04LSLLN Automatic Control

Automotive Engineering

Lecture 20. Robust Margins

Robustness: General Definitions

• Robustness (from Oxford dictionary): the ability to withstand or overcome

adverse conditions.

• Robust control: Designing a controller that reaches a control goal in

presence of uncertainties: unknown dynamics, parameters, disturbances.

• We have already briefly discussed robustness against disturbances: how

to attenuate some sorts of disturbances (constant, ramp signal etc.) or
reduce their influence.

• For instance, to make the output independent of the (near) constant

disturbance, we can include the integrator into the loop.

• Robustness against parametric uncertainties is challenging!

One Example: Robust Stability is Trivial

• We know the equations (derived from physical principles), but some

characteristics are unknown.

• We know that such parameters are within some range.

• For instance, in the mass-spring-damper system we may fail to know

the exact value of damping, and elasticity. But it can be given that K is
from [K1,K2] (a similar range is known for damping).

𝑀𝑀𝑧𝑧̈ 𝑡𝑡 = −𝛽𝛽 𝑧𝑧̇ 𝑡𝑡 − 𝐾𝐾𝐾𝐾 𝑡𝑡 + 𝐹𝐹(𝑡𝑡)

• Can we understand whether the system is stable or not for all values of
the unknown parameters?
• The mass-spring-damper system remains stable for all positive
coefficients! WHY?
For Matrices, The Problem is Very Hard!

• Consider a linear LTI system.

• Some entries (denoted *) of the matrix A are unknown, e.g.,

• Unknown entries lie within known ranges, e.g.

• Such matrices constitute a convex polytope in the space of all matrices

(that’s the reason for term “polytopic”)

• Are all such matrices stable or not?

• A very complicated problem even in the case of 1 unknown entry (solved
only in 1985s). In the case of k unknown elements, the best of known
algorithms uses operations, where n=dim x. Practically
infeasible for large matrices!
Robust Stabilization in Frequency Domain

r y
- C G

• Suppose that G 𝑠𝑠 is not given but belongs to a set of models (= described

by a model with uncertain parameters). For example,
1
G 𝑠𝑠 = , 𝑎𝑎 ∈ [0.9,1.1].
𝑠𝑠(𝑠𝑠+𝑎𝑎)2

The parameters are unknown, but they belong to known intervals.

• Robustness is the property by which the controller 𝐶𝐶 stabilizes the closed-

loop system for all the models in the uncertainty set.
Reminder: Nyquist Criterion

R(s) + Y(s)
- L(s)

• The Nyquist criterion allows us to study the stability properties of the

closed-loop function T(s) from the knowledge of the (open) loop
function L(s).

• Very visual: position of (-1,0) with respect to the curve determines the
stability

• Sometimes, allows to answer the question about robustness in a

predefined uncertainty set (see next slides).
Nyquist Criterion (L13)

• N = number of clockwise
encirclements of the Nyquist
diagram around the point (-1,0).
Im[L(jω)]

(-1,0)
• P = number of poles of L(s) with
real part > 0.

Re[L(jω)]

• Z = number of poles of T(s) with real part ≥ 0

Nyquist stability criterion: Z=N+P

Robust Stability For All Uncertain Transfer Functions (P=0).

0.5
0.4

0.3
0.2

Imaginary Axis
0.1

0
-0.1
-0.2

-0.3
-0.4
-0.5

-1 -0.8 -0.6 -0.4 -0.2 0 0.2

Real Axis

• The closed-loop system is stable (N=0) for all the models in the
uncertainty set, because (-1,0) is always outside the curve.

↓↓↓

The closed-loop system is robustly stable

The controller is robust
No Robust Stability (P=0).

0.5
0.4

0.3
0.2

Imaginary Axis
0.1

0
-0.1
-0.2

-0.3
-0.4
-0.5

-1 -0.8 -0.6 -0.4 -0.2 0 0.2

Real Axis

• Some models in the uncertainty set are unstable (one model, e.g., has
an imaginary pole, because the curve passes through (-1,0)).

↓↓↓

The closed-loop system is not robustly stable

The controller is not robust
Can This Graphical Method Be Used?

• Works in the case of one parameter: you can sweep the range of
parameters (with a small step) and draw all curves.

• How can we understand what happens to the Nyquist curve when we

sweep multiple parameters within their ranges? It is too difficult to draw all
possible curves. So, the problem of robustness in a set of uncertainties is
quite difficult also in the frequency domain!

• Engineers basically consider a “rough” characteristics of robustness,

measuring a “distance” the Nyquist curve is from the point (-1,0).

• Before introducing such characteristics (phase and gain margins), consider

an example.
Example: Small Difference Between Unstable…

r(t) + y(t)
- L(s)
Y (s) L( s )
T (s) = =
R ( s ) 1 + L( s )

2.05
L( s ) =
s(s + 1)
2
N =2 P=0

⇓
Z =2
T(s) unstable
… And Stable Systems

r(t) + y(t)
- L(s)
Y (s) L( s )
T (s) = =
R ( s ) 1 + L( s )

1.95
L( s ) =
s(s + 1)
2

N =0 P=0

⇓
Z =0
T(s) stable with a
low margin (will be
defined later)
Informally: curve
passes very close
to (-1,0).
Example: A Sufficiently Robust System

r(t) + y(t)
- L(s)
Y (s) L( s )
T (s) = =
R ( s ) 1 + L( s )

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

P=0

⇓
N =0

Z =0
T(s) stable, stability is
The Nyquist diagram of L(jω) «sufficiently» robust
is far from the critical point (-1,0) (high margins)
Phase Margin

Definition. Gain crossover frequency:

𝜔𝜔𝑔𝑔𝑐𝑐 > 0: 𝐿𝐿(𝑗𝑗𝜔𝜔𝑔𝑔𝑐𝑐 ) = 1

Imaginary Axis
Definition. Phase margin:
+
𝑚𝑚𝜙𝜙 = ∠𝐿𝐿(𝑗𝑗𝜔𝜔𝑔𝑔𝑔𝑔 ) + 180𝑜𝑜 mϕ
∠L( jω c )
Margin must be in [-180,180] (if 𝜔𝜔𝑔𝑔𝑔𝑔
>180, deduce 360).

• In the most typical example (P=0,

one crossing with the circle): Real Axis
mϕ > 0 => T(s) stable
mϕ < 0 => T(s) is usually unstable

mϕ is the maximum perturbation on the phase of L(s) for which T(s) remains stable.
If the closed-loop system is stable, mϕ measures its robustness properties.

Remark: In general (P>0), a positive mϕ also does not imply stability!

Delay Robustness
Delay block introduces a time lag in a signal: u(t)  u(t-T).
Delays are usually introduced by communication lines.
Phase Margin: MATLAB

𝜔𝜔𝑔𝑔𝑐𝑐

mϕ

• MATLAB command: margin(L);

Gain Margin
Definition. Phase crossover frequency:
𝜔𝜔𝑝𝑝𝑐𝑐 > 0: ∠𝐿𝐿 𝑗𝑗𝜔𝜔𝑝𝑝𝑝𝑝 = −180

Imaginary Axis
Definition. Gain margin (in dB):
mg = 20*log10(1/| 𝐿𝐿 𝑗𝑗𝜔𝜔𝑝𝑝𝑝𝑝 |) 𝜔𝜔𝑔𝑔𝑔𝑔

+
In our example (P=0): positive for stable, negative ∠L( jω c )
for unstable.

The positive factor by which the loop function needs

to be multiplied in order to make the Nyquist curve Real Axis

pass through (-1,0), measured in dB (logarithmic scale)

Margin +20dB means that multiplying the amplitude (at

phase crossover) by 10, the Nyquist curve will pass
through (-1,0). Margin -20dB means that dividing the
amplitude by 10, the curve passes through (-1,0).
Gain Margin in Matlab

𝑚𝑚𝑔𝑔

𝜔𝜔𝑝𝑝𝑝𝑝

• Matlab command: margin(L);

Margins For Standard Architectures?

𝑟𝑟

r + u*
-
P y

O 𝑥𝑥�

K
Think How to Reduce the System to A Simple Closed-Loop!
r +
- e
y
L

For the first architecture, the output (compared to r(t)) is y(t)

𝑟𝑟

For the second architecture, the output (compared to r(t)) must be 𝑲𝑲�
𝒙𝒙(𝒕𝒕)

r + u*
- P y

O 𝑥𝑥�

K
Engineering Observations: Pole Placement Design

• Eigenvalue placement techniques (II order dominant poles):

increasing the damping (reducing the overshoot)

↓↓↓

bringing the closed-loop eigenvalues far from the imaginary axis

↓↓↓

increasing the phase margin (in modulus)

↓↓↓

increasing robustness
Magic Robustness of LQR Controller!

• LQR control:

- LQR state-feedback control: robust (|𝑚𝑚𝜑𝜑 | ≥ 600 , gain margin

either infinite or less than -6dB. In reality, multiplying matrix K
by any factor >0.5, you will preserve stability of the closed-
loop system).

- LQR output-feedback control

 with fast observer: robust (almost same as without observer).
 with slow observer: can be non-robust.