Lecture 7 Stability Introduction

K. J. Åström • Risk for instability is the main drawback of feedback

• Instabilities were frequently encountered in early use of
1. Introduction feedback
2. Stability concepts • Created a pressing need for theory
3. Linear Systems – Understand mechanisms that create instability
4. The Nyquist Theorem – Criteria for stability - beginning of control theory
5. Stability Margins – Ways to avoid instabilities
6. Summary • Linear systems are very special

Theme: Risk for instability is the main drawback of feedback. • Stability problems appear in many different contexts,
Important to understand weaknesses. General concept. Linear buckling, critical speed, oscillations in combustion, etc.
systems are very special. Nyquist’s theorem. Stability margins.

Maxwell’s Observations 1868 Maxwell’s Qualitative Observation

“It will be seen that the motion of a machine with its governor Stable Unstable
1 1
consist in general of a uniform motion, combined with a disturbance
which may be expressed as the sum of several component motions.
0.5 0.5
These components may be of four different kind:
1. The disturbance may continually increase.
0 0
2. It may continually diminish. 0 10 20 30 40 50 0 10 20 30 40 50
Stable
1 Unstable
1
3. It may be an oscillation of continually increasing amplitude.
4. It may be an oscillation of continually decreasing amplitude.
0 0
The first and third cases are evidently inconsistent with the stability
of the motion: and the second and fourth alone are admissible in a
−1 −1
good governor. Stability is mathematically equivalent to the condi- 0 10 20 30 40 50 0 10 20 30 40 50
tion that all roots of an algebraic equation (the characteristic equa-
tion) are in the left half planessible roots, and all the possible parts How to formalize the ideas?
of the impossible roots of a certain equation shall be negative.


c K. J. Åström October, 2000 1
 Stability Concept Lyapunov Stability
Consider a solution x(t, a) to the differential equation The solution x(t, a) is called stable if  x(t, a) − x(t, b) < ε for all
b such that a − b < δ .
dx
= f ( x)
dt The solution is called asymptotically stable if it is stable
and if in addition  x(t, a) − x(t, b) goes to zero as t increases
with initial conditions x(0, a) = a. Investigate what happens to a
towards ∞.
solution x(t, b) with initial condition x(0, b) = b where b is close
to a. Notice that we can only talk about stability of a particular
solution. One solution may be stable and another unstable.
Example: the pendulum.
In control we will require asymptotic stability.
It is convenient to normalize so that the interesting solution is
x(t, a) = 0.

Linear Systems Linear Time Invariant Systems are Very Special

Consider the solutions x(t, a) and x(t, b) to the equations • In general we can only talk about the stability of a specific
solution
dx
= Ax • This means that some solutions may be stable and other
dt
unstable
with initial conditions x(0, a) = a and x(0, b) = b. We have
• Linear time invariant system are very special because if
x( t, a) = e a,
At
x( t, b) = e b
At one solution is stable all other solutions are also stable
• It is thus possible to talk about the stability of a solution
Hence x(t, a) − x(t, b) = e At(a − b)
• This is a very unusual property for linear systems
The solution x(t, a) is stable if the matrix A has all eigenvalues
in the left half plane or on the imaginary axis, and if the
eigenvalues on the imaginary axis are simple. The solution is
asymptotically stable if all eigenvalues of A are in the proper
left half plane.

c K. J. Åström October, 2000 2
 The Characteristic Equation Eigenvectors and Eigenvalues
dx An eigenvector v of a matrix A has the property
The system = Ax has the characteristic equation
dt
Av = λ v
det( sI − A) = 0
where λ is the eigenvalue. This means that the equation
dn y n−1
The system dtn
+ a1 ddtn−1y + ... + an−1 dy
dt
+ an y = 0
( A − λ I )v = 0
has the characteristic equation
has a non-trivial solution, hence det(λ I − A) = 0
sn + a1 sn−1 + . . . + an−1 s + an = 0
Now consider the differential equation
so does the system
dx
= Ax
n− 1 dt
Y ( s) b1 s + . . . + b1 s + bn
= n
U ( s) s + a1 sn−1 + . . . + an−1 s + an
The function x(t) = eλ t v is a solution with the initial condition
x(t) = v

Example Inverted Pendulum Using Matlab

Linearize around x 1 = 0. Does the equation







d x1 0 1 x1 x1 ( t) 1 −t
1 s3 + s2 + s + k = 0
= = c1 e t
+ c2 e
dt x2 1 0 x2 x2 ( t) 1 −1
det ( sI − A) = s2 − 1 have roots in the right half plane for k = 1 or k = 10?
>> roots([1 3 2 1]) >> roots([1 3 2 10])
Characteristic equation has roots s = ±1, solution is unstable! ans = ans =
Linearize around x 1 = π . -2.3247 -3.3089






 -0.3376 + 0.5623i 0.1545 + 1.7316i
d x1 0 1 d x1 d x1 sin t cos t -0.3376 - 0.5623i 0.1545 - 1.7316i
= = c1 + c2
dt x2 −1 0 dt x2 dt x2 cos t − sin t
Algebraic stability conditions (Routh-Hurwitz) were been
det ( sI − A) = s2 + 1
important historically, but are now less important because of
computational tools like Matlab. The commands roots and
Characteristic equation has roots s = ±i, the solution is stable
eigen give numerical solutions. What do we mean by solution
but not asymptotically stable
to a problem?


c K. J. Åström October, 2000 3
 Algebraic Criteria Routh-Hurwitz The Furuta Pendulum
All zeros of polynomial a 0 s + a1 are in left half plane if all
coefficients are positive
Model
All zeros of polynomial a 0 s2 + a1 s + a2 are in left half plane if all
coefficients are positive J p(θ¨ − ω 2 sin θ cos θ ) − mg sin θ = 0

All zeros of polynomial a 0 s3 + a1 s2 + a2 s + a3 are in left half

plane if all coefficients are positive and if θ tilt angle of pendulum
ω rate of rotation of arm
a1 a2 − a0 a3 > 0

Example: The polynomial s 3 + 3s2 + 2s + k has all zeros in LHP Stationary solutions for ω = constant.
if k < 6 because
 mg 
− J pω 2 cos θ + CD sin θ = 0
a1 a2 − a0 a2 = 3 2−1 k= 6−k J pω 2

The Furuta Pendulum Linearization

Stationary solutions ω = const. Model
 mg 
− J pω 2 cos θ + CD sin θ = 0 J p (θ¨ − ω 2 sin θ cos θ ) − mg sin θ = 0
J pω 2
 Introduce x 1 = θ och x2 = θ˙
Two solutions if ω < mg/ J p
dx1
= x2
θ = 0, θ =π dt
dx2 mg 1
  = sin x1 + ω 2 sin 2x1
Four solutions if ω > mg/ J p  g/  7rad/s dt Jp 2

θ = 0, θ = π, θ = θ 0, θ = −θ 0 Linearize around x = x 0
 
where θ 0 = arccos (−mg/ J pω ). Physical interpretation!
2  f  0 1 

A=  = mg
 x x= x0 cos x1 + ω 2 cos 2x1 0 
Jp x = x0

c K. J. Åström October, 2000 4
 Stability of Stationary Solutions Stability of Stationary Solutions ...
Solution x1 = θ = 0 has Solutions θ = ±θ 0 = arccos (−mg/ J pω 2),
   
0 1 0 1
A= mg  mg 2
+ ω2 0 A= −ω 2
0
Jp J pω

The matrix A has one eigenvalue in the RHP, unstable The matrix A has eigenvalues on the imaginary axis if
Solution x1 = θ = π
J pω 2 > mg
 
0 1
A= mg the solution is then stable. Physical interpretation!
− + ω2 0
Jp

The matrix A has eigenvalues on the imaginary axis if mg >

J pω 2 the solution is then stable. If mg < J pω 2 the matrix A
has one eigenvalue in the RHP and the solution is unstable.

Nyquist’s Stability Theorem Conditions for Oscillations

• So far focus on the characteristic equation
• Difficult to see how the characteristic equation is influ-
enced by controller.
• How to change the controller to make unstable system
stable?
• Nyquist’s results was a major paradigm shift
Cut the loop. Let e1 be a sinusoid. If e2 is a sinusoid with the
• Investigate propagation of sinusoids around the loop same amplitude and phase, then the loop can be closed and
• Based on transfer functions (Always useful to have differ- the oscillation will be maintained. The condition for this is
ent ways to look at a problem!)
L( iω ) = −1
• Strong practical implications
• Possibilities to introduce stability margins. where L = PC is the loop transfer function. The Nyquist curve
of L goes through the critical point −1!

c K. J. Åström October, 2000 5
 The Complete Nyquist Curve Nyquist’s Stability Theorem
The complete Nyquist curve is image of the contour C under In the special case when the loop transfer function does not
the map L(s) have poles in the right half plane the closed loop system is
s-plane stable if the complete Nyquist curve does not encircle the
L(s)-plane critical point.
5

4 There is a much more general result which also covers the

3
case where the loop transfer function has poles in the RHP.
2

1
Use pencil and string to determine encirclements in tricky
0 situations.
−1

−2

−3

−4
There is some really beautiful mathematics behind this!
−5
−5 −4 −3 −2 −1 0 1 2 3 4 5

Example 3(s + 1)2

Conditional Stability L(s) =
5 s(s + 6)2
4
400 0.5

3 0.4
300

0.3
2
200

1 0.2

L( s ) = 1
s( s + 1)2 100
0.1

0
0 0

L no poles in RHP. −1 −0.1

−100

No encirclements −0.2
−2
Stable −200
−0.3
−3 −300
−0.4

−4 −400 −0.5
−400 −300 −200 −100 0 100 200 300 400 −9 −8 −7 −6 −5 −4 −3 −2 −1 0 1

−5
−5 −4 −3 −2 −1 0 1 2 3 4 5
Loop transfer function has no zeros in the RHP. No encir-
clements. Closed loop system stable. Notice counterintuitivity.

c K. J. Åström October, 2000 6
 Stability Margins Gain and Phase Margins in Bode Plots
Stability as it has been defined Make a Bode plot of the loop transfer function L = PC
is black and white. In practice
Gain
there is often a need to have 10
1

concepts like degrees of sta-

bility. Some useful concepts 10
0

are
• Gain margin g m (2-6)
−1
10
○ ○
• Phase margin ϕ m (45 -60 )
−1 0
10 10

Phase
• Shortest distance to critical −100

point d (0.5-0.8) −120

Notice d is safe but only one of −140

gm or ϕ m is not! −160

−180

−200
−1 0
10 10

Practical Implications of Nyquist’s Theorem Summary

H. Nyquist Regeneration Theory, Bell System Technical Journal • Stability important in control and in many other fields
11 (1932) 126-147 – Buckling, critical speeds, combustion instability, acous-
• The key to design of feedback amplifiers tics
• Very valuable for design • Stability concepts
• Loop-shaping: a method to twist the Nyquist curve of the • Stability of solutions and stability of systems
loop transfer function L = PC to a desired shape. • Linear systems: Characteristic equation det (sI − A) tells
• Industrial implications, ABB experiences from 1945 all
• Stability margins • Nyquist’s Theorem - a very different view of stability
• Useful to have two different ways to view dynamics
• Stability margins

c K. J. Åström October, 2000 7