DASC IPD462

Lecture 2
The Fundamental Problem in Control
Affine Parametrization

Prof: Angel L. Cedeño

Electronic Engineering Department

Universidad T. Federico Santa María, Chile

March 2023

[email protected] Lecture 2 1 / 36
Outline

1 Fundamental Problem of Control

2 Affine Parameterisation: The stable case

[email protected] Lecture 2 2 / 36
References:

K. J. Åström and R. Murray. Feedback Systems: An introduction for scientists

and engineers. Princeton University Press, 2008.
G. C. Goodwin, S. F. Graebe, and M. E. Salgado. Control System Design.
Prentice Hall, Upper Saddle River, NJ, 2001.
J. Doyle, B. Francis, and A. Tannenbaum. Feedback Control Theory. Macmillan
publishing company, New York, USA, 1992.

Acknowledgement
This course is based on the material of Professor Juan C. Aguëro, who belongs to the
Electronic Engineering Department at Universidad Técnica Federico Santa María.

[email protected] Lecture 2 3 / 36
Outline

1 Fundamental Problem of Control

2 Affine Parameterisation: The stable case

[email protected] Lecture 2 4 / 36
The fundamental problem of control

The central problem in control is to find a technically feasible way to act on a given
process so that the process behaves, as closely as possible, to some desired behavior.
Furthermore, this approximate behavior should be achieved in the face of uncertainty
of the process and in the presence of uncontrollable external disturbances acting on the
process.

Desired behavior
Feasibility
Uncertainty
Action
Disturbances
Approximate behavior
Measurements.

[email protected] Lecture 2 5 / 36
The fundamental problem of control

Plant
Let us assume that the required behavior is specified by a scalar target signal, or ref-
erence r(t), for a particular process variable, y(t) which has an additive disturbance
do (t), and we have available a single manipulated variable, u(t).
Do(s)
+
U(s) + Y (s)
Go(s)

Goal: Y(s) should be close to R(s).

Open-Loop Solution
Do(s)
− U(s) +
R(s) + + Y (s)
G−1
o (s) Go(s)

[email protected] Lecture 2 6 / 36
The fundamental problem of control
Abuse of notation: for clarity, and when to be convenient, we will consider dropping
the dependence on the complex variable s. For instance Go (s) = Go .

Requirements for the Open-Loop Solution

The model Go needs to describe the plant exactly.
The model Go should be formulated such that a bounded Y is produced by a
bounded U.
The model Go needs to be invertible.
The manipulated input U should be realizable and not violate any constraint.

Problems
Uncertainty
Non-linearities
Disturbances and Measurement Noise
Plant might not be invertible

[email protected] Lecture 2 7 / 36
From Open to Closed Loop Architectures

Open loop control with built–in inverse

R + U Y
C Plant
−

Go

Open loop controller

This scheme is suggested for realizing an approximate inverse of a plant model.

Closed loop control

R + U Y
C Plant
−

[email protected] Lecture 2 8 / 36
One Degree of Freedom Closed Loop
Nominal loop:
Di Do
+ +
R E U + Y
+ C Go
− +

+
+
Dn

The error signal E = R − Y − Dn , then

Y = Go (CE + Di ) + Do
Y(1 + Go C) = Go C(R − Dn ) + Go Di + Do
U = CR − CY − CDn
U(1 + Go C) = C(R − Dn − Do ) − CGo Di

[email protected] Lecture 2 9 / 36
One Degree of Freedom Closed Loop
Defining the following filters:
1 Go
So = (Sensitivity) Sio = (Input dist. sensitivity)
1 + Go C 1 + Go C
Go C C
To = (Comp. sensitivity) Suo = (Control sensitivity)
1 + Go C 1 + Go C

Then the input and output of the nominal loop are given by:

Y = To (R − Dn ) + Sio Di + So Do
U = Suo (R − Dn − Do ) − To Di

Alternatively, we can write in matrix form

R
 
Y T Sio So  Di 
   
−To
= o
U Suo Do 
 
−To −Suo −Suo
Dn
[email protected] Lecture 2 10 / 36
Closed Loop Requirements
Y = To (R − Dn ) + Sio Di + So Do
U = Suo (R − Dn − Do ) − To Di
So + To = 1
1 To (jω) ≈ 1 where R(jω) is important. To
2 To (jω) ≈ 0 where Dn (jω) is important. Sio = Go So =
C
3 So (jω) ≈ 0 where Do (jω) is important. To
Suo = CSo =
4 Sio (jω) ≈ 0 where Di (jω) is important. Go

(1), (3), and (4) are, in general, compatible (but not equivalent).
(1) and (2) are, in general, not compatible.
BW{To } Should not be too large with respect to BW{Go }

Tipically:
Reference ↔ low frequencies.
Noise ↔ high frequencies.
Disturbances ↔ low frequencies.

[email protected] Lecture 2 11 / 36
Internal stability

Definition
The nominal loop is internally stable if So , To , Sio , Suo are all stable.

P Bo
Let C = and Go = , then
L Ao
Bo P Ao L
To = So =
Ao L + Bo P Ao L + Bo P
Bo L Ao P
Sio = Suo =
Ao L + Bo P Ao L + Bo P

Theorem
The nominal loop is internally stable iff the roots of Acl = Ao L + Bo P are stable.

[email protected] Lecture 2 12 / 36
Robustness to modelling errors

True plant

G = Go (1 + G∆ )

where G∆ the multiplicative uncertainty.

Theorem
Assume that C is a controller that achieves nominal internal stability. Also assume that
Go C and GC have the same number of unstable poles, then a sufficient condition for
the stability of the true feedback loop by applying the controller to the true plant is that

|To (jω)G∆ (jω)| < 1, ∀ω

[email protected] Lecture 2 13 / 36
Outline

1 Fundamental Problem of Control

2 Affine Parameterisation: The stable case

[email protected] Lecture 2 14 / 36
Open Loop Inversion Revisited

Recall that control implicitly and/or explicitly depends on inversion. This is best
seen in the case of open loop control.
In open loop control the input, U, is generated from the reference signal R, by a
transfer function Q, i.e. U = QR.

R U Y
Q Go

This leads to an input-output transfer function of the following form:

To = QGo .

[email protected] Lecture 2 15 / 36
Open Loop Inversion Revisited

This simple formula highlights the fundamental importance of inversion, as

To (jω) will be 1 only at those frequencies where Q(jω) inverts the model.
A key point is that To = QGo is affine in Q.

Recall that with a conventional feedback controller, C, the closed loop transfer
function has the form
Go C
To =
1 + Go C
Note that the above expression is nonlinear in C.

Comparing the two previous equations, we see that the former affine relationship
holds if we simply parameterise C in the following way
C
Q=
1 + Go C
This is the essence of affine parametrisation.

[email protected] Lecture 2 16 / 36
Affine Parameterisation: The Stable Case

We can rearrange the relationship given on the previous slide to express C in

terms of Q and Go :
Q
C= .
1 − QGo

We will then work with Q as the design variable rather than the C.

Note: the relationship between C and Q is one-to-one and thus there is no loss of
generality in working with Q.

[email protected] Lecture 2 17 / 36
Affine Parameterisation: The Stable Case
Youla’s parametrisation of all stabilising controllers for stable plants
Q
This particular form of the controller, i.e. C = , can be drawn schematically
1 − QGo
as:
Controller Di Do
+ +
R E U + Y
Q Plant
+ +
−

Go
−
+
Ym
+
+
Dn

Q
U = C(R − Ym ) = (R − Ym ) = QE
1 − QGo
E = (R − Ym ) + QGo E = R − Ym + Go U
[email protected] Lecture 2 18 / 36
Stability

A very hard question to answer

Given a stable transfer function Go , describe all controllers, C, that stabilise this
nominal plant.

However, in the Q form this question has a very simple answer, namely all that is
required is that Q be stable.

Lemma: Affine parametrisation for stable systems

Consider a plant having a stable nominal model Go controlled in a one d.o.f.
feedback architecture with a proper controller.

Then the nominal loop is internally stable if and only if Q is any stable proper
transfer function when the controller transfer function C is parameterised as:
Q
C= .
1 − QGo

[email protected] Lecture 2 19 / 36
Stability

Proof
We note that the four sensitivity functions can be written as:

To = QGo
So = 1 − QGo
Sio = (1 − QGo ) Go
Suo = Q

We are for the moment only considering the case when Go is stable. Then, we see that
all of the above transfer functions are stable if and only if Q is stable.

[email protected] Lecture 2 20 / 36
Nominal Design

For the nominal design case recall that

To = QGo
So = 1 − QGo
Sio = (1 − QGo ) Go
Suo = Q.

All of these equations are affine in Q.

This makes design particularly straightforward.

[email protected] Lecture 2 21 / 36
Prototype Control Solution

Specifically, if we look at To , i.e.

To = QGo

Recall that a reasonable design goal is to have To near 1 since this implies that
the system output exactly follows the reference signal.

Thus a prototype controller would seem to be simply to choose:

Q = (Go )−1 .

Unfortunately, (Go )−1 is most likely to be improper in practice.

[email protected] Lecture 2 22 / 36
Design Considerations

Hence we introduce a filter FQ to keep Q proper.

It thus seems that a reasonable choice for Q might be:

Q = FQ (Go )−1

where (Go )−1 is the exact inverse of Go .

Not unexpectedly, we see that inversion plays a central role in this prototype
solution.

NOTE: In this case: To = QGo = FQ (Go )−1 Go = FQ .

[email protected] Lecture 2 23 / 36
Design Considerations

Although the design proposed above is a useful starting point it will have to be refined
to accommodate more detailed design considerations.

In particular, we will investigate the following issues

1 Non-minimum phase zeros
2 Model relative degree
3 Disturbance rejection
4 Control effort
5 Robustness

[email protected] Lecture 2 24 / 36
1. Non-minimum Phase Zeros

Recall that, provided Go is stable, then Q only needs to be stable to ensure closed loop
stability:

This implies that, if Go contains NMP zeros, then they cannot be included in
(Go )−1 .

One might therefore think of replacing the previous equation by:

Q = FQ (Go )i
where (Go )i is a stable approximation to (Go )−1 .

[email protected] Lecture 2 25 / 36
1. Non-minimum Phase Zeros

For example, if one factors Go as:

Bos Bou
Go =
Ao
where Bos and Bou are the stable and unstable factors in the numerator,
respectively, with Bou (0) = 1.

A suitable choice for (Go )i would be

Ao
(Go )i = .
Bos

Note that Bou (0) = 1 in order to have the correct inverse for s = 0 (constant
reference).

[email protected] Lecture 2 26 / 36
2. Model Relative Degree

In order to obtain a proper controller it is necessary that Q be proper

Thus it is necessary that the shaping filter, FQ , ensures that Q is proper.

Conceptually, this can be achieved by including factors of the form (τs + 1)nd ,
where τ ∈ R+ , in the denominator.

[email protected] Lecture 2 27 / 36
3. Disturbance Rejection

Recall, the expressions for the closed loop sensitivity functions

To = QGo
So = 1 − QGo
Sio = (1 − QGo ) Go
Suo = Q.

It would seem that to achieve perfect disturbance rejection at frequency ωi

simply requires that QGo be 1 at ωi .

For example, rejection of a d.c. disturbance requires Q(0)Go (0) = 1.

[email protected] Lecture 2 28 / 36
3. Disturbance Rejection

Consider a stable model Go with input and/or output disturbance at zero

frequency.

Then, a one d.o.f. control loop, giving zero steady state tracking error, is stable if
and only if the controller C can be expressed in the affine form and Q satisfies:
Q = sQ̄ + (Go (0))−1 Qa

where Q̄ is any stable transfer function, and Qa is any stable transfer function
which satisfies Qa (0) = 1.

The simplest choice for Qa is

Qa = 1

Check proof in Goodwin, Graebe and Salgado (Ch. 15).

[email protected] Lecture 2 29 / 36
4. Control Effort

All plants are typically low pass. Hence, any attempt to make Q close to the
model inverse necessarily gives a high pass transfer function from Do to U. This
will lead to large input signals and may lead to controller saturation.

[email protected] Lecture 2 30 / 36
5. Robustness

Finally, we turn to the issue of robustness in choosing Q

Recall that, in order to ensure robustness, the closed loop bandwidth should be
such that the frequency response |To (jω)| rolls off before the effects of
modelling errors become significant.

Thus, in the framework of the affine parametrisation under discussion here, the
robustness requirement can be satisfied if FQ reduces the gain of To (jω) at high
frequencies.

This is usually achieved by including appropriate poles in FQ .

[email protected] Lecture 2 31 / 36
Choice of Q: Summary for the Case of Stable Plants
We have seen that a prototype choice for Q is simply the inverse of the open loop
plant transfer function Go .

However, this ideal solution needs to be modified in practice to take into account:
Non-minimum phase zeros. Internal stability precludes the cancellation of these
zeros. They must therefore appear in To . This implies that the gain of Q must be
reduced at these frequencies for robustness reasons.
Relative degree. Excess poles in the model must necessarily appear as a lower
bound for the relative degree of To , since Q must be proper to ensure that the
controller C is proper.
Disturbance trade-offs. Whenever we roll To off to satisfy measurement noise
rejection, we increase sensitivity to output disturbances at that frequency.
Control energy. Q is typically a high pass transfer function. This will lead to
large input signals and may lead to controller saturation.
Robustness. Modeling errors usually become significant at high frequencies, and
hence to retain robustness it is necessary to attenuate To , and hence Q, at these
frequencies.
[email protected] Lecture 2 32 / 36
Summary for Stable Systems

C = Q (1 − QGo )−1 , where the design is carried out by designing the transfer
function Q.

Nominal sensitivities:

To = QGo
So = 1 − QGo
Sio = (1 − QGo ) Go
Suo = Q

[email protected] Lecture 2 33 / 36
Design Methodology

Determine the design specifications. (What do you want the B.W. of To to be?)
Invert the nominal model Go . (Do you need to take an approximate inverse?)
Recall Q = FQ (Go )i . Determine the relative degree of FQ to ensure Q is proper.
Specify the parameters of FQ to satisfy the design specification.
Design Q.
Convert to C if required.

[email protected] Lecture 2 34 / 36
Summary of results for stable systems

Advantages of affine parametrisation

Observe the following advantages of affine parametrisation:

Nominal stability is explicit.

The known quantity Go and the quantity sought by the control engineer (Q)
occur in the highly insightful relation To = QGo (multiplicative in the frequency
domain).

The simple multiplicative relation QGo provides deep insights into the trade-offs
of a particular problem and provides a very direct means of pushing the design
by shaping Q.

The sensitivities are affine in Q, which is a great advantage for synthesis

techniques relying on numerical minimisation of a criterion.

[email protected] Lecture 2 35 / 36
Summary of results for stable systems

Misconceptions
The following points are important to avoid some common misconceptions:

The associated trade-offs are not a consequence of the affine parametrisation:

they are general and hold for any linear time invariant controller including PID,
pole placement, LQR, H2 , H∞ .

Affine parametrisation makes the general trade-offs more visible and provides a
direct means for the control engineer to make trade-off decisions.

The fact that Q must approximate the inverse of the model at frequencies where
the sensitivity is meant to be small (S ≈ 0 ⇔ T ≈ 1) is perfectly general and
highlights the fundamental importance of inversion in control.

[email protected] Lecture 2 36 / 36