
all t between 0 and any finite T > 0.

I If x(0) can be deduced, then we can reconstruct x(t) exactly because we know
u(t) ⇒ we can find x(t) ∀t
I Thus, it is sufficient to only consider the zero-input (homogeneous) solution to
study observability

y(t) = CeAt x(0)

I Definition: A state x∗ 6= 0 is said to be unobservable if the zero-input solution

y(t), with x(0) = x∗ , is zero for all t ≥ 0
I Equivalent to saying that x∗ is an unobservable state if
At ∗
Ce x =0∀t≥0

Themistoklis Charalambous
ELEC-C8201 - Control and automation
10/26
Example: Let us consider the system
h i h i
−2 0 2
ẋ = x+ u
0 −1 1

y= 3 0 x
T
Is the state x∗ = 0 1 observable?

Answer:
he−2t 0
ih i
0 h0i
CeAt x∗ = 3 0 = 3e−2t 0 =0∀t
0 e−t 1 1
T
So, x∗ = 0 1 is an unobservable state of this system.

Themistoklis Charalambous
ELEC-C8201 - Control and automation
11/26
Theorem

A linear time-invariant (LTI) system is observable iff it has no unobservable

states.

Proof (by a simple example): Let x∗ 6= 0 be an unobservable state and compute the
output of the initial conditions x1 (0) and x2 (0) = x1 (0) + x∗ . Then,

y2 (t) = CeAt x2 (0) = CeAt (x1 (0) + x∗ )

At ∗
CeAt x1 (0) + |
Ce{z x} = y1 (t)
| {z }
=y1 (t) =0

Thus, two different initial conditions give the same output y(t), so it would be
impossible for us to deduce the actual initial condition of the system (x1 (0) or x2 (0))
given y1 (t)

I Testing system observability by searching for a vector x(0) such that CeAt x(0) ∀t
is feasible, but very hard in general – better tests are available.

Themistoklis Charalambous
ELEC-C8201 - Control and automation
12/26
Theorem

A LTI system is observable, iff it has no unobservable states.

I Necessary and sufficient condition for observability is that
 
C
 CA2 
 CA  = n
rank(Po ) , rank 
 .. 

.
CAn−1

Sketch of the proof: If a vector x∗ is an unobservable state, then

CeAt x∗ = 0 ∀t ≥ 0
But all the derivatives ofCeAtexist and for this condition to hold, all derivatives
must be zero at t = 0. Hence,

CeAt x∗ = 0 ⇒ Cx∗ = 0
t=0
d d At ∗
CeAt x∗ =0⇒C e x =0

dt t=0 dt t=0

⇒ CAeAt x∗ = 0 ⇒ CAx∗ = 0
t=0

Themistoklis Charalambous
ELEC-C8201 - Control and automation
13/26
d2

= 0 ⇒ CA2 eAt x∗ = 0 ⇒ CA2 x∗ = 0
CeAt x∗
dt 2 t=0
t=0
..
.
dk

= 0 ⇒ CAk eAt x∗ = 0 ⇒ CAk x∗ = 0
CeAt x∗
dtk t=0 t=0

We only need retain up to the (n − 1)th derivative because of the Cayley-Hamilton

theorem.

I Why does this make sense?

I The requirement for an unobservable state is that for x∗ 6= 0, Po x∗ = 0, which is
equivalent to saying that x∗ is orthogonal to each row of Po
I But if the rows of Po are considered to be vectors and these span the full
n-dimensional space, then it is not possible to find an n-dimensional vector x∗ that is
orthogonal to each of these
I To determine if the n rows of Po span the full n-dimensional space, we need to test
their linear independence, which is equivalent to the rank test

I When the system is not completely observable, but where the states (or linear
combinations of them) that cannot be observed are inherently stable, are classified
as detectable. If a system is completely observable, it is also detectable.
Themistoklis Charalambous
ELEC-C8201 - Control and automation
14/26
Example: Let us consider the system
" # " #
0 1 0 0
ẋ = 0 0 1 x+ 0 u
−a0 −a1 −a2 1

y= 1 0 0 x

Is the system observable?

Answer: First, we find the observability matrix:

" # " #
C 1 0 0
Po = CA = 0 1 0
CA2 0 0 1

Then, we compute the determinant:

1 0 0
det(Pc ) = 0
1
1
0 =
0
= 1 6= 0 (system is observable)
0 0 1
0 1

Themistoklis Charalambous
ELEC-C8201 - Control and automation
15/26
Full state feedback control design

I We consider full-state variable feedback to achieve the desired pole locations of the
closed-loop system

I 1st step: we assume that all the states are available for feedback, i.e., we have
access to the complete state x(t) for all t

I The system input u(t) is given by

u(t) = −Kx(t)
Objective: Determine the gain matrix K.

I The full-state feedback block diagram is illustrated in the figure below

System

ẋ = Ax + Bu

u x

Controller
K

Themistoklis Charalambous
ELEC-C8201 - Control and automation
16/26
I We find the closed-loop system to be
ẋ = Ax + Bu = Ax − BKx = (A − BK)x (1)

I The characteristic equation associated with equation (1) is

det λI − (A − BK) = 0
If all the roots lie in the left half-plane, then the closed-loop system is stable.
Hence, for any initial condition x(0), it follows that
x(t) = e(A−BK)t x(0) → 0 as t → ∞

I Given the pair (A, B), we can always determine K to place all the system
closed-loop poles in the left half-plane if and only if the system is controllable

I The addition of a reference input can be written as

u(t) = −Kx(t) + N r(t)
where r(t) is the reference input.
I When r(t) = 0 for all t > 0, the control design problem is known as the regulator
problem, i.e., we want to compute K so that all initial conditions are driven to 0 in a
specified fashion (as determined by the design specifications)

I When using this state variable feedback, the roots of the characteristic equation
are placed where the transient performance meets the desired response
Themistoklis Charalambous
ELEC-C8201 - Control and automation
17/26
Ackermann’s formula for the state variable feedback matrix

I For a single-input, single-output system, Ackermann’s formula is useful for

determining the state variable feedback matrix

K = k1 k2 ... kn

where u(t) = −Kx(t).

I Given the desired characteristic equation

q(λ) = λn + an−1 λn−1 + . . . + a1 λ + a0

the state feedback gain matrix is given by

K= 0 0 ... 1 Pc−1 q(A)

where

q(A) = An + an−1 An−1 + . . . + a1 A + a0 I

and Pc is the controllability matrix.

Themistoklis Charalambous
ELEC-C8201 - Control and automation
18/26
Example: Consider the second-order system
Y (s) 1
= G(s) = 2
U (s) s
and determine the feedback gain to place the closed-loop poles at s = −1 ± j.

Answer: We require that

q(λ) = [λ − (−1 + j)][λ − (−1 − j)] = λ2 − (−1 + j − 1 − j)λ + 2 = λ2 + 2λ + 2
Hence,
h i2 h i h i h i
0 1 0 1 1 0 2 2
q(A) = A2 + 2A + 2I = +2 + =
0 0 0 0 0 1 0 2
With x1 = y and x2 = ẋ1 , the matrix equation of the system G(s) becomes
h i h i
0 1 0
ẋ = x+ u
0 0 1
The controllability matrix Pc is
1
h i h i h i
0 1 0 −1 0 1
Pc = B AB = ⇒ Pc−1 = =
1 0 −1 −1 0 1 0
Therefore,
h0 1
ih
2
i
2
K= 0 1 Pc−1 q(A) = 0 1 = 2 2
1 0 0 2

Themistoklis Charalambous
ELEC-C8201 - Control and automation
19/26
Observer design

I Generally speaking, only a subset of the states are readily measurable and available
for feedback.

I Even if all the states were available for feedback, these states should be measured
with a sensor or sensor combinations.
I The cost and complexity of the control system increase as the number of required
sensors increases ⇒ even in situations where extra sensors are available, it may not be
cost effective to employ them

I According to Luenberger, the full-state observer for the system

ẋ = Ax + Bu
y = Cx
is given by

x̂˙ = Ax̂ + Bu + L y − C x̂
where x̂ denotes the estimate of the state x

I The matrix L is the observer gain matrix and is to be determined as part of the
observer design procedure

Themistoklis Charalambous
ELEC-C8201 - Control and automation
20/26
I The observer has two inputs, u and y, and one output, x̂, as depicted in the figure
below:
u y

Observer +
x̂ ỹ = y C x̂
x̂˙ = Ax̂ + Bu + Lỹ ⌃

I Objective: provide an estimate x̂ so that x̂ → x as t → ∞

I We do not know x(0) precisely; therefore we must provide an initial estimate x̂(0)
to the observer

I Define the observer estimation error as

e(t) = x(t) − x̂(t)
The observer design should produce an observer with the property that e(t) → 0
as t → ∞
Themistoklis Charalambous
ELEC-C8201 - Control and automation
21/26
I One of the main results of systems theory is that if the system is observable, we
can always find L so that the tracking error is asymptotically stable, as desired.

I Taking the time-derivative of the estimation error

˙
ė(t) = ẋ(t) − x̂(t)

I Using the system model and the observer

ė(t) = Ax + Bu − Ax̂ + Bu + L(y − C x̂)
= (A − LC)(x(t) − x̂(t)) = (A − LC)e(t)

We can guarantee that e(t) → 0 as t → ∞ for any initial tracking error e(0) if the
characteristic equation

det λI − (A − LC) = 0

has all its roots in the left half-plane.

I This can always be accomplished if the system is observable; i.e, if the

observability matrix Po has full rank

Themistoklis Charalambous
ELEC-C8201 - Control and automation
22/26
Ackermann’s formula for the observer gain matrix

I For a single-input, single-output system, Ackermann’s formula is useful for

determining the observer gain matrix
T
L = l1 l2 ... ln

I Given the desired characteristic equation

p(λ) = λn + bn−1 λn−1 + . . . + b1 λ + b0

the observer gain matrix is given by

T
L = p(A)Po−1 0 0 ... 1

where

p(A) = An + bn−1 An−1 + . . . + b1 A + b0 I

and Po is the observability matrix.

Themistoklis Charalambous
ELEC-C8201 - Control and automation
23/26
Integrated full-state feedback and observer

I We now consider the feedback law

u(t) = −K x̂(t) (2)
We need to verify that, when using the feedback control law in equation (2), we
retain the stability of the closed-loop system.

I Consider the observer

x̂˙ = Ax̂ + Bu + L y − C x̂

I Substituting the feedback law in equation (2) and rearranging terms in the
observer yields
x̂˙ = (A − BK − LC)x̂ + Ly

I Computing the estimation error

ė(t) = Ax − BK x̂ − (A − BK − LC)x̂ + Ly
= (A − LC)(x(t) − x̂(t)) = (A − LC)e(t) (3)

I This is the same result as we obtained for the estimation error before, because the
estimation error does not depend on the input
Themistoklis Charalambous
ELEC-C8201 - Control and automation
24/26
I Recall that the underlying system model is given by
ẋ = Ax + Bu
y = Cx
Substituting the feedback law u(t) = −K x̂(t) into the system model yield
ẋ = Ax − BK x̂(t) = Ax − BK(x(t) − e)
= (A − BK)x + BKe (4)

I Combining equations (3) and (4) in matrix form, we have

h i h ih i
ẋ A − BK BK x
= (5)
ė 0 A − LC e

I The characteristic equation associated with equation (5) is

∆(λ) = det λI − (A − BK) det λI − (A − LC)

I If the roots of det λI − (A − BK) = 0 lie in the left half-plane (which they do by
design of the full-state feedback law), and

I if the roots of det λI − (A − LC) = 0 lie in the left half-plane (which they do by
design of the observer),
then the overall system is stable! The fact that the full-state feedback law and the
observer can be designed independently is an illustration of the separation principle
Themistoklis Charalambous
ELEC-C8201 - Control and automation
25/26
Learning outcomes...
...the student will:
I Become familiar with the concepts of controllability and observability.

I Be able to design full-state feedback controllers and observers.

I Appreciate pole-placement methods and the application of Ackermann’s formula.

I Understand the separation principle and how to construct state variable controllers.

System
u x y
ẋ = Ax + Bu C

Observer +
Controller x̂ ỹ = y C x̂
K x̂˙ = Ax̂ + Bu + Lỹ ⌃

Themistoklis Charalambous
ELEC-C8201 - Control and automation
26/26