Chapter 2 Goodwin, Graebe, Salgado©, Prentice Hall 2000

Chapter 2

Introduction to the Principles

of Feedback
Topics to be covered include:
 An industrial motivational example;
 A statement of the fundamental nature of the control problem;
 The idea of inversion as the central ingredient in solving
control problems;
 Evolution from open loop inversion to closed loop feedback
solutions.
Chapter 2 Goodwin, Graebe, Salgado©, Prentice Hall 2000

We will see that feedback is a key tool that can be

used to modify the behaviour of a system.
This behaviour altering effect of feedback is a key
mechanism that control engineers exploit
deliberately to achieve the objective of acting on a
system to ensure that the desired performance
specifications are achieved.
Chapter 2 Goodwin, Graebe, Salgado©, Prentice Hall 2000

A motivating industrial example

We first present a simplified, yet essentially
authentic, example of an industrial control problem.
The example, taken from the steel industry, is of a
particular nature, however the principal elements of
specifying a desired behaviour, modeling and the
necessity for trade-off decisions are generic.
Chapter 2 Goodwin, Graebe, Salgado©, Prentice Hall 2000

Photograph of Bloom Caster

Chapter 2 Goodwin, Graebe, Salgado©, Prentice Hall 2000

Process schematic of an Industrial Bloom Caster

Chapter 2 Goodwin, Graebe, Salgado©, Prentice Hall 2000

Continuous caster. Typical bloom (left) and simplified

diagram (right)
Chapter 2 Goodwin, Graebe, Salgado©, Prentice Hall 2000

Operators viewing the mould

Chapter 2 Goodwin, Graebe, Salgado©, Prentice Hall 2000

The cast strip in the secondary cooling chamber

Chapter 2 Goodwin, Graebe, Salgado©, Prentice Hall 2000

Performance specifications
The key performance goals for this problem are:
 Safety: Clearly, the mould level must never be in danger of
overflowing or emptying as either case would result in
molten metal spilling with disastrous consequences.

 Profitability: Aspects which contribute to this requirement

include:
 Product quality

 Maintenance

 Throughput
Chapter 2 Goodwin, Graebe, Salgado©, Prentice Hall 2000

Modeling
To make progress on the control system design problem,
it is first necessary to gain an understanding of how the
process operates. This understanding is typically
expressed in the form of a mathematical model.
h* : commanded level of steel in mould
h(t ) : actual level of steel in mould
v(t ) : valve position
 (t ) : casting speed
qin(t ) : inflow of matter into the mould
qout (t ) : outflow of matter from the mould
Chapter 2 Goodwin, Graebe, Salgado©, Prentice Hall 2000

Model as simple tank

Tundish
Molten Steel
Valve

Mould Level

Cooling Water
Chapter 2 Goodwin, Graebe, Salgado©, Prentice Hall 2000

Block diagram of the simplified mould level dynamics,

sensors and actuators

These variables are related as shown below:

Chapter 2 Goodwin, Graebe, Salgado©, Prentice Hall 2000

Feedback and Feedforward

We will find later that the core idea in control is that
of inversion. Moreover, inversion can be
conveniently achieved by the use of two key
mechanisms (namely, feedback and feedforward).
Chapter 2 Goodwin, Graebe, Salgado©, Prentice Hall 2000

Figure 2.4: Model of the simplified mould level control with

feedforward compensation for casting speed

Suggested Control Strategy:

Note that this controller features joint feedback and a

preemptive action (feedforward).
Chapter 2 Goodwin, Graebe, Salgado©, Prentice Hall 2000

A first indication of trade-offs

On simulating the performance of the above control
loop for K=1 and K=5, see Figure 2.5, we find that
the smaller controller gain (K=1) results in a slower
response to a change in the mould level set-point. On
the other hand, the larger controller gain (K=5),
results in a faster response but also increases the
effects of measurement noise as seen by the less
steady level control and by the significantly more
aggressive valve movements.
Chapter 2 Goodwin, Graebe, Salgado©, Prentice Hall 2000

Figure 2.5: A first indication of trade-offs: Increased

responsiveness to set-point changes also increases
sensitivity to measurement noise and actuator wear.
Chapter 2 Goodwin, Graebe, Salgado©, Prentice Hall 2000

Question
We may ask if these trade-offs are unavoidable or
whether we could improve on the situation by such
measures as:
 better modelling
 more sophisticated control system design

This will be the subject of the rest of our deliberations.

(Aside: Actually the trade-off is fundamental as we

shall see presently).
Chapter 2 Goodwin, Graebe, Salgado©, Prentice Hall 2000

Definition of the control problem

Abstracting from the above particular problem, we can
introduce:
Definition 2.1:
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
behaviour. Furthermore, this approximate behaviour
should be achieved in the face of uncertainty of the
process and in the presence of uncontrollable external
disturbances acting on the process.
Chapter 2 Goodwin, Graebe, Salgado©, Prentice Hall 2000

Prototype solution to the control

problem via inversion
One particularly simple, yet insightful way of
thinking about control problems is via inversion. To
describe this idea we argue as follows:
 say that we know what effect an action at the input of a
system produces at the output, and
 say that we have a desired behaviour for the system
output, then one simply needs to invert the relationship
between input and output to determine what input action
is necessary to achieve the desired output behaviour.
Chapter 2 Goodwin, Graebe, Salgado©, Prentice Hall 2000

Figure 2.6: Conceptual controller

The above idea is captured in the following diagram:

Chapter 2 Goodwin, Graebe, Salgado©, Prentice Hall 2000

We will actually find that the inverse solution given

on the last slide holds very generally.
Thus, all controllers implicitly generate an inverse of
the process, in so far that this is feasible. However,
the details of controllers will differ with respect to
the mechanism used to generate the required
approximate inverse.
Chapter 2 Goodwin, Graebe, Salgado©, Prentice Hall 2000

High gain feedback and inversion

We next observe that there is a rather intriguing property
of feedback, namely that it implicitly generates an
approximate inverse of dynamic transformations, without
the inversion having to be carried out explicitly.
Figure 2.7:
Realisation of
conceptual controller

The loop implements an approximate inverse of f o,

i.e. u = f r, if
r - h-1u  r
Chapter 2 Goodwin, Graebe, Salgado©, Prentice Hall 2000

Specifically,
uh rz h r f u

or
h 1 u  r  f u

Hence
u  f 1 r  h 1 u

 f 1 r

Provided h 1 u is small, i.e. h is high gain.

Chapter 2 Goodwin, Graebe, Salgado©, Prentice Hall 2000

The above equation is satisfied if h-1u is large. We

conclude that an approximate inverse is generated
provided we place the model of the system in a high
gain feedback loop.
Chapter 2 Goodwin, Graebe, Salgado©, Prentice Hall 2000

Example 2.3
Assume that a plant can be described by the model
dy(t )
 2 y(t )  u(t )
dt
and that a control law is required to ensure that y(t)
follows a slowly varying reference.
One way to solve this problem is to construct an
inverse for the model which is valid in the low
frequency region. Using the architecture in Figure
2.7, we obtain an approximate inverse, provided that
ho has large gain in the low frequency region.
Chapter 2 Goodwin, Graebe, Salgado©, Prentice Hall 2000

Figure 2.8: Tank level control using approximate inversion

Simulating the resultant controller gives the results

below:
Chapter 2 Goodwin, Graebe, Salgado©, Prentice Hall 2000

From open to closed loop

architectures
Unfortunately, the above methodology will not lead
to a satisfactory solution to the control problem
unless:
 the model on which the design of the controller has been
based is a very good representation of the plant,
 the model and its inverse are stable, and
 disturbances and initial conditions are negligible.
We are thus motivated to find an alternative solution
to the problem which retains the key features but
which does not suffer from the above drawbacks.
Chapter 2 Goodwin, Graebe, Salgado©, Prentice Hall 2000

Figure 2.9: Open loop control with built-in inverse

Figure 2.10: Closed loop control

Chapter 2 Goodwin, Graebe, Salgado©, Prentice Hall 2000

 The first thing to note is that, provided the model

represents the plant exactly, and that all signals are
bounded (i.e. the loop is stable), then both schemes are
equivalent, regarding the relation between r(t) and y(t).
The key differences are due to disturbances and different
initial conditions.
 In the open loop control scheme the controller incorporates
feedback internally, i.e. a signal at point A is fed back.
Chapter 2 Goodwin, Graebe, Salgado©, Prentice Hall 2000

 In the closed loop scheme, the feedback signal depends on

what is actually happening in the plant since the true plant
output is used.

We will see later that this modified architecture has

many advantages including:
 insensitivity to modelling errors;
 insensitivity to disturbances in the plant (that are not
reflected in the model).
Chapter 2 Goodwin, Graebe, Salgado©, Prentice Hall 2000

Trade-offs involved in choosing

the feedback gain
The preliminary insights of the previous two sections
would seem to imply that all that is needed to generate
a controller is to put high gain feedback around the
plant. This is true in so far that it goes. However,
nothing in life is cost free and this also applies to the
use of high gain feedback.
For example, if a plant disturbance leads to a non-zero
error e(t), in Figure 2.10, then high gain feedback will
result in a very large control action u(t). This may lie
outside the available input range and thus invalidate the
solution.
Chapter 2 Goodwin, Graebe, Salgado©, Prentice Hall 2000

Another potential problem with high gain feedback is

that it is often accompanied by the very substantial risk
of instability. Instability is characterised by self
sustaining (or growing) oscillations. As an illustration,
the reader will probably have witnessed the high pitch
whistling sound that is heard when a loudspeaker is
placed too close to a microphone. This is a manifestation
of instability resulting from excessive feedback gain.
Tragic manifestations of instability include aircraft
crashes and the Chernobyl disaster in which a runaway
condition occurred.
Chapter 2 Goodwin, Graebe, Salgado©, Prentice Hall 2000

Yet another potential disadvantage of high loop gain

was hinted at in the mould level example. There we
saw that increasing the controller gain lead to
increased sensitivity to measurement noise.
(Actually, this turns out to be generically true).
Chapter 2 Goodwin, Graebe, Salgado©, Prentice Hall 2000

In summary, high loop gain is desirable from many

perspectives but it is also undesirable when viewed
from other perspectives. Thus, when choosing the
feedback gain one needs to make a conscious trade-
off between competing issues.
Chapter 2 Goodwin, Graebe, Salgado©, Prentice Hall 2000

The previous discussion can be summarised in the

following statement:

High loop gain gives approximate inversion which is

the essence of control. However, in practice, the
choice of feedback gain is part of a complex web of
design trade-offs. Understanding and balancing
these trade-offs is the essence of control system
design.
Chapter 2 Goodwin, Graebe, Salgado©, Prentice Hall 2000

Measurements
 Finally, we discuss the issue of measurements (i.e. what it
is we use to generate the feedback signal).

 A more accurate description of the feedback control loop

including sensors is shown in Figure 2.11.
Chapter 2 Goodwin, Graebe, Salgado©, Prentice Hall 2000

Figure 2.11: Closed loop control with sensors

Chapter 2 Goodwin, Graebe, Salgado©, Prentice Hall 2000

Desirable attributes of sensors

 Reliability. It should operate within the necessary range.
 Accuracy. For a variable with a constant value, the
measurement should settle to the correct value.
 Responsiveness. If the variable changes, the measurement
should be able to follow the changes. Slow responding
measurements can, not only affect the quality of control but
can actually make the feedback loop unstable. Loop instability
may arise even though the loop has been designed to be stable
assuming an exact measurement of the process variable.
Chapter 2 Goodwin, Graebe, Salgado©, Prentice Hall 2000

 Noise immunity. The measurement system, including the

transmission path, should not be significantly affected by
exogenous signals such as measurement noise.
 Linearity. If the measurement system is not linear, then at least
the nonlinearity should be known so that it can be
compensated.
 Non intrusive. The measuring device should not significantly
affect the behaviour of the plant.
Chapter 2 Goodwin, Graebe, Salgado©, Prentice Hall 2000

Figure 2.12: Typical feedback loop

In summary, a typical feedback loop (including sensor

issues) is shown below.