SYSTEM THEORY

STATE SPACE ANALYSIS

AND
CONTROL THEORY

Lecture Notes in Control Theory

PhD (Dr. ing.)

David Di Ruscio

Master study in Systems and Control Engineering

Faculty of Technology
Telemark University College

August 1996

August 29, 2017 Report:

Systems and Control Engineering
Faculty of Technology
Telemark University College
Kjølnes ring 56
N-3914 Porsgrunn
Preface

These lecture notes is meant to be used in the Control Theory part of the course
SCE1106 which is to be held for the master study in Systems and Control Engi-
neering. The contents is also basic theory for courses as System Identification and
Advanced Control theory.
The following words should be noted

All this –
was for you, dear reader,
I wanted to write a book
that you would understand.
For what good is it to me
if you can’t understand it ?
But you got to try hard —

This verse is from Kailath (1980).

Contents

Preface i

I State Space Analysis viii

1 Basic System Theory 1

1.1 Models of dynamic systems . . . . . . . . . . . . . . . . . . . . . . . 1
1.2 Linear Time State Space Models . . . . . . . . . . . . . . . . . . . . 2
1.2.1 Proof of the solution of the state equation . . . . . . . . . . . 4
1.3 Linear transformation of state space models . . . . . . . . . . . . . . 5
1.4 Eigenvalues and eigenvectors . . . . . . . . . . . . . . . . . . . . . . 6
1.4.1 Krylovs method used to find the coefficients of the character-
istic equation . . . . . . . . . . . . . . . . . . . . . . . . . . . 7
1.5 Similarity Transformations and eigenvectors . . . . . . . . . . . . . . 8
1.6 Time constant . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 9
1.7 The matrix exponent and the transition matrix . . . . . . . . . . . . 11
1.7.1 Computing the matrix exponent by diagonalisation . . . . . . 11
1.7.2 Parlets method for computing the matrix exponent . . . . . . 12
1.7.3 Matrix exponential by series expansion . . . . . . . . . . . . . 13
1.8 Examples . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 14
1.9 Transfer function and transfer matrix models . . . . . . . . . . . . . 18
1.10 Poles and zeroes . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 19
1.11 Time Delay . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 19
1.12 Linearization . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 20
1.12.1 Calculating the Jacobian matrix numerically . . . . . . . . . 23
1.13 Stability of linear systems . . . . . . . . . . . . . . . . . . . . . . . . 25
1.13.1 Stability of continuous time linear systems . . . . . . . . . . . 25
CONTENTS iii

1.13.2 Stability of discrete time linear systems . . . . . . . . . . . . 25

1.14 State Controllability . . . . . . . . . . . . . . . . . . . . . . . . . . . 26
1.15 State Observability . . . . . . . . . . . . . . . . . . . . . . . . . . . . 27

2 Canonical forms 30
2.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 30
2.2 Controller canonical form . . . . . . . . . . . . . . . . . . . . . . . . 31
2.2.1 From transfer function to controller canonical form . . . . . . 31
2.2.2 From state space form to controller canonical form . . . . . . 33
2.3 Controllability canonical form . . . . . . . . . . . . . . . . . . . . . . 33
2.4 Observer canonical form . . . . . . . . . . . . . . . . . . . . . . . . . 34
2.4.1 From transfer function to observer canonical form . . . . . . 34
2.5 Observability canonical form . . . . . . . . . . . . . . . . . . . . . . 37
2.6 Duality between canonical forms . . . . . . . . . . . . . . . . . . . . 39
2.6.1 Duality between controller and observer canonical forms . . . 39
2.6.2 Duality between controllability and observability canonical forms 40
2.7 Examples . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 41
2.8 Summary . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 43

II Analysis and design of control systems 45

3 Simple PID tuning and model reduction 46

3.1 Feedback systems . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 46
3.2 Standard feedback system . . . . . . . . . . . . . . . . . . . . . . . . 46
3.3 Standard feedback systems with disturbances . . . . . . . . . . . . . 47
3.4 Simple PID tuning rules . . . . . . . . . . . . . . . . . . . . . . . . . 47
3.4.1 The half rule . . . . . . . . . . . . . . . . . . . . . . . . . . . 49
3.4.2 Tuning of PI and PID controllers . . . . . . . . . . . . . . . . 50
3.4.3 Choice of Tc . . . . . . . . . . . . . . . . . . . . . . . . . . . . 52
3.4.4 Modification of the integral term . . . . . . . . . . . . . . . . 52
3.5 PID controller for oscillating process . . . . . . . . . . . . . . . . . . 52
3.6 ID controller for systems with no damping . . . . . . . . . . . . . . . 53
3.7 PI Control of first order process . . . . . . . . . . . . . . . . . . . . . 54
3.8 Integrating process with time delay . . . . . . . . . . . . . . . . . . . 55
CONTENTS iv

3.8.1 The SIMC settings: neglecting the time delay . . . . . . . . . 56

3.8.2 Settings by approximating time delay as inverse response . . 58
3.8.3 Settings by approximating time delay with Pade and Balchen
approximation . . . . . . . . . . . . . . . . . . . . . . . . . . 59
3.9 Re-tuning to avoid oscillations . . . . . . . . . . . . . . . . . . . . . 61
3.10 Controller for special type systems . . . . . . . . . . . . . . . . . . . 61
3.10.1 Pure time delay process . . . . . . . . . . . . . . . . . . . . . 61
3.11 Examples . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 62

4 The basic PID controller 71

4.1 The PI controller . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 71
4.1.1 Frequency domain description of the PI controller . . . . . . 71
4.1.2 Continuous Time domain description of the PI controller . . 71
4.1.3 Discrete Time domain description of the PI controller . . . . 72
4.2 The PID controller . . . . . . . . . . . . . . . . . . . . . . . . . . . . 75
4.2.1 Frequency domain description of the PID controller . . . . . 75
4.2.2 Continuous Time domain description of the PID controller . 77
4.2.3 Discrete Time domain description of the PID controller . . . 77
4.3 Anti windup and constraints . . . . . . . . . . . . . . . . . . . . . . 77
4.4 Bumpless transfer . . . . . . . . . . . . . . . . . . . . . . . . . . . . 80
4.4.1 Bumpless transfer between manual and automatic mode . . . 80
4.4.2 Bumpless transfer between PID parameter changes . . . . . . 82

5 Time delay 83
5.1 Padé approximations to the exponential for eθ . . . . . . . . . . . . . 83
5.1.1 Developing a 1st order Padé approximation . . . . . . . . . . 83
5.1.2 Alternative prof of the 1st order Padé approximation . . . . . 84
5.1.3 Developing a (1, 0) Padé approximation . . . . . . . . . . . . 85
5.1.4 (s, t) Padé approximations . . . . . . . . . . . . . . . . . . . . 86
5.2 Padé approksimasjoner for e−τ s . . . . . . . . . . . . . . . . . . . . . 88
5.3 Balchen approksimasjoner for e−τ s . . . . . . . . . . . . . . . . . . . 89

6 Feedback systems 90
6.1 Description of feedback systems . . . . . . . . . . . . . . . . . . . . . 90
6.2 Reasons for using feedback . . . . . . . . . . . . . . . . . . . . . . . 96
CONTENTS v

7 Direct synthesis and design of standard controllers 97

7.1 On the PID controller formulations . . . . . . . . . . . . . . . . . . . 97
7.2 Controlling a static (steady state) process . . . . . . . . . . . . . . . 99
7.3 Control of a non-minimum phase process . . . . . . . . . . . . . . . . 101
7.4 Controlling of lead-lag systems . . . . . . . . . . . . . . . . . . . . . 103

8 Feed forward control 105

8.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 105
8.2 Feedback with feed forward from the reference . . . . . . . . . . . . 106
8.3 Feed-forward from the disturbance . . . . . . . . . . . . . . . . . . . 108
8.3.1 Design based on a state space model . . . . . . . . . . . . . . 108
8.4 Ratio control . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 109

9 Frequency domain analysis and controller synthesis 111

9.1 Complex numbers . . . . . . . . . . . . . . . . . . . . . . . . . . . . 111
9.2 Frequency response . . . . . . . . . . . . . . . . . . . . . . . . . . . . 112
9.3 Gain margin and Phase margin . . . . . . . . . . . . . . . . . . . . . 115
9.4 Bodes stability criterion . . . . . . . . . . . . . . . . . . . . . . . . . 122
9.5 Ziegler-Nichols method for PID-controller tuning . . . . . . . . . . . 127
9.6 Minimum-phase and non-minimum-phase systems . . . . . . . . . . . 128
9.7 The bandwidth of a feedback control system . . . . . . . . . . . . . . 129

10 Discrete time systems 131

10.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 131
10.2 θ method for discretization . . . . . . . . . . . . . . . . . . . . . . . 131
10.3 Deviation form of the PI controller . . . . . . . . . . . . . . . . . . . 132
10.4 Deviation form of the PID controller . . . . . . . . . . . . . . . . . . 133
10.4.1 The continuous time PID controller . . . . . . . . . . . . . . 133
10.4.2 Deviation form of the PID controller using the explicit Euler
method . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 134
10.4.3 Deviation form of the PID controller using the trapezoid method135
10.5 Discussion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 135

11 Time delay in discrete systems 136

11.1 Modeling of time delay . . . . . . . . . . . . . . . . . . . . . . . . . . 136
11.1.1 Transport delay and controllability canonical form . . . . . . 136
CONTENTS vi

11.1.2 Time delay and observability canonical form . . . . . . . . . 138

11.2 Implementation of time delay . . . . . . . . . . . . . . . . . . . . . . 139
11.3 Examples . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 140

12 Adjustment of PID control parameters 141

12.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 141
12.2 Main Results . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 142
12.2.1 Method 1 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 142
12.2.2 Method 2 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 143
12.2.3 Method 3 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 143
12.3 Examples . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 145
12.3.1 Example 1 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 145
12.3.2 Example 2 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 146
12.4 Concluding Remarks . . . . . . . . . . . . . . . . . . . . . . . . . . . 147
12.5 References . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 148
12.6 Appendix . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 148
12.6.1 A proof of Equation (12.7) . . . . . . . . . . . . . . . . . . . 148
12.6.2 A proof of Equation (12.11) . . . . . . . . . . . . . . . . . . . 149
12.6.3 A proof of Equation (12.20) . . . . . . . . . . . . . . . . . . . 150

III Model based control 151

13 Modified Smith Predictor 152

13.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 152
13.2 The Smith Predictor . . . . . . . . . . . . . . . . . . . . . . . . . . . 152
13.2.1 Transfer Function . . . . . . . . . . . . . . . . . . . . . . . . 152
13.2.2 Process with integrator and time delay . . . . . . . . . . . . . 153
13.3 Modified Smith predictor . . . . . . . . . . . . . . . . . . . . . . . . 154
13.4 Modified Smith Predictor by Åström . . . . . . . . . . . . . . . . . . 155
13.4.1 A Model Based Point of View . . . . . . . . . . . . . . . . . . 156

IV Control of multivariable systems 158

14 Multivariable control 159

14.1 Interaction and pairing of variables . . . . . . . . . . . . . . . . . . . 159
CONTENTS vii

14.1.1 Closed loop u2 → y2 . . . . . . . . . . . . . . . . . . . . . . . 160

14.1.2 Proof of Equation (14.7) . . . . . . . . . . . . . . . . . . . . . 162
14.1.3 Rules for pairing variables . . . . . . . . . . . . . . . . . . . . 163
14.2 Relative Gain Array (RGA) . . . . . . . . . . . . . . . . . . . . . . . 165
14.2.1 Relationship between RGA and ∆ . . . . . . . . . . . . . . . 167

15 Multiple inputs and outputs and control structures 168

15.1 Split Range Control . . . . . . . . . . . . . . . . . . . . . . . . . . . 168

References 169

A Laplace Transformations 170

A.1 Laplace transformation of the exponential . . . . . . . . . . . . . . . 170
A.2 Laplace transformation of triogonometric functions . . . . . . . . . . 171
 Part I

State Space Analysis

Chapter 1

Basic System Theory

1.1 Models of dynamic systems

The aim of this section is not to discuss modeling principles of dynamic systems in
detail. However we will in this introductory section mention that dynamic models
may be developed in many ways. For instance so called first principles methods as
mass balances, force balances, energy balances, i.e., conservation of law methods,
leads to ether non-linear models of the type
ẋ = f (x, u) (1.1)
y = g(x) (1.2)
or linear or linearized models of the type
ẋ = Ax + Bu (1.3)
y = Dx (1.4)
Note also that a linearized approximation of the non-linear model usually exist. We
will in the following give a simple example of a system which may be described by
a linear continuous time state space model

Example 1.1 (Model of a damped spring system)

Assume given an object with mass, m, influenced by three forces. One force F1 used
to pull the mass, one force F2 = kx from the spring and one force F3 = µẋ = µv
that represents the friction or viscous damping.
We define x as the position of the object and ẋ = v as the velocity of the object.
Furthermore the force F1 may be defined as a manipulable control input variable and
we use u as a symbol for this control input, i.e., u = F1 .
from this we have the following force balance
3
X
ma = mv̇ = Fi = F1 − F2 − F3 = −kx − µv + u (1.5)
i=1

The model for the damped spring system consists of two continuous time ordinary
differential equations. Those two ODEs may be written in standard state space form
1.2 Linear Time State Space Models 2

as follows
ẋ A x B
z }| { z }| { 
z }| { z }| {
ẋ 0 1 x 0
= k µ + 1 u (1.6)
v̇ −m −m v m

Modeling from first principles, e.g., as the in the damped spring example above,
often leads to a standard linear continuous time state space model on the form
ẋ = Ax + Bu (1.7)
where x ∈ Rn is the state vector, u ∈ Rr is the control input vector, A ∈ Rn timesn

is state matrix and B ∈ Rn timesr is the control input matrix.

1.2 Linear Time State Space Models

An important class of state space models is the time invariant linear and continuous
time state space model of the form

ẋ = Ax + Bu, x(0) = x0 , (1.8)

y = Dx, (1.9)
where u ∈ Rr is the control vector, x ∈ Rn is the state vector, y ∈ Rm is the mea-
surements vector and x0 = x(t0 ) ∈ Rn is the initial value of the state vector, which
usually is assumed to be known. Time invariant means that the model matrices A,
B and D are constant matrices, i.e. time invariant.
It can be shown that the exact solution of the state equation (1.8) at time t0 ≤ t
is given by
Z t
A(t−t0 )
x(t) = e x(t0 ) + eA(t−τ ) Bu(τ )dτ. (1.10)
t0

As we see, the solution consists of two parts. The first part represents the
autonomous response (homogenous solution) driven only by initial values different
from zero. The second term represents the in homogenous solution driven by the
control variable, u(t).
In order to compute the first term we have to compute the matrix exponential
eA(t−t0 ) . This matrix exponential is defined as the transition matrix, because it
defines the transition of the state from the initial value, x(t0 ), to the final state x(t)
in an autonomous system ẋ = Ax with known initial state x(t0 ). The transition
matrix is defined as follows
def
Φ(t) = eAt . (1.11)
Using this definition of the transition matrix we see that the solution (1.10) can be
written as follows
Z t
x(t) = Φ(t − t0 )x(t0 ) + Φ(t − τ )Bu(τ )dτ. (1.12)
t0
1.2 Linear Time State Space Models 3

The second term in the solution (1.10) (or equivalent as in (1.12)) consists of a
convolutional integral. This integral must usually be computed numerically, e.g. it
is usually hard to obtain an analytically solution. However, an important special
case is the case where the control u(τ ) is constant over the integration interval
t0 < τ ≤ t.

x(t) = Φ(t − t0 )x(t0 ) + ∆u(t0 ), (1.13)

where ∆ is shown to be
Z t Z t−t0
∆= eA(t−τ ) Bdτ = eAτ Bdτ (1.14)
t0 0

Note also that

∆ = A−1 (eA(t−t0 ) − I)B, (1.15)

when A is non singular. It is this solution which usually is used in order to compute
the general solution to the state equation. Hence, the control input u(t) is assumed
to be constant over piece wise identical intervals ∆t = t − t0 .
The constant interval ∆t is in control theory and control systems defined as the
sampling time in the digital controller. If we now are putting t = t0 + ∆t in the
solution (1.13) then we get

x(t0 + ∆t) = Φ(∆t)x(t0 ) + ∆u(t0 ), (1.16)

where ∆ is given by

∆ = A−1 (eA∆t − I)B. (1.17)

The solution given by (1.16) and (1.17) is the starting point for making a discrete
time state space model for the system. In digital control systems discrete time mod-
els are very important. Discrete time models are also very important for simulation
purposes of a dynamic system.
Consider now the case where we let t0 in (1.16) and (1.17) take the discrete time
values

t0 = k∆t ∀ k = 0, 1, . . . , (1.18)

We then have a discrete time model of the form

x((k + 1)∆t) = Φ(∆t)x(k∆t) + ∆u(k∆t), (1.19)

It is however common to use the short hand notation

xk+1 = Φxk + ∆uk . (1.20)

We have here defined

xk = x(k∆t) = x(t0 ) (1.21)

1.2 Linear Time State Space Models 4

Note also that we usually are using symbols as A and B also for discrete time state
space models, e.g., so that the model (1.20) is written as

xk+1 = Axk + Buk . (1.22)

we will usually using the symbols A and B also for discrete time models. However,
in cases where there can be conflicts symbols as Φ and ∆ are used.
It is important to note that the steady state solution to a continuous time model
ẋ = Ax + Bu can be found by putting ẋ = 0. I.e., the steady state solution when
time approach infinity (t → ∞) is given by

x = −A−1 Bu. (1.23)

Here the system matrix A is assumed to be non singular.

In a stable system, the transients and dynamic responses will die out as time
approach infinity, and al variables will be constant as function of time. Therefore is
also the derivative of the states with time equal to zero, i.e.
dx
ẋ = =0 (1.24)
dt
Note also that the steady state solution of a continuous time model and the discrete
time model should be the same. This is obvious

1.2.1 Proof of the solution of the state equation

It can be shown that the homogenous solution to the state equation ẋ = Ax + Bu

(with known initial state x(t0 )) when u = 0 is of the form

x(t) = eA(t−t0 ) z (1.25)

because dx/dt = AeA(t−t0 ) z = Ax.

The solution of the in homogenous differential equation can be found by assuming
that the vector z is time variant. We then have from (1.25) that

ẋ = AeA(t−t0 ) z + eA(t−t0 ) ż (1.26)

We also have from the state equation that

ẋ = Ax + Bu = AeA(t−t0 ) z + Bu (1.27)

where we have used that x is given as in (1.25). Comparing (1.26) and (1.27) shows
that

eA(t−t0 ) ż = Bu. (1.28)

This gives

ż = e−A(t−t0 ) Bu (1.29)
1.3 Linear transformation of state space models 5

where we have used that (eA )−1 = e−A . This gives the following solution for the
vector z, i.e.,
Z t
z(t) = z0 + e−A(τ −t0 ) Budτ. (1.30)
t0

We are putting (1.30) in (1.25). This gives

Z t
A(t−t0 ) A(t−t0 )
x(t) = e z=e z0 + eA(t−τ ) Budτ. (1.31)
t0

Putting t = t0 shows that x(t0 ) = z0 and the final solution is found to be

Z t
A(t−t0 )
x(t) = e x(t0 ) + eA(t−τ ) Budτ. (1.32)
t0

This section is meant to show how the solution to the continuous time state equation
can be proved.

1.3 Linear transformation of state space models

Let x be the state vector in the state space realization (A, B, D) such that

ẋ = Ax + Bu, x(0) = x0 (1.33)

y = Dx (1.34)

An equivalent realization of the system defined by (1.33) and (1.34) can be found
by choosing another basis for the state space (choosing another state). The state
vector x can be transformed to a new coordinate system. This can be done by
defining a non singular transformation matrix T ∈ Rn×n and the following linear
transformation,

x = T z ⇔ z = T −1 x (1.35)

An equivalent realization of the system (1.33) and (1.34) is then given by

ż = T −1 AT z + T −1 Bu, z(0) = T −1 x0 (1.36)

y = DT z (1.37)

These two state space realizations can be shown to be identical and represent the
same system. The two systems has the same transfer function from the input, u to
the output, y.
An infinite number of non singular transformation matrices T can be chosen.
This leads to an infinite number of state space model realizations. Some of these
realizations has special properties, e.g., state space models with special properties
can be found by choosing T properly.
1.4 Eigenvalues and eigenvectors 6

1.4 Eigenvalues and eigenvectors

Consider given a matrix A ∈ Rn×n . The characteristic polynomial of A is then

defined as

p(A) = det(λI − A) = det(A − λI) (1.38)

n n−1
= λ + pn λ + · · · + p2 λ + p1 (1.39)

where the n coefficients p1 , p2 , . . . , pn−1 , pn are real values. These coefficients can be
found by actually expressing the determinant. The characteristic equation is defined
as

p(A) = det(λI − A) = 0. (1.40)

The n roots of the character polynomial and equivalently, the n solutions to the
characteristic equation, λi ∀ i = 1, . . . , n, is defined as the eigenvalues of the matrix
A. The characteristic equation has always n solutions and the matrix A ∈ Rn×n has
always n eigenvalues. The eigenvalues can be real or complex. If the eigenvalues are
complex, then they will consists of complex conjugate pair, i.e., if λk = α + jβ is an
eigenvalue, then λk+1 = α − jβ will also be an eigenvalue. The matrix A is said to
have multiple eigenvalues if two or more of the eigenvalues are identical.
The spectrum of the eigenvalues of the matrix A is defined as all the eigenvalues,
i.e. the collection σ(A) := {λ1 , λ2 , · · · , λn } of al the eigenvalues is the spectrum of
A.
The spectral radius of the matrix A is defined from the eigenvalue with the
largest absolute value, i.e., ρ(A) = max |λi | ∀ i = 1, . . . , n.
The following theorem is useful in linear algebra and system theory

Teorem 1.4.1 (Cayley-Hamilton) Given a matrix A ∈ Rn×n and the character-

istic polynomial

p(A) = det(λI − A) = λn + pn λn−1 + · · · + p2 λ + p1 . (1.41)

The Cayley-Hamilton theorem states that the matrix A satisfies its own characteristic
polynomial, i.e., such that

An + pn An−1 + · · · + p2 A + p1 I = 0 (1.42)

where I is the identity matrix with the same dimension as A.

Proof The prof is stated only for the case in which the eigenvector matrix M of A
is non-singular. An eigenvalue decomposition of A gives that A = M ΛM −1 where Λ
is a diagonal matrix with the eigenvalues λi ∀ i = 1, . . . , n on the diagonal. Putting
this into (1.42) gives

M (Λn + pn Λn−1 + · · · + p2 Λ + p1 I)M −1 = 0. (1.43)

This gives n equations, i.e.,

λni + pn λn−1
i + · · · + p2 λi + p1 = 0 ∀ i = 1, . . . , n (1.44)
1.4 Eigenvalues and eigenvectors 7

which is of the same form as the characteristic equation (1.41). 4

The Cayley-Hamilton theorem can be used for, e.g.:

• to find an expression for the matrix inverse A−1 , i.e.

1 n−1
A−1 = − (A + pn An−2 + · · · + p2 I) (1.45)
p1

• to find an expression for the power An as a function of An−1 , · · · , A, i.e.

An = −(pn An−1 + · · · + p2 A + p1 I) (1.46)

• to find a way of computing the coefficients p1 , · · · , pn of the characteristic

polynomial by Krylovs method. This is presented in the next section.
• develop the controllability and observability matrices of an linear dynamical
system, i.e. from the matrix pairs (A, B) and (A, D).

1.4.1 Krylovs method used to find the coefficients of the charac-

teristic equation

We will in this section study a method which can be used to compute the coefficients,
p1 , · · · , pn , in the characteristic polynomial of a n × n matrix A. This method is
referred to as Krylovs method, Krylov (1931).
If we multiply Equation (1.42) from right with a vector b ∈ Rn , then a linear
system of equations can be defined as
p
z }| {

Cn
p1
z }| {  p2 
b Ab · · · An−1 b  .  = −An b. (1.47)
 
.
 . 
pn
This equation, Cn p = −An b, can be solved with respect to the vector p of coefficients.
We have that

p = −Cn−1 An b (1.48)

if the vector b is chosen in such a way that the matrix pair (A, b) is controllable,
i.e.. in this case in such a way that the controllability matrix Cn is invertible (non-
singular).
An arbitrarily random vector b is here usually sufficient. Note that the solution
p generally is independent of the choice of b as long as the matrix Cn is invertible.
Note also that (1.42) can directly be written as the linear system of equations

Ap = −vec(An ) (1.49)

where
2 ×n
vec(An−1 ) ∈ Rn
 
A= vec(I) vec(A) · · · . (1.50)
1.5 Similarity Transformations and eigenvectors 8

The solution is given by

p = −(AT A)−1 AT vec(An ) (1.51)

An advantage of this method is that we do not have to chose the vector b. A

disadvantage is that this last method is much more computing expense than the
first method in which an arbitrarily vector b is chosen.

1.5 Similarity Transformations and eigenvectors

Assume given a non-singular matrix

T ∈ Rn×n and a matrix A ∈ Rn×n . The matrix B defined by

B = T −1 AT (1.52)

is then said to be similar to A. In particular, the eigenvalues of B is identical to the

eigenvalues of A. The equation (1.52) is defined as a similarity transformation.
If we are putting the transformation matrix T equal to the eigenvector matrix,
M , of the matrix A, then we have that

Λ = M −1 AM (1.53)

where Λ is the eigenvalue matrix of the system (matrix A). The eigenvalue matrix
Λ os a diagonal matrix with the eigenvalues on the diagonal. This can equivalently
be written as

AM = M Λ (1.54)

where
 
M= m1 m2 · · · mn (1.55)

The columns, m1 , · · · , mn in the eigenvector matrix M is the eigenvectors corre-

sponding to the eigenvalues λ1 , · · · , λn .
Λ is a diagonal matrix with the eigenvalues λ1 , · · · , λn on the diagonal. Hence,
Equation (1.54) can then be written as n linear equations which can be used to
compute the eigenvectors, i.e.,

Am1 = λ1 m1
Am2 = λ2 m2
.. (1.56)
.
Amn = λn mn

If the matrix A and the eigenvalues λ1 , · · · , λn are known, then, the eigenvectors
and eigenvector matrix (1.55) can be found by solving the linear equations (1.56).
1.6 Time constant 9

1.6 Time constant

Consider the case in which a 1st order differential equation

ẋ = ax, (1.57)
with known initial value x(t0 ) is given. The time constant of the system is then
defined as
1
T =− . (1.58)
a
The solution to Equation (1.57) can then be written as
1
x(t) = ea(t−t0 ) x(t0 ) = e− T (t−t0 ) x(t0 ). (1.59)
We see that the solution x(t) at time instant t = t0 + T is given by
1
x(t0 + T ) = e− T (t0 +T −t0 ) x(t0 ) = e−1 x(t0 ) ≈ 0.37x(t0 ). (1.60)
I.e., the response have fallen 63% after t = T Time Constant time units. See also
Example 1.5 for illustration.

Example 1.2 (Time response of a 1st order system and the time constant T )

In connection with the Time Constant consider a transfer function model of a 1st
order system y(s) = hp (s)u(s) where the transfer function is
K
hp (s) = . (1.61)
Ts + 1
The corresponding continuous state space model description is
1 K
ẋ = − x + u, (1.62)
T T
where K is the steady state system gain and T is the Time Constant.
The solution of the state eq. (1.62) is then given by the general solution in eq.
(1.10), i.e.
Z t
t−t t−τ K
− T0
x(t) = e x(t0 ) + e− T u(τ )dτ, (1.63)
t0 T
where t0 is the initial time. For simplicity of illustration assume that a unit step
u = 1 ∀ t0 ≤ t is feed into the system, initial time t0 = 0, the system gain K = 1
and zero initial state x(t0 ) = 0. The solution of eq. (1.63) is then
1 t − t−τ 1
Z
t−τ t
x(t) = e T dτ = [T e− T ]t0 = 1 − e− T (1.64)
T t0 T
We observe that the final state at time t = ∞ is x(∞) = 1 and that the value of the
state at time equal to the Time Constant, i.e. t = T is
x(T ) = 1 − e−1 ≈ 0.63. (1.65)
I.e. the value x(T ) of the state after t = T time units is equal to (the common) 63%
of the final value of the state x(∞) = 1. see Figure 1.1 for illustration of the step
response.
1.6 Time constant 10

Unit step response of 1st order system with T=10 and K=1
1

0.9

0.8

0.7
1−e−1

0.6
x(t)

0.5

0.4

0.3

0.2

0.1

0
0 10 20 30 40 50 60
Time 0 ≤ t ≤ 6T
T

Figure 1.1: Time response of 1st order system ẋ = ax + bu where x(0) = 0, a = − T1 ,

b= K T , gain K = 1 and with Time Constant T = 10. Notice that after time equal
to the time constant T then the response have reached 1 − e−1 ≈ 0.63 (63 %) of the
final value.

Definition 1.1 (Time constants)

Given a linear dynamic system, ẋ = Ax + Bu, where x ∈ Rn is the state vector of
the system. The system matrix A has n eigenvalues given by
λi = λi (A) ∀ i = 1, . . . , n. (1.66)
If the eigenvalues are all real, distinct and have negative values (stable system), then
the system will have th n time constants given by
1
Ti = − i = 1, . . . , n. (1.67)
λi

Note also that the connection with the eigenvalues in a discrete time system
xk+1 = φx + δu, (1.68)
and the continuous equivalent
1
ẋ = − x + bu, (1.69)
T
then is given by
1
φ = e− T ∆t (1.70)
which gives that
∆t
T =− . (1.71)
ln φ
Methods in system identification can be used to identify discrete time models from
known input and output data of a system. Usually there are the parameters φ and
δ which are estimated (computed). The relationship (1.71) is therefore very useful
in order to find the time constant of the real time system.
1.7 The matrix exponent and the transition matrix 11

1.7 The matrix exponent and the transition matrix

We have earlier in this section shown that that the transition matrix are involved
in the exact solution of a linear time invariant dynamical system. Consider the
autonomous system

ẋ = Ax (1.72)

with known initial value x0 = x(t0 = 0). Then the solution is given by

x(t) = Φ(t)x(0) (1.73)

where the transition matrix Φ(t) is given by

Φ(t) = eAt . (1.74)

As we see, the problem of computing the transition matrix Φ(t), is the same problem
as computing the matrix exponent

F = eA . (1.75)

1.7.1 Computing the matrix exponent by diagonalisation

Let f (A) be an analytical matrix function of A which also should contain the eigen-
value spectrum of A. A more general formulation of the similarity transformation
given in (1.52) is then defined as

f (B) = T −1 f (A)T (1.76)

Assume now that we want to compute the matrix exponent eA . As we have shown
in Equation (1.53 the matrix A can be decomposed as

A = M ΛM −1 (1.77)

when the eigenvector matrix M is invertible. Using (1.77), (1.76) and f (A) = eA
gives

eA = M eΛ M −1 (1.78)

As we see, when the eigenvector matrix M and the eigenvalue matrix Λ of the matrix
A are known, then the matrix exponential eA can simply be computed from (1.78).
Equation (1.78) can be proved by starting with the autonomous system

ẋ = Ax (1.79)

with known initial state vector x(0). This system has the solution

x(t) = eAt x(0). (1.80)

Transforming (1.79) by using x = M z gives

ż = Λz (1.81)
1.7 The matrix exponent and the transition matrix 12

with initial state z(0) = M −1 x(0). The canonical (transformed) system (1.81) have
the solution

z(t) = eΛt z(0). (1.82)

Using the transformation x = M z and putting this into (1.82) gives

x(t) = M eΛt z(0) = M eΛt M −1 x(0). (1.83)

Comparing the two solutions (1.83) and (1.80) gives (1.78).

Note that in some circumstances there may be simpler to compute the transition
matrix or matrix exponential f (A) by solving the equation system

f (A)M = M f (Λ) (1.84)

because we in this case do not explicitly have to compute the matrix inverse M −1 .
For some problems M is not invertible. This may be the case for systems which
have multiple eigenvalues. We are referring to Parlet (1976) for a more detailed
description of matrix functions and computing methods.

1.7.2 Parlets method for computing the matrix exponent

It can be shown, Parlet (1976), that the matrix exponent F = eA and the system
matrix A commutes, i.e. the following is satisfied

F A = AF (1.85)

If the matrix A has a special structure, e.g., upper or lower triangular, then Equation
(1.7.2) can with advantage be used in order to compute the unknown elements in
the transition matrix.
Note that the matrix exponential of an upper triangular matrix
 
a11 a12
A= , (1.86)
0 a22

is given by

ea11 f12
 
F = eA = . (1.87)
0 ea22

The unknown coefficient f12 can then simply be found from equation ().

Example 1.3 (computing the transition matrix)

Given an autonomous system described by

ẋ = Ax, (1.88)

with the initial state x0 = x(0). The system matrix A is given by

 
λ1 α
A= . (1.89)
0 λ2
1.7 The matrix exponent and the transition matrix 13

We want to compute the transition matrix

F = eAt (1.90)

by Parlets method. First we find immediately that

 λt 
e 1 f
F = . (1.91)
0 eλ2 t

We now have to find the unknown constant f12 in the transition matrix. This can
be done from the equation system

AF = F A. (1.92)

This gives four equations but only one of them gives information of interest,

λ1 f12 + αeλ2 t = eλ1 t α + f12 λ2 . (1.93)

Solving with respect to f12 gives

e λ1 t − e λ2 t
f12 = α (1.94)
λ1 − λ2
As we see, this method can simply be used for system matrices which have a tri-
angular structure, and in which the eigenvalues are distinct and not identical to
zero.

1.7.3 Matrix exponential by series expansion

It can be shown that the matrix exponential F A can be expressed as an infinite

Taylor series
1
eA = I + A + A2 + · · · (1.95)
2
The transition matrix can be expressed in the same way, e.g.,
1
eAt = I + At + A2 t2 + · · · (1.96)
2
This is in general not a good method for computing the transition matrix, because it
will in general lead to numerical problems when computing powers of A like A9 , A10 ,
etc. especially when A contains small values. This is so due to the finite precision of
the computer. Note that the machine precision of a 32 bit computer is eps = 1/252 .
The series method is however very useful for computing the transition matrix of
many simple systems. This will be illustrated in the following example.

Example 1.4 (computing transition matrix)

Given an autonomous system described by

ẋ = Ax, (1.97)
1.8 Examples 14

where the initial state x0 = x(0) is given and the system matrix is given by
 
0 α
A= . (1.98)
0 0

The transition matrix for this system is simply found from the two first terms of
the Taylor series (1.96) because A is so called nil-potent, i.e., we have that A2 = 0,
A3 = 0 and so on. We have
 
1 αt
Φ(t) = I + At = . (1.99)
0 1

1.8 Examples

Example 1.5 (autonomous response and time constant)

Given an autonomous system

ẋ = ax (1.100)

where the initial state is x0 = x(t0 = 0) = 1 and the system parameter a = − T1

where the time constant is T = 5. The solution of this differential equation is
1 1
x(t) = e− T t x0 = e− 5 t . (1.101)

Let us now plot the solution in the time interval 0 ≤ t ≤ 25. Note that the state
will have approximately reached the steady state value after 4T (four times the time
constant). The solution is illustrated in Figure 1.2.
Autonom respons
1

0.9

0.8

0.7

0.6

0.5
x

0.4
e−1=0.37

0.3

0.2

0.1

0
0 5 10 15 20 25
Tid [s]
T=5

Figure 1.2: Time response of autonomous system ẋ = ax where x0 = 0 and a = − T1

and with Time Constant T = 5.
1.8 Examples 15

Autonom respons
1

0.9

0.8

0.7

0.6

0.5
x

0.4
e−1=0.37

0.3

0.2

0.1

0
0 5 10 15 20 25
Tid [s]
T=5

Figure 1.3: Time response of autonomous system ẋ = ax where x0 = 0 and a = − T1

and with Time Constant T = 5.

Example 1.6 (computation of matrix exponent)

Given the system matrix
 
−3 1
A= . (1.102)
2 −2

The eigenvalue matrix Λ and the corresponding eigenvector matrix can be shown to
be as follows
   
1 1 −1 0
M= , Λ= (1.103)
2 −1 0 −4

Find the matrix exponent F = eA .

We have the relation F = M eΛ M −1 , which is equivalent with

F M = M eΛ (1.104)

which gives

e−1 0
     
f11 f12 1 1 1 1
= (1.105)
f21 f22 2 −1 2 −1 0 e−4

From this we have the four equations

  −1
e−4
 
f11 + 2f12 f11 − f12 e
= (1.106)
f21 + 2f22 f21 − f22 2e−1 −e−4
1.8 Examples 16

Taking element 1, 1 minus element 1, 2 on the left hand side of Equation (1.106)
gives

3f12 = e−1 − e−4 (1.107)

Putting the expression for f12 into element 1, 2 on the left hand side gives f11 , i.e.,

f11 − f12 = e−4 (1.108)

which gives
1
f11 = (e−1 + 2e−4 ) (1.109)
3
Taking element 2, 1 minus element 2, 2 on the left hand side of (1.106) gives

3f22 = 2e−1 + e−4 (1.110)

Putting the expression for f22 into e.g., element 2, 2 gives f21 . This gives the final
result
 −1
e + 2e−4 e−1 − e−4

A 1
F =e = (1.111)
3 2e−1 − 2e−4 2e−1 + e−4

Note that the transition matrix could have been computed similarly, i.e.,
 −t
e + 2e−4t e−t − e−4t

At Λt −1 1
Φ(t) = e = M e M = (1.112)
3 2e−t − 2e−4t 2e−t + e−4t

Example 1.7 (computation of transition matrix for upper triangular system)

Consider given an autonomous system described by the matrix differential equation

ẋ = Ax, (1.113)

where the initial state x0 = x(0) is given and where the system matrix is given by
 
λ1 α
A= . (1.114)
0 λ2

The transition matrix Φ(t) = eAt will have the same upper triangular structure as
A and the diagonal elements in Φ(t) is simply e−λ1 t and e−λ2 , i.e.,
 λt 
e 1 f12
Φ(t) = . (1.115)
0 eλ2 t

The unknown element f12 can now simply be computed from Parlets method, i.e.,
we solve the equation

ΦAt = AtΦ (1.116)

or equivalent

ΦA = AΦ. (1.117)
1.8 Examples 17

This gives the equation

eλ1 t α + f12 λ2 = λ1 f12 + αeλ2 t . (1.118)

Solving for the remaining element f12 gives

eλ1 t − eλ2 t
f12 = α . (1.119)
λ1 − λ2
Note that this method only can be used when the system has distinct eigenvalues, i.e.
when λ1 6= λ2 .
However, it can be shown that in the limit when λ2 → λ1 = λ that

f12 = αteλ1 t (1.120)

Example 1.8 (Set of higher order ODE to set of first order ODE)
Consider a system described be the following couple of coupled differential equations

ÿ1 + k1 ẏ1 + k2 y1 = u1 + k3 u2
ẏ2 + k4 y2 + k3 ẏ1 = k6 u1

where u1 and u2 is defined as the control inputs and y1 and y2 is defined as the
measurements or outputs
We now define the outputs and if necessary the derivatives of the outputs as
states. Hence, define the states

x1 = y1 , x2 = ẏ1 , x 3 = y2 (1.121)

This gives the following set of 1st order differential equations for the states

ẋ1 = x2 (1.122)
ẋ2 = −k2 x1 − k1 x2 + u1 + k3 u2 (1.123)
ẋ3 = −k5 x2 − k4 x3 + k6 u1 (1.124)

and the following measurements (outputs) variables

y1 = x 1 (1.125)
y2 = x 3 (1.126)

The model is put on matrix (State Space) form as follows

      
ẋ1 0 1 0 x1 0 0  
 ẋ2  =  −k2 −k1 0   x2  +  1 k3  u1
u3
ẋ3 0 −k5 k4 x3 k6 0
 
    x1
y1 1 0 0 
= x2 
y2 0 0 1
x3
and finally in matrix form as follows

ẋ = Ax + Bu (1.127)
y = Dx (1.128)
1.9 Transfer function and transfer matrix models 18

1.9 Transfer function and transfer matrix models

Laplace transforming ẋ gives

L(ẋ(t)) = sx(s) − x(t = 0). (1.129)

Similarly, the Laplace transform of a time dependent variable x is defined as

L(x(t)) = x(s). (1.130)

Using (1.129) and the definition (1.130) in the state space model

ẋ = Ax + Bu, x(t = 0) = 0, (1.131)

y = Dx + Eu, (1.132)

gives the Laplace transformed model equivalent

x(s) = (sI − A)−1 Bu(s), (1.133)

y(s) = Dx(s) + Eu(s). (1.134)

We can now write (1.133) and (1.134) as a transfer matrix model

y(s) = H(s)u(s), (1.135)

where H(s) is the transfer matrix of the system

H(s) = D(sI − A)−1 B + E. (1.136)

For single-input and single-output systems then H(s) will be a scalar function of
the Laplace variable s. In this case we usually are using a small letter, i.e., we are
putting h(s) = H(s). Note also that we have included a direct influence from the
input u to the output y in the measurement (output) equation. This will be the
case in some circumstances. However, the matrix or parameter E is usually zero in
control systems, in particular E = 0 in standard feedback systems.
Note also that when the eigenvalue decomposition A = M ΛM −1 exists then we
have that the transfer matrix can be expressed and computed as follows

H(s) = D(sI − A)−1 B + E (1.137)

−1 −1
= DM (sI − Λ) M B + E. (1.138)

Finally note the following important relationship. The properties of a time de-
pendent state space model when t → ∞, i.e. the steady state properties, can be
analyzed in a transfer function Laplacian model by putting s = 0. The transient
behavior when t = 0 is analyzed by letting s → ∞.
For Single Input and Single Output (SISO) systems we often write the plant
model as
ρ(s)
hp (s) = D(sI − A)−1 B + E = , (1.139)
π(s)
where ρ(s) is the zero polynomial and π(s) is the pole polynomial
1.10 Poles and zeroes 19

1.10 Poles and zeroes

For a Single Input Single Output ( SISO) system we may write the transfer function
model from the control input, u, to the output measurement, y,as follows

y(s) = h(s)u(s), (1.140)

where the transfer function may be written as

ρ(s)
h(s) = , (1.141)
π(s)
where ρ(s) is the zero polynomial and π(s) is the pole polynomial.
The poles of the system is the roots of the denominator pole polynomial, i.e.,

π(s) = 0. (1.142)

The zeroes of the system is given by the roots of the numerator zero polynomial, i.e.

ρ(s) = 0. (1.143)

Remark that the n poles is the n roots of the pole polynomial π(s) = 0 and
that these n poles is identical to the n eigenvalues of the A matrix and that the
pole polynomial also may be deduced from the system matrix A of an equivalent
observable and controllable state space model, i.e.,

π(s) = det(sI − A) = 0. (1.144)

For the poles to be equal to the eigenvalues of the A matrix the linear state space
model ẋ = Ax + Bu and y = Dx have to be a minimal realization, i.e., the model is
both controllable and observable.

1.11 Time Delay

A time delay (or dead time) in a system may in continuous time domain be described
as follows. Suppose a variable y(t) is equal to a variable x(t) delayed τ ≥ 0 time
units. Then we may write

y(t) = x(t − τ ) (1.145)

where y(t) = 0 when 0 ≤ t < τ and y(t) = x(t − τ ) when time t ≥ τ .

The Laplace plane model equivalent of the time delay is

y(s) = e−τ s x(s) (1.146)

Notice that the exponential e−τ s is an irrational function and that we can not do
algebra with such functions, and that rational approximations to the exact delay
have to be used if a delay should be used in algebraic calculations.
Numerous approximations exist and may be derived from the series approxima-
tion of the exponential.
1.12 Linearization 20

1.12 Linearization

In many cases the starting point of a control problem or model analysis problem is
a non-linear model of the form

ẋ = f (x, u), (1.147)

y = g(x, u). (1.148)

Here x ∈ Rn is the state vector, u ∈ Rr is the control input vector and y ∈ Rm is

the output or measurements vector. The functions f (·, ·) ∈ Rn and g(·, ·) ∈ Rm may
be non-linear smooth functions of x and u. Note also that the initial state is x(t0 )
which should be given ore known before the state space model can be simulated in
time.
In this case it may be of interest to derive a linear model approximation to
(1.147) and (1.148).
The two first (linear terms) of a Taylor series expansion of the right hand side
of (1.147) around the points x0 and u0 gives
 
∂f  ∂f 
f (x, u) ≈ f (x0 , u0 ) + ∂x T  (x − x 0 ) + ∂uT
 (u − u0 ). (1.149)
x0 ,u0 x0 ,u0

Define the deviation variables

∆x = x − x0 , (1.150)
∆u = u − u0 . (1.151)

Also define the matrices


∂f
A= ∈ Rn×n (1.152)

∂xT

x0 ,u0

which also is named the Jacobian matrix. Similarly, define


∂f 
B = ∂u T  ∈ Rn×r (1.153)
x0 ,u0

Putting (1.149), (1.150) and (1.151) into the state equation (1.147) gives the
linearized state equation model
˙ = A∆x + B∆u + v,
∆x (1.154)

where

v = f (x0 , u0 ) − ẋ0 . (1.155)

Usually the points x0 and u0 is constant steady state values such that

x˙0 = f (x0 , u0 ) = 0. (1.156)

Hence, a linearized state equation is given by

˙ = A∆x + B∆u.
∆x (1.157)
1.12 Linearization 21

Similarly the output equation (1.148) can be linearized by approximating the

right hand side by the first two terms of a Taylor series expansion, i.e.,
 
∂g  ∂g 
y ≈ g(x0 , u0 ) + ∂x T  (x − x 0 ) + ∂uT
 (u − u0 ). (1.158)
x0 ,u0 x0 ,u0

Now defining

y0 = g(x0 , u0 ) (1.159)
∆y = y − y0 (1.160)

∂g
D = (1.161)

∂xT

x ,u
 0 0
∂g
E = (1.162)

∂uT

x0 ,u0

gives the linearized output equation

∆y = D∆x + E∆u. (1.163)

Usually the deviation variables are defined as

x := x − x0 , (1.164)
u := u − u0 . (1.165)

Hence, a linear or linearized state space model, given by (1.157) and (1.163), is
usually written as follows.

ẋ = Ax + Bu, (1.166)
y = Dx + Eu. (1.167)

One should therefore note that the variables in a linearized model may be deviation
variables, but this is not always the case. One should also note that only linear
models can be transformed to Laplace plane models. Note also that the initial state
in the linearized model is given by ∆x(t0 ) = x(t0 ) − x0 .

Example 1.9 (Linearization of a pendulum model)

An non-linear model for a pendulum can be written as the following second order
differential equation, i.e.,
b g
θ̈ + 2
θ̇ + sin(θ) = 0, (1.168)
mr r
where θ is the angular position (deviation of the pendulum from the vertical line, i.e.
from the steady state position). m = 8 is the mass of the pendulum, r = 5 is the
length of the pendulum arm, b = 10 is a friction coefficient in the base point and
g = 9.81m/s2 is the acceleration of gravity constant.
The second order model can be written as a set of 1st order differential equations
by defining the states

x1 = θ angular position (1.169)

x2 = θ̇ angular velocity (1.170)
1.12 Linearization 22

from this definitions we have that ẋ1 = θ̇ = x2 which gives the state space model

ẋ1 = x2 , (1.171)
g b
ẋ2 = − sin(x1 ) − x2 , (1.172)
r mr2
which is equivalent to a non-linear model

ẋ = f (x) (1.173)

with the vector function

   
f1 x2
f (x) = = (1.174)
f2 − gr sin(x1 ) − b
x
mr2 2

Linearizing around the steady state solution x1 = 0 and x2 = 0 gives

ẋ = Ax, (1.175)

where the Jacobian is given by

" #
∂f1 ∂f1    
0 1 0 1

∂f ∂x ∂x
A = ∂xT  = ∂f2 ∂f2 1 2 = = (1.176)

g
0 ∂x1 ∂x2
b
− r cos(x1 ) − mr2 0
− gr − mrb 2
0

Putting into the numerical values we obtain

 
0 1
A= (1.177)
−1.962 −0.050

Note that the linearized model could have been obtained more directly by using that
sin(x1 ) ≈ x1 for small angles x1 .

Example 1.10 (Simulation of a non-linear pendulum model)

The nonlinear state space pendulum model

ẋ1 = x2 , (1.178)
g b
ẋ2 = − sin(x1 ) − x2 , (1.179)
r mr2
with g = 9.81, r = 5, m = 8 and b = 10 can be simply simulated in MATLAB by
using an ODE solver, e.g.,

>> sol=ode15s(@ fx_pendel, 0:0.1:50,[1;0]);

>> plot(sol.x,sol.y)

Here sol is an object where sol.x is the time axis and sol.y is the states. The ode15s
function simulate the pendulum model over the time horizon t = t0 : h : tf , i.e.
from the initial time t0 = 0 and to the final time tf = 50 with step length (sampling
interval) ∆t = 0.1. Try it! The file fx pendel is an m-file function given in the
following
1.12 Linearization 23

function fx=fx_pendel(t,x)
% fx_pendel
% fx=fx_pendel(t,x)
% Modell av pendel.

m=8; g=9.81; b=10; r=5;

fx=zeros(2,1);
fx(1)=x(2);
fx(2)=-b*x(2)/(m*r^2)-g*sin(x(1))/r;

1.12.1 Calculating the Jacobian matrix numerically

In some circumstances the Jacobian matrix A = dfdx(x) dg(x)

T , and matrices dxT , may be
hard and difficult to calculate analytically. In these cases it may be of grate interest
to calculate the derivatives numerically.
Notice that we will in this section assume that the vector f (x) ∈ Rm and vector
x ∈ Rn . The Jacobian matrix is in this case defined as
 
df1 (x) df1 (x)
dx 1
· · · dx n
df (x) h df (x) df (x) df (x)
i 
.. .. .. 
A= T
= dx · · · dx · · · dx
= 
. . .
 (1.180)
dx 1 i n 
dfm (x) dfm (x)

dx1 · · · dxn

This may for instance be done by using an approximation to the derivative, e.g.
as the simple approximation (or similar)

df (xi ) f (xi + h) − f (xi )

≈ ∀ i = 1, 2, ..., n (1.181)
dxi h
for some small number h. Hence, for each elements in vector f (x) ∈ Rm , say for
generality each of the m elements in in vector f (x), we loop through all n elements
in x using this approximation. This means that we for each element 1 ≤ j ≤ m in
f (x) we calculate

dfj (x) fj (xi + h) − fj (xi )

aji = ≈ ∀ i = 1, 2, ..., n (1.182)
dxi h
However, this may in MATLAB be done effectively by calculating one column of the
Jacobian matrix A at at time, hence

f (xi + h) − f (xi )
A(:, i) ≈ ∀ i = 1, 2, ..., n (1.183)
h
where A(:, i) is MATLAB notation for the entire i − th column of the A matrix.
The cost of this procedure for numerically calculating the Jacobian matrix are n + 1
function evaluations. This numerical procedure to calculate the Jacobian matrix is
implemented in the following MATLAB jacobi.m function.
1.12 Linearization 24

function A=jacobi(fx_fil,t,x)
%JACOBI Function to calculate the Jacobian matrix numerically
% A=jacobi(’fx_file’,t,x)
% PURPOSE
% Function to calculate the Jacobian matrix A=df/dx of the nonlinear
% function fx=f(t,x), i.e. the linearization of a possible non-linear
% function fx=f(t,x) around values t and x.
% ON INPUT
% fx_file - m-file to define the function fx=f(t,x) with
% syntax: function fx=fx_file(t,x) where fx is,
% say m dimensional
% t - time instant.
% x - column vector with same dimension as in the function
% fx=f(t,x), say of dimension n
% ON OUTPUT
% A - The jacobian matrix A=df/dx of dimension (m,n)
h=1e-5;
n=length(x);
xh=x;
fx=feval(fx_fil,t,x);
m=length(fx);
A=zeros(m,n);
for i=1:n
xh(i)=x(i)+h;
fxh=feval(fx_fil,t,xh);
xh(i)=xh(i)-h;
A(:,i)=(fxh-fx)/h;
end
% END JACOBI

In order to illustrate how MATLAB may be used to numerically calculating the

Jacobian matrix we use the pendulum example in Example 1.10 which gives the
same result as the analytic expression in Example 1.9 and Equation 1.177.

>> A=jacobi(’fx_pendel’,0,[0;0])

A =

0 1.0000
-1.9620 -0.0500

>>
1.13 Stability of linear systems 25

1.13 Stability of linear systems

1.13.1 Stability of continuous time linear systems

Given a linear continuous time system

ẋ = Ax + Bu, (1.184)
y = Dx. (1.185)

• If the eigenvalues of A is complex, then

√ they are existing in complex conjugate
pairs, i.e. λi = α ± βj where j = −1 is the imaginary number. Hence, one
pair of complex conjugate eigenvalues results in the eigenvalues λi = α + βj
and λi+1 = α − βj. This is a property of the eigenvalues of a real matrix A.
• The system is stable if the real part of the eigenvalues, λi (A) ∀ i = 1, 2, . . . , n,
are negative. The eigenvalues of a stable system are located in the left part of
the complex plane.
• If one ore some of the eigenvalues are located on the imaginary axis we say
that the system is marginally stable.
• If one (or more) of the eigenvalues are zero we say that we have an integrator
(or have several integrators) in the system.
• If one ore more of the eigenvalues are located in the right half plane, i.e. with
positive real parts, then the system is unstable.

The transfer function/matrix model equivalent to the state space model is y =

h(s)u where the transfer function/matrix is given by
h(s) = D(sI − A)−1 B. (1.186)
The pole polynomial of a transfer function model is given by det(sI − A) and its
roots, the solution of the pole polynomial det(sI − A) = 0 is the system poles.
The poles of a transfer function model and the eigenvalues of the A matrix
coincide for controllable and observable systems, i.e. the poles are identical to the
observable and controllable eigenvalues in the state space model equivalent. This
also implies that possible unobservable and/or uncontrollable states/eigenvalues in
a state space model can not be computed from the transfer function model.

1.13.2 Stability of discrete time linear systems

Given a linear discrete time system

xk+1 = Axk + Buk , (1.187)
yk = Dxk . (1.188)

• The discrete time linear system described by Eq. (1.187) is stable if the eigen-
values of the system matrix A is located inside the unit circle in the complex
plane. This is equivalent to check wether the eigenvalues has magnitude less
than one, i.e., |λi | < 1 ∀ i = 1, . . . , n.
1.14 State Controllability 26

• If one (or more) of the eigenvalues have magnitude equal to one, we have an
integrator (or integrators) in the system (if the eigenvalue are real).

Example 1.11 Given a linear continuous time system ẋ = Ax + Bu and y = Dx

with model matrices
   
−α −β 1  
A= , B= , D= 1 0 . (1.189)
β −α 0

The eigenvalues of the system matrices are simply given by a complex conjugate
pair of eigenvalues, i.e., λ1,2 = −α ± βi. This system is then stable for all α > 0.
The eigenvalues may simply be found as the elements on the diagonal of the A
matrix for this example, but in general as the roots of the characteristic equation
det(λI − A) = λ2 + 2αλ + α2 + β 2 = 0.
The transfer function model is y = h(s)u with
s+α
h(s) = D(sI − A)−1 B = (1.190)
s2 + 2αs + α2 + β 2

1.14 State Controllability

The question of how we can (and if there is possible to) find a suitable control
input u(t) that will take the system from an initial state x(t0 ) to any desired final
state x(t1 ) in a finite (often very small) time, is answered by the theory of state
controllability.

Definition 1.2 (State Controllability)

A system described by a state space model ẋ = Ax + Bu with initial state x(t0 ) given
is controllable if there, for an arbitrarily finite time t1 > t0 exist a control function
u(t) defined over the time interval t0 ≤ t ≤ t1 , such that the final state, x(t1 ), can
be arbitrarily specified.

There exist a few algebraic definitions which can be used for the analysis of state
controllability. Such a theorem is defined via the so called controllability matrix.

Teorem 1.14.1 (Controllability matrix)

A system described by a state space model ẋ = Ax + Bu is state controllable if the
controllability matrix

Cn = B AB A2 B · · · An−1 B
 
(1.191)

has full rank, i.e.,

rank(Cn ) = n. (1.192)

Note that for single input systems, i.e., r = 1, then Cn ∈ Rn×n which implies that
Cn should be invertible and that det(Cn ) 6= 0 in order for the system to be state
controllable.
1.15 State Observability 27

Remark 1.1 (Diagonal form and controllability)

Consider a state space model ẋ = Ax + Bu and y = Dx + Eu and its diagonal
canonical form

ż = Λz + M −1 Bu (1.193)
y = DM z + Eu (1.194)

where Λ is a diagonal
 matrix with the eigenvalues λi ∀ i = 1, . . . n of A on the
diagonal and M = m1 · · · mn is the corresponding eigenvector matrix. Note
the relationship Ami = λi mi between the ith eigenvalue, λi , and the ith eigenvector,
mi .
The system is controllable if no rows in the matrix M −1 B is identically equal to
zero.
One should also note that there also is a dual phenomena, observability. The
system is observable if no columns in the matrix DM is identically equal to zero.

1.15 State Observability

Definition 1.3 (State Observability)

A system described by a state space model ẋ = Ax + Bu and y = Dx with initial
state x(t0 ) is observable if there, from knowledge of known inputs, u(t), and outputs,
y(t), over a time interval t0 ≤ t ≤ t1 , is possible to compute the (initial) state vector,
x(t0 ).

Teorem 1.15.1 (Observability matrix)

Define the observability matrix
 
D
 DA 
 
 DA2 
Oi =   ∈ Rmi×n , (1.195)
 .. 
 . 
DAi−1

The pair (D, A) is observable if and only if the observability matrix Oi for i = n has
rank n, i.e. rank(On ) = n.
If rank(D) = rD ≥ 1 and n − rD + 1 > 0, then we have that the pair (D, A) is
observable if and only if the reduced observability matrix On−rD +1 have rank n. For
single output systems we use i = n and On ∈ n × n.
4

Example 1.12 (Observability of continuous autonomous system) Consider

a single output autonomous system

ẋ = Ax, (1.196)
y = Dx (1.197)
1.15 State Observability 28

from this we have that

y = Dx (1.198)
ẏ = DAx (1.199)
2
ÿ = DA x (1.200)
..
. (1.201)
(n−1)
y = DAn−1 x (1.202)
(n−1) (n−1)
dn−1 y
where y denotes the n − 1th derivative of y, i.e., y =
dtn−1
. From these n
equations we define the following matrix equation

y0|n = On x (1.203)

where
 
y

 ẏ 

 ÿ 
y0|n =
 ..

 (1.204)

 . 

(n−1)
y

and where On ∈ Rn×n observability matrix as defined in the above definition, Equa-
tion (1.195). If the observability matrix, On , is non-singular then we can compute
the state vector x(t) as

x = On−1 y0|n (1.205)

An n × n matrix On is non-singular if rank(On ) = n.

Example 1.13 (Observability of discrete autonomous system) Consider a sin-

gle output autonomous system

xk+1 = Axk , (1.206)

yk = Dxk (1.207)

from this we have that

yk = Dxk (1.208)
yk+1 = DAxk (1.209)
2
yk+2 = DA xk (1.210)
..
. (1.211)
n−1
yk+n−1 = DA x (1.212)

From these n equations we define the following matrix equation

yk|n = On xk (1.213)
1.15 State Observability 29

where
 
yk
 yk+1 
 
yk|n =  yk+2 (1.214)
 

 .. 
 . 
yk+n−1

and where On ∈ Rn×n observability matrix as defined in the above definition, Equa-
tion (1.195). If the observability matrix, On , is non-singular then we can compute
the state vector xk as

xk = On−1 yk|n (1.215)

An n × n matrix On is non-singular if rank(On ) = n.

Chapter 2

Canonical forms

2.1 Introduction

A state space model realizations (A, B, D) can be represented in an infinite number

of realizations (coordinate systems). Let x be the state vector for the realization
(A, B, D). Then, the transformed state z = T −1 x will be the state of the realization
(T −1 AT , T −1 B, DT ). Hence, there exists an infinite number of such non-singular
transformation matrices T .
The number of parameters in the model (A, B, D) are n2 + nr + mn. However,
the number of free independent parameters are much less. It can be shown that the
minimum number of free and independent parameters are nm+nr. This means that
there may exist a transformation matrix, T , that transforms a realization (A, B, D)
to a canonical form with special structure and with a minimum number of free and
independent parameters. In such canonical form models there also is a number
of fixed parameters which is equal to zero and one, in addition to the nm + nr
independent system parameters. Hence, canonical forms are realizations of a given
realization (A, B, D) with a special structure and with a minimal number of free
parameters and as many ones and zeros as possibile.
Canonical form state space realizations is important in, omong others, the fol-
lowing problems:

• Analysis of the characteristic properties of the system and the model.

Computations of eigenvalues and thereby analysis of stability and time con-
stants. The characteristic equation which often is the starting point for com-
putations of eigenvalues are easily obtained from the canonical model realiza-
tion. It can also be analysis of controllability and reachability, observability
and constructibility.

• Controller design. A pole placement controller can be easily designed if the

model is on a so called controller canonical form.

• System identification. If we want to fit a model to real system input and output
data, then it may be more easy to find the best model parameters such that
the model output is as similar to the real output data, if we have as few free
2.2 Controller canonical form 31

parameters as possible. Canonical form models is a central step in classical

system identification methods such as e.g. the prediction error method. It has
been a tremendous research activity on canonical forms.

• Choice of state variables. Canonical form models have a close relationship

with the choice of state variables. As an example, we can transform a given
physical state variable, x, to a so called controllable canonical form by using,
z = Cn−1 x where Cn is the controllability matrix of the system. z is here
representing the transformed state space.

We are in the following not considering diagonal and block diagonal forms, i.e.,
so called eigenvalue canonical form. The most common canonical forms in system
theory are as follows:

1. Controller canonical form. (No: regulator kanonisk form)

2. Controllability canonical form. (No: styrbarhets kanonisk form)

3. Observer canonical form. (No: estimator kanonisk form)

4. Observability canonical form. (No: Observerbarhets kanonisk form)

2.2 Controller canonical form

2.2.1 From transfer function to controller canonical form

Only single input single output (SISO) systems are considered in this section. In
order to illustrate the controller canonical form consider a transfer function model
with n = 3 states, i.e.,

b2 s2 + b1 s + b0
G(s) = (2.1)
s3 + a2 s2 + a1 s + a0
y(s) = G(s)u(s) (2.2)

The following 3rd order state space model is equivalent.

   
0 1 0 0
ẋ =  0 0 1 x +  0 u (2.3)
−a0 −a1 −a2 1
 
y = b0 b1 b2 x (2.4)

In order to show this we can write down a block diagram for the system based on the
two model representations and show that the resulting block diagram is the same.
Another method for proving the equivalence is to compute the transfer function
model from the state space model.
The above state space model realization is not unique. For instance rows and
columns in the model matrices can be interchanged. This can be easily done by a so
2.2 Controller canonical form 32

called permutation matrix. Consider the following permutation matrix, P , defined

by
 
0 0 1
P = 0 1 0  (2.5)
1 0 0

Note the following properties for permutation matrices, i.e., P = P −1 = P T .

The resulting state space model obtained from the transformation z = P x is
given by
   
−a2 −a1 −a0 1
ż =  1 0 0 z +  0 u (2.6)
0 1 0 0
 
y = b2 b1 b0 z (2.7)

It should be noted that it this last state space form which is referred to as controller
canonical form.
Those two canonical forms, i.e., the formulation given by (2.54) and (2.4) and the
formulation given by (2.12) and (2.7) are essentially identical state space canonical
forms.

Example 2.1 (2nd order system)

Given a SISO process described by a 2nd order transfer function model

b1 s + b0
G(s) = (2.8)
s2
+ a1 s + a0
y(s) = G(s)u(s) (2.9)

An equivalent 2nd order linear state space model is of the form

   
0 1 0
ẋ = x+ u (2.10)
−a0 −a1 1
 
y = b0 b1 x (2.11)

 by using the transformation z = P x with permutation matrix P =

and equivalently

0 1
P −1 = we obtain the controller canonical form
1 0
   
−a1 −a0 1
ż = z+ u (2.12)
1 0 0
 
y = b1 b0 z (2.13)

The simplest is to prove this by taking the Laplace transformation of the state space
model Eq. (2.11).
2.3 Controllability canonical form 33

2.2.2 From state space form to controller canonical form

Single input multiple output (SIMO) systems

In the following a method for the transformation of a SIMO controllable realization

given by (A, B, D) to so called controller canonical form is presented.
From the system matrix, A, we find the characteristic polynomial
det(sI − A) = sn + a1 sn−1 + · · · an−1 s + an
from the coefficients a1 , · · · , an−1 in the characteristic polynomial we obtain the
following upper triangular matrix with ones on the main diagonal.
 
1 a1 · · · an−1
.. 
 0 1 ...

. 
M = . 
..
 (2.14)
 .. .

a1 
0 0 ··· 1
The matrix M has 1 on the main diagonal, a1 on the diagonal above, and so on.
The matrix, M is referred to as a Toeplitz matrix because of its structure. Matrices
which elements are constant along its diagonals are called Toeplitz matrices.
Consider now the transformation matrix
T = Cn M (2.15)
where Cn is the controllability matrix for the pair (A, B) given by
Cn = B AB · · · An−1 B
 
(2.16)
The resulting state space model obtained by the transformation
z = T −1 x (2.17)
is on so called controller canonical form. The structure of the model will be as
illustrated in (2.12) and (2.7).
The system output matrix DT in the regulator canonical form will in general be
a full matrix, i.e., with parameters different from zeros and ones. It will be mn free
parameters in the output matrix DT and n free parameters in the controller canon-
ical form system matrix T −1 AT . The input matrix T −1 B will contain only zeros
and ones. The number n of free parameters in T −1 AT is identical to the coefficients
a1 , · · · , an−1 , an in the characteristic polynomial. It is a number of mn + n free and
independent parameters in a controller canonical state space model realization.

2.3 Controllability canonical form

Single input systems

If the system is controllable, then a non singular trnaformation matrix can be ob-
tained directly from the controllability matrix. Define the transformation matrix
T = Cn (2.18)
2.4 Observer canonical form 34

where Cn is the controllability matrix for the pair (A, B).

The resulting transformed state space model by using z = T −1 x will have the
following structure
Aco Bco
z}| { z }| {
0 0 −a0 1
ż =  1 0 −a1  z +  0  u
0 1 −a2 0
y = DT z (2.19)

where Aco = T −1 AT and Bco = T −1 B. This is referred to as controllable canonical

form. The system matrix Aco is said to be on controllable canonical form. Note that
the coefficients of the characteristic polynomial det(λI − A) can be obtained from
the right column in Aco .

Multiple input systems

The method described above will often also work for systems with multiple inputs.
Note that the controllability matrix, Cn , is an (n × nr) matrix. A transformation
matrix T can often be taken directly as the n first columns in Cn . However, a
problem which may occur is that the n first columns is linearly dependent and
that the transformation matrix T thereby becomes singular. However, when the
system is controllable then we can obtain n linearly independent columns from the
controllability matrix Cn ∈ Rn×nr , and take the transformation matrix T as those
columns.

Remark 2.1 It is worth noticing that for SISO systems, a controllable canonical
form state space model can be obtained by first constructing the extended controlla-
bility matrix

Cn+1 = B AB · · · An−1 B An B
 
(2.20)

and thereby in MATLAB notation

Aco = (Cn+1 (:, 1 : n))−1 Cn+1 (:, 2 : n + 1) (2.21)

Hence, the system matrix Aco is on controllable canonical form.

2.4 Observer canonical form

2.4.1 From transfer function to observer canonical form

Given the transfer function model as presented in the above Section, 2.2.1, i.e.,

b2 s2 + b1 s + b0
G(s) = (2.22)
s3 + a2 s2 + a1 s + a0
y(s) = G(s)u(s)
2.4 Observer canonical form 35

A state space model on observer canonical form can be developed from the transfer
function model as presented in the following. Equation (2.22) can be written as
follows
y(s) b2 s−1 + b1 s−2 + b0 s−3
=
u(s) 1 + a2 s−1 + a1 s−2 + a0 s−3

by multiplying the denominator and nominator in Equation (2.22) with s−3 . This
gives

y(s) = (b2 u(s) − a2 y(s))s−1 + (b1 u(s) − a1 y(s))s−2 + (b0 u(s) − a0 y(s))s−3 (2.23)

Let us make the following definitions

def
sx1 (s) = b2 u(s) − a2 y(s) + x2 (s) (2.24)
def
sx2 (s) = b1 u(s) − a1 y(s) + x3 (s) (2.25)
def
sx3 (s) = b0 u(s) − a0 y(s) (2.26)
def
y(s) = x1 (s) (2.27)

Putting the definitions, (2.24)-(2.26), into the expression given by Equation (2.23)
gives that y(s) = x1 (s), as is exactly the same as the definition, Equation (2.27).
This means that the model given by the definitions, Equations (2.24)-(2.26), is an
equivalent representation of the original transfer function model, Equation (2.22).
Transforming the model given by Equations (2.24)-(2.27) to the state space (inverse
Laplace transformation) gives

ẋ1 = −a2 x1 + x2 + b2 u
ẋ2 = −a1 x1 + x3 + b1 u
ẋ3 = −a0 x1 + b0 u
y = x1

The last state space model can simply be written on so called observer canonical
form, Hence we have
   
−a2 1 0 b2
ẋ =  −a1 0 1  x +  b1  u
−a0 0 0 b0
 
y = 1 0 0 x (2.28)

Example 2.2 (Sinus)

Given a process described by a sinusoid movement. This process may be e.g. a
disturbance, like waves and temperatures. Let

y(t) = sin(ωt) (2.29)

A sinusoid disturbance is common in many processes. This example is therefore of

importance when modeling disturbances. We want to write a state space description
2.4 Observer canonical form 36

of the sinusoid process. The sinusoid time domain description of the process can in
the Laplace plane be written as follows, see e.g. (se Edgar et al (1989), s. 47)
ω
y(s) = u(s)
ω 2 + s2
where u(s) = 1, is an unit impulse at time t = t0 , e.g. at time t0 = 0. This Laplace
formulation can be written as

s2 y(s) = ωu(s) − ω 2 y(s)

| {z }
sx2
| {z }
s 2 x1

here, we define

s2 y(s) = s2 x1 ⇒ y = x1
sx2 = ωu(s) − ω 2 y(s) ⇒ ẋ2 = ωu(t) − ω 2 y(t)
s2 x1 = sx2 ⇒ ẋ1 = x2

This gives the state space model

ẋ A x B
z }| {
 z
 }| { z }| { z }| {
ẋ1 0 1 x1 0
= + u(t) (2.30)
ẋ2 −ω 2 0 x2 ω
 
  x1
y = 1 0 (2.31)
| {z } x2
D

where u(t) is a unit impulse given by


1 for t = t0
u(t) = (2.32)
0 for al t > t0

Example 2.3 (Shifted sinus)

Given a process described by a sinusoid wave

y(t) = sin(ωt + φ). (2.33)

The corresponding Laplace transformation is (se e.g. Edgar et al (1989), s. 47)

s sin(φ) + ω cos(φ)
y(s) = u(s)
s2 + ω 2
Using (2.11) gives an equivalent state space model
   
0 1 0
ẋ = x+ u (2.34)
−ω 2 0 1
 
y = ω cos(φ) sin(φ) x (2.35)

where u(t) is a unit impulse. In this example we have used the results in Example
2.1.
2.5 Observability canonical form 37

Example 2.4 (Observability canonical form: 2nd order system)

Given a 2nd order system described by the transfer function model
b1 s + b0
y= u. (2.36)
s2 + a1 s + a0
Multiply eq. (2.36 with the de-numerator on the right hand side and write
s(s + a1 )y + a0 y = b1 su + b0 u (2.37)
Collect all terms propertional with s on the left hand side and write
x2
z }| {
s (sy + a1 y − b1 u) y + a0 y = b0 u (2.38)

Simply define y = x1 and

x2 = sy + a1 y − b1 u (2.39)
This gives from eq. (2.39)
ẋ1 = −a1 x1 + x2 + b1 u (2.40)
and from eq. (2.38) we obtain
ẋ2 = −a0 x1 + b0 u (2.41)
Eqs. (2.40) and (2.41) gives the linear state space model equivalent
ẋ A x B
z }| {
 z
 }| { z }| { z }| {
ẋ1 −a1 1 x1 b1
= + u (2.42)
ẋ2 −a0 0 x2 b0
 
  x1
y = 1 0 . (2.43)
| {z } x2
D

Hence, eq. (2.43) is a 2nd order (n = 2 states) state space model on observer
canonical form.

2.5 Observability canonical form

The observability canonical form is commonly used in system identification, i.e.,

how to find the model parameters such that the model fits the real system data.
The name have its origin because a transformation matrix, T , is constructed from
n linearly independent rows in the observability matrix, On
Given a realization (A, B, D). The observability matrix for the pair (A, D) is
defined by
 
D
 DA 
 
2
OL =  DA (2.44)
 

 .. 
 . 
DAL−1
2.5 Observability canonical form 38

OL is denoted an extended observability matrix when L > n and as an observability

matrix when L = n. An algorithm for transforming an observable state space model
(A, B, D) to observable canonical form is as follows.

Algorithm 2.5.1 (observability canonical form)

Step 1: Construct a non singular transformation matrix, T , from n linearly depen-
dent rows in the observability matrix OL where L ≥ n. We can chose

T = OL (1 : n, :)

when this choice is non singular. This will always hold for SISO observable systems.
Step 2: An observability canonical form state space model is then given by

Ac = (OcT Oc )−1 OcT OL AT −1 (2.45)

Bc = (OcT Oc )−1 OcT OL B (2.46)
−1
Dc = DT (2.47)

where

Oc = OL T −1 (2.48)

Note that Oc = I for single output systems. This is not the case for multiple output
systems.

This Algorithm 2.5.1 works for multiple output and multiple input (MIMO)
systems. The critical step is step 1. For systems with multiple outputs it is not sure
that the n first rows in OL is linearly independent. In such cases one have to search
through OL in order to obtain n linearly dependent rows, and use this rows as a
basis for T .

Example 2.5 (From transfer function model to OCF)

To illustrate the observability canonical form (OCF) we here use a 2nd order system
described with the transfer function model
b0
y= u (2.49)
s2 + a1 s + a0
From (2.49) we express with definitions

s2 y = −a0 y − a1 sy + b0 u (2.50)
| {z }
sx2
| {z }
s 2 x1

Hence, from (2.50) we have the following definitions

s2 x1 = sx2 ⇒ ẋ1 = x2
sx2 = −a0 y − a1 sy + b0 u ⇒ ẋ2 = −a0 x1 − a1 x2
s2 y = s2 x1 ⇒ y = x1
2.6 Duality between canonical forms 39

From this we find the following 2nd order state space model on observability canonical
form
   
0 1 0
ẋ = x+ u (2.51)
−a0 −a1 b0
 
y = 1 0 x (2.52)
 
x1
where x = is the state vector.
x2

Example 2.6 (Transfer function model and OCF)

Consider a system described with the following transfer function model
b0 + sb1
y= u (2.53)
s2 + a1 s + a0
The equivalent observability canonical form state space model is
   
0 1 b1
ẋ = x+ u (2.54)
−a0 −a1 b0
 
y = 1 0 x (2.55)

The simplest way to prove this is to go from the state space form (2.55) to the
transfer function model (2.53).

2.6 Duality between canonical forms

We are in the following going to discuss the principle of duality between canonical
form and state space representations. The importance of this is that it normally
is enough to learn one method of how to transform a given state space model to
canonical form. The reason for this is because the duality principle can be used to,
e.g., transform an observer canonical form to a controller canonical form. There is
a duality between the observer and controller canonical forms.

2.6.1 Duality between controller and observer canonical forms

A closer study of the controller and observer canonical forms, and the controllability
and observability canonical forms, shows that those are what we call dual canonical
forms.
Let xc be the state vector for the system in controller canonical form, (index c
for controller form) with realization (Ac , Bc , Dc ), i.e.,

ẋc = Ac xc + Bc u (2.56)
y = Dc xc (2.57)

The transformation matrix, T , witch transforms a realization (A, B, D) to controller

canonical form is given by

Tc = Cn M (2.58)
2.6 Duality between canonical forms 40

where M is an upper triangular Toeplitz matrix formed from the coefficients in the
characteristic polynomial.
Similarly, let xo be the state vector for a system on observer canonical form,
(index o for observer form) with realization (Ao , Bo , Do ), i.e.,
ẋc = Ao xo + Bo u (2.59)
y = Do xo (2.60)
The transformation matrix, T , that transforms a given realization (A, B, D) to so
called observer canonical from is then given by
To = On−1 (M T )−1 (2.61)
where M is the Toplitz matrix.
Comparing the two canonical forms we find the following connections:

Ac = ATo Bc = DoT Dc = BoT (2.62)

This phenomena is what we call controller and observer duality.

2.6.2 Duality between controllability and observability canonical

forms

In the same way, there is a duality between controllability and observability canonical
forms.
As we have shown,
Tco = Cn (2.63)
where Cn is the controllability matrix of the system, is the transformation matrix
which transform a given realization (A, B, D) to controllable canonical form, given
by
ẋco = Aco xco + Bco u (2.64)
y = Dco xco (2.65)
where index co denotes controllability form. In the same way
Tob = On−1 (2.66)
is the transformation matrix x = T xob which transforms a given realization (A, B, D)
to so called observability canonical form, given by

ẋco = Aob xob + Bob u (2.67)

y = Dob xob (2.68)
where index ob stand for observability form. Comparing the two canonical forms
shows that:
Aco = ATob Bco = Dob
T T
Dco = Bob (2.69)
2.7 Examples 41

2.7 Examples

Example 2.7 (Synthesis of controller based on COF)

Given a state space model on COntroller canonical Form (COF)
A B
z }| { z }| {
−a1 −a0 1
ẋ = x+ u (2.70)
1 0 0

Problem: We want ti find a controller of the form Vi ø

u = −G(xs − x) + us (2.71)
 
G = g1 g2 (2.72)

in such a way that the closed loop system has the eigenvalues (poles) given by s1 and
s2 .
Solution: The closed loop system is described by

us
 
 
ẋ = (A + BG)x + B − GB (2.73)
xs
 
−a − 1 + g1 −a2 + g2
A + BG = (2.74)
1 0

Note that the closed loop system matrix, A+BG, also is on controller canonical form.
An expression for the eigenvalues can be found from the characteristic equation, i.e.
 
 s + a1 − g1 a2 − g2 
|sI − (A + BG)| =    = s2 + (a1 − g1 )s + a2 − g2 = 0 (2.75)
−1 s 

If the eigenvalues of the closed loop system should be s1 and s2 then the closed loop
characteristic equation must be være

(s − s1 )(s − s2 ) = s2 − (s1 + s2 )s + s1 s2 = 0 (2.76)

Comparing the coefficients in the two polynomials gives

a2 − g2 = s1 s2 (2.77)
a1 − g1 = −(s1 + s2 ) (2.78)

This gives the coefficients g1 and g2 in the controller as a function of the prescribed
eigenvalues s1 and s2 and the system parameters a1 and a2 . We have

g2 = a2 − s1 s2 (2.79)
g1 = a1 + s1 + s2 (2.80)

Example 2.8 (Four canonical forms)

Given a system

ẋ = Ax + Bu (2.81)
y = Dx + Eu (2.82)
2.7 Examples 42

where
   
−4 3 2 1
A =  2 −6 1  B= 2 
(2.83)
 1 2 −8
 3
D = 2 −1 1

We will in this example compute all the four canonical forms which are described in
this section. Nota that the results would have been the same if the given state space
form had been a discrete time system. A MATLAB script for the computations of
the canonical forms is given in Example 2.9. Running the script within MATLAB
gives the following results.
Controller canonical form (cof ):
   
−18 −94 −113 1
Akrf =  1 0 0  Bkrf = 0 

(2.84)
 0 1  0 0
Dkrf = 3 58 279

Observer canonical form (obf ):

   
−18 1 0 3
Akef =  −94 0 1  Bkef =  58 
(2.85)
 −113 0 0 279
Dkef = 1 0 0

Controllable canonical form (cf ):

   
0 0 −113 1
Aksf =  1 0 −94  Bksf =  0 
(2.86)
 0 1 −18 0
Dksf = 3 4 −75

Observability canonical form (of ):

   
0 1 0 3
Akof =  0 0 1  Bkof =  4 
(2.87)
 −113 −94  −18 −75
Dkof = 1 0 0

Note the following connections:

Akof = ATksf T
Bkof = Dksf T
Dkof = Bksf
(2.88)
Akef = ATkrf T
Bkef = Dkrf T
Dkef = Bkrf

Example 2.9 (Matlab script for computation of canonical forms)

2.8 Summary 43

disp(’ SYSTEM MODELL) MATRICES:’)

A=[-4,3,2;2,-6,1;1,2,-8]
B=[1;2;3]
D=[2,-1,1]

disp(’1. CONTROLLER CANONICAL FORM (krf):’)

disp(’...HIT e key to continue’), pause
Qs=ctrb(A,B);
alfa=poly(A); M=eye(3); M(1,2)=alfa(2); M(2,3)=alfa(2); M(1,3)=alfa(3);
Tkrf=Qs*M;
Akrf=inv(Tkrf)*A*Tkrf, Bkrf=inv(Tkrf)*B, Dkrf=D*Tkrf

disp(’2. OBSERVER CANONICAL FORM (kef):’)

disp(’...hit a key to continue’), pause
Qo=obsv(A,D);
Tkef=inv(Qo)*inv(M’);
Akef=inv(Tkef)*A*Tkef, Bkef=inv(Tkef)*B, Dkef=D*Tkef

disp(’3. CONTROLLABILITY CANONICAL FORM( ksf):’)

disp(’...hit a key to continue’), pause
Tksf=Qs;
Aksf=inv(Tksf)*A*Tksf, Bksf=inv(Tksf)*B, Dksf=D*Tksf

disp(’4. OBSERVABILITY CANONICAL FORM (kof):’)

disp(’...hit a key to continue’), pause
Tkof=inv(Qo);
Akof=inv(Tkof)*A*Tkof, Bkof=inv(Tkof)*B, Dkof=D*Tkof

2.8 Summary

It is worth noticing that the problem of computing canonical forms is sensitive for
numerical errors. This is mostly a problem in MIMO systems. Even for controllable
and observable systems, the transformation matrix, T , constructed from independent
columns in the controllability matrix (ore rows in the observability matrix) may be
close to singular.
Note also that the system must be controllable for the controller and controllable
canonical forms to exists. Similarly, the system must be observable for the observer
and observability canonical forms to exists.
The parameters in a canonical form may be very sensitive to perturbations. This
can results in a poor choice for using the canonical form in a prediction error system
identification method. In some circumstances there may be impossible to find a
canonical form. This problem may occur for MIMO systems and is usually not a
problem for SISO systems.
2.8 Summary 44

There exists other essentially unique state space model realizations with good
numerical properties. The most common is balanced realization, input normal real-
ization and output normal realization. Those realizations are computed through a
singular value decomposition. This is not a topic here.
 Part II

Analysis and design of control

systems
Chapter 3

Simple PID tuning and model

reduction

3.1 Feedback systems

Intuitively, feed-forward control gives perfect control in case of an exact model.

However, models are never exact and the control system will have to involve feed-
back.
The main reasons for using feedback control are in case of unknown disturbances,
model uncertainty and unstable plants. Feedback systems are studied in the follow-
ing.

3.2 Standard feedback system

Consider a standard feedback system as illustrated in Figure 3.1.

r - i e- u- y -
hc (s) hp (s)
−6

Figure 3.1: Standard feedback system. Plant described by a transfer function model
hp (s) and controller transfer function hc (s).

In Figure 3.1 y is the measured plant output assumed for simplicity to be equal
to the desired output. Often in the feedback path there is an implementation of the
function ym = hm (s)y where ym is the measurement of the desired output y. Here
we assume, as explained, that the measurement system is exact and that hm (s) = 1
and that ym = y. In Figure 3.1 r is the specified reference for the plant output. The
error difference e = r − y is the input signal to the controller represented by the
transfer function hc (s). Here the true plant is represented with a transfer function
model hp (s). In the following we will study this standard feedback system.
3.3 Standard feedback systems with disturbances 47

3.3 Standard feedback systems with disturbances

Usually, in addition to the input u the system is also influenced by an external

disturbance v as illustrated in Figure 3.2
v
?
hv (s)

r - i e- u- ? y
hc (s) hp (s) - i -
−6

Figure 3.2: Standard feedback system with disturbance v at the output. Plant
described by a transfer function model hp (s), disturbance model hv (s) and controller
transfer function hc (s).

Hence, the plant may be influenced both from the control input u and an external
disturbance v. The plant is illustrated with the dashed
v box in Figure 3.3.
Plant ?
HH
H hv (s)

r - i e- u- y-
hp (s) - i
?
hc (s)
−6

Figure 3.3: Standard feedback system with disturbance v at the output. Plant
described by a transfer function model hp (s), disturbance model hv (s) and controller
transfer function hc (s). Plant indicated with the dashed box.

A more general situation is as illustrated in Figure 3.4 where the measurement

system is illustrated in the dashed box and where the actual measurement ym is
related to the desired output y as ym = hm (s)y + w where w is measurement noise,
and hm (s) a model for the measurement system. Usually, as described above we
assume hm (s) = 1. The measurement noise w is usually high frequent noise but
may also represent drifts etc.

3.4 Simple PID tuning rules

We will in this section present a simple and direct method for PID controller tun-
ing and design. The method is very simple and intuitive and leads to a practical
and robust method which also give very good results. The method is based on an
approximate model of 1st or 2nd order with time delay or inverse response. If the
3.4 Simple PID tuning rules 48
v
Plant Measurement
HH ? HH w
H
H hv (s)

r - i e- u- y - ? ym
hp (s) - i hm (s) - i -
?
hc (s)
−6

Figure 3.4: Standard feedback system with disturbance v at the output. Plant
described by a transfer function model hp (s), a disturbance model hv (s) and con-
troller transfer function hc (s). Plant indicated with the dashed box. The actual
measurement ym is related to the desired output y as ym = hm (s)y + w where w is
measurement noise.

process is approximated with a 2nd order model with time delay then a PID con-
troller is the result, but if a 1st order model with time delay is used a PI controller
is the result. Hence, the resulting controller is defined by the starting point which
is the approximate model. Only stable models is considered,
Consider an approximate 2nd order model with time delay given by the following
transfer function
e−τ s
hp (s) = k . (3.1)
(1 + T1 s)(1 + T2 s)
or with inverse response as follows
1 − τs
hp (s) = k , (3.2)
(1 + T1 s)(1 + T2 s)

Note that (3.2) is an approximation of the exact time delay model (3.1) because,

e−τ s ≈ 1 − τ s. (3.3)

The method can still be used if the second time constant, T2 = 0. Then we have
an approximate model of the form

e−τ s 1 − τs
hp (s) = k ≈k . (3.4)
1 + T1 s 1 + T1 s

If the starting point is a more complicated higher order model or you have
identified a higher order model from data by system identification methods such as,
e.g. DSR, then a 1st or 2nd order model approximation can be constructed by model
reduction techniques. System identification methods can possibly also be used to
construct a lower order model directly.
In the above low order models the parameter τ represents the effective time delay
or the inverse response time and the time constant T1 ≥ T2 ≥ 0 is the dominant time
constant. Note that many higher order models with time delay can be approximated
by a 1st or 2nd order model with time delay. A method for model reduction which
is to be used in this section is the half-rule.
3.4 Simple PID tuning rules 49

3.4.1 The half rule

As an introductory example of the half-rule consider the system (3.1) or (3.2) and
assume that this system do not have dominant 2nd order dynamics, then this model
can with advantage be approximated with a 1st order model with inverse response
as given in the following model (3.5).
1 − τs
hp (s) = k , (3.5)
1 + T1 s
where we from the half rule find the parameters

T1 := T1 + 12 T2 , (3.6)
1
τ := τ + 2 T2 . (3.7)

The half rule says that the neglected smallest time constant T2 is distributed evenly
over the remaining time constant T 1 and the time delay or inverse response time τ .
The model (3.5) is a very good starting point for the tuning of a PI controller.
If the system has dominant 2nd order dynamics in such a way that it is not wise to
neglect T2 , then we use a 2nd order model of the form (3.2) and designing a PID
controller.

Example 3.1 (Model reduction by the half rule)

Given a 4th order process y = hp (s)u where the process is described by the transfer
function, hp , given by
e−τ s
hp (s) = k (3.8)
(1 + T1 s)(1 + T2 s)(1 + T3 s)(1 + T4 s)
where T1 ≥ T2 ≥ T3 ≥ T4 > 0 is the time constants and τ is the time delay.
Using the half rule gives a 1st order model with time delay approximation
1 − τs
hp (s) = k , (3.9)
1 + T1 s
where
1
T1 := T1 + T2 , (3.10)
2
1
τ := τ + T2 + T3 + T4 , (3.11)
2
or by a 2nd order model approximation
1 − τs
hp (s) = k , (3.12)
(1 + T1 s)(1 + T2 s)
where

T1 := T1 , (3.13)
1
T2 := T2 + T3 , (3.14)
2
1
τ := τ + T3 + T4 . (3.15)
2
3.4 Simple PID tuning rules 50

Example 3.2 (Model reduction by the half rule)

Given a 4th order process y = hp (s)u where the process is described by the transfer
function, hp , given by
1
hp (s) = 2 . (3.16)
(1 + 6s)(1 + 4s)(1 + 2s)(1 + 1s)
This model may be approximated by a 1st order+time delay model

e−5s
h1a
p (s) = 2 , (3.17)
1 + 8s
i.e., the time constant is T = 6 + 21 4 = 8 and the time delay is τ = 12 4 + 2 + 1 = 5.
e−5s
Remark: a better 1st order approximation at high frequencies will be h1a
p (s) = 2 1+10s .

The model may be approximated by a 2nd order+time delay model

e−2s
h2a
p (s) = 2 , (3.18)
(1 + 6s)(1 + 5s

where the first time constant is T1 := T1 = 6, the second is T2 := T2 + 12 T3 and the

time delay τ = 12 T3 + T4 = 2.
The step responses are illustrated in Figure 3.5. The step response is implemented
in the MATLAB m-file ex2 half.m.

Model response and 1st and 2nd order approximations

2
Model response
1st order approx
2nd order approx
1.5

0.5

−0.5

−1

−1.5
0 5 10 15 20 25 30 35 40 45
Time

Figure 3.5: Step response of model hp (s) and the 1st and 2nd order model approx-
imations in Example 3.2 The time delay is approximated with an inverse response
term in the responses, i.e. e−τ s ≈ 1 − τ s.

3.4.2 Tuning of PI and PID controllers

Let us specify the response from the reference signal, r, to the measurement output,
y, as follows
y 1 − τs
= . (3.19)
r 1 + Tc s
3.4 Simple PID tuning rules 51

r e u y
hc (s) hp (s)

Figure 3.6: Standard feedback system with process hp (s) and controller hc (s).

here Tc is a user specified time constant for the set-point response. The choice for
Tc will be discussed later but it is natural to chose it related to the time delay
or inverse response time τ . It is nothing to do with a transport delay or inverse
response phenomena in a process (when the process already is designed), hence the
time delay ore inverse response must also be in the set point response yr . This is the
reason for the numerator polynomial 1 − τ s in (3.19).
We also know from the block diagram in Figure 3.6 of the control system that
y hp hc
= , (3.20)
r 1 + hp hc

where hc (s) is the transfer function for the controller, which at this stage is unknown.
Solving for the controller gives
y
1 r 1 1
hc (s) = y = 1 . (3.21)
hp (s) 1 − r hp (s) y −1
r

Putting the transfer functions (3.19) and (3.2) (or (3.5)) into the expression (3.21)
for the controller, hc (s). This gives the following controller
1 T1 1 + T1 s
hc (s) = (1 + T2 s). (3.22)
k Tc + τ T1 s
This is a so called cascade formulation of a PID controller.
1 + Ti s
hc (s) = Kp (1 + Td s), (3.23)
Ti s
where the parameters in the cascade PID controller are given by
1 T1
Kp = , (3.24)
k Tc + τ
Ti = T1 (3.25)
Td = T2 . (3.26)

Derivative action is usually only recommended for systems with dominant 2nd order
dynamics, i.e. with a large time constant T2 , say large relative to the time delay and
in case where T2 > τ is a rule of thumb. Notice also in connection with this that in
case of high frequent noise on the output, it make sense to skip derivative action in
order not to amplify noise on the control action and to the system.
Note that if we instead have used (3.5)) in (3.21) then a PI controller
1 + Ti s
hc (s) = Kp , (3.27)
Ti s
3.5 PID controller for oscillating process 52

with parameters as given in (3.24) and (3.25) have been the result.
At this stage the only unknown parameters in the controller settings is the time
constant Tc for the set point response. We can for many processes chose Tc equal
(ore grater) then the time delay or the inverse response time τ . Hence, a simple
choice is simply

Tc = τ. (3.28)

This gives the proportional constant

1 T1
Kp = , (3.29)
2k τ

3.4.3 Choice of Tc

Locking at the formula (3.27) for Kp shows us that it may be natural to force the
term TcT+τ
1
to be positive. Hence we must chose Tc + τ > 0. This means that Tc can
be chosen relatively within wide ranges −τ < Tc < ∞. However, an optimal setting
for Tc will be a trade of between:

1. Fast response in order to controlling disturbances. This is favored by a small

value of Tc .

2. Stability and robustness and using a small amount of control u. This is favored
by a small value of Tc .

A good trade of between these wishes is as presented in Skogestad (2002) to

chose as in (3.28). However, the choice

Tc ≥ τ. (3.30)

Then from (3.24) we have that

1 T1
Kp ≤ Kpmax = (3.31)
2k τ

3.4.4 Modification of the integral term

3.5 PID controller for oscillating process

Given a process described by

e−τ s
hp (s) = k , (3.32)
τ02 s2 + 2τ0 ξs + 1

where ζ is the relative damping, τ0 = ω1 determines the speed of response, and where
ω is the natural frequency. Remark that when ζ = 1 we have two real poles/time
constants and a pole polynomial τ02 s2 + 2τ0 ξs + 1 = (τ0 s + 1)2 .
3.6 ID controller for systems with no damping 53

Note that ξ < 1 gives oscillating process dynamics. The poles of the system is
given by
p
−ξ ± ξ 2 − 1
s= . (3.33)
τ0
We see that the poles becomes complex when ξ 2 − 1 < 0 and that the dynamics may
be oscillating in this case. It is possible with oscillating dynamics when ξ is negative
and with |ξ| < 1 but in this cases we must have τ0 < 0 in order for the system to be
stable.
The period time of the oscillations, Pu , is given by
2π
Pu = p τ0 . (3.34)
1 − ζ2

Note that with significant oscillations we have ζ 2 << 1 and Pu = 2πτ0 .

Putting (3.32) and (3.19) into (3.21) gives the following controller
1
hc (s) = Kp (1 + + Td s) (3.35)
Ti s
where
2τ0 ξ
Kp = (3.36)
k(Tc + τ )

Ti = 2τ0 ξ (3.37)

τ0
Td = (3.38)
2ξ
Hence, as we see (3.35) is an ideal PID controller.
Notice that the above PID controller is valid for under-damped processes where
the relative damping is 0 < ξ < 1. When xi = 1 we have two real poles and both a
PI and a PID controller may be used. The case without damping, i.e. when xi = 0
needs to be analyzed separately as in the next subsection.

3.6 ID controller for systems with no damping

In this section we study pure oscillating systems, that is systems with no damping
and where ξ = 0. This could be the case for a spring or a pendulum without friction.
A linearized process model for such a process could be
e−τ s
hp (s) = k . (3.39)
τ02 s2
+1
Putting (3.39) and (3.19) into (3.21) gives the following controller

1 τ02
hc (s) = + s, (3.40)
k(Tc + τ )s k(Tc + τ )
3.7 PI Control of first order process 54

which is a pure ID (Integral and Derivative) controller.

Notice that the PID controller Eq. (3.35) with parameters as in Eqs. (3.36)-
(3.38) may be written as

Kp
hc (s) = Kp + + Kp Td s, (3.41)
Ti s
where the proportional gain is as in (3.36) and

Kp 1
= , (3.42)
Ti k(Tc + τ )

τ02
Kp Td = (3.43)
k(Tc + τ )

Hence, this result is consistent with the above result for systems without damping
and ξ = 0.

3.7 PI Control of first order process

In some simple cases the process may be modeled by a 1st order process
1
hp (s) = k , (3.44)
1 + Ts
where k is the gain and T the time constant.
For such processes one may add a fictive time delay, τ , due to un-modeled effects
and time delay due to sampling, and it make sense to add a time delay τ ≥ ∆t2 where
∆t is the sampling interval.
However, we will in this section use the simple controller tuning method directly
from the 1st order model (3.44).
Let us specify the closed loop response as
y 1
(s) = , (3.45)
r 1 + Tc s
where Tc is a user specified time constant for the closed loop response from r to y.
Putting (3.44) and (3.45) into (3.21) gives the following controller

T 1 + Ts 1 + Ti s
hc (s) = = Kp (3.46)
kTc T s Ti s
which is a PI controller with
T
Kp = (3.47)
kTc

Ti = T (3.48)
3.8 Integrating process with time delay 55

Hence, as we see (3.35) is an ideal PI controller.

The prescribed time constant, Tc , should with advantage be specified smaller
than the time constant in the process. It make sense to specify Tc up to ten times
T
faster, hence chose 10 ≤ Tc ≤ T . Notice that smaller Tc (faster set-point response)
implies higher controller proportional gain Kp which implies higher control action in
u and saturation may be a problem. Notice also that physical insight of the process
also should be used when specifying time constants

3.8 Integrating process with time delay

Consider a system described by the transfer function

e−τ s
hp (s) = k , (3.49)
s
which is an integrator process with time delay. Note that we may approximate Eq.
(3.49) as an inverse response with τ ≥ 0. Using the method for a process y = hp (s)u
leads to a P-controller with proportional gain
1 1
Kp = = , (3.50)
k(Tc + τ ) 2kτ

where the last equality is obtained by the simple rule of thumb, Tc = τ .

Consider now a feedback system with a disturbance v occuring at the input side
of the plant as illustrated in Figure 3.7.
v
r - i e- u ?
i hp (s) y -
hc (s) - -
−6

Figure 3.7: Standard feedback system with disturbance at the input. Plant described
by a transfer function model hp (s) and controller transfer function hc (s).

A P-controller will give set-point error for disturbances at the input, i.e. for
systems y = hp (s)(u + v) because the response from the disturbance to the output
then is given by
y y
r
(s) v
(s)
z }| { z }| {
hc hp hp
y= r+ v. (3.51)
1 + hc hp 1 + hc hp

Locking at the response from the disturbance, v, to the output, y, for a process
−τ s
hp = k e s and a P-controller, i.e., hc = Kp gives,
−τ s
y ke s
(s) = −τ s . (3.52)
v 1 + Kpk e s
3.8 Integrating process with time delay 56

In steady state we have yv (s = 0) = 1

Kp and that

1
y=r+ v. (3.53)
Kp

This implies that we need a PI-controller for integrating+time delay processes in

order to eliminate the offset from load disturbances v at the input, i.e., we need a
controller in which yv (s = 0) = 0. Note that load disturbances at the output will
be removed by using a P-controller, i.e., for systems y = hp (s)u + v and integrating
plus time delay process as in Eq. (3.49).

Table 3.1: PI-controller settings for an integrating plus time delay process hp (s) =
−τ s
k e s with gain, k, and time delay τ ≥ 0. Setting 1 is the Skogestad IMC (SIMC)
setting. Settings 2 is suggested by Haugen and settings 3 are proposed in this note.
The process model, hp , used for the derivation of the settings and the corresponding
relative damping factor ζ used are indicated in the table for completeness.

Kp Ti hp (s) ζ
1 k
1 2kτ 8τ s 1√
1 k 2
2 2kτ 4τ s √2
1 3
3 2kτ 6τ k 1−τ
s
s
2

3.8.1 The SIMC settings: neglecting the time delay

In practice, for the reason of eliminating load disturbances v at the input, i.e., for
systems y = hp (s)(u + v) and in case of un-modeled effects we use a PI controller
for integrating processes. The Skogestad PI settings are
1
Kp = , Ti = 4(Tc + τ ) (3.54)
k(Tc + τ )

and with the time constant for the setpoint response, Tc = τ we obtain the SIMC
settings
1
Kp = , Ti = 8τ. (3.55)
2kτ

This may be proved as follows. Consider a 1st order system with time delay, and
with a large time constant, T , i.e. we may write the model as

e−τ s k 0 e−τ s e−τ s

hp (s) = k 0 = ≈ k (3.56)
1 + Ts T T1 + s s
0
with k = kT . For systems with large time constants and neglecting the time delay
we obtain the transfer function
1
hp (s) ≈ k (3.57)
s
3.8 Integrating process with time delay 57

which is used for the derivation of the SIMC PI-settings.

The pole polynomial for the disturbance and set-point response is obtained from
1 + Ti s k 1 Kp k
1 + hc hp = 1 + Kp = 2 (s2 + (1 + Ti s))
Ti s s s Ti
1 Kp k Ti 2
= ( s + Ti s + 1)) (3.58)
s2 Ti Kp k

This gives a pole polynomial on standard second order form as

Ti 2
π(s) = s + Ti s + 1 = τ02 s2 + 2ζτ0 s + 1 (3.59)
Kp k

By comparing the coefficients in the pole polynomial and the corresponding coeffi-
cients in the standard second order polynomial we may find relations between Kp
and Ti . We have
Ti
τ02 = , 2ζτ0 = Ti (3.60)
Kp k

This gives (2ζτ0 )2 = Ti2 and

1
Ti = 4ζ 2 (3.61)
Kp k

Using the setting for the proportional gain, i.e.,

1 T
Kp = = 0 . (3.62)
k(Tc + τ ) k (Tc + τ )

where k 0 is the gain and T the time constant in the 1st order process. Note that the
0
slope is k = kT in case of an integrating process. Hence we have

Ti = 4ζ 2 (Tc + τ ). (3.63)

Putting ζ = 1 gives real roots and a pole polynomial π(s) = (1 + τ0 s)(1 + τ0 s) =

τ02 s2 + 2τ0 s + 1. Furthermore using the settings Kp = k(Tc1+τ ) = 2kτ
1
gives the SIMC
setting Ti = 4(Tc + τ ) = 8τ when Tc = τ . Note also that this gives an approximate
1
time constant τ0 = 2ζ Ti = 4τ for the responses.
Unfortunately, as pointed out by Haugen (2010) the response of eliminating
disturbances in v is slow by this settings and the integral time constant T√i may be
reduced by a factor of two, i.e. by allowing oscillations and requiring ζ = 22 ≈ 0.7.
This gives

Ti = 4τ (3.64)
√
This setting gives a Buttherworth pole polynomial π(s) = τ02 s2 + 2τ0 s + 1 with
τ0 = 2τ .
3.8 Integrating process with time delay 58

3.8.2 Settings by approximating time delay as inverse response

Instead of neglecting the time delay as in the derivation of the SIMC PI settings we
will in this section deduce alternative settings for the integral time constant Ti .
Let us study the disturbance response in case of a PI controller. We have
−τ s
y hp ke s kse−τ s
(s) = = −τ s = Kp k
(3.65)
v 1 + hc hp 1 + Kp 1+T is e
Ti s k s s2 + Ti (1 + Ti s)e−τ s
Approximating the delay as an inverse response term we get
y ks(1 − τ s) Ti s(1 − τ s)
(s) = Kp k
= Ti 2
(3.66)
v s2 + + Ti s)(1 − τ s) Kp Kp k s + (1 + Ti s)(1 − τ s)
Ti (1

The poles are given by the roots of the pole polynomial, i.e.,
Ti 2 1
π(s) = s + (1 + Ti s)(1 − τ s) = Ti ( − τ )s2 + (Ti − τ )s + 1
Kp k Kp k
= τ02 s2 + 2τ0 ζs + 1 (3.67)
Comparing the coefficients with the standard second order form polynomial we find
Ti 1
τ02 = − Ti τ = Ti ( − τ) (3.68)
Kp k Kp k
and
2τ0 ζ = Ti − τ (3.69)
Hence, from 4ζ 2 τ02 = (Ti − τ )2 we find the following 2nd order polynomial for the
relationship between Ti and Kp as a function of the relative damping coefficient ζ,
i.e.,
1
Ti2 − (4ζ 2 ( − τ ) + 2τ )Ti + τ 2 = 0. (3.70)
Kp k

1
With the setting Kp = 2kτ for the proportional gain we obtain τ02 = Ti τ . Re-
quiring ζ = 1 gives
4Ti τ = (Ti − τ )2 (3.71)
and
Ti2 − 6τ Ti + τ 2 = 0 (3.72)
with the positive solution
√
32
6+ √
Ti = τ = (3 + 2 2) ≈ 6τ. (3.73)
2
This gives very good set-point and disturbance responses.
√
2
Putting ζ = 2 gives
√
Ti = (2 + 3) ≈ 4τ, (3.74)
which is approximately the same setting as in Eq. (3.64).
3.8 Integrating process with time delay 59

Example 3.3 (PI control of integrating plus delay process)

Given an integrator plus time delay system described by the transfer function

e−τ s
hp (s) = k , (3.75)
s
with gain k = 1 and time delay τ = 1. The results by using a PI controller with
settings as in Table 3.1 are illustrated in Figure 3.8, which shows set-point and
disturbance rejection responses after a unit step in the reference, r = 1 at time
t = 0, and a unit step in the disturbance from v = 0 to v = 1 at time t = 40.
The simulations is done by the MATLAB m-file function ex3b half.m. As we
see the SIMC settings gives a relatively slow response from both the set-point and
the disturbance. The Buttherworth setting (2) results in the fastest responses but
have small oscillations. The settings derived in this note (3) results in nice, smooth
response approximately as fast as the response of the Butterworth settings.

PI settings for integrating pluss delay process

3
1. SIMC
2. Buttherworth
3. This note

2.5

1.5

0.5

0
0 10 20 30 40 50 60 70 80
Time

−τ s
Figure 3.8: PI control of integrating pluss delay process, hp (s) = k e s . PI-controller
hc (s) = Kp 1+T is
Ti s with settings as in Table 3.1.

3.8.3 Settings by approximating time delay with Pade and Balchen

approximation

Let us study the disturbance response in case of a PI controller. We have

−τ s
y hp ke s kse−τ s
(s) = = −τ s = Kp k
. (3.76)
v 1 + hc hp 1 + Kp 1+T
Ti s
is e
k s s2 + Ti (1 + Ti s)e−τ s
3.8 Integrating process with time delay 60

Approximating the delay with an approximation

1 − αs
e−τ s ≈ , (3.77)
1 + αs
where α = τ2 gives the 1st order Pade approximation. An alternative is to use the
Balchen approximation, i.e., α = 2τ
π . This gives

y ks 1−αs
1+αs Ti s 1−αs
1+αs
(s) = Kp k
= Ti 2
. (3.78)
v s2 + Ti (1
1−αs
+ Ti s) 1+αs Kp Kp k s + (1 + Ti s) 1−αs
1+αs

which is equivalent with

y Ti s(1 − αs)
(s) = Ti 2
, (3.79)
v Kp Kp k s (1 + αs) + (1 + Ti s)(1 − αs)

and
y Ti s(1 − αs)
(s) = T 1
. (3.80)
v Kp α K k s + Ti ( K k − α)s2 + (Ti − α)s + 1
i 3
p p

Hence, we have the pole polynomial

Ti 3 1
π(s) = α s + Ti ( − α)s2 + (Ti − α)s + 1. (3.81)
Kp k Kp k

We may now find a relationship between the controller parameters by specifying

√ the
polynomial coefficients. One choice is a Buttherworth configuration with ζ = 22 in
a prescribed 3rd order pole polynomial

π(s) = (1 + τ0 s)(τ02 s2 + 2ζτ0 s + 1). (3.82)

We will instead for the sake of increased robustness in the resulting feedback system
chose ζ = 1 and three multiple real time constants, i.e. a prescribed pole polynomial

π(s) = (1 + τ0 s)(τ02 s2 + 2τ0 s + 1 = (1 + τ0 s)3 = τ03 s3 + 3τ02 + 3τ0 + 1. (3.83)

Comparing the coefficients in polynomials (3.81) and (3.83) we find

Ti 1
τ03 = α , 3τ02 = Ti ( − α), 3τ0 = Ti − α. (3.84)
Kp k Kp k

Interestingly, from this, by using that 3τ03 = Ti ( K1p k − α)τ0 we find the linear ex-
pression for Ti , i.e.,
1 1 1
3α =( − α) (Ti − α), (3.85)
Kp k Kp k 3

and the integral time constant

α( K10p k + α)
Ti = 1 . (3.86)
Kp k −α
3.9 Re-tuning to avoid oscillations 61

1 1
Let us use the standard SIMC setting for the proportional gain, Kp = k(Tc +τ ) = 2kτ
for the simple choice Tc = τ . We find
41 τ
Ti = τ ≈ 6.83τ, α = . (3.87)
6 2
This setting is approximately the same as we found by using the approximation
e−τ s ≈ 1 − τ s. Furthermore using the Balchen approximation we find
20 + π2 2τ
Ti = τ ≈ 9.64τ, α = . (3.88)
π−1 π
This integral time setting gives better margins, i.e., a gain margin GM ≈ 3, a
phase margin P M = 49.3 and a maximal time delay error dτmax ≈ 1.68. The
corresponding SIMC PI settings with Ti = 8τ gives GM ≈ 2.96, P M = 46.86 and
dτmax ≈ 1.58.

3.9 Re-tuning to avoid oscillations

Some feedback loops are wrong tuned so that the plant output is influenced by
inherent oscillations. Suppose the PI controller parameters Kpold , Tiold is specified
by the operator and that the feedback loop have inherent oscillations. We will in
the following suggest how to re-tune the PI controller to avoid oscillations.
From eq. (3.61) we have that
4 Kp Ti
= . (3.89)
k ξ2

Using eq. (3.89) for two different PI controller tunings Kpold , Tiold and Kpnew , Tinew
we obtain
ξ2
Kpnew Tinew = new Kpold Tiold . (3.90)
ξ2
old
It make sense to chose the relative damping in the re-tuned feedback loop as ξnew ≥
1 to avoid oscillations. Hence, we have to ensure
1
Kpnew Tinew ≥ Kpold Tiold . (3.91)
ξ2
old

3.10 Controller for special type systems

The method for standard controller design presented in this section may be used
directly for simple processes.

3.10.1 Pure time delay process

Consider a pure time delay process

hp (s) = ke−τ s ≈ k(1 − τ s). (3.92)

3.11 Examples 62

The method leads in this case to a pure integral I controller

1
hc (s) = (3.93)
Ti s
with Ti = k(Tc + τ ).
y 1−τ s
The set-point response from r to y is specified as r = 1+Tc s for the above control
systems.

3.11 Examples

Example 3.4 (Effect of feedback control: SISO system)

Given a system described by the transfer function
1−s
hp (s) = . (3.94)
1+s
This system have a zero for s = 1 and a pole s = −1. Since the pole is negative and
located in the left half plane the system is stable. The zero is positive and located
in the rigt half part of the complex plane. The system is therefore said to be a
non-minimum phase system. Some limitations in the feedback system is therefore
to be expected. Typically, there will be limitations in the set-point response from the
reference, r, to the measurements output, y. Typically we can not use large values
of the proportional constant, Kp .
Assume that we want to use a simple proportional feedback controller given by

u = Kp (r − y), (3.95)

where r is the reference signal for y and Kp is the proportional constant. The closed
loop system is therefore described by.

y hp (s)hr (s) Kp (1 − s)
= hcl (s) = = , (3.96)
r 1 − (−1)hp (s)hr (s) (1 − Kp )s + 1 + Kp

where we have used negative feedback. As we see, the closed loop system also have a
zero at s = 1. Hence, the open loop zero is not influenced by the feedback as expected.
Zeroes are not influenced by feedback. However, the pole for the closed loop system
is
1 + Kp
scl = − . (3.97)
1 − Kp

It is natural to demand that the closed loop system is stable, i.e., scl < 0. Hence we
have that

−1 < Kp < 1. (3.98)

This means that the speed of response is limited. In this example we got problems
with the inverse response for large values of Kp , since the system is non-minimum
3.11 Examples 63

phase. The system have an inverse response since the gain at time zero (t = 0) is
given by
−Kp
hcl (s = ∞) = =→ −∞ når Kp → 1 (3.99)
1 − Kp
In addition, the system have a positive steady state gain given by
Kp
hcl (s = 0) = (3.100)
1 + Kp
This means that the inverse response goes to infinity as Kp → 1. However, the
response got faster since the pole of the closed loop system got more negative, i.e.
scl → −∞ as Kp → 1.
The problem is here that we can not get both fast response and a small inverse
response. This is illustrated in Figure 3.9. We also se from Figure 3.10 that the
amount of control signal u increases to infinity when Kp → 1.

SISO inverse response example

0
y

−5

−10

−15

−20

−25
0 0.05 0.1 0.15 0.2 0.25 0.3 0.35 0.4
Continuous time

1−s
Figure 3.9: Step response simulation of a system with hp (s) = 1+s and u = Kp (r−y)
for varying proportional coefficient 0.8 < Kp < 0.96. Note that symbols g = Kp and
r = y 0 is used in the figure.

Example 3.5 (Inverse response in state space and transfer function models)

Given a system described by the state space model

1 T +τ
ẋ = − x + k u, (3.101)
T T2
kτ
y = x− u. (3.102)
T
3.11 Examples 64

SISO inverse response example: input

25

u
20

15

10

0
0 0.05 0.1 0.15 0.2 0.25 0.3 0.35 0.4
Continuous time

Figure 3.10: The input u = Kp (r − y) after a step response simulation of a system

with hp (s) = 1−s
1+s for varying proportional coefficient 0.8 < Kp < 0.96.Note that
symbols g = Kp and r = y 0 is used in the figure.

This model is equivalent with the following transfer function model

y 1 − τs
=k . (3.103)
u 1 + Ts
This system have an inverse response because of the zero in the right half plane, i.e.,
the zero s0 = τ1 . Note also that the inverse response term 1 − τ s in the numerator
is an approximation to a pure time delay since e−τ s ≈ 1 − τ s. The transfer func-
tion model (3.103) is a good starting point for PI controller synthesis by, e.g. the
Skogestad (2002) tuning method.

Example 3.6 (PI-control of a non-minimum phase SISO system)

Given a system described by the transfer function model y = hp (s)u where the trans-
fer function is given by
1 − 2s 1 − 2s
hp (s) = = . (3.104)
s2 + 3s + 2 (s + 1)(s + 2)

The system is to be controlled by a PI controller given by

1 + Ti s
hc (s) = Kp . (3.105)
Ti s
A commonly used choice is to chose the integral time Ti equal to the dominant time
constant in the system in order to cancel it and thereby simplify the loop transfer
function , h0 = hp hc . The system has two poles s1 = −1 and s2 = −2. This gives
3.11 Examples 65

the two time constants T1 = − s11 = 1 and T2 = − s12 = 12 . Hence, the loop transfer
function is with Ti = T1 = 1 given by
1 − 2s 1 + Ti s Kp 1 − 2s
h0 (s) = hp hc = Kp = , (3.106)
(s + 1)(s + 2) Ti s Ti s(s + 2)
where we have chosen Ti = 1. We can now find expressions for Kp by demanding
the closed loop system to be stable. The system from the reference, r, to the output,
y, must be stable. The transfer function from the reference r to the output y in a
feedback system (with negative feedback) is given by
Kp
1−2s p K
y h0 Ti s(s+2) Ti (1 − 2s)
= = K 1−2s
= K Kp
(3.107)
r 1 + h0 1 + Tip s(s+2) s2 + 2(1 − Tip )s + Ti

It can be shown that the roots of a 2nd order polynomial , s2 + a1 s + a0 = 0,

has roots in the left half plane (stable system) if the coefficients are positive, i.e.,
stable system if a1 > 0 and a0 > 0. This can be shown by study the polynomial,
(s + λ1 )(s + λ2 ) = s2 + (λ1 + λ2 )s + λ1 λ2 = 0 which has roots s1 = −λ1 and
s2 = −λ2 . If the roots are in the left half plane, i.e., s1 < 0 and s2 < 0 then we must
have that λ1 > 0 and λ2 > 0. This implies that the coefficients must be positive, i.e.,
a0 = λ1 λ2 > 0 and a1 = λ1 + λ2 > 0.
We obtain the following demands for Kp :
Kp Kp
2(1 − ) > 0 and > 0. (3.108)
Ti Ti
This gives
Kp
0< < 1. (3.109)
Ti

We have simulated the closed loop system for different values of Kp after a posi-
tive and unit step change in the reference r. The results are presented in Figure 3.11.
As we see, the response got more oscillations and overshot for increasing values of
Kp < 1. At the same time the response got a larger inverse response starting at time
t = 0. Inverse responses is typical when controlling non-minimum phase systems
with zeroes in the right half plane.
As we see it is difficult to at the same time obtain fast response, small overshot
and small inverse response. the reason for this problems is the zero in the right half
plane.
Some trial and errors gives the following reasonable choice
Kp = 0.42, Ti = 1. (3.110)
This PI controller settings gives a gain margin, GM = 2.8 [dB], and a phase margin
P M = 71◦ .

Example 3.7 (PI-control of non-minimum phase SISO system)

Consider a system described by the transfer function
1 − 2s 1 − 2s
hp (s) = = . (3.111)
s2 + 3s + 2 (s + 1)(s + 2)
3.11 Examples 66

PI−regulering av ikke−minimum−fase prosess

2.5

K =0.9
p
2

1.5

1
K
p

0.5

K =0.1
p
0

−0.5

−1

−1.5
0 2 4 6 8 10 12 14 16 18 20
Continuous time

Figure 3.11: Unit step response (in the reference) simulation of a control system with
1−2s
process model, hp (s) = (s+1)(s+2) , and PI-controller hc (s) = Kp 1+T is
Ti s with Ti = 1
and for varying proportional coefficients in the interval, 0.1 ≤ Kp ≤ 0.9. The Figure
is generated by the MATLAB script siso zero ex.m.

The frequency response of the system is given by

hp (jω) = |hp (jω)|ej∠hp (jω) , (3.112)

where the phase and magnitude are given by

ω
∠hp (jω) = −(arctan(2ω) + arctan(ω) + arctan( )), (3.113)
2

√
1 + 4ω 2
|hp (jω)| = √ √ . (3.114)
1 + ω2 4 + ω2
The phase crossover frequency (critical frequency), ω180 , is then given by the fre-
quency where the phase is −180◦ , i.e., ∠hp (jω180 ) = −π. The critical gain, Kcu , is
then the gain such that Kcu |hp (jω180 )| = 1. The parameters Kcu and ω180 can for
example be found by using the MATLAB function margin. We obtain

ω180 = 1.8708, (3.115)

Kcu = 1.5. (3.116)

We may now simply find the parameters in a PI controller

1 + Ti s
hc (s) = Kp . (3.117)
Ti s
3.11 Examples 67

by using the Ziegler-Nichols method. Hence,

Kcu 2π Pu
Kp = = 0.68, Pu = = 3.36, Ti = = 2.79. (3.118)
2.2 ω180 1.2
It can by simulations be shown that the response in, y, after a step in the reference,
r, is relatively poor by this PI controller settings.
The closed loop behavior may be further investigated by the vlosed loop transfer
function from r to y, i.e.,
Kp
y h0 Ti (1 − 2s)(1 + Ti s)
= = K Kp
. (3.119)
r 1 + h0 s + (3 − 2Kp )s2 + (Kp − 2 Tip + 2)s +
3
Ti

Example 3.8 (Inverse response and model reduction with the half rule)
Given a system described by the transfer function model y = hp (s)u where the trans-
fer function is given by
1 − 2s 1 − τs
hp (s) = =k , (3.120)
(s + 1)(s + 2) (1 + T1 s)(1 + T2 s)
where
1 1
k = , τ = 2, T1 = 1, T2 = . (3.121)
2 2
A good 1st order reduced model for PI controller synthesis is
1 − τs
hp (s) = k , (3.122)
1 + T1 s
1
where k = 2 and τ and T1 is found by using the half rule, i.e.,
1 1 9
τ := τ + T2 = 2 + = , (3.123)
2 4 4
1 1 5
T1 := T1 + T2 = 1 + = . (3.124)
2 4 4
A good PI controller setting is then obtained from the Skogestad (2002) method by
5
Ti = T1 = ≈ 1.25, (3.125)
4
and
1 T1 5
Kp = = ≈ 0.56. (3.126)
2 kτ 9

Simulation results of the set point responses for the various PI control tunings
are shown in Figures (3.12) and (3.13).

Example 3.9 (PI control of 3rd order process)

Given a 3rd order process
k
hp (s) = , (3.127)
(1 + T1 s)(1 + T2 s)(1 + T3 s)
3.11 Examples 68

Ex. PID control of process hp=k/((1+T1s)(1+T2s)(1+T3s))

1.2

1
0.8

2
0.6

3
0.4
y

1. Skogestad Tuning. Kp= 0.56, Ti=1.25

0.2
2. Prøv og feil. Kp=0.42, Ti=1
3. Ziegler−Nichols. Kp=0.68, Ti=2.79
0

−0.2

−0.4

−0.6
0 2 4 6 8 10 12 14 16 18 20
Time t [sec]

1−2s
Figure 3.12: PI control of 2nd order process, hp (s) = 0.5 (1+s)(1+0.5s) . PI-controller
hc (s) = Kp 1+T is
Ti s with parameters as in Example 3.8. The figure is generated with
the MATLAB script main piex2.m.

Ex. PID control of process hp=k/((1+T1s)(1+T2s)(1+T3s))

2.4

2.2

2
1

1.8

2
1.6

1.4
u

3
1.2

0.8

0.6

0.4
0 2 4 6 8 10 12 14 16 18 20
Time t [sec]

1−2s
Figure 3.13: PI control of 2nd order process, hp (s) = 0.5 (1+s)(1+0.5s) . PI-controller
hc (s) = Kp 1+T is
Ti s with parameters as in Example 3.8. The figure is generated with
the MATLAB script main piex2.m.
3.11 Examples 69

where k = 0.5, T1 = 8, T2 = 5 and T3 = 3. For simulation purpose we use a state

space equivalent, e.g. the following

− T11 1
0
   
T1 0
ẋ =  0 − T12 1 
T2 x+ 0 u (3.128)
1 k
0 0 − T3 T3
y = x1 (3.129)

Using the Skogestad method gives the following PI controller parameters

Ti = T = 10.5 (3.130)

1 10.5
Kp = Tτ = ≈ 1.91 (3.131)
2k 5.5
where T and τ is the parameters in the reduced model, i.e. a first order model with
inverse response
1 − τs
hp (s) = k (3.132)
1 + Ts
The parameters in the reduced model is obtained from the half rule, i.e.
1
T = T1 + T2 = 10.5 (3.133)
2

1
τ = T2 + T3 = 5.5 (3.134)
2

In order to compare the the Skogestad tuning with other PI control tuning rules
we chose the PI control tunings in Table 9.3 in Balchen et al (2003). The Balchen
tunings are

Ti = T1 = 8 (3.135)

1 8
Kp = T1 T2 = 2 = 3.2 (3.136)
k 5

The tunings proposed in Balchen et al (2003) gives to much proportional action,

both because of a small Ti and a large Kp (compared to the Skogestad tuning). The
tunings proposed by Skogestad gives a more reasonable set-point response. Note that
the proportional action is less because of both a larger Ti and a smaller Kp . Note
that a set-point response without overshoot can be obtanied by reducing Kp , e.g. the
settings Kp = 1.3 and Ti = 10.5 gives a rather nice response. The set-point responses
are illustrated in Figures (3.14) and (3.15).
3.11 Examples 70

Ex. PID control of process hp=k/((1+T1s)(1+T2s)(1+T3s))

1.5

Balchen Table 9.3 PI tuning

Skogestad PI tuning

1
y

Kp=1.3 and Ti=10.5

0.5

0
0 10 20 30 40 50 60 70 80 90
Time t [sec]

k
Figure 3.14: PI control of 3rd order process, hp (s) = (1+8s)(1+5s)(1+3s) . PI-controller
hc (s) = Kp 1+T is
with parameters as in Example 3.9. The figure is generated with
Ti s
the MATLAB script main piex1.m.

Ex. PID control of process hp=k/((1+T1s)(1+T2s)(1+T3s))

5

4.5

Balchen Table 9.3. PI tuning

3.5

3
Skogestad PI tuning
u

2.5

1.5
Kp=1.3
Ti=10.5
1

0.5
0 10 20 30 40 50 60 70 80 90
Time t [sec]

k
Figure 3.15: PI control of 3rd order process, hp (s) = (1+8s)(1+5s)(1+3s) . PI-controller
hc (s) = Kp 1+T is
with parameters as in Example 3.9. The figure is generated with
Ti s
the MATLAB script main piex1.m.
Chapter 4

The basic PID controller

4.1 The PI controller

4.1.1 Frequency domain description of the PI controller

The PI controller can in the Laplace plane be written as follows

u(s) = hc (s)e(s), (4.1)
where the controller input deviation, e(s), is given by
e(s) = r(s) − y(s), (4.2)
and the controller transfer function is given by
1 1 + Ti s
hc (s) = Kp (1 + ) = Kp . (4.3)
Ti s Ti s

4.1.2 Continuous Time domain description of the PI controller

In order to simulate and implementing the PI controller a state space formulation

of the controller is needed. From (4.1) and (4.3) we have that
Kp 1
u(s) = e(s) + Kp e(s). (4.4)
Ti s
The PI controller have one internal state. And a state space model can be formulated
in at least three ways. Define z as the controller state. Then we can chose the state
in the following ways:

1. Choosing
Kp
z= e(s) =⇒ u(s) = z(s) + Kp e(s), (4.5)
Ti s
gives the following continuous state space model formulation
Kp
ż = e, (4.6)
Ti
u = z + Kp e. (4.7)
4.1 The PI controller 72

2. Choosing
Kp 1
z= e(s) =⇒ u(s) = z(s) + Kp e(s), (4.8)
s Ti
gives the following continuous state space model formulation
ż = Kp e, (4.9)
1
u = z + Kp e. (4.10)
Ti
3. Choosing
1 Kp
z = e(s) =⇒ u(s) = z + Kp e, (4.11)
s Ti
gives the following continuous state space model formulation
ż = e, (4.12)
Kp
u = z + Kp e. (4.13)
Ti
Hence, three different continuous state space formulations for the PI controller can be
formulated. Alternative one given by (4.6) and (4.7) is the most common alternative
since the controller state, z, has a more intuitive explanation because z = u in steady
state. In steady state, that is when t → ∞ and the system is stable then we have
that ż = 0 which gives e = r − y = 0 and hence z = u. The approach of assuming
stability and putting the derivatives of the states equal to zero when t → ∞ is a
very simple way of steady state analysis of a system.
Using that the solution of a linear differential equation is as presented in Eq.
(1.10) we may also formulate the PI controller by using the that the solution of the
differential Eq. (4.6) is
Kp t
Z
z = z(t0 ) + edt, (4.14)
Ti t0
and then
t
Kp
Z
u = Kp e + u(t0 ) + edt, (4.15)
Ti t0

since z(t0 ) = u(t0 ).

4.1.3 Discrete Time domain description of the PI controller

In order to implement the PI controller in a digital computer we use an explicit

Euler approximation of, ż, i.e.,
zk+1 − zk
ż ≈ (4.16)
∆t
where ∆t > 0 is the sampling interval, in order to develop a discrete time state space
model for the controller as follows
Kp
zk+1 = zk + ∆t ek , (4.17)
Ti
uk = zk + Kp ek . (4.18)
4.1 The PI controller 73

Example 4.1 (Inverse response and model reduction with the half rule)
Given a system modeled by the transfer function model y = hp (s)u where

1 − 2s
hp (s) = (4.19)
(s + 1)(s + 2)

Simulate the system with a PI feedback system with settings Kp = 0.56 and Ti = 1.25.
Write a MATLAB script for the solution.

% ex1_pi.m
% Purpose: Illustrate the implementation of a PI controller for
% controlling the process, h_p(s)=(1-2s)/(s+1)(s+2)
%
h_p=tf([0,-2,1],[1,3,2]); % Process transfer function
sys=ss(h_p); % State space model
A=sys.a; B=sys.b; D=sys.c;

Ti=1.25; Kp=0.56; % PI controller parameters

r=1; % The reference signal

Dt=0.1; t=0:Dt:20; % Time horizon for simulation

N=length(t);

hc=Dt*Kp/Ti; % Controller param. to save computing time

z=0; x=[0;0]; % Initial values for the states
for k=1:N
y=D*x; % The measurement output
e=r-y; % The controller input

u=z+Kp*e; % The controller output

z=z+hc*e; % Updating the controller state

Y(k,1)=y; U(k,1)=u; % Storing variables

x=x + Dt*(A*x+B*u); % Puting the control to the process

end

plot(t,Y), grid, xlabel(’Time’), ylabel(’y(t)’)

title(’Output response of PI controlled system’)

Example 4.2 (PI control of a Chemical Reactor)

A chemical isothermal reactor (Van de Vusse) is studied in this example. The rela-
tionship from the feed flow rate u into the reactor to the concentration of the product
y at the outlet of the reactor is modeled by the following non-linear state space model.

x˙1 = −k1 x1 − k3 x21 + (v − x1 )u, (4.20)

x˙2 = k1 x1 − k2 x2 − x2 u, (4.21)
y = x2 , (4.22)
4.1 The PI controller 74

where the reaction rate coefficients are given by k1 = 50, k2 = 100, k3 = 10. The
concentration of the by-product into the reactor, v, is treated as an unknown constant
or slowly varying disturbance with nominal value v s = 10. Choosing a steady state
control us = 25 gives the steady states xs1 = 2.5 and y s = xs2 = 1.
A linearized model gives the transfer function model y = hp (s)u where
1 − τs
hp (s) = K (4.23)
(1 + T s)2
2 1
with gain K = 125 , inverse response time constant τ = 250 and a double time
1
constant T = 125 .
We want to use a PI controller so we are approximating the 2nd order model
with a 1st order time delay model (using the half rule for model reduction)
1 − τs
hp (s) = K (4.24)
1 + Ts
where the new model parameters are T := T + 12 T = 250
3 1
≈ 83.3 , new inverse response
1 1
time constant τ := τ + 2 T = 125 and the gain K is unchanged.
The SIMC PI controller parameters using the simple choice Tc = τ gives Kp =
1 T 1
2K τ ≈ 62.5 and Ti = T = 75 .
The chemical reactor is controlled with a PI controller with these tuning param-
eters and implemented in the following MATLAB m-file. Se Figure 4.1 for a step
response simulation.

% main_reacsim_pid.m
% Function to simulate PI control of a non-linear
% chemical reactor model using the explicit Euler method.
% Purpose: Supplement to Exercise 3

clear all
global u
% Nominal variables
u=25;
x=[2.5;1];

dt=0.0005; % Sampling interval

tf=0.2; % Final time instant
t=0:dt:tf; % Time horizon
N=length(t); % Number of discrete time instants

% PI controller parameters
Kp=62.5; Ti=1/125;
Kp=46.9; Ti=1/83.3; % SIMC tuning rules
Td=Ti/3;% Test

z=u; % Initial controller state

r=1.0; % Step change in r from 1.0 to 1.02 at t=0
4.2 The PID controller 75

% Vectors to store the variables and allocate space

U=zeros(N,1); Y=zeros(N,1); X=zeros(N,2); R=zeros(N,1);

% The main simulation loop using explicit Euler method

y_old=x(2);
for k=1:N
y=x(2); % The measurement

if k >= 41; r=1.02; end

% The PI controller
e=r-y;
u=Kp*e+z - Kp*Td*(y-y_old)/dt; y_old=y;
z=z+dt*Kp*e/Ti;

U(k,1)=u; % Save u
Y(k,1)=y; % Save output y
X(k,:)=x’; % Save the states
R(k,1)=r; % The reference r

fx=fx_chemreac(0,x); % Calculate the right hand side in dot(x)=f(x,u). u enters as gl

x=x+dt*fx; % Calculate x_(k+1) using explicit Euler
end

% Plotting results
figure(2)
subplot(311), plot(t,U), grid, title(’Control input u’)
subplot(312), plot(t,X(:,1)), grid, title(’State x_1’)
subplot(313), plot(t,Y,’-r’,t,R,’b’), grid, title(’Output y=x_2 and reference r’)
xlabel(’Continuous time 0 \leq t \leq t_f’)
% END main_reacsim_pid.m

4.2 The PID controller

4.2.1 Frequency domain description of the PID controller

The PID controller can be written as follows

u(s) = hc (s)e(s), (4.25)
where the controller input deviation, e(s), is given by
e(s) = r(s) − y(s), (4.26)
and the so called cascade (series) controller transfer function is given by
1 + Ti s
hc (s) = Kp (1 + Td s). (4.27)
Ti s
4.2 The PID controller 76

Control input u
27

26.5

26

25.5

25
0 0.02 0.04 0.06 0.08 0.1 0.12 0.14 0.16 0.18 0.2

State x1
2.65

2.6

2.55

2.5
0 0.02 0.04 0.06 0.08 0.1 0.12 0.14 0.16 0.18 0.2

Output y=x2 and reference r

1.04

1.02

0.98
0 0.02 0.04 0.06 0.08 0.1 0.12 0.14 0.16 0.18 0.2
Continuous time 0 ≤ t ≤ t
f

Figure 4.1: Simulated responses from the chemical reactor example in Example 4.2.
The Figure is generated by the above m-file main reacsim pid.m

Note that this can be expressed as

Ti + Td 1 Ti + Td 1 Ti
hc (s) = Kp ( + + Td s) = Kp (1 + + Td s).(4.28)
Ti Ti s Ti (Ti + Td )s Ti + Td

Note that the cascade formulation (4.27) is equivalent to the ideal form PID con-
troller with transfer function
1
hc (s) = Kp0 (1 + + Td0 s). (4.29)
Ti0 s

in which
Ti + Td Td
Kp0 = Kp = Kp (1 + ) (4.30)
Ti Ti

Td
Ti0 = Ti + Td = Ti (1 + ) (4.31)
Ti

Ti 1
Td0 = Td = Td (4.32)
Ti + Td 1 + TTd i

Often the derivative time, Td , is much smaller than the integral time, Ti . Hence, in
this case Kp ≈ Kp0 , Ti ≈ Ti0 and Td ≈ Td0 .
4.3 Anti windup and constraints 77

4.2.2 Continuous Time domain description of the PID controller

The ideal form of the PID controller output can be written as follows
Kp
u(s) = Kp e(s) + e(s) + Kp Td se(s), (4.33)
Ti s
where e(s) is the controller input, usually e = r − y. Choosing
Kp
z= e(s), (4.34)
Ti s
gives the time domain description of the PID controller
Kp
ż = e (4.35)
Ti
u = Kp e + z + Kp Td ė (4.36)

4.2.3 Discrete Time domain description of the PID controller

Similar as in the discrete implementation of the PI controller we now need an discrete

approximation for, ė. At the present discrete time, k, only yk and past outputs,
yk−1 , yk−2 , · · · and so on is available. Hence, it is reasonable to use the following
approximation
ek − ek−1 rk − yk − (rk−1 − yk−1 )
ė ≈ = , (4.37)
∆t ∆t
where ∆t is the sampling time and where we have used that e = r − y. In order not
to get abrupt changes in the controller output, u, for step changes in the reference
we usually are assuming that the reference signal is constant when implementing
the PID controller, such that rk − rk−1 = 0. Hence, we are using
yk − yk−1
ė ≈ − . (4.38)
∆t
Hence, we have the following discrete implementation of the PID controller
Kp
zk+1 = zk + ∆t ek , (4.39)
Ti
yk − yk−1
uk = zk + Kp ek − Kp Td (4.40)
∆t
For noisy systems in which the output can be modelled as y = Dx + w where w
is white noise (Normally gaussian distributed random numbers) a large derivative
time constant, Td , will increase the noise level on the control input, u. A rule of
thumb is to chose a small Td .

4.3 Anti windup and constraints

In practical implementations of the PI and PID controller there may be reasons for
handling constraints on the control input, u. These constraints may be physical
4.3 Anti windup and constraints 78

limitations of the controller organ. Hence, the practical and physical control input,
u(t), should lie within hard bound constraints as

umin ≤ u(t) ≤ umax . (4.41)

When the controller output, u(t), is within the bounds (4.41) then we are using the
controller state equation

u = Kp e + z + Kp Td ė (4.42)
Kp
ż = e. (4.43)
Ti
However, when the controller are saturated, i.e. when the controller output, u, are
outside bounds, i.e. when

u < umin , (4.44)

then we are using

u = umin (4.45)
ż = 0, (4.46)

and when

u > umax , (4.47)

then we are using

u = umax (4.48)
ż = 0, (4.49)

in order not to integrate z when the controller are saturated. Note that the control
input, u, is computed at all time instants by Equation (4.42) and checked if it is
within bounds. This is the principle of Anti windup and it is implemented in the
computer with pure logic.

Example 4.3 (PI controller implementation: anti windup and constraints)

Given the same system as in Example 4.2. In this example we have the additional
constraints that

0 ≤ u ≤ 1.75 (4.50)

A MATLAB script with anti windup and constraints implementation is presented in

the following.

% ex1_pi_anti.m
% Purpose: Illustrate the implementation of a PI controller for
% controlling the process, h_p(s)=(1-2s)/(s+1)(s+2)
% -Implementation of Anti-Windup
%
4.3 Anti windup and constraints 79

h_p=tf([0,-2,1],[1,3,2]); % Process transfer function

sys=ss(h_p); % State space model
A=sys.a; B=sys.b; D=sys.c;

Ti=1.25; Kp=0.56; % PI controller parameters

r=1; % The reference signal

Dt=0.1; t=0:Dt:20; % Time horizon for simulation

N=length(t);

v=0.7; u_max=1.5; u_min=0;

z=0; x=[0;0]; % Initial values for the states
for k=1:N
y=D*x; % The measurement output
e=r-y; % The controller input

u=z+Kp*e; % The controller output

Z(k,1)=z;

if u>u_max % Handeling constraints and Anti-windup

u=u_max;
z=z;
%z=u-Kp*e;
%z=z+Dt*Kp*e/Ti +Dt*(u_max-u)/Ti;
elseif u<u_min;
u=u_min;
z=z;
else
z=z+Dt*Kp*e/Ti; % Updating the controller state
end

Y(k,1)=y; U(k,1)=u; % Storing variables

up=u+v;
x=x + Dt*(A*x+B*up); % Puting the control to the process
end

U_max=u_max*ones(size(t’));
subplot(311), plot(t,Y), grid, xlabel(’Time’), ylabel(’y(t)’)
title(’Output response of PI controlled system’)
subplot(312), plot(t,[U U_max]), grid, xlabel(’Time’), ylabel(’u(t)’)
subplot(313), plot(t,Z), grid, xlabel(’Time’), ylabel(’z(t)’)
4.4 Bumpless transfer 80

4.4 Bumpless transfer

4.4.1 Bumpless transfer between manual and automatic mode

The problem or need for a bumpless transfer between manually process operations
and closed loop feedback control is of practical importance. This topic will be
discussed in this section. It is important to make sure that the manually setting for
the control coincide with the controller output at the time of switching.
Suppose that the process is operating manually in open loop with a control input,
uoperator , specified by the operator, i.e. the process input is

u = uoperator (4.51)

and that at a specified time instant, t = ton , we want to switch to closed loop
feedback PID control.

u = Kp e + z + Kp Td ė (4.52)
Kp
ż = e (4.53)
Ti
As we have seen, the PI and PID controllers need an initial value for the controller
state, z. The change from manually to feedback control will be bumpless if the
initial controller state is computed by

z = uoperator − Kp e − Kp Td ė (4.54)

when t = ton and switching to closed loop PID feedback control. At time instants
ton < t we use the controller equations (4.58) and (4.59).
In order to switch bump-less from manually to automatic control we can use the
following scheme.

if t = ton

u = uoperator (4.55)
z = uoperator − Kp e − Kp Td ė (4.56)
(4.57)

else

u = Kp e + z + Kp Td ė (4.58)
Kp
ż = e (4.59)
Ti

% ex1_pi_bump.m
% Purpose: Illustrate the implementation of a PI controller for
% controlling the process, h_p(s)=(1-2s)/(s+1)(s+2)
% -Implementation of Anti-Windup
%
4.4 Bumpless transfer 81

h_p=tf([0,-2,1],[1,3,2]); % Process transfer function

sys=ss(h_p); % State space model
A=sys.a; B=sys.b; D=sys.c;

Ti=1.25; Kp=0.56; % PI controller parameters

r=1; % The reference signal

Dt=0.1; t=0:Dt:20; % Time horizon for simulation

N=length(t);
t_on=10;

v=0.5; % Constant disturbance

Hd=-D*inv(A)*B; % Steady state gain
us=inv(Hd)*(r-Hd*v); % Steady state control
xs=-inv(A)*B*(us+v); % Steady state

u_max=1.75; u_min=0;
x=xs; % Initial values for the states
for k=1:N
y=D*x; % The measurement output
e=r-y; % The controller input 9

if k <= t_on % Manual control

u=us;
elseif k==t_on % Swich to automatic for k>=t_on
z=us-Kp*e;
u=z+Kp*e;
else % Automatic control

u=z+Kp*e; % The controller output

Z(k,1)=z;

if u>u_max % Handeling constraints and Anti-windup

u=u_max;
%z=z;
z=u-Kp*e;
elseif u<u_min;
u=u_min;
z=z;
else
z=z+Dt*Kp*e/Ti; % Updating the controller state
end
end
Y(k,1)=y; U(k,1)=u; % Storing variables

up=u+v;
x=x + Dt*(A*x+B*up); % Puting the control to the process
4.4 Bumpless transfer 82

end

U_max=u_max*ones(size(t’));
subplot(311), plot(t,Y), grid, xlabel(’Time’), ylabel(’y(t)’)
title(’Output response of PI controlled system’)
subplot(312), plot(t,[U U_max]), grid, xlabel(’Time’), ylabel(’u(t)’)
subplot(313), plot(t,Z), grid, xlabel(’Time’), ylabel(’z(t)’)

4.4.2 Bumpless transfer between PID parameter changes

Chapter 5

Time delay

5.1 Padé approximations to the exponential for eθ

Approximations to the exponential eθ are used in many circumstances. They are

among others used in order to compute approximations to the transition matrix
eA∆t and for discretizations of continuous time state space models. Another topic
is to obtain rational approximations to the time delay e−τ s .
We will in this section present the theory behind what we call Padé approxima-
tions to il eθ . The starting point is the following series expansion of the exponential
1 1
eθ = 1 + θ + θ2 + θ3 + hot (5.1)
2 6
where hot is short for “higher order terms”.

5.1.1 Developing a 1st order Padé approximation

Assume that we want to approximate eθ with

1 + b1 θ
eθ ≈ (5.2)
1 + a1 θ
where b1 and a1 are scalar coefficients. Equation (5.2) is referred to as a 1st order
Padé approximation. Some times also the term (1, 1) Padé approximation is used
about this approximation. The reason for this is thet this approximation have a 1st
order nominator and a 1st order denominator polynomial in θ.
We may now find the coefficients a1 and b1 such that the error between (5.1)
and (5.2) is minimized, i.e. to minimize the error ε = eθ − (1 + b1 θ)/(1 + a1 θ) with
respect to a1 and b1 . In order to find b1 and a1 we need two equations. Those
equations may be found by putting the coefficients in the terms for θ and θ2 in a
series expansion for the error ε to specified values, e.g. eual to zero. The error term
wil by this approach be of 3rd order, i.e. the error is proportional to θ3 .
The method we now are to present is based on that eθ may be written as
1 + b1 θ
eθ = + c3 θ3 + hot (5.3)
1 + a1 θ
5.1 Padé approximations to the exponential for eθ 84

where hot stands for “higher order terms”. Equation (5.3) may be proved by series
expansion of (1 + b1 θ)/(1 + a1 θ) by using long division and combining this with a
series expansion for eθ . This gives the error
1 + b1 θ
ε = eθ − = c1 θ + c2 θ + c3 θ3 + hot (5.4)
1 + a1 θ
If we chose c1 = 0 and c2 = 0 in Equation (5.4) we deduce Equation (5.3). We
assume that there eqists a1 and b1 which gives c1 = 0 and c2 = 0.
Putting (5.1) in the left hand side of Equation (5.3) and multiplying on both
sides with 1 + a1 θ gives,
1 1
(1 + a1 θ)(1 + θ + θ2 + θ3 + hot) = 1 + b1 θ + c3 θ3 + a1 c3 θ4 + hot (5.5)
2 6
This last expression may be written as a polynomial, i.e. with terms of θ, θ2 , θ3 ,
and so on, Vi får
1 1 1
(1 + a1 − b1 )θ + ( + a1 )θ2 + ( + a1 − c3 )θ3 + hot = 0 (5.6)
2 6 2
where we have included the term a1 c3 θ4 in the hot term. We may now find three
equations by putting the coefficients in front of θ, θ2 and θ3 equal to zero. This
gives the three equations

1 + a1 − b1 = 0 (5.7)
1
+ a1 = 0 (5.8)
2
1 1
+ a1 − c3 = 0 (5.9)
6 2
which gives

a1 = − 12 b1 = 1
2
1
c3 = − 12 (5.10)

We have now deduced the following approximation

1 + 12 θ
eθ ≈ Φ1,1 (θ) = (5.11)
1 − 12 θ
1 3
The error in this approximation is given by c3 θ3 = − 12 θ . This is usually a good
approximation when θ < 1.
Equation (5.11) is referred to as a (1, 1) ( ore 1st order) Padé approximation to
the exponential eθ because both the nominator and the denominator to the approx-
imation is of 1st order.

5.1.2 Alternative prof of the 1st order Padé approximation

we will in this section present an alternative development of the 1st order Padé
approximation. Assume that eθ are to be approximated with
1 + b1 θ
eθ ≈ (5.12)
1 + a1 θ
5.1 Padé approximations to the exponential for eθ 85

where b1 and a1 are two scalar coefficients.

By using long-division ore from a formula collection book we find the series for
(1 + x)−1 as
1
= 1 − x + x2 − x3 + x4 − · · · (5.13)
1+x
Note that this equation may be proved by using long-division. Using this formula
we find
1 + b1 θ
= (1 + b1 θ)(1 − a1 θ + a21 θ2 − a31 θ3 + · · · ) (5.14)
1 + a1 θ
This gives
1 + b1 θ
= 1 + (b1 − a1 )θ + a1 (a1 − b1 )θ2 + (a21 (b1 − a1 ) + c3 )θ3 − c3 θ3 + · · ·(5.15)
1 + a1 θ
Note that Equation (5.15) may be proved directly by using long-division.
We now demand that the series for eθ given by Equation (5.1) should be as equal
to the series for equation (5.15) as possibile. We therefore are putting the coefficients
in the terms θ and θ2 equal to the coefficients in the corresponding coefficients in
Equation (5.1). The coefficients in the θ3 term gives an expression for the main
error. Rhis gives

b1 − a1 = 1 (5.16)
1
a1 (a1 − b1 ) = (5.17)
2
1
a21 (b1 − a1 ) + c3 = (5.18)
6
Which gives a1 = − 21 , b1 = 1 + a1 = 12 and c3 = 61 − a21 = − 12
1
. Note that we have
used Equation (5.15) in order to find the coefficient c3 in the error term. The result
is
1 + 12 θ
eθ ≈ Φ1,1 (θ) = (5.19)
1 − 12 θ

which is equivalent with (5.11). The error in the approximation is given by

1 + 12 θ
ε = eθ − = c3 θ3 + hot (5.20)
1 − 12 θ

5.1.3 Developing a (1, 0) Padé approximation

We start the discussion by the following example of using long-division.

Example 5.1 (long-division)

We will shortly illustrate long-division of the term 1/(1 + x). We have
1 1+x−x
= (5.21)
1+x 1+x
5.1 Padé approximations to the exponential for eθ 86

which is equivalent with

1 1
=1−x (5.22)
1+x 1+x
By repetition use of Equation (5.22) we obtain
1 1 1
= 1 − x(1 − x ) = 1 − x + x2 (5.23)
1+x 1+x 1+x
Putting again equation (5.22) on the right hand side gives
1 1 1
= 1 − x + x2 (1 − x ) = 1 − x + x2 − x3 (5.24)
1+x 1+x 1+x
Continuation of this process gives Equation (5.13).

From Example 5.1 we have that long-division of 1/(1 + aθ) gives

1
= 1 − aθ + a2 θ2 − a3 θ3 + hot (5.25)
1 + aθ
We find the following expression for the error
1 1 1
ε = eθ − = (1 + a)θ + ( − a2 )θ2 + ( + a3 )θ3 + hot (5.26)
1 + aθ 2 6
The error is minimized by choosing a = −1. We therefore have the following ap-
proximation
1
eθ ≈ Φ1,0 (θ) = (5.27)
1−θ

where the main error is (1/2 − a)θ2 = −1/2θ2 . Φ1,0 (θ) is referred to as a (1, 0) Padé
approximation, because the approximation have a 1st order denominator and a 0
order nominator polynomial in θ.

5.1.4 (s, t) Padé approximations

We will in this section introduce the notation Φs,t (θ) as a short for an (s, t) Padé
approximation to the exponential eθ , i.e., a t-order nominator polynomial and a
s-order denominator polynomial. Hence, in general we may express eθ as

eθ = Φs,t (θ) + cs+t+1 θs+t+1 + hot (5.28)

where
1 + b1 θ + · · · + bt θt
Φs,t (θ) = (5.29)
1 + a1 θ + · · · + as θs

and Φs,t (θ) is defined as a (s, t) Padé approximation to eθ .

Similar to the developement of the (1, 0) and (1, 1) Padé approximations we may
find the following (s, t) Padé approximations.
5.1 Padé approximations to the exponential for eθ 87

(s,t) Φs,t (θ) Main error

1 2
(0,1) 1+θ 2θ
(0,2) 1 + θ + 12 θ2 1 3
6θ
1
(1,0) 1−θ − 12 θ2
1+ 12 θ 1 3
(1,1) 1− 21 θ
− 12 θ
1+ 12 θ+ 12
1 2
θ 1 5
(2,2) 1 1 2
1− 2 θ+ 12 θ
− 720 θ

Padé approximations are used in many circumstances. Of particular interest in

control-theory is the problem of obtaining discrete state space models from contin-
uous time state space models, and for simulation of dynamic systems.
Another important application of Padé approximations is to obtain rational ap-
proximations to the time delay in the frequency domain, e−τ s , which is a Laplacian
description of a time delay. Padé approximations for e−τ s will be presented in the
next section.
Common methods for simulations of dynamic systems is the explicit Euler method,
implicit Euler method and the trapezoid method. All those methods may be deduced
from the Padé approximations. This will be illustrated in the following Example 5.2.

Example 5.2 (Discretization)

Consider a simple dynamic system
ẋ = −ax (5.30)
An exact discrete time model is given by
xk+1 = Φxk (5.31)
−a∆t
Φ = e (5.32)
where ∆t is the step length.
The explicit Euler method:
Using the explicit Euler method gives
xk+1 − xk
≈ −axk (5.33)
∆t
This gives
xk+1 = Φ0,1 xk (5.34)
Φ0,1 = 1 − a∆t (5.35)
Note that the (0, 1) Padé approximation for Φ = e−a∆t is identical to using the
explicit Euler method for discretization.
Implicit Euler method:
Using the implicit Euler method for discretization gives
xk+1 − xk
≈ −axk+1 (5.36)
∆t
This gives
xk+1 = Φ1,0 xk (5.37)
−1
Φ1,0 = (1 + a∆t) (5.38)
5.2 Padé approksimasjoner for e−τ s 88

Note that the approximation to the transition matrix we have found by using implicit
Euler method is identical to a use (1, 0) Padé approximation to Φ.
Trapes metoden:
Vi benytter Trapes metoden for diskretisering og får
xk+1 − xk 1
≈ (−axk+1 − axk ) (5.39)
∆t 2
Dette gir

xk+1 = Φ1,1 xk (5.40)

(5.41)

∆t
1− 2 a
Φ1,1 = ∆t
(5.42)
1+ 2 a

Denne approksimasjonen, dvs. Trapes metoden for diskretisering, er identisk med

(1, 1) Padé approksimasjonen til Φ.
Fordelen med Trapes metoden og implisitt Eulers metode for diskretisering er at
den resulterende diskrete modellen er numerisk stabil for alle ∆t > 0, dvs. skrit-
tlengden kan velges vilkårlig stor. Dette er ikke tilfelle med eksplisitt Eulers metode
for diskretisering.

5.2 Padé approksimasjoner for e−τ s

Av spesiell interesse i reguleringsteknisk sammenheng er Φ1,1 (−τ s) og Φ2,2 (−τ s)

approksimasjonene til transferfunksjonen e−τ s , som beskriver en transportforsinkelse
i Laplace-planet.
Φ1,1 (−τ s) approksimasjonen refereres også til som en 1.ordens Padé approksi-
masjon fordi approksimasjonen har 1. ordens teller og 1. ordens nevner polynom
i s. En 1. ordens Padé approksimasjonen finner vi fra ligning (5.11) ved å sette
θ = −τ s.
1. og 2. ordens Padé approksimasjoner er gitt ved

1 − 12 τ s
e−τ s ≈ (1. ordens Padé approksimasjon) (5.43)
1 + 12 τ s

1 − 12 τ s + 1 2 2
12 τ s
e−τ s ≈ (2. ordens Padé approksimasjon) (5.44)
1 + 21 τ s + 1 2 2
12 τ s

der s er Laplace-plan operator og τ er transportforsinkelsen.

Det kan vises at en n. ordens Padé approksimasjon for transferfunksjonen e−τ s
kan skrives slik.
1 − a1 s + a2 s2 − a3 s3 + · · · + (−1)n an sn
e−τ s ≈ (5.45)
1 − b1 s + b2 s2 − b3 s3 + · · · + bn sn
5.3 Balchen approksimasjoner for e−τ s 89

5.3 Balchen approksimasjoner for e−τ s

I Balchen (1990) er det utledet approksimasjoner til transferfunksjonen e−τ s som

i mange tilfeller gir bedre resultater for bruk i forbindelse med stabilitetsanalyse
sammenlignet med tilsvarende Padé approksimasjoner.
Noen Balchen approksimasjoner er som følger

1 − π2 τ s
e−τ s ≈ (1. ordens Balchen approksimasjon) (5.46)
1 + π2 τ s

3 1 2 2
−τ s 1− 2π τ s + π2
τ s
e ≈ 3 1 2 2 (2. ordens Balchen approksimasjon) (5.47)
1+ 2π τ s + π2
τ s

1 − 0.504τ s + 0.1013τ 2 s2 − 0.0108s3

e−τ s ≈ (3. ordens Balchen approksimasjon)
(5.48)
1 + 0.504τ s + 0.1013τ 2 s2 + 0.0108s3
der s er Laplace-plan operator og τ er transportforsinkelsen.
Chapter 6

Feedback systems

6.1 Description of feedback systems

v
 Prosess
?
Hd w
Måling
ys
d +? +?
r - m - u y- m y - m y-
m
Hc - H
p
- H
m +
+ − +
6

Figure 6.1: Standard feedback system. A model description of the control is included
in Hp . Normally H : m = I in this section. r, is the reference, u is the control, v
is the disturbance, y is the process output, w is measurements noise and ym is the
process output measurements.

We are in this section going to investigate and describe the fundamental prop-
erties of a feedback system. The description is given for multiple input and output
(MIMO) systems. We are normally using capital letters for transfer matrices in
MIMO (multivariable) systems and lowercase letters for transfer functions in SISO
(monovariable) systems.
Consider a system described by the linear continuous time system state space
model

ẋ = Ax + Bu + Cv (6.1)
y = Dx (6.2)

where y is a vector of output variables, u is a vector of control variables and x is

a vector of state variables. v is a vector of disturbance variables (disturbances).
6.1 Description of feedback systems 91

However, note that v may contain variables not used as control variables. v may
contain both measured and unmeasured variables.
The measurement, ym , of the process output, y, is described by

ym = y + w (6.3)

where we for simplicity have used Hm = I. Se figure 6.1. w is a vector of measure-

ments noise. This is usually a high frequence noise.
Note that in feedback systems as shown in Figure 6.1 we usually want to control
y and not the measurement ym .
From the state space model description of the process we find the following
Laplace plane model description of the process
Hp Hd
z }| { z }| {
y = D(sI − A)−1 B u + D(sI − A)−1 C v (6.4)

This model with Hm = I is illustrated in Figure 6.1. Comparing the notation used
in Figure 6.1 (i.e. y = y d + y s ) we find that

y d = Hp u (6.5)
s d
y = H v (6.6)
d s
y = y +y (6.7)

As we see, the process output y can be divided into two parts, or influenced from
two contributions. One contribution y d from the manipulable control variables u
and one contribution y s driven from the disturbances v. y d may be defined as the
deterministic (known) contribution to y, i.e. the contribution driven from the known
control variables u. y s is denoted the stochastic (high frequency) contributions to y,
i.e. contributions driven from unmeasured and measured stochastic disturbances v.
Figure 6.1 is also illustrated a controller of the form

u = Hc (r − ym ) = Hc (r − y − w) (6.8)

Combining Equation (6.4) and (6.8) gives

y = yd + ys
= Hp Hc (r − y − w) + Hd v
⇓
(I + Hp Hc )y = Hp Hc r + Hd v − Hp Hc w (6.9)

We find that the response of the closed loop system is given by

T S y s T
z }| { z }| { z}|{ z }| {
−1 −1 −1
y = (I + Hp Hc ) Hp Hc r + (I + Hp Hc ) Hd v − (I + Hp Hc ) Hp Hc w (6.10)

where we have marked the central closed loop system transfer matrices, T , and S.
6.1 Description of feedback systems 92

The following definitions are commonly used in control theory

H0 = Hp Hc loop transfer matrix (function)

S = (I + Hp Hc )−1 Sensitivity function
(Norwegian: avviksforhold og sensitivitets-funksjon)(6.11)
T = (I + Hp Hc )−1 Hp Hc complementary sensitivity function
(Norwegian: følgeforhold)

We see that S is the closed loop transfer function (matrix) from y s = Hd v to the
output y. y s is the output from the disturbance model Hd and the contribution
to the output y which is influenced from the disturbance v. y s have a direct and
additive influence on the process output y. It is natural to demand S to be small in
order for the disturbance v to have a small influence on the process output.
T is the closed loop transfer function (matrix) from the reference r to the output
y. For good tracking behavior it will be natural to demand T ≈ I. Then we will
have that y ≈ r when S ≈ 0.
It is important to note the following relationship between S and T , i.e.,

S+T =I (6.12)

This relationship, equation (6.12). can be proved as follows

S+T = (I + H0 )−1 + (I + H0 )−1 H0

⇓
S+T = (I + H0 )−1 (I + H0 ) = I QED.

where we have used the expressions for S and T as given in (6.11).

We will in the following discuss the importance of the relationship S + T = I.
We have three signals involved in the control system, i.e., r, y s = Hd v and w, which
influences upon the process output y. we have that

y = T r + Sy s − T w (6.13)

Ideally we want that y = r. This is the case if T = I, S = 0 and w = 0. This

indicates that we have a problem if there are much measurements noise on y, i.e.
when w 6= 0.
The relationship S +T = I is important because it shows that the design problem
of a feedback system is a trade off between on the one side, good tracking properties
(fast response from r to y and good tracking when T ≈ I) and on the other side,
ability to reduce the influence of the disturbances w upon y which is obtained when
T ≈ 0.
From the identity S + T = I we see that these demands is in conflict with each
other.
The reason for this is that the measurements noise w have the same influence
upon the output y as the reference r. This means that the transfer function from
the measurements noise w to the process output y is given by −T (s). We also have
that the transfer function from the reference r to the process output y is given by
T (s).
6.1 Description of feedback systems 93

Ideally we want S small in order to have a small influence upon the output y
from the disturbance y s . We also usually want T small in order for the influence
from the noise w upon the output y to be small. We see from the identity (6.12) that
this hardly may be fulfilled. Those demands is in conflict with each other because
we have that S + T = I.
It is common practice to demand the feedback system to have good properties
to reduce the influence from the disturbances y s = Hd v upon the output y which
is the case when S ≈ 0). We also want good tracking properties. This means that
we usually want S ≈ 0 og T ≈ I. From the identity S + T = I we see that this is
possible, at least for some frequencies. However, there are at the same time difficoult
to reduce the influence on y from measurements noise. In steady state, when s = 0,
it is usually easy to obtain S = 0 and T = I. See Figure 6.2 for typically amplitude
and phase behavior of H0 , S and T as a function of the frequency ω. Nothe that
s = jω.

Remark 6.1 Let us have a look on the expression for the control u which influences
the system. We have from Figure 6.1 that

u = Hc (r − y − w) (6.14)

We also have that the output is described by

y = T r + SHd v − T w (6.15)

Putting this into the expression for the control gives

u = Hc Sr − Hc SHd v − Hc Sw (6.16)

where we have used that I − T = S and T − I = −S

Remark 6.2 Note that if the system is monovariable (i.e. when y, r, v, w and u
are scalar variables) then we have that
T S
z }| { z }| {
hp hc 1 hp hc
y= r+ hd v − w (6.17)
1 + hp hc 1 + hp hc 1 + hp hc

Note that all terms have the same nominator 1 + hp hc . We define the character-
istic equation of the system as

1 + hp hc = 1 + h0 = 0 (6.18)

The characteristic equation describes the stability properties of the system, here the
stability of the closed loop system. The term 1 + h0 = 0 gives us a polynom in the
Laplace operator s, the characteristic polynomial. If h0 (s) contains transprt delay
terms of the type e−τ s then we can use approximations to the exponential. The
roots of this characteristic polynomial is the poles of the closed loop system. The
closed loop poles should lie in the left part of the complex plane in order for the
closed loop system to be stable.
6.1 Description of feedback systems 94

This indicates that stability is a property of the system and not influenced of
the external signals r, v and w. The reason for this is that there only is the process
model hp and the controller hc which influences the characteristic polynomial and
thereby the stability.

Remark 6.3 Note the following alternative expressions for the complimentary sen-
sitivity function given in (6.11). Equation (6.11) is rewritten for the sake of com-
pleteness.
T = (I + Hp Hc )−1 Hp Hc (6.19)
−1
T = Hp Hc (I + Hp Hc ) (6.20)
−1
T = Hc (I + Hc Hp ) Hp (6.21)
Equation (6.20) can be proved as follows: from Figure 6.1 we have )with v = 0,
w = 0 and Hm = I)
e = r − y = r − Hp Hc e
⇓ (6.22)
e = (I + Hp Hc )−1 r
Putting this into the expression y = Hp Hc e gives
T
z }| {
y = Hp Hc (I + Hp Hc )−1 r (6.23)
Equation (6.21) can be proved as follows: From Figure 6.1 we have that (with v = 0,
w = 0 and Hm = I)
u = Hc e = Hc (r − y) = Hc r − Hc Hp u
⇓ (6.24)
u = (I + Hc Hp )−1 Hc r
Putting this into the expression y = Hp u we obtain
T
z }| {
y = Hc (I + Hc Hp )−1 Hp r (6.25)

Remark 6.4 As shown above, the output y of a linear system can be divided into
two contributions. This can also be shown from the state space model given in (6.1)
and (6.2).
deterministisk part stochastisk part
z }| { z }| {
ẋd = Axd + Bu ẋs = Axs + Cv
y d = Dxd y s = Dxs (6.26)
x= xd+ xs
y = yd + ys
This follows from the properties of linear systems.

Remark 6.5 In some circumstances symbols N (s) for S(s) is used and M (s) for
T (s) is used.
6.1 Description of feedback systems 95

Amplitude of S T and H_0

1
10
H_0
Log Magnitude

0 T S
10

−1
10 S

−2
10 −2 −1 0 1
10 10 10 10

Phase of S T and H_0

100
S
Phase (degrees)

0 T

−100 H_0

−200

−300 −2 −1 0 1
10 10 10 10
Frequency (radians/sec)

Figure 6.2: The figure shows typical amplitude and phase frequency plot for S,
T og H0 . The frequency response is obtained by putting s = jω in the transfer
function. The process model used here is hp (s) = 3(1 − 2s)/(5s + 1)(10s + 1), with
PI-controller hc (s) = Kp (1 + Ti s)/(Ti s) with Ti = 12.7 and Kp = 1.136 (Ziegler-
Nichols parameters). The feedback system has negative feedback.
6.2 Reasons for using feedback 96

6.2 Reasons for using feedback

There may be different reasons for using feedback control and some particular im-
portant reasons for using feedback is as listed in the following items:

1. Unknown disturbances.

2. Modeling errors. In case of a perfect model, feed-forward control is an intuitive

option.

3. Unstable processes/systems.

Unknown (and not measured) disturbances are commonly in the process industry.
Such process disturbances may be slowly varying trends and drifts, but more high
frequency types of disturbances which are filtered through the process may also be
present. Such process disturbances may be viewed as a stochastic disturbance ys on
the process output y (se Figure 6.1). This contribution is additive for linear systems.
Slowly varying process disturbances may e.g. be caused from different properties
of raw products, changes in the process environments as temperature and pressure
etc.
The influence of unknown process disturbances may be a great problem. Distur-
bances may lead to saturations in the manipulable control variables u. Even with
feedback and in case of unknown slowly varying disturbances, it may happen that
the best one can achieve is that the process output y are varying around the refer-
ence r. This may be satisfactory because some times it is the mean one earn money
from, i.e. a satisfactory mean quality. However, great variations in quality should
be avoided.
Usually we only have approximate process models hp and these models are never
perfect. With perfect models y = hp u we can use feed-forward control and notice
that y = r gives an ideal feed-forward controller u = h1p r. Such an ideal feed-forward
controller is unrealistic due to modeling errors and the solution is to use feedback
control (or a combination of feedback and feed-forward control).
Feedback is a grate tool to stabilize unstable processes and systems and some
times such feedback solutions are remarkably simple and effective.
Chapter 7

Direct synthesis and design of

standard controllers

7.1 On the PID controller formulations

An ideal PID controller may be written as follows

1
hc (s) = Kp (1 + + Td s) (7.1)
Ti s
where Kp is the proportional constant, Ti is the integral time and Td is the derivative
time.
Consider now a controller (in cascade form) of the form

(1 + T1 s)(1 + T2 s)
hc (s) = k (7.2)
T3 s
We will now show that the cascade form controller (7.2) may be written as an ideal
PID controller. From Equation (7.2) we have

1 + (T1 + T2 )s + T1 T2 s2
hc (s) = k
T3 s
T1 + T2 1 T1 T2
= k( + + s)
T3 T3 s T3
T1 + T2 1 T1 T2
= k (1 + + s) (7.3)
T3 (T1 + T2 )s T1 + T2

Comparing Equation (7.3) with the ideal form PID controller (7.1) shows that the
two formulations are equivalent if
T1 + T2
Kp = k (7.4)
T3
Ti = T1 + T2 (7.5)
T1 T2
Td = (7.6)
T1 + T2
7.1 On the PID controller formulations 98

This means that a cascade controller given by Equation (7.2) is identical to an ideal
PID controller as in Equation (7.1) if the PID controller parameters Kp , Ti and Td
are chosen as in (7.4)-(7.6).
We will now show that under certain circumstances there are possible to find the
controller parameters T1 , T2 and T3 in the cascade formulation as a function of the
controller parameters Kp , Ti and Td in the ideal PID controller.
From (7.4) and (7.5) we have directly that
k
T3 = Ti (7.7)
Kp

From (7.4) we have

T2 = Ti − T1 (7.8)

Putting this expression for T2 into (7.6) gives

T1 (Ti − T1 )
Td = (7.9)
Ti
This last expression may be written as a 2nd order equation for T1 . We have that

T12 − Ti T1 + Ti Td = 0 (7.10)

This equation have two solutions. It seams reasonable to chose the solution with
the largest value in order also to avoid an negative value on T1 . Hence,
p
Ti + Ti (Ti − 4Td )
T1 = (7.11)
2
In order to obtain a real solution we must have that Ti − 4Td ≥ 0.

Remark 7.1 (Ideal and cascade PID controllers) Hence, the conclusion is then
that an ideal PID controller given by Equation (7.1) may be written as a cascade PID
controller as in Equation (7.2) if Ti ≥ 4Td . Hence, there are an exact relationship
between the ideal PID (7.1) and the cacscade PID (7.2) when Ti ≥ 4Td .

A useful approximation may be used when Ti  Td . In this case we see from

Equation (7.11) that T1 ≈ Ti . From (7.7) we find that T3 = Ti if k = Kp . From
(7.6) and (7.5) we have that Td = T1TTi 2 ≈ T2 when T1 ≈ Ti . This means that we
have the following useful remark regarding this approximation.

Remark 7.2 (Approximation between the ideal and cascade PID controllers)
The following approximation between the cascade and ideal PID controllers
1 (1 + Ti s)(1 + Td s)
hc (s) = Kp (1 + + Td s) ≈ Kp (7.12)
Ti s Ti s
may be used when

Ti  Td (7.13)
7.2 Controlling a static (steady state) process 99

Example 7.1 (PID controller relationship)

Given an ideal PID controller as in Equation(7.1). Assume controller settings Ti =
10 and Td = 1. Kp is not specified. We want to find an alternatice cascade PID
controller as in Equation (7.2).
Using the formulas given in Equations (7.11) and (7.8) we find that
p
10 + 10(10 − 4)
T1 = ≈ 8.873 (7.14)
2
T2 = 10 − 8.873 = 1.127 (7.15)

From equation (7.7) we find

T3 = Ti = 10 (7.16)

where we have put k = Kp . We then have the following exact relationship between
the ideal PID controller and the cascade PID controller.
1 (1 + 8.873s)(1 + 1.127s)
hc (s) = Kp (1 + + s) = Kp (7.17)
10s 10s

7.2 Controlling a static (steady state) process

If we want the output, y, of a steady state process to be close to a specified reference

signal, r, then we may simply use a feedback control strategy with an integral (I)
controller in the feedback loop. Real world processes are not static but they may
under certain circumstances approximated as a steady state (static) process. In
practice, often a PI controller is used instead of an ideal Integral controller.
Some processes may be viewed and modeled as a steady state (static) process
relative to the dynamics introduced by the feedback system. This may be the case
for systems with small time constants, i.e. very fast systems, e.g. for systems with
time constants smaller than the controller sample time.
An example is a Thermo Mechanical Pulping (TMP) process in the paper indus-
try. The TMP process is used in order to produce paper pulp for making paper, from
small wood chips. A TMP process is often controlled such that the consistency of
the pulp obtaine a specified reference value, because this may give a specified quality
of the paper. The for instance, u, be the dilution water flow rate to the refiner and
let y be the measured consistency in the blow-line out of the TMP refiner. The
consistency, y, is a measure of the dryness of the pulp paper mass and a reasonable
value for the consistency is about 50%. The dynamics in the TMP refiner from
u to the output y may be neglected compared compared to the sampling time in
the control system. The dominant time constant in the process is in believed to be
smaller than one second. On tye other hand the sampling time in the digital control
system is about ∆T = 0.50 [s] to ∆T = 1.0 [s]. Hence, the process may be viewed
as a static system compared to the dynamics in the feedback system.
Another example is when designing local controllers for the control actuator
itself. An example is a flow controller where u is the position of the valve and y is
the mass flow out of the valve. In this case it may be reasonable to model y as a
7.2 Controlling a static (steady state) process 100

static function of u, i.e. y = g(u) where g(·) is a function describing the behavior of
the valve.

Example 7.2 (Integral (I) control of a static process)

We want in this example to investigate the behavior of a closed loop system of a
static process controlled with an I-controller.
We have the controller and the process transfer functions

hc (s) = T1i s I controller

hp = k static process

The transfer function from the reference r to the output y is given by

y hp hc (s) k T1i s k
(s) = = 1 = T s+k (7.18)
r 1 + hp hc (s) 1 + k Ti s i

This may be written as

y 1
(s) = (7.19)
r 1 + Tc s
where Tc is given by
Ti
Tc = (7.20)
k
Tc is the time constant of the closed loop system after an change in the reference,
and Tc may be chosen in order to obtain Ti . Note that Tc in general is different from
the integral time Ti in the controller. In steady state we have that yr (s = 0) = 1. The
integrator in the controller ensures that we have zero steady state error, i.e. y = r
when t → ∞.
We also see that we obtain ideal control when Tc = 0 and that this may be
approximated with a small value on the integral time constant Ti .

Example 7.3 (PI control of a static process)

We want in this example investigate the closed loop properties of a static system
controlled with a feedback PI controller. We have

hc (s) = Kp 1+T
Ti s
is
PI controller
hp = k static process model

The transfer function from the reference, r, to the output,y, is given by

y hp hc (s) kKp 1+T

Ti s
is
kKp (1 + Ti s)
(s) = = 1+Ti s
= (7.21)
r 1 + hp hc (s) 1 + kKp Ti s Ti + kKp (1 + Ti s)
s

and
y kKp (1 + Ti s) 1 + Ti s
(s) = = T (1+kK )
(7.22)
r Ti (1 + kKp )s + kKp 1 + i kKp p s
7.3 Control of a non-minimum phase process 101

This may be written as follows

y 1 + Ti s
(s) = (7.23)
r 1 + Tc s
where Tc is given by
1 + kKp 1
Tc = Ti = Ti (1 + ) (7.24)
kKp kKp
Tc is the time constant of the closed loop system and note that Tc is different from
Ti .
At low frequencies s = 0 we have in steady state that yr (s = 0) = 1. As we see,
1
we obtain ideal control when | kKp | 1, i.e., when | kK p
|≈ 0. Then we have that
Tc ≈ Ti and y/r ≈ 1. This may be achieved by using a large proportional constant
Kp . We also note that the closed loop system is stable for kKp > −1 fordi Ti > 0.
We may now specify the time constant Tc of the closed loop system and then
obtain the PI controller settings parameters Kp and Ti from the above equations.
This solution strategy is not unique. We may in addition for instance also specify Ti
in order to define Kp . This indicates that there is not necessarily with two parameters
in the controller. The reason for this is that this static process may be reasonably
controlled by a simple I-controller with only one controller parameter Ti .
At the end you should also note the following state space model formulation of
the closed loop system. From the theory of canonical forms (ore inverse Laplace
transformations) we find that
1 1
ẋ = − x+ r (7.25)
Tc Tc
Tc − Ti Ti
y = x+ r (7.26)
Tc Tc
As we see we have a direct proportional feed-through influence from the reference, r,
to the output, y. Hence, it works like a feed-forward control action from the reference
upon the output y.

7.3 Control of a non-minimum phase process

We are referring to Section 9.6 for a description of non-minimum phase systems.

A common property of controlling non-minimum phase systems with standard con-
trollers as P, PI ore PID controllers there exist an upper value for the proportional
constant Kp in order to ensure stability of the closed loop system. We will in this
section illustrate this by examples.

Example 7.4 (P control of a non-minimum phase system)

We want to analyse the closed loop system of a non-minimum phase system con-
trolled by a P-controller. We have
hc (s) = Kp P controller
1−T1 s
hp (s) = k 1+T 2s
non-minimum phase process
7.3 Control of a non-minimum phase process 102

where T1 > 0 is an inverse response time constant and T2 > 0 is the time constant
in the process. The transfer function from the reference r to the output y is given by
1−T1 s
y hp hc (s) Kp k 1+T 2s
(s) = = (7.27)
r 1 + hp hc (s) 1 + Kp k 1−T 1s
1+T2 s

This gives
y Kp k 1 − T1 s
(s) = (7.28)
r 1 + Kp k 1 + Tc s

where the time constant of the closed loop system is given by

T2 − Kp kT1
Tc = (7.29)
1 + Kp k

We see that the closed loop system have a pole s = −1/Tc . In order to ensure
stability of the closed loop system we may demand the pole to be negative. The
closed loop system is therefore stable if Tc > 0. If we assume that the denominator
in the expression for Tc , i.e., 1 + Kp k > 0 (which gives a lower limit −1/k < Kp )
we obtain the following demand for ensuring stability.

Tc > 0 ⇒ − k1 < Kp < Kcu (7.30)

where the upper critical value for Kp is given by

T2 1
Kcu = (7.31)
T1 k
Hence, if T1 is large then we must chose a small value on Kp , and if T1 is small
then Kp may be chosen larger. A limit solution is obtained when T1 → 0, hence

lim Kcu = ∞ (7.32)

T1 →0

Another limit conclusion is obtained by letting T1 → ∞, hence.,

lim Kcu = 0 (7.33)

T1 →∞

The zero in the right half complex plane given by s = 1/T1 , gives constraints for the
choice of Kp . We may conclude and state that the stability properties of the closed
loop system is influenced strongly by the process zero in the right half plane. This is
usually always the case when controlling non-minimum systems with standard PID
feedback controllers.

Example 7.5 (PI control of a non-minimum phase system)

We want in this example investigate the behavior of a standard feedback system, with
a non-minimum phase system controlled by a PI controller. We have

hc (s) = Kp 1+T
Ti s
is
PI controller
hp (s) = k 1−T1s
1+T2 s non-minimum phase process
7.4 Controlling of lead-lag systems 103

where T1 > 0 and T2 > 0. The transfer function from the reference, r, to the output,
y, is given by,

y hp hc (s) Kp 1+T i s 1−T1 s

Ti s k 1+T2 s
(s) = = (7.34)
r 1 + hp hc (s) 1 + Kp 1+T i s 1−T1 s
Ti s k 1+T2 s

We will now for simplicity chose Ti = T2 . This gives

y kKp (1 − T1 s) 1 − T1 s
(s) = = (7.35)
r (T2 − kKp T1 )s + kKp 1 + Tc s
where the time constant of the closed loop system is given by
T2 − Kp kT1
Tc = (7.36)
Kp k
Let us assume that Kp k > 0.
If k > 0 then this means that a lower limit for the proportional constant is
0 < Kp . This only means that Kp should be positive.
We have the following important demand for the upper limit for Kp in order to
ensure stability of the closed loop system.

Tc > 0 ⇒ 0 < Kp < Kcu (7.37)

where the upper limit for Kp is given by

T2 1
Kcu = (7.38)
T1 k
As we see, if T1 is large then we must specify Kp small, and if T1 is small then we
may chose Kp large. A limiting case when T1 → 0 is given by

lim Kcu = ∞ (7.39)

T1 →0

Another limiting case when T1 → ∞ is given by

lim Kcu = 0 (7.40)

T1 →∞

Hence, the zero in the right half plane, s = 1/T1 , gives limitations to the size of
the proportional constant Kp . The stability properties of the closed loop controlled
system is limited by the positive zero in the rhp. Note that another choice than
putting Ti = T2 may have been used in the above discussion.

7.4 Controlling of lead-lag systems

Example 7.6 (PI control of lead-lag process)

We want in this example to design a PI controller of a lead-lag type process. We
have
hc (s) = Kp 1+T
Ti s
is
PI controller
1+T1 s
hp (s) = k 1+T2 s lead-lag process
7.4 Controlling of lead-lag systems 104

The transfer function from the reference r to the output, y, is given by the compli-
mentary sensitivity function T (s).

hp hc (s) Kp 1+T i s 1+T1 s

Ti s k 1+T2 s
T (s) = = (7.41)
1 + hp hc (s) 1 + Kp 1+T i s 1+T1 s
Ti s k 1+T2 s

This gives

y(s) Ti T1 s2 + (T1 + T2 )s + 1
= T (s) = (7.42)
r(s) Tc2 s2 + 2Tc ξs + 1

where the parameters Tc and ξ are defined from.

Ti (T2 + Kp kT1 ) Ti + Kp k(Ti + T1 )

Tc2 = , 2Tc ξ = . (7.43)
Kp k Kp k

If we now are specifying values for the time response parameters Tc and ξ then we
will obtain two equations for finding the PI controller parameters Kp and Ti .
ξ is the relative damping of the closed loop system. If ξ < 1 then the system is
under damped and the poles will be complex conjugate. If ξ < 1 then we will obtain
oscillations in the response from r to y. Tc = 1/ωn where ωn is referred to as the
undamped resonance frequency. We are finally referring to tables which shows the
relationship between Laplace functions and time domain descriptions for an exact
time domain equivalent to T (s).
Chapter 8

Feed forward control

8.1 Introduction

Feed-forward control is maybe the most intuitive way of controlling a process. Con-
sider a process to be controlled and that the process is given by the model

y = hp (s)u (8.1)

and that we want the output, y, to follow a specified reference, r. We will in this
case obtain perfect control, i.e. y = r by the feed-forward controller
1
u= r (8.2)
hp (s)

if the model, hp (s), is an exact description of the controller. This feed-forward

controller, which consists of the inverse of the process model, is usually not useful
in practice due to modeling errors. Hence, we need additional feedback in order to
obtain y = r in steady state.
Feed-forward control may also be effective from measured ore estimated distur-
bances, v. Hence, the main variables to use in feed-forward control is the following:

• Feed-forward control from disturbances, v.

• Feed-forward control from references, r.

Usually a feed-forward controller involves the inversion of the process model

hp (s). However, when the inverse of the process model does not exists, due to time
delay or inverse responses in the process and the model, a model approximation
which have an inverse may be used. For instance a model of lead-lag type is often
used.
Feed-forward control from known references, r, is also denoted tracking control
problems. Feed-forward control from references is also involved in a Model Predic-
tive Control (MPC) solution, as well as in Linear Quadratic (LQ) optimal tracking
problems.
8.2 Feedback with feed forward from the reference 106

8.2 Feedback with feed forward from the reference

let the system be described by the model

y = Hp u + Hd v (8.3)

Assume that vi in addition to normal feedback from the output want a feed forward
controller from the reference. A control input to the process can in this case be
generated by

u = Hc (r − y) + uf (8.4)

where Hc is the controller transfer function (ore transfer matrix). The signal uf is
the feed forward which is a function of the reference signal. The feed forward control
is specified to be

uf = Hfr r (8.5)

where Hfr is the transfer function/matrix from the reference to the feed forward
signal.
Putting the control given by (8.4) into (8.3) gives

y = Hp Hc (r − y) + Hp uf + Hd v (8.6)

Ideally we want y = r. Putting this into equation (8.6) gives the following expression
for the feed forward signal, uf .

r = Hp uf + Hd v
⇓ (8.7)
uf = Hp−1 r − Hp−1 Hd v

where we have assumed that the transfer function model is invertible, i.e. we have
assumed that Hp−1 exists. The ideal feed forward controller from the reference is
therefore expressed by

Hfr = Hp−1 . (8.8)

where we have assumed that the process model is non-singular, i.e. Hp−1 exists.
As we see, we have also derived the ideal feed forward control from the distur-
bance, v, i.e., Hfv = Hp−1 Hd . Another assumption for using it is that the disturbance
is measured and known, ore estimated. We will in the following only analyze the
feed forward control

uf = Hp−1 r (8.9)

Note that even if the process model Hp (s) can be inverted theoretically, it is not
sure that the inverse Hp−1 is a rational and implementable function. However, one
can often use a realizable approximation to the inverse Hp−1 . This will be further
studied lather.
8.2 Feedback with feed forward from the reference 107

The closed loop system with feed forward control is described by

y = (I + Hp Hc )−1 (Hp Hc + I)r + (I + Hp Hc )−1 Hd v (8.10)
As we see, one may have perfect control with feed forward control from the reference
because y = r also without feedback control. This can be seen by putting Hc = 0
in (8.10).
If we have a perfect and non singular model model Hp and the inverse Hp−1 is
rational and implementable, then we will have perfect control, i.e. y = r. (assuming
also that the control is not saturated). This is only theoretically because we always
have modeling errors in practice, often also the inverse Hp−1 is not rational. However,
approximations may often give good and reasonable improvements over classical feed
forward control.
For a SISO system we have a similar to Equation (8.10)
hp hc + 1 hd
y= r+ v (8.11)
1 + hp hc 1 + hp hc
The stability of the feedback system with the feed-forward controller is given by
the characteristic equation. The characteristic equation is found by putting the
denominator in Equation (8.11) equal to zero. Hence, we have
1 + hp hc = 0 (8.12)
For a system with feedback only we have the following transfer function from the
reference, r, to the output, y, i.e.,
hp hc hd
y= r+ v (8.13)
1 + hp hc 1 + hp hc
As we see, a feedback system with feed-forward control, have the same characteristic
equation as a feedback system without feed-forward control. This means that the
feed-forward signal does not influence upon the stability properties of the system.
This may also be shown to hold for systems with feed-forward from the disturbance,
v.

Example 8.1
Given a stable process described by
k
hp (s) = (8.14)
1 + Ts
The ideal feed-forward from the reference is then given by
1
hrf (s) =
(1 + T s) (8.15)
k
This solution is impractical due to the derivative term which will amplify high-
frequent measurement of process noise. A more realistic and practical solution is
to instead use the approximation
1 1 + Ts
hrf (s) = (8.16)
k 1 + Tf s
where Tf may be locked upon as a filter time constant. This approximation to the
ideal feed-forward control is known as a lead-lag feed-forward controller.
8.3 Feed-forward from the disturbance 108

8.3 Feed-forward from the disturbance

8.3.1 Design based on a state space model

Consider a system described by a linear time invariant state space model

ẋ = Ax + Bu + Cv (8.17)
y = Dx (8.18)
We will in the following discussion assume that v is a measured process disturbance
vector. We also assume that the control signal,u, is of the form
u = G(r − y) + uf (8.19)
uf = Gf v (8.20)
where G is a constant feedback matrix (P type controller) and that uf = Gf v is the
feed-forward part. Gf is a constant feed-forward matrix.
Putting the control signal, u, into the state space model gives
ẋ = Ax + BG(r − y) + Buf + Cv (8.21)
A simple method of designing the ideal feed-forward signal uf = Gf v is tho chose
uf such that the two last terms in (8.21) becomes equal to zero. Hence, we have
Buf + Cv = 0
⇓ (8.22)
uf = −B −1 Cv
where we have assumed that B is a non-singular matrix. The ideal feed-forward
controller is then given by
uf = Gf v, Gf = −B −1 C. (8.23)
This controller will cancel the influence of the disturbances upon the system.
Note also thet the feed-forward part of the controller does not influence upon the
stability properties of the feedback system. This may be shown as follows Putting
the feed-forward controller (8.23) into the state equation for the closed system(8.21)
gives
ẋ = (A − BGD)x + BGr, (8.24)
y = Dx. (8.25)
As we see, it is only the feedback matrix, G, which influences upon the stability of
the system.

Remark 8.1 The static feed-forward control (8.23) may off-course also be used
if the feedback part of the controller is dynamic instead of static. A linearized
controller with dynamic feedback part and feed-forward control may be described
by the state space model
ż = Au z + Bu (r − y), (8.26)
u = Du z + Eu (r − y) + uf , (8.27)
8.4 Ratio control 109

where Au , Bu , Du and Eu are controller matrices. z is the state vector in the

controller. As we see, the control, u, is defined in terms of two input signals, the
control deviation e = r − y and the feed-forward signal uf = Gf v. A PI controller
may be written on this form.

8.4 Ratio control

So called ratio control is a special type of feed-forward control which is widely used
in for instant the process industry. An example is for example a chemical reactor
where it is important to feed the reactor with two substances, say substance A and
substance B, in a specified ratio so that the feed to the reactor is blended with the
correct ratio.
In ratio control the objective is for instance to hold the ratio of two variables
close to or at a specific value. Consider as an example a process with two mass flow
variables, qA , and qB , and that we want to hold the ratio
qB
k= (8.28)
qA
constant or close to constant also for varying qA and qB . Hence, the problem is to
hold the ratio, k, constant instead of the two individual variables qA and qB . In
industry, usually flow variables and flow controllers are involved when using ratio
control. Hence we want qA = kqB where k is the a constant ratio.
There are some common solutions to the problem of ratio control. We will discuss
two solutions to ratio control:

Method 1. One standard solution to ratio control problem is the use one flow
controller, for say flow qB , and letting the set-point to this flow controller be
taken as rB = kqA , i.e., as a factor k of a measurement of flow qA . The entire
ratio control system is then influenced by manipulating the valve for the qA
flow manually. This solution is illustrated in Figure 8.1.

Method 2. A safer solution Method 1 above is to modify the method by using two
flow controllers where the entire ratio control system is influenced from one
set-point, rA , to a flow controller for flow qA , and the set-point for the flow
controller of flow qB is taken similar as in Method 1 above, i.e. rB = kqA .

The reader should with advantage sketch block diagrams for the above methods for
ratio control.

Example 8.2 (Feed-forward control with steady state models) Consider a bi-
nary distillation column for the separation of two products A and B. The steady state
mass balances are

F = B + D, (8.29)
F xF = BxB + DxD . (8.30)
8.4 Ratio control 110
kqA uB
hc (s)
−
qB
qB

P lant
k
uA

qA qA
qA

Figure 8.1: Simple ratio control structure.

where the feed flow rate F and the fraction of product A in the feed flow rate is xF
are considered as disturbances (ore load variables). The flow rate of distillate is D
and the composition of product A in D is xD , similarly, the flow rate of the bottom
product is B and the composition of product A in B is xB .
Consider the case in which all variables are measured. Then we may, e.g., solve
for the distillate flow rate, D, as
xF − xD
D=F . (8.31)
xD − xB

Hence, Equation (8.31) may be used as a feed-forward controller from measured

disturbances F and xF . Note also that this feed-forward control strategy is nonlinear
due to the product of F and xF . Here the top and bottom compositions xD and xB ,
respectively, are measured output variables.
Chapter 9

Frequency domain analysis and

controller synthesis

9.1 Complex numbers

√
The imaginary number j is defined as j = −1 and j 2 = −1. Numbers of the type
c = a + jb, where a and b are arbitrarily real numbers are called complex numbers.
Moreover, a is called the real part, and b the imaginary part of the complex number
c, respectively. This is defined formally as a = Re(c) and b = Im(c).
A complex number, c, can be written in so called polar form as follows
c = a + jb =| c | ejφ , (9.1)
where the magnitude, | c |, (or length, modulus) to the complex number c is defined
as
p
| c |= a2 + b2 , (9.2)
and the phase angle φ is given by
b
φ = arctan( ). (9.3)
a
The Euler formula is important and useful in connection with complex numbers,
and is defined as follows
ejφ = cos(φ) + j sin(φ). (9.4)
Note also that the ratio between two complex numbers, c = a + jb, and f = d + je,
can be written on polar form as follows
√ b √
c a + jb a2 + b2 ej arctan( a ) a2 + b2 j(arctan( b )−arctan( e ))
= =√ e = √ e a d . (9.5)
f d + je d2 + e2 ej arctan( d ) d2 + e2
Similarly, the product of two complex numbers c = a + jb, and f = d + je, can be
written on polar form as follows
b e
p p
cf = (a + jb)(d + je) = a2 + b2 ej arctan( a ) d2 + e2 ej arctan( d )
b e
p p
= a2 + b2 d2 + e2 ej(arctan( a )+arctan( d )) . (9.6)
9.2 Frequency response 112

Also the following rules regarding calculus of exponentials are use-full. Given two
scalars φ1 and φ2 . Then

eφ1 eφ2 = eφ1 +φ2 , (9.7)

e φ1
= eφ1 −φ2 . (9.8)
e φ2

9.2 Frequency response

We start this section by defining the frequency response of a system.

Definition 9.1 (Frequency response)

h(jω) is defined as the frequency response of a system h(s). By putting s = jω in the
transfer function h(s) of a system, we obtain the frequency response h(jω), where
0 ≤ ω ≤ ∞ [ rad
s ] is the frequency.

The frequency response of a system describes how the output of a system behaves
after a sinusoid signal with frequency ω [rad/s] is feed into the input of the system.
Assume that a sinusoid signal

u(t) = u0 sin(ωt + α) (9.9)

is feed into the system. Here u0 is the magnitude of the input signal, ω [ rad
s ] is a
specified frequency and α is a constant.
It can be shown that if this sinusoid input signal is exciting the system for long
time, i.e. to all transients has died out and that there is steady state behavior
(t → ∞), then the output response of the system will be described by

y(t) = y0 sin(ωt + β) (9.10)

where y0 is the magnitude and β is a constant describing the phase displacement

from the input signal. It can be shown that the phase displacement is given by

φ=β−α [rad]. (9.11)

It can be shown that the ratio y0 /u0 between the magnitude of the output, y0 ,
and the magnitude of input signal, u0 , can be defined directly from the frequency
response h(ω) of the system, i.e.,
y0
u0 = | h(jω) | (Magnitude. No: amplitudeforholdet)
(9.12)
φ = ∠h(jω) (Phase angle. No: fasevinkelen)

From the theory of complex numbers we can show that the frequency response h(jω)
can be written as

h(jω) =| h(jω) | ej∠h(jω) (9.13)

which is the polar form of a complex number.

9.2 Frequency response 113

Note that, uy00 =| h(jω) |, is the system gain at frequency ω, i.e. the system gain
if the system is excited with a sinusoid signal with frequency ω). In particular, at
low frequencies, i.e. when (s = jω = 0 or ω = 0, we have that the steady state gain
is given by uy00 =| h(j0) |.
As we see, the magnitude | h(jω) | and the phase angle (or phase shift) ∠h(jω)
depend on the frequency ω. This dependency, i.e. the magnitude | h(jω) and the
phase shift ∠h(jω), may be plotted in so called Bode plots.
The most important results is summed up in the following Theorem.

Teorem 9.2.1 (Frequency response)

Given a linear system described by

y = h(s)u (9.14)

where h(s) is the transfer function of the system. If we feed a time varying sinusoid
signal at the input of the system, i.e. an input given by

u(t) = u0 · sin(ωt + α) (9.15)

where u0 is a constant input amplitude, ω [rad/s] is a specified frequency of oscilla-

tion and α [rad] is a constant phase displacement. The stationary output response
is then given by

y(t) =| h(jω) | u0 · sin(ωt + α + φ) (9.16)

where

φ = ∠h(jω) (9.17)

The magnitude | h(jω) | and the phase shift (phase angle) ∠h(jω) is defined from
the frequency response of the system as follows

h(jω) =| h(jω) | ej∠h(jω) (9.18)

Example 9.1 (Frequency response)

Given a process y = h(s)u where

−s + 4 1 − 14 s
h(s) = = 0.5 (9.19)
s+8 1 + 18 s

The frequency response of the system is defined as

1 − j ω4
h(jω) = 0.5 =| h(jω) | ej∠h(jω) (9.20)
1 + j ω8

where
1 + ( ω4 )2
p
| h(jω) |= 0.5 p (9.21)
1 + ( ω8 )2
9.2 Frequency response 114

ω ω
∠h(jω) = arctan(− ) − arctan( ) (9.22)
4 8
Assume that the system input, u, is excited with a signal

u(t) = sin(ωt) (9.23)

From the frequency response theory we have that the stationary output response for
the frequency ω = 0.3 [rad/s] is given by

ω = 0.3 → y(t) = 0.501 sin(0.3t − 6.44◦ 180

π
◦) (9.24)

where the magnitude and the phase shift is given by

| h(j0.3) |= 0.501, ∠h(j0.3) = −6.44◦ . (9.25)

Similarly, the stationary output response for a higher frequency ω = 10 [rad/s] is

given by

ω = 10 → y(t) = 0.841 sin(10t − 119.54◦ 180

π
◦) (9.26)

where the magnitude and the phase shift is given by

| h(j10) |= 0.841, ∠h(j0.3) = −119.54◦ . (9.27)

This is illustrated in Figures 9.1 and 9.2.

Frequency response: h(s)=(−s+4)/(s+8) and u(t)=sin(0.3 t)

1 u(t)

0.8

0.6
y(t)
0.4

0.2

−0.2

−0.4

−0.6

−0.8

−1

0 10 20 30 40 50 60 70 80 90 100
Time [sec]

Figure 9.1: The Figure shows the output response y(t) for a system h(s) = (s −
4)/(s + 8) excited with a sinusoid input signal u(t) = sin(ωt) where the frequency is
ω = 0.3 [rad/s]. Se Example 9.1 for details.
9.3 Gain margin and Phase margin 115

Frequency response: h(s)=(−s+4)/(s+8) and u(t)=sin(10 t)

1 u(t)
y(t)
0.8

0.6

0.4

0.2

−0.2

−0.4

−0.6

−0.8

−1

0 0.2 0.4 0.6 0.8 1 1.2 1.4 1.6 1.8 2

Time [sec]

Figure 9.2: The Figure shows the output response y(t) for a system h(s) = (s −
4)/(s + 8) excited with a sinusoid input signal u(t) = sin(ωt) where the frequency is
ω = 10 [rad/s]. Se Example 9.1 for details.

9.3 Gain margin and Phase margin

The following definitions are important in the theory of frequency analysis and design
of feedback control systems

• Gain margin, GM (No: forsterkningsmargin)

• Phase Margin, P M (No: fasemargin)
• Phase crossover frequency, ω180 (No: fase kryssfrekvens)
• Gain crossover frequency, ωc (No: forsterknings kryssfrekvens)

The starting point for defining the gain margin and phase margin is the loop
transfer function H0 (s) = H0 (jω). We are only locking for the case when s takes
the values along the imaginary axis, i.e. we are putting s = jω.

Gain margin

The Gain margin GM is defined as follows

def 1
GM = Gain Margin (9.28)
| H0 (jω180 ) |
where ω180 is the phase crossover frequency, i.e. the frequency where the phase angle
of the loop transfer function becomes −180◦ the first time.
9.3 Gain margin and Phase margin 116

Phase crossover frequency

ω180 is defined as the phase crossover frequency. The definition of ω180 is as follows

∠H0 (jω180 ) = −180◦ (9.29)

The phase crossover frequency ω180 is formally computed as follows

ω180 = solω [∠H0 (jω) + 180◦ = 0] (9.30)

where solω [f (ω) = 0] means solving the function f (ω) = 0 with respect to ω.

Phase Margin

The Phase Margin P M is defined as

def
P M = ∠H0 (jωc ) + 180◦ Phase Margin (9.31)

where ωc is the gain crossover frequency, i.e. the frequency where the magnitude of
the loop transfer function is equal to one, i.e. | H0 (jωc ) |= 1.

Gain crossover frequency

the gain crossover frequency, ωc , is defined as follows

| H0 (jωc ) |= 1 (9.32)

the gain crossover frequency, ωc , is formally computed as follows

ωc = solω [| H0 (jω) | −1 = 0] (9.33)

where solω [f (ω) = 0] means the solution of the function f (ω) = 0 with respect to ω.
The gain Margin, GM , is a measure of how much the system gain can be in-
creased at the phase crossover frequency, ω = ω180 , before the system becomes
unstable. Such an increase of the system gain can be the result of an increasing
proportional constant, Kp , in a PID controller or an uncertain process gain.
The Gain Margin, GM , is often presented in decibel [DB] units. For a given
absolute value of the Gain Margin we get it in decibel by the conversion

GM [DB] = 20 log(GM ) (9.34)

For example, a Gain Margin of 2 corresponds to GM = 6.02 [DB].

The Phase Margin, P M , is the extra negative phase angle which can be tolerated
before the system becomes unstable. The Phase Margin, P M is the extra phase
which can be added to ∠H0 (jωc ) before the phase becomes −180◦ and the system
becomes unstable.
Typical demands for a feedback control system is a Gain Margin GM > 2 (which
is equivalent to GM > 20 log(2) = 6.02 dB) and a Phase Margin P M > 30◦ − 45◦ .
9.3 Gain margin and Phase margin 117

Gm=6.848 dB, (w= 1.095) Pm=49.34 deg. (w=0.4899)

20

0 1/GM
Gain dB

−20

w_c w_180
−40 −2 −1 0 1
10 10 10 10
Frequency (rad/sec)

−90
Phase deg

PM
−180

−270

w_c
−360 −2 −1 0 1
10 10 10 10
Frequency (rad/sec)

Figure 9.3: Bode plot for H0 (s) = 0.5(1 − s)/((1 + 0.1s)(1 + s)). The Gain margin
GM and the Phase Margin P M is indicated. For this system we have GM = 2.2,
P M = 49.34◦ , ωc = 0.4899 [rad/s] and ω180 = 1.095 [rad/s]. The figure shows
the magnitude in [dB] and 1/GM is illustrated. Transformation to [dB] is done by
1/GM = −20 · log(GM ) [dB]. Here the MATLAB Control System Toolbox function
marginis used.
9.3 Gain margin and Phase margin 118

Example 9.2 (loop transfer function)

Consider given a Proportional and Integral (PI) controller and a process repre-
sented/modeled by a 1st order process with time delay, i.e.
1 + Ti s
hc = Kp (9.35)
Ti s
k
hp = e−τ s (9.36)
1 + Ts
Let us for simplicity chose the integral time constant equal to the time constant of the
process, Ti = T . Hence, we have the following loop transfer function (h0 = hp hc ),
1
h0 (s) = hp hc = k e−τ s (9.37)
s
Kp k
k = (9.38)
T
We will in the following show that it is, in particular, very simple to derive Gain
Margin (GM) and Phase Margin (PM) for a process with a loop transfer function of
this form.

Example 9.3 (Magnitude and Phase of integrator with time delay)

Given a Proportional (P) controller and a process represented by an integrator and
a transport delay, i.e.,

hc = Kp , (9.39)
1 −τ s
hp = e . (9.40)
s
Note that it is usually sufficient to use a P controller in order to control integrating
processes. However, a PI controller may also be used if there are modelling errors.
This gives the following loop transfer function
1
h0 (s) = Kp e−τ s (9.41)
s
We will in the following derive the magnitude and the phase of the frequency response
of the system. The frequency response of the process is defined by putting s = jω in
the transfer function h0 (s), i.e.,
1 −τ jω
h0 (jω) = Kp e (9.42)
jω
The magnitude and phase is given by
Kp
| h0 (jω) | = (9.43)
ω
π
∠h0 (jω) = −( + τ ω) (9.44)
2
This means that the frequency response, h0 (jω), can be written in polar form as
follows
Kp −j( π +τ ω)
h0 (jω) =| h0 (jω) | ej∠h0 (jω) = e 2 (9.45)
ω
9.3 Gain margin and Phase margin 119

Example 9.4 (Gain and Phase margin)

Consider the following loop transfer function
1
h0 (s) = Kp e−τ s (9.46)
s
The magnitude and phase of the frequency response, h0 (jω), was derived in Example
9.3.
The gain crossover frequency, ωc , is given by
Kp
ωc = solω [| H0 (jω) | −1 = 0] = solω [ ω − 1 = 0]
⇓ (9.47)
ωc = Kp
The phase crossover frequency is given by
ω180 = solω [∠H0 (jω) + 180 = 0] = solω [−( π2 + τ ω) + 180◦ = 0]
⇓ (9.48)
π
ω180 = 2τ
Hence, the magnitude and phase of the frequency response is given by
Kp 2τ Kp
| h0 (jω180 ) | = = (9.49)
ω180 π
π π
∠h0 (jωc ) = −( + τ ωc ) = −( + τ Kp ) (9.50)
2 2
Hence, we have the following Gain Margin (GM) and Phase Margin (PM).
π
GM = (9.51)
2τ Kp
π
PM = − τ Kp (9.52)
2
Note that if we specify the closed loop feedback system to have a Gain Margin, GM =
2, then we should specify a proportional constant
π
Kp = (9.53)
4τ
This controller gives a Phase Margin
π
PM = = 45◦ (9.54)
4

Remark 9.1 Assume that there are an uncertainty, dτ , in the transport delay, τ ,
which is used in the process model. Let τp = τ + dτ be the true transport delay in
the process. Hence, τ is the modelled transport delay and dτ the uncertainty (or
the error in the transport delay). The Phase Margin (PM) can be used to identify
how much uncertainty, dτ , which can be tolerated before the closed loop feedback
system becomes unstable.
If P M is the Phase Margin computed with respect to the model with transport
delay τ , and P Mp is the Phase Margin for the real process with transport delay τp ,
then we have that

P Mp = P M − dτ ωc (9.55)
9.3 Gain margin and Phase margin 120

The system becomes unstable if we have an uncertainty dτ which results in a zero

Phase Margin, i.e., P Mp = 0. Hence, we have the following upper limit for the
uncertainty in the transport delay
PM
dτmax = (9.56)
ωc
This means that the closed loop feedback system can tolerate an uncertainty in the
transport delay, dτ , given by
PM
dτ < = dτmax (9.57)
ωc
where P M and ωc is computed with respect to the loop transfer function, h0 (s) =
hp (s)hc (s). Note also that if the unit of gain crossover frequency, ωc is given in
[rad/s] then we must have that the unit of the Phase Margin P M is given in [rad].
The unit in the uncertainty, dτ , is then consistent and given in [s]. Remember also
that 1 [rad] = 180◦ /π ≈ 57.3◦ .

Gm=3.922 dB, (w= 0.7854) Pm=32.7 deg. (w=0.5)

50
Gain dB

0 1/GM

w_c
−50 −1 0 1 2
10 10 10 10
Frequency (rad/sec)

0
w_180
−90
Phase deg

−180 PM

−270

−360 −1 0 1 2
10 10 10 10
Frequency (rad/sec)

Figure 9.4: Bode plot for H0 (s) = 0.5e−2s /s. The Gain Margin (GM ) and the
Phase Margin (P M ) are indicated in the Figure. The closed loop feedback system
has an GM = π/2, P M = π/2 − 1 = 32.7◦ , ωc = 0.5 [rad/s] and ω180 = π/4 [rad/s].
Nothe that the figure presents the magnitude in [dB] but that we have illustrated
1/GM . Consistent transformation to dB is that 1/GM = −20 · log(GM ) [dB]. We
have used the MATLAB Control System Toolbox function margin. Using Equation
(9.57) we find that this system are tolerating an uncertainty dτ = 0.726[s]in the
transport delay, τ , before the system becomes unstable.
9.3 Gain margin and Phase margin 121

Example 9.5 (Design og PI regulator)

We will in this example suggest settings for the proportional constant, Kp , and the
integral time, Ti , in a PI controller. The process is the same as in Example 9.2, but
with specific umerical values.
1 + Ti s
hc = Kp (9.58)
Ti s
2
hp = e−s (9.59)
1 + 4s
We demand that the gain Margin should be GM = 2. We also want the Phase
Margin to be P M > 30◦ . We note that there is a transport delay τ = 1 in the
process.
For simplicity, let us chose the integral time constant equal to the time constant
in the process, i.e.,
Ti = T = 4 (9.60)
This may often be a reasonable setting for the integral time constant. Hence, we
have the following simple loop transfer function
1
h0 (s) = k e−s (9.61)
s
1
k = Kp (9.62)
2
We are now using the results found for GM and P M as found in Example 9.4. 9.4.
π π
GM = = (9.63)
2τ k Kp
π π Kp
PM = − τk = − (9.64)
2 2 2
the demands for the gain Margin GM = 2 and the Phase Margin P M > 30◦ gives
π
GM = = 2, (9.65)
Kp
π Kp
PM = − > 30◦ . (9.66)
2 2
From the demand GM = 2 we find that
π
Kp = . (9.67)
2
Moreover, we find that the system has a Phase Margin
π
P M = = 45◦ (9.68)
4
As we see, that we can not demand a Gain Margin GM and at the same time
demand a given Phase Margin P M . If the Phase Margin is not acceptable we can
chose another value for Ti and recompute the calculations.
See Figures 9.6 and 9.7 for time response. simulations of the closed loop feedback
system. We have also compared the PI controller settings found in this example with
the settings found from Ziegler-Nichols method for PI controller synthesis.
9.4 Bodes stability criterion 122

Amplitude of S T and H_0

1
10
h_0
Log Magnitude

0 T S
10

S
−1
10

−2
10 −1 0 1
10 10 10

Phase of S T and H_0

100
S
Phase (degrees)

0 T

−100 h_0

−200

−300

−1 0 1
10 10 10
Frequency (radians/sec)

Figure 9.5: The figure shows magnitude and phase shift plot of S, T and H0 for
data as given in Example 9.5. Process hp (s) = 2e−s /(4s + 1), PI-controller Hc (s) =
Kp (1 + Ti s)/(Ti s) with Ti = 4 and Kp = π/2. This system have GM = 2 and
P M = 45◦ . The closed loop system have negative feedback.

9.4 Bodes stability criterion

The transfer function of a feedback controlled system is given by

h0 (s) 1
y= r+ ys (9.69)
1 + h0 (s) 1 + h0 (s)

where negative feedback is assumed and r is the reference signal and y s the distur-
bance. We have earlier shown that the stability of a feedback system is given by the
characteristic equation which is defined from

1 + h0 (s) = 0 (9.70)

The characteristic equation is obtained by putting the denominator of the transfer

function equal to zero.
As we see, if the denominator of the transfer function is identical equal to zero,
then the output y will be undefined (infinity). This indicates that a value of the loop
transfer function, h0 (s) = −1, is a critical point. This is the basic for the classical
Bode stability criterion and also the similar Nyquist stability criterion.
The frequency response of the loop transfer function, h0 (jω), is obtained by
putting s = jω in the transfer function h0 (s). We have

h0 (jω) =| h0 | ej∠h0 (9.71)

9.4 Bodes stability criterion 123

Step response for (1) Z−N tuning and (2) GM=2, Ti=4 tuning
1.4
Ziegler−Nichols tuning

1.2

0.8

0.6

0.4

0.2

−0.2
0 1 2 3 4 5 6 7 8 9 10
Time [sec]

Figure 9.6: The figure shows the closed loop time response in the output y after a
positive unit step response at time t = 0 in the reference. The process and data
is as in Example 9.5, i.e. the process is hp (s) = 2e−s /(4s + 1) and PI controller
hc (s) = Kp (1 + Ti s)/(Ti s). Negative feedback is used. Two PI controller parameter
settings are shown. The first is a Ziegler-Nichols settings. The critical gain and
frequency is found to be Kcu = 3.467 and ωcu = 1.715. The ultimate period is then
Pu = 2π/ωcu . This gives the Z-N PI controller parameters Kp = Kcu /2.2 = 1.576
and Ti = Pu /1.2 = 3.052. The closed loop system with Z-N settings have the
margins GM = 1.91 and P M = 38.6. The other PI settings are simply found by
putting Ti = T = 4and cancelling a pole in the loop transfer function, and thereby
computing Kp so that GM = 2. This gives Ti = 4 and Kp = π/2 . The closed
loop system with this settings have a phase margin P M = 45◦ . As we see, the
Z-N setting gives more overshot on the output compared to the second PI controller
setting. The second setting is therefore to be chosen. Se Figure 9.7 for the response
in the control input, u.
9.4 Bodes stability criterion 124

Response in input for (1) Z−N tuning and (2) GM=2, Ti=4 tuning
2.5

Ziegler−Nichols tuning
2
1.

1.5

u(t)
0.5
2.

0
0 1 2 3 4 5 6 7 8 9 10
Time [sec]

Figure 9.7: The figure shows the closed loop time response in the input u after a
positive unit step response at time t = 0 in the reference. The process and data
is as in Example 9.5, i.e. the process is hp (s) = 2e−s /(4s + 1) and PI controller
hc (s) = Kp (1 + Ti s)/(Ti s). Negative feedback is used. Two PI controller parameter
settings are shown. The first is a Ziegler-Nichols settings. The critical gain and
frequency is found to be Kcu = 3.467 and ωcu = 1.715. The ultimate period is then
Pu = 2π/ωcu . This gives the Z-N PI controller parameters Kp = Kcu /2.2 = 1.576
and Ti = Pu /1.2 = 3.052. The closed loop system with Z-N settings have the
margins GM = 1.91 and P M = 38.6. The other PI settings are simply found by
putting Ti = T = 4and cancelling a pole in the loop transfer function, and thereby
computing Kp so that GM = 2. This gives Ti = 4 and Kp = π/2 . The closed
loop system with this settings have a phase margin P M = 45◦ . As we see, the Z-N
setting gives more overshot on the input and more control energy compared to the
second PI controller setting. The second setting is therefore to be chosen. Se Figure
9.6 for the response in the control input, y.
9.4 Bodes stability criterion 125

Assume now that the frequency response of the loop transfer function in a point (i.e.
for a given critical frequency ω = ωcu ) has magnitude equal to 1 and phase shift
equal to −180◦ . Hence | h0 |= 1 and ∠h0 = −π). This means that
h0 (jω) = 1 · e−jπ (9.72)
Putting this into the denominator in the transfer function T (jω) = h0 (jω)/(1 +
h0 (jω) we find that
h0 (jω) + 1 = 1 · e−jπ + 1 = 1 · cos(−π) + 1j · sin(−π) + 1 = −1 + 0j + 1 = 0(9.73)
This means that the output response of the closed loop system y = T r+Sy s becomes
undefined, and the system becomes unstable if the phase of the loop transfer function
becomes −180◦ and the magnitude of the loop transfer function becomes 1. This
means that the closed loop feedback system becomes unstable if |h0 (jω180 )| = 1,
which also is equivalent that the system becomes unstable for a Gain Margin, GM =
1. On this basis we formulate the Bodes stability criterion as follows

Definition 9.2 (Bode stability criterion)

A closed loop feedback system is unstable if the frequency response, h0 (jω), of the
loop transfer function, h0 (s), have a magnitude greater than 1, i.e. | h0 (jω) |> 1 at
the critical frequency.
A closed loop system is stable if the frequency response, h0 (jω), of the loop trans-
fer function, h0 (s), have a magnitude less than 1, i.e. | h0 (jω) |< 1 at the critical
frequency. This is formulated as follows
stable closed loop system ⇔ | h0 (jω180 |< 1 (9.74)
where the critical frequency ω180 is the phase crossover frequency where the phase
shift of the frequency response of the loop transfer function h0 (jω) is −180◦ i.e.
∠h0 (jω180 ) = −180◦ .

Remark 9.2 The Bode stability criterion have two assumptions. The first one
is that the Bode stability criterion only is valid for open loop stable systems, that
means that the process hp (s) is stable. The second assumption is that Bodes stability
criterion only is valid for systems where the phase shift ∠h0 (jω) passes −180◦ once.
If ∠h0 (jω) passes −180◦ several times or the open loop system is unstable, then
Nyquists stability criterion may be used.

Remark 9.3 A great advantage with Bodes stability criterion is that it can be used
for stability analysis of systems with transport delays. Bodes stability criterion can
handle transport delay terms of the type e−τ s in the transfer function. We does
not need Padé approximations to e−τ s which is the case when we are to compute
the roots of the characteristic equation, ore the eigenvalues of the corresponding A
matrix (for continuous systems).

Example 9.6 (Stability)

Given a system described by the transfer function
e−s
h(s) = 2k (9.75)
1 + 4s
9.4 Bodes stability criterion 126

The frequency response is then given by putting s = jω, i.e.,

e−jω
h(jω) = 2k (9.76)
1 + j4ω
This can be written on polar form, expressed in terms of the magnitude and the
phase shift

e−jω 2k
h(jω) = 2k p =p e−j(ω+arctan(4ω) (9.77)
1 + (4ω)2 ej arctan(4ω) 1 + (4ω)2

Hence, the magnitude and phase characteristics of the system is given by

2k
∠h(jω) = φ = −(ω + arctan(4ω)), | h(jω) |= p . (9.78)
1 + (4ω)2

As we see, the proportional constant, k, does nor exists in the phase characteristic.
This is typically also the case for systems controlled by P , P I or P ID controllers.
We will first compute the phase crossover frequency, ω180 . Hence, we have to
solve

φ = −(ω + arctan(4ω)) = −π (9.79)

with respect to ω. This can for instance be done by the Newton-Raphsons method
for the solution of non-linear equations f (ω) = 0. here a trial and error method is
used.
⇒ φ = −2.91 180 ◦
ω = 1.5 π = −166
ω = 1.7 ⇒ φ = −3.12 π = −179◦
180
◦ (9.80)
ω = 1.715 ⇒ φ = −3.141 180
π = −179.96
⇒ φ = −3.146 180 ◦
ω = 1.72 π = −180.2

This means that the phase crossover frequency, ω180 , is given by

ω180 ≈ 1.715. (9.81)

The system is at the stability limit when the magnitude, | h(jω) |), is equal to one
at the frequency ω = ω180 . The critical and largest gain, k, this system can tolerate
before being unstable is therefore given from
2k
| h(jω) |) = p = 1. (9.82)
1 + (4ω)2

Using ω = ω180 gives

p
2
1 + 16ω180
Kcu = = 3.466 (9.83)
2
Hence, the system is stable if

−Kcu < k < Kcu (9.84)

Note that the result only says something about the magnitude of the gain k.
9.5 Ziegler-Nichols method for PID-controller tuning 127

Remark 9.4 The Bode stability criterion is often used to compute the critical gain
Kcu at the critical (phase crossover) frequency ω180 .
The critical gain Kcu and the phase crossover frequency ω180 is the basic pa-
rameters for using the Ziegler-Nichols method for the PID controller settings. In
particular, the Ziegler-Nichols methog gives the following PI controller parameters
in terms of Kcu and ω180
Kcu 2π Pu 2π 1 1
Kp = , Pu = , Ti = = ≈ 5.25 (9.85)
2.2 ω180 1.2 1.2 ω180 ω180
where Pu is the oscillating period, ore ultimate period. Hence, Pu is the periode
time of an oscillation in the output y(t). Note that the symbol ωcu sometimes is
used instead of ω180 , when the magnitude is 1 at the same time.

Remark 9.5 Note also that when the gain of the system is identical to Kcu then
the gain crossover frequency, ωc , will be equal to the phase crossover frequency ω180 .
Hence, the frequency in which ωc = ω180 is referred to as the critical frequency.

9.5 Ziegler-Nichols method for PID-controller tuning

The Ziegler-Nichols (ZN) method for PID controller tuning can be used both as an
experimental and as an analytical model based method for PID controller tuning.
We will here start by presenting the model based ZN method.

Definition 9.3 (Critical gain, Kcu )

Assume that the PID controller is substituted with a proportional P controller.

Note that a PID controller is reduced to a P controller by letting Ti = ∞ and Td = 0.
The frequency response of the loop transfer function is then given by
h0 (jω) = Kp hp (jω) (9.86)
The critical gain, Kcu , is the largest gain which results in
|h0 (jω180 )| = 1, (9.87)
which gives
1
Kcu = . (9.88)
|hp (jω180 )|
This means that if the system, hp (s), is controlled with a P controller with Kp = Kcu
then the closed loop system will have a gain margin GM = 1 and the system is at
the stability limit.

The Ziegler-Nichols method for PID controller tuning is then a function of the
parameters Kcu and ω180 . Note also that the ultimate period is given by
2π
Pu = (9.89)
ω180
9.6 Minimum-phase and non-minimum-phase systems 128

Table 9.1: ZN-method.

Controller Kp Ti Td
Kcu
P 2 ∞ 0
Kcu 2π 1
PI 2.2 1.2 ω180 0
3Kcu Pu
PID 5 2 0.12Pu

9.6 Minimum-phase and non-minimum-phase systems

Minimum-phase system

It can be shown that among all systems with frequency response, h(jω), which have
the same magnitude | h(jω) | there exist one system with the smallest negative
phase shift ∠h(jω). This system is called a minimum phase system.
A necessary and sufficient condition for a system to be a minimum-phase system
is that all zeroes and poles of the transfer function, h(s), lies in the left half plane,
i.e. to the left of the imaginary axis in the complex plane.

Non-minimum-phase system

Zeroes in the right half plane (RHP) and transport delay terms gives more phase
shift to the transfer function h1 (s) compared to a transfer function h2 (s) without
zeroes in the RHP and without transport delay terms. At the same time the transfer
function h1 (s) may have the same magnitude as h2 (s). If the transfer functions h1 (s)
and h2 (s) have the same magnitude for all frequencies ω where s = jω, then the
system represented by h1 (s) will be a non-minimum-phase system.
An example is the system (or process) h1 (s) = e−τ s which have magnitude
| h1 (jω) |= 1 and phase shift ∠h1 (jω) = −τ ω. This system have the same magnitude
as the system h2 (s) = 1 which has zero phase shift, i.e. ∠h1 (jω) = 0. Hence, h1 (s)
is a non-minimum phase system.
Another example is the system h(s) = a−s
a+s which have a zero s = 1 in the right
half of the complex plane. The magnitude is | h(jω) |= 1 and the phase shift is
∠h(jω) = −2 arctan(aω). This system is a non-minimum phase system because the
system h(s) = 1 have the same magnitude but less phase shift, i.e. ∠h1 (jω) = 0.
A non-minimum-phase system is often recognized because the output, y(t), from
such a system have an negative inverse response (or time delay) after a positive unit
response is feed into the input, u(t). See Figure 9.8 for an illustration of the output
response of a non-minimum-phase system. It should be remarked that it usually is
”harder” to control a non-minimum-phase system compared to a minimum-phase
system.
9.7 The bandwidth of a feedback control system 129

Step response for non−minimum−phase and minimum−phase systems

2

1.5
h_1(s)

1
Amplitude

0.5

h_2(s)
0

−0.5
0 5 10 15 20 25 30
Time (secs)

Figure 9.8: Unit step response plot of a minimum-phase system h1 (s) = 2/(4s + 1)
and a non-minimum-phase system h2 (s) = 2e−5s /(4s + 1). The response of the
non-minimum-phase system is recognized from the negative inverse response. The
MATLAB Control Systems toolbox function step is used.

9.7 The bandwidth of a feedback control system

The bandwidth of a feedback control system is defined as the frequency region

in which the control system is effective. Let the bandwidth be described by the
frequency region [ω1 , ω2 ].
Assume that the system is excited by a disturbance with frequency ω. This
means that the control system can control and take care of the disturbances with
frequencies ω1 < ω < ω2 .
A control system is normally effective if the system is excited by low frequent
noise. Hence, we say that the control system is effective for low frequencies. This is
in particular the case for control systems with integral actions. The lower frequency
in the bandwidth is therefore put to ω1 = 0.
A definition of the bandwidth is therefore, ωB = [0, ω2 ] or simply ωB = ω2 .
A common definition of the upper frequency, ω2 , in the bandwidth is the gain
crossover frequency ωc . The gain crossover frequency, ωc , is the frequency in which
the magnitude of h0 passes 1 the first time. It can be shown that ωc only is an
approximation to ωB if the bandwidth is defined as the frequency region in which
the control system is effective in order to control disturbances.
The influence of the disturbance, y s , on the process output, y, is described by
y = Sy s where S = 1/(1 + h0 ) is the sensitivity function. If the control system is to
9.7 The bandwidth of a feedback control system 130

take care of a disturbance y s with frequency ω then we must have that | S(ω) |< 1 at
this frequency, otherwise the disturbance influence on the output would have been
amplified. From Figure 6.2 we see that | S |< 1 for frequencies somewhat less than
ωc . For the system as illustrated in Figure 6.2 and most common control systems
we have that the exact bandwidth is somewhat less than ωc . Hence we have that
the bandwidth is

ωB < ωc (9.90)

This means also that the control system is not effective if the system is excited by
disturbances with a frequency greater than the bandwidth ωB . An upper limit for
the bandwidth of a control system is therefore ωc .
Chapter 10

Discrete time systems

10.1 Introduction

we have already discussed discrete time versions of continuous time models and
in particular discrete time versions of the standard PID controller as discussed in
Chapter 4. However, we will in this section discuss and present the deviation form of
the PID controller. Often incremental form or velocity form of the PID controller is
used instead of deviation form. The point is that only the deviation ∆uk is computed
at each new sample at time, k, and then use uk = uk−1 + ∆uk as the actual control
signal. This version of the PID controller have some advantages and is usually the
version implemented in practice and commercial PID controllers.

10.2 θ method for discretization

Consider a system described by a possibly non linear model

ẋ = f (x). (10.1)

Consider the following method of obtaining a discrete version of (10.1)

xk+1 − xk
= θf (xk ) + (1 − θ)f (xk+1 ). (10.2)
∆t
This is dentoted a so called θ method for discretization. One should here note that
we may chose θ to obtain different famous discrete integration methods:

• θ = 1 gives the explicit Euler method.

• θ = 0 gives the implicit Euler method.

1
• θ= 2 gives the trapezoid method.

One should here note that the explicit Euler method may be numerically unstable
and that there is an upper limit for the step length (sampling time) parameter, ∆t.
The explicit Euler method is a good choice when the step length parameter is chosen
”small” and is accurate enough for most control systems analysis and design.
10.3 Deviation form of the PI controller 132

On the other hand the implicit methods are numerically stable for all choices of
step length parameter, ∆t. However, the solutions may be inaccurate for large step
length parameters ∆t.

10.3 Deviation form of the PI controller

As an introduction to the deviation form of a PID controller in the next section we

here discuss a deviation formulation of the PI controller. We also for the sake of
simplicity only use the explicit Euler method for discretization.
A conventional PI-controller can be written as
1 + Ti s Kp 1
u = Kp (r − y) = Kp (r − y) + (r − y). (10.3)
Ti s Ti s
Defining the PI-controller state, z, as
1
z = (r − y). (10.4)
s
Hence, the PI controller can in continuous time be written as
ż = r − y, (10.5)
Kp
u = Kp (r − y) + z. (10.6)
Ti
We will in the following assume that the reference is constant, rk = r. However, note
that in the final controller we use varying reference signal. A discrete formulation
of the PI controller is then
zk+1 − zk = ∆t(r − yk ), (10.7)
Kp
uk = Kp (r − yk ) + zk , (10.8)
Ti
where ∆t is the sampling interval. A deviation formulation of the PI controller is
then found as follows
Kp Kp
uk − uk−1 = Kp (r − yk ) + zk − (Kp (r − yk−1 ) + zk−1 )
Ti Ti
Kp
= −Kp (yk − yk−1 ) + (zk − zk−1 ). (10.9)
Ti
From (10.7) we have that zk − zk−1 = ∆t(r − yk−1 ). Substituting this into (10.9)
gives
uk = uk−1 + G1 (yk − yk−1 ) + G2 (yk−1 − r). (10.10)
where
Kp
G1 = −Kp , G2 = − ∆t. (10.11)
Ti
Furthermore, using that y = Dx + w gives
uk = uk−1 + G1 D∆xk + G2 (yk−1 − r). (10.12)
The above discussion shows that the PI controller is exactly of the same structure
as a state feedback controller (LQ optimal controller). The controller takes feedback
from the deviation state vector ∆xk = xk − xk−1 while the PI-controller only uses
feedback from the output deviation ∆yk = D∆xk . But notice that yk = ∆yk + yk−1 .
10.4 Deviation form of the PID controller 133

10.4 Deviation form of the PID controller

The PID controller may in discrete time be formulated on so called deviation form.
Also incremental form or velocity form is used on this version of the discrete PID
controller. We have that the PID controller may be expressed as

uk = uk−1 + ∆uk , (10.13)

where ∆uk is the computed control deviation (increment) which usually is computed
as

∆uk = g0 ek + g1 ek−1 + g2 (yk − 2yk−1 + yk−2 ), (10.14)

where g0 , g1 and g2 are related to the PID controller parameters Kp , Ti and Td . One
should noe that different methods for discretization gives different formulas for g0 ,
g1 and g2 .
Note also that the computed control deviation in (10.14) is based on known
present and past variables and that ek = rk − yk and ek−1 = rk−1 − yk−1 .

10.4.1 The continuous time PID controller

We here review the PID controller in the time domain.

The PID controller may be written

u = Kp e + z + Kp Td se (10.15)

where
Kp Kp
z= e, ⇒ sz = e. (10.16)
Td s Ti
Inverse Laplace-transformation gives

u = Kp e + z + Kp Td ė, (10.17)
Kp
ż = e. (10.18)
Ti
It make sense to put e(t0 ) = r − y = 0 and ė(t0 ) = 0 at startup. We then have that
the initial value for the controller state may be chosen as

z(t0 ) = u0 , (10.19)

where u0 is a nominal control or working point for the process when turning on
the controller system in automatic mode . The nominal control may be found by
analyzing the steady state behavior of the process. We have

y0 = ku0 = r, (10.20)

where k is the gain. this gives the following initial value.

r(t0 )
z(t0 ) = . (10.21)
k
10.4 Deviation form of the PID controller 134

10.4.2 Deviation form of the PID controller using the explicit Euler
method

We have two possibilities for implementing the derivation ė. The first possibility is
to use
ek − ek−1
ė ≈ . (10.22)
∆t
the second possibility and the most common choice is to not take the derivative of
the reference signal, i.e. using
yk − yk−1
ė = −ẏ ≈ − . (10.23)
∆t
using the last choice and the explicit euler method for discretization of the controller
state space model (10.18) gives
Kp Td
uk = Kp ek + zk − (yk − yk−1 ), (10.24)
∆t
Kp
zk+1 = zk + ∆t ek . (10.25)
Ti
r0
with initial value z0 = u0 = k.
This discrete state space model for the PID controller may be used directly.
However, a formulation on deviation form may be derived as follows (compute the
deviation ∆uk = uk − uk−1 ). This gives
Kp Td
uk − uk−1 = Kp ek + zk − (yk − yk−1 )
∆t
Kp Td
− (Kp ek−1 + zk−1 − (yk−1 − yk−2 )) (10.26)
∆t
which may be written as
Kp Td
uk − uk−1 = Kp ek + zk − zk−1 − Kp ek−1 − (yk − 2yk−1 + yk−2 ) (10.27)
∆t
Using Eq. (10.25) we have that
Kp
zk − zk−1 = ∆t ek−1 . (10.28)
Ti
This gives
∆t Kp Td
uk = uk−1 + Kp ek − Kp (1 − )ek−1 − (yk − 2yk−1 + yk−2 ) (10.29)
Ti ∆t
Hence, this is of the form
uk = uk−1 + g0 ek + g1 ek−1 + g2 (yk − 2yk−1 + yk−2 ) (10.30)
where
∆t Kp Td
g0 = Kp , g1 = −Kp (1 − ), g2 = − . (10.31)
Ti ∆t
The PID controller written on the form in Eq. (10.30) with parameters as in Eq.
(10.31) is denoted the velocity form, deviation form or incremental form, and often
used in practical implementations.
10.5 Discussion 135

10.4.3 Deviation form of the PID controller using the trapezoid

method

Using the trapezoid method for integrating the controller state space model (10.18)
gives
zk+1 − zk 1 Kp 1 Kp
= ek + ek+1 (10.32)
∆t 2 Ti 2 Ti
As we see, it is not possible to formulate an implementable dicrete state space model
for the PID controller of the same form as when the Explicit Euler method was used,
as in Equations (10.24) and (10.25). The reason for this is that we do not know
ek+1 = rk+1 − yk+1 which in this last case is needed in order to compute and update
the controller state zk+1 .
However, we may use the trapezoid method in order to formulate the PID con-
troller on deviation (incremental) form. Using that

∆t Kp
zk − zk−1 = (ek−1 + ek ) (10.33)
2 Ti
and putting this into (10.27) gives

∆t ∆t
uk = uk−1 + Kp (1 + )ek − Kp (1 − )ek−1
2Ti 2Ti
Kp Td
− (yk − 2yk−1 + yk−2 ) (10.34)
∆t
This gives the controller formulation

uk = uk−1 + g0 ek + g1 ek−1 + g2 (yk − 2yk−1 + yk−2 ) (10.35)

where
∆t ∆t Kp Td
g0 = Kp (1 + ), g1 = −Kp (1 − ), g2 = − . (10.36)
2Ti 2Ti ∆t

10.5 Discussion

We have in this chapter discussed the deviation form of the discrete time PID con-
troller. This form of the PID controller may with advantage be used when im-
plementing the PID controller in practice. This form of the controller also have
some advantages, i.e., the controller is insensitive to constant offset values on the
control action due to the fact that only the deviation ∆k = uk − uk−1 matters.
Anti-Windup is also not a topic in the deviation form of the controller because the
controller integrator state is eliminated from the controller.
Chapter 11

Time delay in discrete systems

11.1 Modeling of time delay

We will in this section discuss systems with possibly time delay. Assume that the
system without time delay is given by a proper state space model as follows

xk+1 = Axk + Buk , (11.1)

yk− = Dxk + Euk , (11.2)

and that the output of the system, yk , is identical to, yk− , but delayed a delay τ
samples. The time delay may then be exact expressed as

yk+τ = yk− . (11.3)

Discrete time systems with time delay may be modeled by including a number
of fictive dummy states for describing the time delay. Some alternative methods are
described in the following.

11.1.1 Transport delay and controllability canonical form

Formulation 1: State space model for time delay

We will include a positive integer number τ fictive dummy state vectors of dimension
m in order for describing the time delay, i.e.,
x1k+1 = Dxk + Euk 


x2k+1 = x1k


.. (11.4)
. 

xτk−1

xτk+1 =


The output of the process is then given by

yk = xτk (11.5)

We se by comparing the defined equations (11.4) and (11.5) is an identical description

as the original state space model given by (11.1), (11.2 and (11.3). Note that we in
11.1 Modeling of time delay 137

(11.4) have defined a number τ m fictive dummy state variables for describing the
time delay.
Augmenting the model (11.1) and (11.2) with the state space model for the delay
gives a complete model for the system with delay.
x̃k+1 Ã x̃k B̃
z }| { z
 }| { z }| { 
z }| {
x A 0 0 ··· 0 0 x B

 x1 

 D 0 0 ···
 0 0   x1 
 

 E 


 x2 
 =  0 I 0 ···
 0 0   x2 
  +
 0  uk

(11.6)
 ..   .. .. .. . . .. ..   ..   .. 
 .   . . . . . .  .   . 
xτ k+1
0 0 0 ··· I 0 xτ k
0

x̃k
z }| {
x
D̃

 x1 

z }| {  x2 
yk = 0 0 0 ··· 0 I  .. (11.7)
 

 . 
xτ −1
 
 
xτ k

hence we have

x̃k+1 = Ãx̃k + B̃uk (11.8)

yk = D̃x̃k (11.9)

where the state vector x̃k ∈ Rn+τ m contains n states for the process (11.1) without
delay and a number τ m states for describing the time delay (11.3).
With the basis in this state space model, Equations (11.8) and (11.9), we may
use all our theory for analyse and design of linear dynamic control systems.

Formulation 2: State space model for time delay

The formulation of the time delay in Equations (11.6) and (11.7) is not very com-
pacter. We will in this section present a different more compact formulation. In
some circumstances the model from yk− to yk will be of interest in itself. We start
by isolating this model. Consider the following state space model where yk − ∈ Rm
s delayed an integer number τ time instants.
xτk+1 Aτ xτk Bτ
z }| { z }| { z }| { z }| {
x 1 0 0 0 ··· 0 0 x 1 I

 x2  

 I 0 0 ··· 0 0   x2 
 
 0 
 
3
x  0 I 0 ··· 0 0   x3 
=  +  0  yk− (11.10)
   
  
 ..   .. .. ..
.. .. ..   ..   .. 
 .   . . . . . .  .   . 
xτ k+1 0 0 0 ··· I 0 xτ k 0
11.1 Modeling of time delay 138

xτk
z }| {
x
Dτ

 x1 

z }| {  x2 
yk = 0 0 0 ··· 0 I  .. (11.11)
 

 . 
xτ −1
 
 
xτ k

which may be written as

xτk+1 = Aτ xτk + B τ yk− (11.12)

yk = Dτ xτk (11.13)

where xτk ∈ Rτ m . the initial state for the delay state is put to xτ0 = 0. Note here that
the state space model (11.10) and (11.11) is on so called controllability canonical
form.
Combining (11.12) and (11.13) with the state space model equations (11.1) and
(11.2), gives an compact model for the entire system, i.e., the system without delay
from uk to yk− , and for the delay from yk− to the output yk .

x̃k à x̃k B̃
z
 }| { z
 { z }| { z }| {
}|
x A 0 x B
= + uk (11.14)
xτ k+1
τ
B D A τ τ
x k Bτ E
x̃k
D̃
z }| {
z }| { 
x
yk = 0 Dτ (11.15)
xτ k

Note that the state space model given by Equations (11.14) and (11.15), is identical
with the state space model in (11.6) and (11.7).

11.1.2 Time delay and observability canonical form

A simple method for modeling the time delay may be obtained by directly taking
Equation (11.3) as the starting point. Combining yk+τ = yk− with a number τ − 1
fictive dummy states, yk+1 = yk+1 , · · · , yk+τ −1 = yk+τ −1 we may write down the
following state space model
xτk+1 Aτ xτk Bτ
z }| { z
 }| { z }| { z }| {
yk+1 0 I 0 ··· 0 0 yk 0
 yk+2   0 0 I ··· 0 0   yk+1
    0 
.. .. .. .. ..   yk+2   ..  −
     
 yk+3 =
  ..  +  .  yk (11.16)
 . . . . . .  
 ..     ..   
 .   0 0 0 ··· 0 I  .
   0 
yk+τ 0 0 0 ··· 0 0 yk+τ −1 I
11.2 Implementation of time delay 139

xτk
z
 }| {
yk
Dτ  yk+1 
z }| {  
yk = I 0 0 ··· 0 
 yk+2 
 (11.17)
 .. 
 . 
yk+τ −1

where xτk ∈ Rτ m .
The initial state for the time delay is put to xτ0 = 0. Note that the state space
model (11.16) and (11.17) is on observability canonical form.

11.2 Implementation of time delay

The state space model for the delay contains a huge number of zeroes and ones when
the time delay is large, ie when the delay state space model dimension mτ is large.
In the continuous time we have that a delay is described exact by yk+τ = yk− . It
can be shown that instead of simulating the state space model for the delay we can
obtain the same by using a matrix (array or shift register) of size nτ × m where we
use nτ = τ as an integer number of delay samples.
The above state space model for the delay contains nτ = τ state equations which
may be expressed as
−
x1k = yk−1
x2k = x1k−1
..
. (11.18)
xτk−1 = −2
xτk−1
τ −1
yk = xk−1

where we have used yk = xτk . This may be implemented efficiently by using a matrix
(or vector x. The following algorithm (or variants of it) may be used:

Algorithm 11.2.1 (Implementing time delay of a signal)

Given a vector yk− ∈ Rm . A time delay of the elements in the vector yk− of nτ time
instants (samples) may simply be implemented by using a matrix x of size nτ × m.
At each sample, k, (each call of the algorithm) do the following:

1. Put yk− in the first row (at the top) of the matrix x.

2. Interchange each row (elements) in matrix one position down in the matrix.

3. The delayed output yk is taken from the bottom element (last row) in the matrix
x.
11.3 Examples 140

yk = x(τ, 1 : m)T
for i = τ : −1 : 2
x(i, 1 : m) = x(i − 1, 1 : m)
end
x(1, 1 : m) = (yk− )T

Note that this algorithm should be evaluated at each time instant k.

4

11.3 Examples

Example 11.1 Delay τ = 2 samplews

Given a system where the undelayed output yk− is delayed τ = 2 samples. The model
for the undelayed part from input uk to output yk− is given by

xk+1 = Axk + Buk , (11.19)

yk = Dxk . (11.20)

For the delay we define τ = 2 dummy states

x1k+1 = yk− , (11.21)

x2k+1 = x1k , (11.22)
yk = x2k . (11.23)
Chapter 12

Adjustment of PID control

parameters 1

Abstract

Methods for the adjustment of PID control parameters, such that the loop transfer
function, the sensitivity transfer function or the complementary sensitivity transfer
function is tuned to a prescribed amplitude and phase, are presented.
Keywords: PID control parameters; automatic tuning.

12.1 Introduction

Many good methods for on-line identification of process parameters needed for tun-
ing PID controllers are available.
A method for the adaption of PID controller parameters in closed loop systems
has been presented in Balchen et al (1989). The method is based on an algorithm
where the phase angle ∠N (jω) is prescribed, and the corresponding frequency and
amplitude of the sensitivity transfer function N (jω) are estimated. The propor-
tional gain is adjusted by an integral controller with setpoint equal to a prescribed
amplitude of N . The integral and derivative times are adjusted proportionally to
the inverse of the frequency ω.
A different scheme has been presented in Schei (1991), where the amplitude of the
complementary sensitivity transfer function M (jω) and the frequency of oscillation
are estimated from a describing function approach. The phase angle is fixed to
∠M (jω) = −90◦ . The controller parameters are updated at each oscillation period
in such a way that the amplitude of the loop transfer function h0 (jω) attains a
specified value. The integral and derivative times are adjusted in the same way as
in Balchen et al (1989).
1
This chapter is based on the paper
Di Ruscio (1992). Adjustment of PID control parameters. Modeling, Identification and Control,
Vol. 13, No. 4, pp. 189-197.
12.2 Main Results 142

However, the methods for the adjustment of the proportional gain in these works
may be improved.
This note is concerned with the problem of adjusting the PID controller param-
eters in such a way that some process performance specifications are reached. Such
specifications may be prescribed values on the amplitude and phase of the loop,
sensitivity or the complementary sensitivity transfer function.
The paper is organized as follows: Section 2 presents the main results. Two
examples are presented in Section 3 and some concluding remarks follow in Section
4.

12.2 Main Results

The main results in this section are stated in Subsection 2.1, 2.2 and 2.3. Subsec-
tion 2.1 deals with the loop transfer function h0 (jω), Subsection 2.2 considers the
sensitivity transfer function N (jω) and Subsection 2.3 is considered with the com-
plementary sensitivity transfer function M (jω) for the adjustment of PID controller
parameters. All methods are based on the idea that the integral time Ti , and the
derivative time Td are functions of the frequency, ie. Ti = Ti (ω) and Td = Td (ω).
Ti and Td are usually set proportional to ω −1 . The problem is then reduced to the
adjustment of the proportional gain Kp .

12.2.1 Method 1

The following iteration formula for the adjustment of Kp such that the amplitude
of the loop transfer function h0 is tuned to a prescribed value |hs0 | is proposed.
|hs0 | |hs | − |hn |
Kpn+1 = Kpn n = Kpn + Kpn 0 n 0 (12.1)
|h0 | |h0 |
where n is the index of iteration. The formula, Equation (12.1), has a quadratic
convergence rate near the solution. This can be seen by using the Newton method
for solving
f¯(Kp ) = |hs0 | − |h0 | (12.2)
to zero, which results in the following iteration scheme,
∂ f¯ −1 ¯
Kpn+1 = Kpn − ( ) fn (12.3)
∂Kp n
where the gradient of f¯ with respect to Kp is given by
∂ f¯ ∂|h0 | 1
=− =− |h0 | (12.4)
∂Kp ∂Kp Kp
Substituting Equation (12.4) into Equation (12.3) gives the iteration formula, Equa-
tion (12.1).
The adjustment formula, Equation (12.1), can be shown to be the same as the
procedure for adjusting Kp described in an unclear and complicated way in Schei
(1991).
12.2 Main Results 143

12.2.2 Method 2

A method for the adjustment of Kp such that the amplitude of the sensitivity transfer
function N is tuned to a prescribed value |N0 | can be derived as follows.
An iteration formula for tuning

f (Kp ) = |N0 | − |N | (12.5)

to zero, is
∂|N | −1
Kpn+1 = Kpn + ( ) fn (12.6)
∂Kp n
The gradient of the amplitude of N with respect to Kp can be shown to be

∂|N | 1
=− |N |(1 − <e(N )) (12.7)
∂Kp Kp

See the Appendix for a proof. Substituting Equation (12.7) into Equation (12.6)
gives the iteration formula
|N0 | − |N n |
Kpn+1 = Kpn − Kpn (12.8)
|N |(1 − <e(N n ))
n

A problem with this formula is that the gradient (12.7) may be zero in the frequency
range considered. This difficulty is avoided by the following approximation.
|N0 | n
n |N0 | − |N |
Kpn+1 = Kpn = Kp
n
+ Kp (12.9)
|N n | |N n |
This iteration scheme is found to be reasonable during simulations, and will hereafter
be referred to as Method 2. Equation (12.9) is found from Equation (12.8) with
<e(N ) = 2. Other approximations, with different convergence properties, may be
proposed by fixing <e(N ) to a constant near the singular point.

12.2.3 Method 3

A method for the adjustment of Kp such that the amplitude of the complementary
sensitivity transfer function, M , is tuned to a prescribed value |M0 | can be derived
from the same procedure as Method 2. We have the iteration formula
|M0 | − |M n |
Kpn+1 = Kpn + Kpn (12.10)
|M n |(1 − <e(M n ))
where we have used that the gradient of the amplitude of M with respect to Kp can
be shown to be
∂|M | 1
= |M |(1 − <e(M )) (12.11)
∂Kp Kp

and that the iteration formula, Equation (12.10), is a Newton method for tuning

f˜(Kp ) = |M0 | − |M | (12.12)

12.2 Main Results 144

to zero. See the Appendix for a proof of Equation (12.11).

If the phase angle ∠M (jω) is prescribed, |M | and ω estimated, then the iteration
formula, Equation (12.10), for adjusting Kp may be justified. Note that <e(M ) ≈ 1
only for ω ≈ 0 which means that the iteration formula, Equation (12.10), will have
no singularities in the main frequency range.

Remark 1
All proposed methods for the adjustment of Kp may be modified to

∂ fˆ −1 ˆ
Kpn+1 = Kpn − λn ( ) fn (12.13)
∂Kp n

where fˆ is the performance function to be tuned to zero. The rate of convergence

of the proposed methods may be improved by choosing a suitable parameter λn .

Remark 2
Assume a PID controller with Ti = Ti (ω) and Td = Td (ω) and that ∠N (ω, Kp )
and |N (ω, Kp )| are estimated or determined from a model. ω and Kp may then be
determined directly by tuning the following set of equations to zero
f1 (ω, Kp ) = ∠N0 − ∠N (ω, Kp ) (12.14)
f2 (ω, Kp ) = |N0 | − |N (ω, Kp )| (12.15)
where ∠N0 and |N0 | are prescribed phase and amplitude of the sensitivity transfer
function, respectively. A Newton method of the same form as Equation (12.13) may
be used, ie.
∂f −1
θn+1 = θn − Λn () f (12.16)
∂θ n n
 T  T
where θ = ω Kp and f = f1 f2 . The matrix Λn is introduced to
improve the convergence. The Jacobian matrix is given by
" ∂f ∂f1
# " ∂∠N ∂∠N #
∂f ∂ω
1
∂Kp ∂ω ∂K
= ∂f2 ∂f2 = − ∂|N | ∂|Np| (12.17)
∂θ ∂ω ∂Kp ∂ω ∂Kp

This iteration scheme is based on the total differential of ∠N and |N |, ie.

∂∠N ∂∠N
d∠N = dω + dKp (12.18)
∂ω ∂Kp
∂|N | ∂|N |
d|N | = dω + dKp (12.19)
∂ω ∂Kp
The partial derivatives of ∠N and |N | with respect to Kp may be determined ana-
|
lytically. ∂|N
∂Kp is given by Equation (12.7). It can further be shown that

∂∠N 1
= |N | sin(∠N ) (12.20)
∂Kp Kp
A proof of Equation (12.20) is given in the Appendix. The partial derivatives of
∠N and |N | with respect to ω may be adjusted numerically during the iterations,
or determined analytically if the process model is known.
12.3 Examples 145

12.3 Examples

12.3.1 Example 1

A PI controller is tuned for the process model

2
hu = (12.21)
(1 + s)(1 + 4s)

Assume that |N | and ω are estimated, the phase angle ∠N (jω) prescribed, and that
the following parameters are specified.
4
|N0 | = 1.5, ∠N0 = 30◦ , Ti = (12.22)
ω
Tuning results for Methods 2 and 3 are shown in Figure (12.1). The final results are
Kp = 1.94, Ti = 3.4 and ω = 1.18. For this example Method 1 did not converge.

2 Proportional gain 10 Integral time

b 8 b
1.5
c 6
1
c
4
0.5
2

0 0
0 5 10 15 0 5 10 15
Iterations Iterations

2 Amplitude of N(jw) 1.5 Frequency w

b b
1.5
c 1
c
1
0.5
0.5

0 0
0 5 10 15 0 5 10 15
Iterations Iterations

Figure 12.1: Kp , Ti , |N | and w during tuning of the PI controller, for b, Method 2

and c, Method 3.

For this example, Figure (12.1) shows that Methods 2 and 3 have approximately
equal rates of convergence. Now, assume that |M | and ω are estimated, the above
specifications are then identical to
4
|M0 | = 0.81, ∠M0 = −111.7◦ , Ti = (12.23)
ω
12.3 Examples 146

Simulation results based on Method 3 for estimation of |M | and Method 2 for esti-
mation of |N | are shown in Figure (12.2).

Proportional gain 10 Integral time

2
c 8
1.5 b
6 b
1
4
c
0.5 2

0 0
0 5 10 15 0 5 10 15
Iterations Iterations

2 Amplitude of M and N 1.5 Frequency w

b c
1.5
1 b
1 c
0.5
0.5

0 0
0 5 10 15 0 5 10 15
Iterations Iterations

Figure 12.2: Kp , Ti , |N |, |M | and w during tuning of the PI controller. b - Method

2 and c - Method 3.

Figure (12.2) shows that Method 3 based on estimation of |M | has considerable

faster rate of convergence compared to Method 2, based on estimation of |N |. How-
ever, the convergence may be improved by the use of the modified scheme, Equation
(12.13), with a suitable parameter λn .

12.3.2 Example 2

A PI controller is tuned for the process model

2
hu = e−s (12.24)
(1 + s)(1 + 4s)

Assume that |N | and ω are estimated, the phase angle ∠N (jω) prescribed, and that
the following parameters are specified.
5
|N0 | = 2, ∠N0 = 0◦ , Ti = (12.25)
ω
Tuning results for Methods 1, 2 and 3 are shown in Figure (12.3). All methods
converged for this example. The final results are Kp = 1.24, Ti = 5.52 and ω = 0.91.
12.4 Concluding Remarks 147

Note that the rate of convergence is considerably faster than for Example 1. This
example indicate faster convergence for processes with a time delay, because the
frequency, ω, is nearly constant during the adaption of the controller parameters.
Compare with Example 1.

2 Proportional gain Integral time

10
1.5 a

1 5
b

0.5 c

0 0
0 2 4 6 0 2 4 6
Iterations Iterations

2.5 Amplitude of N(jw) 1.5 Frequency w

a
2
b 1
1.5
c
1
0.5
0.5

0 0
0 2 4 6 0 2 4 6
Iterations Iterations

Figure 12.3: Kp , Ti , |N | and w during tuning of the PI controller. a - Method 1, b

- Method 2 and c - Method 3.

12.4 Concluding Remarks

Three methods based on gradient informations for the adjustment of the propor-
tional gain in an PID controller are presented. The integral time constant, and the
derivative time constant, are set proportional to the inverse of the frequency.
A method for the determination of PID parameters presented in an unclear and
complicated way in Schei (1991) is shown to be a simple Newton method for the
adjustment of Kp such that the amplitude of the loop transfer function h0 (jω) is
tuned to a specified value.
Simple analytical expressions for the gradient of the amplitude of the sensitivity
transfer function N (jω) and the complementary sensitivity transfer function M (jω)
with respect to the proportional gain Kp are presented. These expressions are the
basis for two Newton methods for the adjustment of the proportional gain.
It is shown that the proportional gain should be adjusted from gradient infor-
12.5 References 148

mations of |M | rather than gradient informations of |N |. This statement is justified

from the fact that the gradient of |N | with respect to Kp may be zero in the main
frequency range, which means that the Newton method will be singular. However,
an approximate iteration formula based on |N |, which may be used, is presented.

12.5 References

Åstrøm K. J., T. Hägglund (1984). Automatic Tuning of Simple Regulators with

Specifications on Phase and Amplitude Margins. Automatica, vol. 20, no. 5,
pp. 645 - 651.

Balchen J. G., K. Telnes and D. Di Ruscio (1989). Frequency response adaptive

control of a refrigeration cycle. Modeling, Identification and Control, vol. 10,
no. 1, pp. 3 - 11.

Schei T. S. (1991). A New Method for Automatic Tuning of PID Control Param-
eters. Proceedings 1th ECC - European Control Conference, Grenoble, July
2-5, 1991, pp. 1522 - 1527.

12.6 Appendix

12.6.1 A proof of Equation (12.7)

The sensitivity transfer function is given by

1
N (jω) = (12.26)
1 + h0 (jω)

Define h = |h0 (jω)|, φ = ∠h0 (jω) and h0 (jω) = hejφ to simplify the notation. We
have
1 1 + he−jφ 1 + he−jφ
N (jω) = = = (12.27)
1 + hejφ (1 + hejφ )(1 + he−jφ ) 1 + h(ejφ + e−jφ ) + h2

which may be written

1 + hcos(−φ) + jhsin(−φ)
N (jω) = = <e(N ) + j=m(N ) (12.28)
1 + h2 + 2hcos(φ)

The amplitude of N (jω) is given by

1
[(1 + hcos(−φ))2 + (hsin(−φ))2 ] 2
|N | = (12.29)
1 + h2 + 2hcos(φ)
or
1
|N | = p (12.30)
1+ h2 + 2hcos(φ)
12.6 Appendix 149

The gradient of |N (jω)| with respect to Kp is

∂|N | 1 1 ∂h ∂φ
=− 3 [(2h + 2cos(φ)) − 2hsin(φ) ] (12.31)
∂Kp 2 (1 + h2 + 2hcos(φ)) 2 ∂Kp ∂Kp
For a PID controller we have
∂h 1
= h (12.32)
∂Kp Kp
and
∂φ
=0 (12.33)
∂Kp
when ω is constant. Substituting Equations (12.32), (12.33) and (12.30) into Equa-
tion (12.31) yields
∂|N | 1 1 + hcos(φ)
=− |N |(1 − ) (12.34)
∂Kp Kp 1 + h2 + 2hcos(φ)
which is seen to be equal to
∂|N | 1
=− |N |(1 − |N |cos(∠N )) (12.35)
∂Kp Kp
by the use of Equation (12.28), and Equation (12.7) is proved.

12.6.2 A proof of Equation (12.11)

The complementary sensitivity transfer function is given by

h0 (jω) h2 + hcos(φ) + jhsin(φ)
M (jω) = = = <e(M ) + j=m(M ) (12.36)
1 + h0 (jω) 1 + h2 + 2hcos(φ)
The amplitude of M (jω) is
h
|M | = p (12.37)
2
1 + h + 2hcos(φ)
The gradient of |M (jω)| with respect to Kp is
∂h ∂h ∂φ
∂|M | ∂Kp (1 + h2 + 2hcos(φ)) − h[(h + cos(φ)) ∂K p
− hsin(φ) ∂K p
]
= 3 (12.38)
∂Kp (1 + h2 + 2hcos(φ)) 2
which, by the use of Equations (12.32) and (12.33), may be shown to be
∂|M | 1 h2 + hcos(φ)
= |M |(1 − ) (12.39)
∂Kp Kp 1 + h2 + 2hcos(φ)
or
∂|M | 1
= |M |(1 − |M |cos(∠M )) (12.40)
∂Kp Kp
and Equation (12.11) is proved.
Note that, <e(N ) + <e(M ) = 1, which may be substituted into Equations (12.7)
and (12.11) for the derivation of alternative expressions.
12.6 Appendix 150

12.6.3 A proof of Equation (12.20)

From Equation (12.28) we have

h sin(−φ)
tan(∠N ) = (12.41)
1 + h cos(−φ)

The partial derivative of the right hand side of Equation (12.41) with respect to Kp
is given by

∂ tan(∠N ) 1 tan(∠N )
= |N | (12.42)
∂Kp Kp cos(∠N )

The chain rule on the left hand side of Equation (12.41) gives

∂ tan(∠N ) 1 ∂∠N
= 2
(12.43)
∂Kp cos (∠N ) ∂Kp

Combining Equations (12.42) and (12.43) gives

∂∠N 1
= |N | sin(∠N ) (12.44)
∂Kp Kp

and Equation (12.20) is proved.

 Part III

Model based control

Chapter 13

Modified Smith Predictor

13.1 Introduction

The Smith predictor is frequently used for controlling processes with long-dead times.
Unfortunately, the Smith predictor cannot compensate for load disturbances upon
an integrator processes with a long dead time.
When working with a course in process control the author constructed a modified
Smith predictor for controlling an integrator process with long dead time. We will
in this chapter discuss the Smith predictor and the necessary modifications in order
to control an integrator processes with a long dead time.

13.2 The Smith Predictor

Assume that a process model transfer function hm (s) is available for the true process
hp (s). The process model can be factorized into the product of a rational part
(which can be inverted) h− +
m (s) and a non-rational part hm (s) (non-invertible part,
e.g. transport delay). We have

hm (s) = h− +
m (s)hm (s) (13.1)

The Smith predictor is illustrated in Figure (13.1).

13.2.1 Transfer Function

Consider the Smith predictor in Figure (13.1). The transfer function model from
the two inputs, i.e. the reference and disturbance signals, to the process output is

y(s) = hr (s)r(s) + hd (s)d(s), (13.2)

where
(1 + hc (h−
m − hm ))hv
hd (s) = − , (13.3)
1 + hm hc + (hp − hm )hc
13.2 The Smith Predictor 153

d - hv (s)

+?
r - m - m - h (s) u - h (s) - m y -
c p
+ 6− + 6− +

+ ?
- h− (s) - h+ (s) - m
m m
−

Figure 13.1: The Smith predictor for controlling processes with long dead times.

and
hc hp
hr (s) = . (13.4)
1+ h−
m hc + (hp − hm )hc

13.2.2 Process with integrator and time delay

Assume that the disturbance (load) transfer function hv (s) is equal to the process
transfer function hp (s) = hm (s). I.e. consider a process:

1
hp (s) = hm (s) = hv (s) = e−τ s . (13.5)
s
The Smith predictor is constructed by putting:
1
h−
m (s) = , (13.6)
s
−τ s
h+
m (s) = e . (13.7)

The transfer function from the disturbance d to the process output is:

(1 + hc (h−
m − hm ))hv
y(s) = d(s) (13.8)
1 + h−m hc

Inserting gives

(1 + hc ( 1s − 1s e−τ s )) 1s e−τ s
y(s) = d(s) (13.9)
1 + 1s hc

series expansion of e−τ s gives

(1 + hc ( 1s − 1s (1 − τ s + 21 τ 2 s2 + · · · ))) 1s e−τ s
y(s) = d(s) (13.10)
1 + 1s hc
13.3 Modified Smith predictor 154

which can be written as

(1 + hc (τ − 12 τ 2 s + · · · ))e−τ s
y(s) = d(s) (13.11)
s + hc
Assume a constant disturbance (load) step change with amplitude ∆d at time zero.
We have:
1
d(s) = ∆d (13.12)
s
Using the final value theorem we find that:
(1 + hc (τ − 12 τ 2 s + · · · ))e−τ s
lim y(t) = lim sy(s) = lim ∆d (13.13)
t→∞ s→0 s→0 s + hc
1
= lim ( + τ )∆d (13.14)
s→0 hc (s)
For a PI controller we have that
1
lim =0 (13.15)
s→0 hc (s)
Using this we find that
lim y(t) = τ ∆d (13.16)
t→∞

From this we conclude that the off-set error is different from zero and that the error
is large when the time delay is large. Hence, the Smith predictor cannot compensate
for constant load changes when the process is a pure integrator with dead time.

13.3 Modified Smith predictor

The Smith predictor cannot reject load disturbances for processes with integration
(process with integrator and time delay).
Consider a process with a pure integration with time delay, i.e.
kp −τ s
hp (s) = e . (13.17)
s
The process model is specified as:
hm (s) = h−
m (s)e
−τ s
. (13.18)
kp
Instead of putting h−
m (s) = s the rational part of the process model is specified to
be
k
h−
m (s) = . (13.19)
s+a
It can now be shown that
lim y(t) = 0 (13.20)
t→∞

for a constant load step change v(s) = ∆v/s.

However, there is one drawback. The term e−τ s will occur in the denominator
term of the transfer functions hr (s) and hd (s) (the characteristic equation) when
hm (s) 6= hp (s).
13.4 Modified Smith Predictor by Åström 155

13.4 Modified Smith Predictor by Åström

The modified Smith predictor by Åström et al is illustrated in Figure 13.2.

Consider in the following that the process gain is kp = 1, see Figure 13.2. The
set-point response is given by
Y (s) k −τ s
Hr (s) = = e (13.21)
R(s) s+k
Note that the set-point response is independent of the choice of M (s). The transfer
function M (s) will be discussed later. The set-point response is simply a delayed
first order response. The controller gain k is chosen in order to shape the time
constant k1 of the 1.st order response. Note also that

lim Hr (s) = 1 (13.22)

s→0

Hence, we have zero steady state offset, i.e. y = r in steady state.

The load response is given by
1 −τ s
Y (s) se
Hd (s) = = (13.23)
D(s) 1 + M (s) 1s e−τ s

The problem is now to shape the transfer function M (s) so that HD (s) rejects
(steady state) load disturbances, i.e.,

lim Hd (s) = 0 (13.24)

s→0

It is also of interest to chose M (s) so that to remove the non-rational term e−τ s
from the denominator term of HD (s), in order to increase the bandwidth of the load
response.
The following transfer function with three adjustable parameters, k1 , k2 and k3
is proposed:

k4 + 1s k3
M (s) = (13.25)
1 + k1 + 1s k2 + s12 k3 − ( 1s k4 + s12 k3 ) − ( 1s k4 + 1
k )e−τ s
s2 3

where k4 = k2 + kr τ .
With this choice, the load (disturbance) response is given by

e−τ s (s2 (1 + k1 ) + k2 s + k3 − (k4 s + k3 )e−τ s )

Hd (s) = (13.26)
s(s2 (1 + k1 ) + sk2 + k3 )
The denominator is a third order polynomial in s.
The modified Smith predictor in Figure 13.2 has some remarkable properties. It
follows that the load response is decoupled from the set-point response. The signal
dˆ can be interpreted as an estimate of the (constant) input disturbance d. We have
ˆ = d.
lim d(t) (13.27)
t→∞
13.4 Modified Smith Predictor by Åström 156

d Process


r +  ũ + u ?+
 y
- k kp −τ s
 s e
- - - - -
 + 
6− 6−
dˆ

1
s
 M (s)
6e

ŷ - 
 y
- e−τ s 
− +

Figure 13.2: The Modified Smith predictor for controlling processes with an inte-
grator and long dead time proposed by Åström et al. (1994).

13.4.1 A Model Based Point of View

The modified Smith predictor, Figure 13.2, have a feedback from the load (distur-
bance) estimate. It is of interest to compare this strategy with a model based control
strategy. I.e. by using a state estimator. The states could be y and d where the
load d is modeled by a pure integrator.
From Figure 13.2 we have that the estimate of y is taken as
k −τ s
ŷ = e r (13.28)
s+k
In the time domain we have
ŷ˙ = −k ŷ + kr (13.29)
where we for simplicity has neglected the time delay.
Using a state estimator (Kalman filter) we would have used the following esti-
mator for the output y.
ŷ˙ = u + dˆ + K1 (y − ŷ) (13.30)
which can be written as
ŷ˙ = −K1 ŷ + u + dˆ + K1 y (13.31)
where K1 is the Kalman filter gain. We want to compare the two estimators, Equa-
tions (13.29) (13.31). When are they similar ?
From Figure 13.2 we have that
ũ = u + dˆ (13.32)
sk
Ũ (s) = R(s) (13.33)
s+k
13.4 Modified Smith Predictor by Åström 157

Hence in steady state we have (the first point)

ˆ
lim ũ(t) = lim (u(t) + d(t)) =0 (13.34)
t→∞ t→∞

and from (13.22 (the second point)

lim y(t) = r (13.35)

t→∞

By replacing u + dˆ and y in the state estimator (Kalman filter) Equation (13.31)

with the steady state values (i.e., u + dˆ = 0 and y = r), we have

ŷ˙ = −K1 ŷ + K1 r (13.36)

Intuitively, the estimator for y used in the modified Smith predictor is similar to a
state estimator (Kalman filter) updated from the steady state values of the inputs
to the estimator.
 Part IV

Control of multivariable systems

Chapter 14

Multivariable control

14.1 Interaction and pairing of variables

u1 - y1-

Hp
r2 - - h u2 - y2-
c22

−6

Figure 14.1: Multivariable control with one loop open.

We will in this section study the control of systems with two control input vari-
ables, u1 and u2 and two output measurements variables y1 and y2 . Such a system
is referred to as a (2 × 2) MIMO control system, where MIMO stands for Multiple
Input and Multiple Output. Often only MIMO systems are used to denote such a
control system.
The process described in Figure 14.1 is described by the transfer matrix model

y = Hp u (14.1)

where Hp is a 2 × 2 transfer matrix. y and u are 2 × 1 vectors. We have

y Hp u
z }| { z }| { z }| {
y1 h11 h12 u1
= (14.2)
y2 h21 h22 u2

which is equivalent with

y1 = h11 u1 + h12 u2 (14.3)

y2 = h21 u1 + h22 u2 (14.4)
14.1 Interaction and pairing of variables 160

When studying the transfer function from u1 to y1 we see that there is a cross
coupling from u2 represented by the transfer function h12 . If h12 = 0 then we see that
there is no cross coupling and the relationship between u1 and y1 is decoupled from
u2 . In the same way we see that there is a cross coupling from u1 to y2 , represented
with the transfer function h21 . I.e., u1 is influencing the transfer function from u2
to y2 .
For a 2 × 2 multivariable control system we have four possible pairings of inputs
and outputs variables. Those possible pairings are defined as follows:

u1 → y1 (using u1 for controlling y1 )

u2 → y1 (using u2 for controlling y1 )
(14.5)
u1 → y2 (using u1 for controlling y2 )
u2 → y2 (using u2 for controlling y2 )

For a 2 × 2 MIMO control system we have four different single input and single
output control system strategies, which are defined as follows

u1 = hc11 (r1 − y1 ) (loop u1 − y1 closed)

u2 = hc22 (r2 − y2 ) (loop u2 − y2 closed)
(14.6)
u1 = hc12 (r2 − y2 ) (loop u1 − y2 closed)
u2 = hc21 (r1 − y1 ) (loop u2 − y1 closed)

We will in this section study two problems of interests in connection with control of
MIMO systems. These problems are defined below.

Problem 14.1 (Interaction)

Assume that we want to design a controller for the loop u1 − y1 in Figure 14.1. We
have in this case two questions of interests. How will the interaction from the cross
coupling in the process influence the transfer function from u1 to y1 ? How will the
interaction from the controller hc22 for the closed loop system u2 − y2 influence the
transfer function from u1 to y1 .

Problem 14.2 (Pairing of input and output variables)

Assume that we want to use single loop controllers in order to control a MIMO
process. We have one central question of interest. Which control input is to be used
in order to control a particular output variable? We want to find out if it make sense
to use u1 or u2 in order to control y1 , and if it make sense to use u1 or u2 in order
to control y2 .

The solution to the pairing problem is therefore to chose the two best control strate-
gies.

14.1.1 Closed loop u2 → y2

We are using the system illustrated in Figure 14.1 as the starting point.
14.1 Interaction and pairing of variables 161

If u2 is described by a feedback from y2 , i.e., u2 = hc22 (r2 − y2 ), then we have

that the resulting transfer function from u1 to y1 will be given by
h12 h21 h12 hc22
y1 = h11 (1 − )u1 + r2 (14.7)
h11 h22 + h11 /hc22 1 + h11 hc22
See Section 14.1.2 for proof. This can be written as
h12 hc22
y1 = h11 (1 − ∆11 )u1 + r2 (14.8)
1 + h22 hc22
where ∆11 is given by
h12 h21
∆11 (s) = T22 (s) (14.9)
h11 h22
where T22 (s) is given by
h22 hc22
T22 (s) = (14.10)
1 + h22 hc22
T22 (s) is the complementary sensitivity function for the closed loop u2 − y2 if there
was no cross coupling, i.e. if h21 = 0. See Equation (14.18) below for a proof.
The transfer function ∆11 represent the interaction (from cross couplings in the
system and the controller hc22 ) on the transfer function from u1 to y1 . ∆11 describes
the interaction from the closed loop and the interaction from the closed loop in the
process. As we see, the resulting transfer function from u1 to y1 is dependent on the
controller hc22 in the closed loop u2 − y2 as well as the cross couplings h12 and h21
in the process.
If ∆11 is small then the interaction is small. We also see that there is no cross
coupling if h12 = 0 and/or h21 = 0, hence in this case ∆11 = 0.
We have shown that ∆11 which describes the interaction is dependent on the
controller hc22 .
Assume now that we want to close the loop u1 − y1 with a controller u1 =
hc11 (r1 − y1 ). The term ∆11 shows that the two controllers not should be controlled
independent of each other when the term h12 h21 /(h11 h22 ), which represents cross
couplings in the process, is large.
Consider now the case in which the controller hc22 have integral action, i.e. the
case when hc22 e.g. is a PI or a PID controller. We can in this case show that ∆11 is
independent of the controller hc22 at low frequencies. We have that T22 (s = 0) = 1.
This means that
h12 h21
∆11 (s = 0) = (s = 0) (14.11)
h11 h22
Often it is sufficient to study ∆11 (s = 0). The magnitude ∆11 (s = 0) describes if it
make sense to use u1 in order to control y1 . It is clear that u1 − y1 is a good pairing
if ∆11 (s = 0) is small, i.e. when ∆11 (s = 0) ≈ 0.
If T22 (s) ≈ 1 then we have the following approximation
h12 h21
∆11 (s) ≈ (14.12)
h11 h22
14.1 Interaction and pairing of variables 162

An advantage with this approximation is that it is independent of the controller

hc22 . If the model is known, then we can plot ∆(s) as a function of the frequency
ω where s = jω, e.g. in a Bode diagram. Hence, u1 − y1 will be a good pairing
strategy if the magnitude of ∆11 (s) is small.
As we see from Equation (14.8) the reference r2 for the closed loop u2 − y2
influences upon the output y1 via the transfer function h12 hc22 /(1 + h22 hc22 ). This
is a phenomena which often is omitted when studying single loop control of MIMO
systems. The contribution from r2 is influencing like a disturbance on the process
output y1 . This contribution may be important if h12 is large.

14.1.2 Proof of Equation (14.7)

We will in this section derive the results described in Section 14.1.1. The starting
point is Figure 14.1 with loop u1 − y1 open and loop u2 − y2 closed. The process is
described by

y1 = h11 u1 + h12 u2 , (14.13)

y2 = h21 u1 + h22 u2 , (14.14)

with the closed loop u2 − y2 described by the feedback

u2 = hc22 (r2 − y2 ), (14.15)

where hc22 is the controller transfer function.

If we are putting Equation (14.15) into Equations (14.13) and (14.14) then we
get the following description of the total system with loop u2 − y2 closed and loop
u1 − y1 open,

y1 = h11 u1 + h12 hc22 (r2 − y2 ), (14.16)

y2 = h21 u1 + h22 hc22 (r2 − y2 ). (14.17)

We know get the transfer function for the closed loop from Equation (14.17), i.e.,
T22
z}| {
h21 h22 hc22
y2 = u1 + r2 . (14.18)
1 + h22 hc22 1 + h22 hc22
The resulting transfer function from u1 to y1 is obtained by putting Equation (14.18)
into (14.16). This gives
h21 h22 hc22
y1 = h11 u1 + h12 hc22 (r2 − u1 − r2 ), (14.19)
1 + h22 hc22 1 + h22 hc22
which gives
h12 h21 hc22 h12 hc22
y1 = (h11 − )u1 + r2 . (14.20)
1 + h22 hc22 1 + h22 hc22
Equation (14.20) can be described as
h12 h21 h12 hc22
y1 = h11 (1 − T22 (s))u1 + r2 , (14.21)
h11 h22 1 + h22 hc22
14.1 Interaction and pairing of variables 163

where T22 (s) is the complementary sensitivity function for the closed loop u2 − y2 ,
i.e.,
h22 hc22
T22 (s) = . (14.22)
1 + h22 hc22

14.1.3 Rules for pairing variables

The system in Figure 14.1 is to be controlled by two single loop controllers. The
four possible control strategies is as shown in 14.6. In the same way as the develop-
ment in Section 14.1.2 and the result in Section 14.1.1 we can develop the following
expressions, ∆11 is represented for the sake of completeness.

T22
z }| {
h12 h21 h22 hc22 h12 h21
∆11 (s) = , ∆11 (s) ≈ . (14.23)
h11 h22 1 + h22 hc22 h11 h22

T11
z }| {
h21 h12 h11 hc11 h12 h21
∆22 (s) = , ∆22 (s) ≈ . (14.24)
h22 h11 1 + h11 hc11 h11 h22

T21
z }| {
h22 h11 h12 hc21 h22 h11
∆21 (s) = , ∆21 (s) ≈ . (14.25)
h21 h12 1 + h12 hc21 h21 h12

T12
z }| {
h11 h22 h21 hc12 h11 h22
∆12 (s) = , ∆12 (s) ≈ . (14.26)
h12 h21 1 + h21 hc12 h12 h21
We have here assumed that the controllers hc,i,j have integral action. This means
that Ti,j (s = 0) = 1 and that Ti,j (s) ≈ 1 at low frequencies. This is the background
for the above approximations. Note that ∆11 (s = 0) = ∆22 (s = 0) and that ∆21 (s =
0) = ∆12 (s = 0).
We form the following matrix of interactions
 
∆11 (s) ∆12 (s)
∆(s) = (14.27)
∆21 (s) ∆22 (s)

Foe s = 0 we have the following steady state interaction matrix, ∆(s = 0), i.e.,
   
∆11 ∆12 ∆11 ∆12
∆(s = 0) = = (14.28)
∆21 ∆22 ∆12 ∆11
This last expression is often valid approximately also in the dynamic case, i.e. if
1/hc,i,j  1), and at low frequencies when s ≈ 0. From the elements in the interac-
tion matrix ∆ we can find the pairings which make sense
14.1 Interaction and pairing of variables 164

It make sense to write down the following table

u u2
 1
y1  ∆11 ∆12
 (14.29)
y2  ∆21 ∆22

We are now assuming that the transfer function Ti,j (s) ≈ 1 and we therefore are
taking the approximative expressions for ∆ij (s) as the starting point. We have
the following rules for pairing of variables in single loop control of MIMO systems,
Balchen (1988).

Regel 14.1 (Pairing of variables based on the ∆ interaction matrix)

1. If | ∆i,j |≈ 0 then there is small interaction. It is therefore favourable to

control yi with uj .

2. If 0 ≤| ∆i,j |≤ 0.8 then there is some interaction. Then there is acceptable to

control yi with uj .

3. If 0.8 ≤| ∆i,j |≤ 1 then there is a strong interaction. uj − yi is therefore an

undesired pairing.

4. If | ∆i,j |≥ 1 then the pairing uj − yi should not be used.

Remark 14.1 The above rules for pairing of input and output variables is generally
valid when the elements ∆ij (s) varies with s, i.e. as a function of the frequency ω
where s = jω. A good approximation is that the transfer functions Ti,j (s = jω) ≈ 1
for frequencies less than the bandwidth of the system. For a 2×2 system it therefore
is sufficient to study the magnitude for
h12 h21
∆11 (s) = ∆22 (s) = (14.30)
h11 h22
and
h11 h22
∆21 (s) = ∆12 (s) = (14.31)
h12 h21

Remark 14.2 One should always first study ∆ under stationary behavior, i.e. first
put s = 0. Such an analysis is sufficient for most systems. Loosely spoken, this is
the case for systems which is easy to control. For a 2 × 2 system it normally holds
to study
h12 h21
∆11 (0) = ∆22 (0) = (s = 0) (14.32)
h11 h22
and
h11 h22
∆21 (0) = ∆12 (0) = (s = 0) (14.33)
h12 h21
14.2 Relative Gain Array (RGA) 165

Remark 14.3 The above rules for defining the pairings of input and output vari-
ables is valid also for m × m MIMO systems, i.e. for systems with m inputs and m
outputs. We are referring to Balchen (1988) for a description.

Example 14.1 (Control of a distillation column)

A linearized steady state model for a distillation column is given as follows
y Hp u
z }| { z }| { z }| {
y1 0.9033 −0.9137 u1
= (14.34)
y2 0.9366 −0.9262 u2
where
y1 − composition of bottom product [·]
y2 − composition of top product [·]
(14.35)
u1 − mass flow of reflux to the top of the column [kg/s]
u2 − mass flow of steam from the reboiler [kg/s]

From the elements hi,j in the steady state transfer matrix Hp (s = 0) we have that
h12 h21
∆11 (0) = ∆22 (0) = (s = 0) ≈ 1.02 (14.36)
h11 h22
and
h11 h22
∆21 (0) = ∆12 (0) = (s = 0) ≈ 0.98 (14.37)
h12 h21
We are now forming the following table from the interaction matrix ∆, i.e.,
u u2
 1
y1  1.02 0.98
 (14.38)
y2  0.98 1.02

This means that the following pairings should be chosen

u1 → y2 (using u1 for controlling y2 )
(14.39)
u2 → y1 (using u2 for controlling y1 )
remark that the interaction indices ∆12 = ∆21 = 0.98, which gives us the pairings
u1 → y2 and u2 → y1 , is very close to one. This is typically for distillation columns.
A distillation column is an example of a process with large interactions and strong
cross couplings between the input and output variables.

14.2 Relative Gain Array (RGA)

the starting point is a linear system described by the transfer matrix model

y = Hp u (14.40)

where y is a m × 1 vector, u is a r × 1 vector and Hp is a m × r transfer matrix. We

will in the following assume that that there is an equal number of input and output
variables, i.e. we assume that m = r. This gives a square transfer matrix Hp .
14.2 Relative Gain Array (RGA) 166

The RGA matrix Λ is defined simply as

Λ = Hp × (Hp−1 )T (14.41)
where × denotes element by element multiplication. We have assumed that Hp is
non singular. For a 2 × 2 we have that
   
λ11 λ12 λ11 1 − λ11 1
Λ= = where λ11 = (14.42)
λ21 λ22 1 − λ11 λ11 1 − 12 h21
h
h11 h22

Regel 14.2 (Pairing of variables based on RGA)

The starting point is the elements λij in the RGA matrix Λ.

1. Chose the pairing uj → yi for which the corresponding RGA element λij is
positive and so close to 1 in magnitude as possible.
2. The pairing uj → yi must be avoided if the RGA element is negative, i.e. when
λij < 0.

The RGA matrix have a number of properties. First we note that Λ is a sym-
metric matrix. It can also be shown that the elements in each row or column in Λ
sums to 1, i.e.
m
X
λij = 1 for i = 1, · · · , m. (14.43)
j=1

The reason for that one should not use pairings which coincides with negative
RGA elements is that such a pairing strategy is structural unstable. There may exists
stable control structures which corresponds to negative RGA elements. However,
the total system may be unstable if one of the control loops are broken ore put in
manual. Such a control system is defined to be structural unstable.
It is of importance to denote this difference of a structural unstable control
system and a unstable control system, in connection with pairing of variables by
using RGA analysis ore Balchen index analysis, for single loop control of MIMO
systems.

Example 14.2 (Controlling a distillation column)

We are taking Example 14.1 as the starting point. The following steady state model
for the distillation column are given, i.e.,
y Hp u
z }| { z }| { z }| {
y1 0.9033 −0.9137 u1
= (14.44)
y2 0.9366 −0.9262 u2
The RGA matrix is then given by
 
−43.72 44.72
Λ = Hp × (Hp−1 )T = (14.45)
44.72 −43.72
The correct pairing strategy is therefore u1 → y2 and u2 → y1 . The corresponding
RGA elements is positive for those pairings, i.e. λ12 = λ21 = 44.72.
14.2 Relative Gain Array (RGA) 167

14.2.1 Relationship between RGA and ∆

Consider a 2 × 2 system. In this case we have the following relationship between the
elements in the RGA matrix and the elements in the ∆ matrix.
1 h12 h21
λ11 = , ∆11 = (14.46)
1 − ∆11 h11 h22

1 h11 h22
λ12 = , ∆12 = (14.47)
1 − ∆12 h12 h21

1 h22 h11
λ21 = , ∆21 = (14.48)
1 − ∆21 h21 h12

1 h12 h21
λ22 = , ∆22 = (14.49)
1 − ∆22 h11 h22
Note that λ11 = λ22 , λ12 = λ21 , ∆11 = ∆22 and ∆12 = ∆21 . The pairing uj → yi
should be avoided if λij < 0. The similar should be avoided when ∆ij > 1.
Chapter 15

Multiple inputs and outputs

and control structures

15.1 Split Range Control

We will in this section discuss the case in which we have multiple inputs (manipulated
variables, control inputs) and only one output (one measurement) to be controlled.
In this case we have a redundant number of input manipulable variables.
A common situation is a standard feedback system of a SISO system with one
input u1 and one output y and where the input (and/or the output) becomes sat-
urated. This means e.g. that the input is bounded as umin 1 ≤ u1 ≤ umax
1 . In
this situation it may be useful and/or necessary to add an extra input u2 (or more
inputs) in order to overcome the saturation and in order to achieve the goal y = r
where r is a desired specified reference for the output y. A strategy could then be
to use u1 for control when y ∈ [ymin , y1 ] and a second manipulable input u2 when
y ∈ [y1 , ymax ]
In the split-range control strategy there are multiple controllers which acts in
parallel rather than in series as in the cascade controller strategy.
If a model of the MIMO system is available, an optimal controller strategy, with
e.g. integral action, is believed to be very useful in this situation.
References

Balchen, J. G. (1984). Reguleringsteknikk, Bind 1, Tapir.

Balchen, J. G. and K. I. Mumme (1988). Process Control, Structures and Applica-

tions. Van Nostrand Reinhold Company Inc., New York.

Balchen, J. G. (1990). Rational transfer function approximations to transport

delay. Modeling, Identification and Control, Vol. 11, No. 3, pp. 127-140.

Di Ruscio, D. (1992). Adjustment of PID control parameters. Modeling, Identifi-

cation and Control, Vol. 13, No. 4.

Kailath, T. (1980). Linear Systems, Prentice-Hall, Inc.

Mataušek, M. R and A. D. Micić (1996). A modified Smith predictor for control-

ling a process with an integrator and long dead time. IEEE Transactions on
Automatic Control, Vol. 41, No. 8.

Parlet, B. N. (1976). A Recurrence among the Elements of Functions of Triangular

Matrices. Linear Algebra and its Applications, Vol. 14, pp. 117-121.

Petkov P. H., N. D. Christov and M. M. Konstantinov (1991). Computational

Methods for Linear Control Systems. Prentice Hall International (UK) Ltd.

Skogestad, S. and I. Postlethwaite (1996). Multivariable Feedback Control, Wiley.

Åström, K. J., Hang, C. C. and B. C. Lim (1994). A new Smith predictor for con-
trolling a process with an integrator and long dead time. IEEE Transactions
on Automatic Control, Vol. 39, No. 2.
Appendix A

Laplace Transformations

In practice we need only a few basic rules about Laplace transformations and the
relationship to differential equations. Some of these will be described in this ap-
pendix.

A.1 Laplace transformation of the exponential

Consider the 1st order autonomous scalar differential equation

ẋ = −ax, x(t = 0) = x0 (A.1)

where we assume that a ≥ 0 is a positive scalar number. The solution to Eq. (A.1)
is

x(t) = e−at x0 , x(t = 0) = x0 (A.2)

From the theory of Laplace transformations we have that

L(ẋ) = sx − x(t = 0) = sx − x0 (A.3)

Then the Laplace transformation of Eq. (A.1) yields

1
x(s) = x0 . (A.4)
s+a
Using the property of linearity of Laplace transformations

L(ax(t) + by(t)) = aL(x) + bL(y) = ax(s) + by(s), (A.5)

we find from the laplace transformation of the solution Eq. (A.2) and Eq. (A.4)
that the Laplace transformation of the exponential is given by
1
L(e−at ) = . (A.6)
s+a
A.2 Laplace transformation of triogonometric functions 171

A.2 Laplace transformation of triogonometric functions

Consider the Laplace transformation of the exponential of a complex number. Using

Eq. (A.6) we obtain

1 s + jω s ω
L(ejωt ) = = 2 2
= 2 2
+j 2 . (A.7)
s − jω s +ω s +ω s + ω2
Using and comparing with the Laplace transformation of the Euler formula

ejωt = sin(ωt) + j cos(ωt), (A.8)

we find that
s
L(sin(ωt)) = , (A.9)
s2 + ω 2

ω
L(cos(ωt)) = . (A.10)
s2 + ω2
Using the identity

sin(ωt + φ) = sin(ωt) cos(φ) + cos(ωt) sin(φ), (A.11)

gives

cos(φ)s + ω sin(φ)
L(sin(ωt + φ)) = . (A.12)
s2 + ω 2