2.

13 Controllability Observability and Stability

CONTENTS of Linear Time Invariant Discrete Data System 2.33
2.13.1 Introduction 2.33
2.14 Controlability of Linear Time-
CHAPTER - 1: INTRODUCTION 1.1 - 1.4
Invarient Discrete-Data System 2.35
1.1 Introduction 1.1 2.14.1 Definition of Controlability 2.35
1.2 Basic Elements of a Discrete-Data. Control System 1.2 2.14.2 Theorems on Controllability 2.36
1.3 Advantages of .Discrete-Data Control Systems 1.3. 2.15 Observability of Linear Time-
Two Marks Questions and' Answers 1.4 ' Invariant Discrete - Data Systems 2.39
2.15.1 Definition of Observability 2.39
CHAPTER - 2: THE STATE-VARIABLE' TECHNIQUE 2.1 - 2.96
2.15.2 Theorem on Dbserv ability 2.39'
2.1 Introduction 2.1 '2.16 Stability of Linear Digital' Control Systems-
2.2 State Equations and State Transit Definitions and Theorems 2.-41
Equations of Continuous-Data Systems 2.1 2.16.1 Definitions on Stability 2.42
2.3 The State. Transition Matrix 2.17 Stability Tests of Discrete - Data Systems 2.43
Solutions' of Homogeneous State Equations 2.3
2.17.1 The Bilinear Transformation Extension of the
2.4 Properties of State Transition Matrise 2.5 ,Routh -Hurwitz Criterion 2.44
.2.5 Solution of Non-homogeneous State Equation 2.5 2.17.2 Jury's Stability Test 2:49
.2.6 State Equations of Discrete-data 2.17.3 The Second Method of Liapunov 2.53
. ~ystem with Sample' and Hol~ 2.6
2.18 Pole-Placement Design and State Observers 2.59
2.7 State .Equations of. Digital Bystems with All Digital Elements 2.8 2.18.1 Introduction 2.59
2.8 The State Trans~tio'~ Equations 2.9
2.18.2 Stability Improvement by State Feedback 2.60
2.8.1 The Recursive. Method 2.9
2.18.3 Necessary and Sufficient Conditions
2.8.2 The 'Z-Transform 'Method 2'.12 for Arbitrary Pole-Placement 2.64
.2.9 Methods of .computing stateTransition Matrix 2.16 2.18.4 ,State Regulator Design 2.68
2.9.1 The. Cayley-Hamilton Theorem Metho,d 2.17 2.18.5 Design of State Observers 2.70
2.9:2, .The Z-Transform Method 2.18 2.18.6 Full-Order State Observer 2.71
.2.9.3 Computing State Transition Matrix f(T) 2.21 2.18.7 Reduced - Order State Observer 2.77
2.10 State Diagram of Digital Systems. 2.24 2.18.8 Digital Control System with State Feedback 2.80
2.11 Decomposition of Discrete-Data Transfer Function 2.27 2.18.9 State Regulator Design 2.80
2.11.1. Direct Decomposition 2.28 2.18.10 Design of State Observers 2.83
2.11.2 Cascade Decomposition 2.30 2.18.11 Prediction Observer 2.84
2.11.3 Parallel Decomposition 2.31 2.18.12 Current Observer 2.85
2.12 State 'Diagram .of Discrete Data 2.18.13 R~duced - Order Observer 2.86
System with Zero-Order Holds 2.32 Two Marks Questions and Answers 2.90
2.12.1 State Diagram of. the Zero-Order Hold 2.32
 C'HA,PTER - 3: I'NTRO:DUCTION TO
SYSTE,M IDENTIFICATIO,N 3.1 - 3.6 6.5.3 Developement of Closed Loop Transfer Function 6.15
3.1 Nee.d, for Modeling Dynamic Systems 3.1 6.5.4 Design of Sample Data Control System 6.20
3.2 Mathematical Modeling and System Identification 3.3 6.5.5 Dead .Beat Control Algorithm 6.24
'3.3 System Identification Steps 3.4 ' 6.5.6 Dahlin Algorithm 6.32
Two Marks Questions and Answers 3.5 6.5.7 Ringing and the Placemerit of Poles 6.39
6.5.8 Digital PID Controller 6.43
CHAPT.ER - 4: ,NON PARAME,TRIC METHODS 6.5.8.1 'Position Algorithm 6.43
,OF SYSTEM IDENTI-FiCATION 4.1 - 4.16
6.5.8.2 Velocity Algorithm 6.44
4.1 Introduction 4.1 6.5.8'.3 Features of Digital PID Controller 6.45
4.2 ,Transient Analysis 4.2 6.5.9 Smith Predictor Algorithm 6.48
4.3 'Frequency Analysis 4~,5 , 6.5.10 Internal Model Control 6.54
4.4', Correlation analysis 4.8 6.5.11 Digital Feed Forward Control 6.59
4.5 Spectral analysis 4.11 Two Marks Questions and, Answers 6.62
,Two Marks Questions and Answers 4.15
CHAPTER -7: MULTI-LOOP REGULATORY CONTROL 7.1 7.50
CHAPTER - 5: PA,RAMETRIC 'METH'ODS O'F 7.1 Introduction 7.1
SYSTEM liDENTIFICATION 5.1 - 5.12
7.2 Multiloop Control 7.2
5.1 Introduction 5.1
7.3 Process Interaction 7.3
5.2 Least Squares Estimation . 5.1
7.4 'Interaction of Control Loops 7.11
5.3 Recursive, Identification Method 5.5
7.5 Steady State Effective Gain of
5.4 ' Recursive Least Squares Estimation 5.7 , Two Input and Two Output Processes 7.16
Two Marks Questions and Answers 5.10 7.6 Relative Gain Array and Selection
of Variable Pair with Minimum Interaction 7.23
CHAPTER - 6:DIGITALC'O'NTROLLER D'E'SlGN 6.1 -6.64
7.7 Multi Loop PID Controller 7.40
6.1 Review of Z-Tr~nsfqrm.ation 6.1 7.8 Biggest LOG-Modulus Tuning (BLT) Method 7.41
6.2 Z- Transfonnof Various Function 6.2 7.9 Decoupler 7.43
6.3 The inverse Z- Transform 6.5 Two Marks Questions and Answers 7.50,
6.4 Modified' Z-transform 6.8
,6.4.1 Evalution of Modified Z-Transform 6.8 CHAPTER - 8: MULTIVARIABLE REGULATORY CONTROL 8.1 - 8.16
6.4.2 Modified 'Z-Transforms of some Simple Function 6.9 8.1 Introduction 8.1
6.5 Pulse Transfer Functions ,6.10 8.2 Multivari-able PID' Controller 8.3
6.5.1 Complex Series Representation of the Sampler 6.10 8'.3 Model 'Predictive Control (MPC) 8.7
6.5.2 Development of Pulse Transfer Function 6.12 ,8.4 Dynmic Matrix Controller (nMC) 8.11
 1.2 computer Control. ofProcesses )
CHAPTER 1.2 BASIC ELEMENTS OF A DISCRETE - DATA CONTROL SYSTEM

1 Figure 1.1 shows the basic elements ofa typical closed - loop control
system with sampled data. The sampler simply represents a device or
operate that outputs a pulse train. No information is transmitted between '.
two consecutive pulses. Figure 1.2 illustrates typical input and output -of
a sampler. A continuous input signal e(t) is sampled 'by the sampler, and
the output is a sequence of pulses. In the illustrated case, the sampler is
assumed to have a uniform sampling rate. The magnitudes of the pulses
at the sampling instants represent the values of the input signal ett) at
INTRODUCTION the corresponding instants,
In general, the sampling scheme has many variations, some of these
are the periodic, cyclic-rate, multirate random, and pulse width-modulated
1.1 . INTRODUCTION samplings. The most eom~on type of sampling found in practical systems
is the single-rate periodic sampling, such as that shown in Figure 1.2.
In recent years significant progress has been made in 'the analysis
and design of discrete data and digital control _systems. There systems c(t)
r(t)
have gained popularity and importance in industry due to in part to the
advances in digital computers for controls and,more recently, in
miroprocessors (MP) and digital signal processor.
Discrete - data and digital control system differ from the
.Fig. 1.1: Closed - loop sampled-date. control system
conventional. continuous - data ~n analog ,systems in that the signals in
one ormore parts of these systems are in the form of either-pulse trains e(t) e~ (t)

or numerical codes. The terms sampled data control systems, discrete-

data control systems, and digital control systems have all been used loosely
and inter changeably in the control systems literature.
\\
Strictly speaking, sampled data are pulse-amplitude modulated
signals. A pulse - amplitude modulated signal-is often .presented in the
,\ -.....
\
\
form of a pulse train with signal information carried by the amplitudes of 010-------+----1-----...
T 2T 3T 4T \
the pulses.
"
Digital data usually are those signals generated by digital computers
or digital transducer, they are often in somekind of digitaly coded form. (a). Continuous data input to sample (b) Discrete data output of the sampler
It will be shown later that practical systems found in industry often Fig. 1.2
contain analog, .sampledvas well as digital data. Therefore, in this text
we shall-use the termdiscretedata systems in-a broad sense to describe The filter located between the sampler and controlled process is
all. systems having some form of digital or sampled signals. used for .the 'purpose ofsmoothing.since most controlled processes, such
as the ones involving a conventional ac or de motor, are naturally designed
and constructed to receive analog signals.
(Introduction (1.3 ) . ComputerControlof Processes J
The block diagram of a typical digital control system is shown in Integrated Control of Industrial Systems
Figure 1.3. The existence of digitally coded signals, such as binary-coded Feedback control' is only one of the functions of a computer. In fact,
signals, in certain parts of the system requirethe use of digital-to-analog most of the information transfer between the process and computer
tD/A) and analog-to-digital (AID) converters. exploits the logical decision -making.
The digital computer block in Figure 1.3 can be a special-purpose Example:
digital computer, a microprocessor' or a.digitalsignal processor. Although
there are. basic difference between the hardware structure and Production planning,scheduling, optimization, operations control
components between sampled-data and digital control system. . etc, may now be implemented.

We shall show that, from an analytical stand ..p oint, both types of Future Generation Control Systems
systems ,are treated by the same analytical tools, The study of emerging applications 'shows that Artificial Intelligence
Digital __- - _ (AI) will affect the design and. application of control systems, as profoundly
code Digital..; To-
input Analogconverter Output as the impact of micro processors in the last two decades. It is clear that
future generation control systems will have a significant AI component,
the list of, applications of computer-base control will continue to' expand.
Analog - To-
Digital converter

Fig. 1.3: A typical control system

TWO MARKS OUESTIONSAND ANSWERS
111 State the advantage of digital control.
1.3 ADVANTAGES OF DISCRETE-DATACONTROL SYSTEMS .. Flexibility
... Wide selection of control Algorithms.
Flexibility
.. .Integrated control of Industrial.' system.
An important advantage offered by digital control is in the flexibility , .. Future generation control .systems.
of its modifying controler characteristics, or in other words." in adaptability
of the controller if plant dynamics change with operating conditions. 2. Draw the block diagram of a typical digital control. svstem,

The ability to 'redesign' the contoller by changing software (ratherthan Digital- To-!
Output.
hardware' is a.n important feature of digital control against analog control). , Analog converter

Wide Selection of Control Algorithms

Analog - To-
Implementation of advancedcontrol techniques was earlier constrained Digital converter
by the limitations of analog controlers and the high costof digital computers.
However, "with the advent of inexpensive' digital computers. 3. What is meant by sampling of signals.

However'; with the advent of inexpensive digital computers with In signal processing, sampling is the reduction of a continuous-
virtually limitless computing power, the techniques of modern control theory. time signal to' discrete time signal. Sample is a value or set of
.may now be put to 'practice. For example, in multivariable control systems value, at a point in timaand/or space.
with more than one input' and one output modern .techniques for optimizing
system performance or reducting interaction between feedback loops can
now be implemented.
 2.2 ComputerControl of Processes )

The q outputs of the systems are related to the state variables and

2 .inputs through output equations which are of the form,

Ck (t)
.

gk[Xl (t), X2
.:W i

(t), ..~' Xn (t), U 1 (t), U 2 (t), ... , Up (t), t] .. (2.2)

for K'= 1,2, ... q

Similar remarks can be made for gk as for (;..
The state equations and output equations are called dynamic
equations of the system.
THE STATE-VARIABLE TECHNIQUE It is customary to write dynamic equations in vector-matr-ix form.
Infact, the compact matrix notation is considered one or-the advantages
of the state-variable method. In matrix form,
2.1 INTRODUCTION State equation
In the state-variable formulation, a continuous-data system is
<t> [xtt), U (t), t] ... (2.3)
represented by a set of first-order differential equations, called state
equations. For a discrete-data system when all the variables are. defined
in discrete time KT, the state equations .are in the form of first-order u1 (t) - - - - + I 1------. C1(t)
u2(t) - - -+I Linearsystem ~-.... ~(t)
difference equations, are in the form of first-order difference equations.
As mentioned earlier, a discrete-data .control system ofteri contains
continuous-data as well as digital-data components, and the state
equations of the system will generally consist' of both first-order difference Fig. 2.1:A linear system'with p inputs,q outputs andn state variables .
'as well as first order differential equations. For this reason, we shall begin
Output Equation
by reviewing the state equations and their solutions of linear continuous-
data systems. Cft) = g [xtt), net), t] u. (2.4)

where x(t) is an n x 1 column matrix and is called the state vector; that is.
2.2 STATE EQUATIONS AND STATE'TRANSIT EQUATIONS OF CONTINUOUS
DATA SYSTEMS
Consider that an nth-order continuous data system with 'p'; inputs ... (2.5)
and 'q' outputs, as shown in Figure 2.1 is characterized by the following
set of n first-order' differential equations (state equations)
The input vector U(t) is a P x 1 column matrix and is denoted as
dXl(t)= f .
[Xl (t), X2 (t), Xn (t), U 1 (t), U 2 (t), ', Up (t), t] (2.1)
i
dt
i = 1,2, ... n
U(t) -- ... (2.6)
where Xl (t), X2(t), ... , Xn (t) are state variables, U 1 (t), U 2 (t) ... Up (t) are
input variables, and h denotes the i th functional relationship. III general,
n can be linear or. a .nonlinear function.
 2.4 ComputerControl of Processes ) ,
fTheState-variable Technique : 2.3
To determine state transition matrixott), we take_laplace transform
The output vector C(t) is defined as
on both sides -of Equation (2,.jO ), we get
S xes) - x(D) = A xes)
Solving, Xes) from last equation, we get
OCt) ... (2.7)
xes). = (SI - Ar" x(o) ... (2.13)
Cq(t) - where we assume that matrix (81 - A) is. non singular. Taking the
inverse laplace transform.
Which isa q x 1 column matrix or vector.
On both sides of the last equation yields
If the system is linear and time-in variant the dynamic equations xtt) = L-1 [eSl - A)=-l] xeD) t ~ 0 ..(2.1)
are written as
Comparing Equations (2.12 ) .and (2-.10 ), the state transit matrix is
, dx(t) written
A x(t) + B U (t) ... (2.8)
dt ... (2.14)
and An alternative method of solving the homogeneous state equation is
C(t) = D xCt) + E U (t) ... (2.9) to assume a solution, as in the classical method of solving linear ordinary
differential equations, and. then substitude it in the state equation to show
where A is an n x n matrix, B is an n x p matrix, D is a q~ xn matrix,
that it is indeal a correctsolution.
-E is a q x P matrix, all with constant coefficients.
Let the solution of equation (2.10) be
2a3 -lH,ESTATE TRANSITIDNMATRIH SOLUTIONS OF HOMOGENEOUS STATE x(t) = eAt xeD) (2.15)
EDUATIDNS - for t :2: 0, where eAt represents the power series of the matrix At, '
The state transition matrix, <I>(t) is defined as . n x n matrix that 1 2 2
-satisfies the homogenous state equation, = I+At+ + ... + ... ... (2.16)
2iAt

dx(t) = A x (t) Taking the derivative of last equation with respect to t, we get,
... (2.10)
- dt
d(e At ) -
Thus 4>Ct) statisfies the equation ---
dt
=' Ae At

Which has same form as Eq (2.11). In addition to equation (2.14)

d~(t) = A<f>(t) . . ... (2.11)
dt we obtained and other expression for S'rM
Let x (0) denote the initial state at t =0 then <p (t) is defined by
(2.16)
matrix equation
x (t) =$(t) x(O) ... (2.12) For non. Z8YO initial line, to, the solution of the homogeneous state
Which is the solution of the homogeneous- state equation for t 2 O. equ-ation is write .
xrt) = $ (t - to) x(to) ... (2.17)
 (, 2.5
2.6
[The State-Variable Technique
The state transition equation for t ~ to is particularly useful for the
2.4 PROPERTIES OF STATE TRANSITIO~J I\.~ATRISE analysis of discrete-data system, since the transition of state is carried
Several useful properties of the state transition matrii $(t) are state. out from one discrete time instant to the next, and the intial time to .is
as follows: advanced each to the next time instant. Once the state transition equation
is determined, the output vector crt) is written
~(O) = I (idc~tity matrix)
. t
, <p(t) is nonsingular for finite elements in A
C(t) = D<p(t-to)x(to) + JD<Ii(t-'t)BUCr)d't + EU(t) ... (2.22)
4>-1 (t) <1>< ~t) to

<!>(tl - t2) <I> (t2 - ta) <!>(t 1 - ta) for any t l , tz, ta' 2.'6 STATEEDUATIONSO,f 'OIS.CRETE-DAJASYST,EMWITHSAMPLE,AN:D -
HOLD
2.5 SO,LUTION OF NON HOMOGENEOUS STATE EOUATION
A. typical multichannel discrete data system with sample-and-hold
The state transition equation (S'I'H) devices or zero-order holds (Zoh) is shown in Figure 2.2.
The. n~n homogeneous state equation
e1(t) - - _ X
u1{t) c1 (t)
dx(t) . , Linear system
- - = A x(t) + BU(t) ... (2.18)
u2(t)
dt e2(t) ,- _-X ,~(t),

is now solved usingthe laplace transfer method. Taking the laplace

transform on both sides of last equation, we have
x., x2,xn
s xes) - x (0) ='A xes) + BU(s) u*p(t)
cq(t)
Solving xes) in the last equation, we get
xes) = (81 - A)-l xeD) + (81 - Ar? BU(s) ... (2.19) , Fig. 2.2: A multlehennel digital system with sample and held device

Using the definition of state transition matrix in .equation (2.14 ) The outputs of the ZOH are described by
and the convolution integral, the last equation is written as UiCt):::: Ui (kT) = ei (kT) kT ~ t <k +i ... (2.23)
t
where k = 0,1,2,and 1,2, ... p
x(t) Ht)Jl:(O) + fHt-'t)BUCr)d't t~O .. (2.21)
Let the dynamics of the linear system be 'represented by the state
o
equation in equation (2.-8). The state transition equation is given by
This; solution of state equation' is called state transition equation. equation (2.21) for all t and to. Since the inputs are constantbetween any
By setting the initial time at t" = to and using x (to) as initial state we , two consecutive sampling periods, the input vector U(T) in equation (2.21)
can show by using equation (2.20) and some of properties of q,(t) that the has, the properly
state transition equation is written generally as IJ( 't)" == TJ(kT) for kT ~ t < (k + 1) T ... (2.24)
Thus, in equation-(2.21) Utt) can be placed outside the integnal, and
t .

x(t) = <p(t-to)x(to) + f~(t- 't)BU('t)d't ...t .~O ...( 2.21) equation (2.21) becomes
to ~
 2.8 computer Control of Processe:!}
.(The state..\ta.riable Technique . 2.7
T

t e(t) JHT-t]Bdt ... (2.32)

x(t) == cp(t-to)x(to) + fHt-t)BU(kT)dt (2.25) o
to
Whichis valid for the time interval, kT :s; t s (k + t) t where we havereplaced m by 'to Equation (2.29) represents a. set of
first order difference equations, refe~red to as the discrete state ~quation
Setting to =' kT, equation (2.25) becomes
of sample data system shown in Figure (2.2). These state equations,
t . however, describe the dynamics of the system only at the sampling
x(t) = cp(t-kT)x(kT) + fCP(t-1:)Bd1: U(kT) ... (2.26) instants. In other words, by setting to = kT and t = (k + l)T in equation
, kT (2.26), all Informa'tion on the system between the sampling instants is
lost.
For kT ~ t ::; (k + l)T. Thelastequation describes the state vector x(t)
at all times between the sampling instants kT and (k +.l)T, k = 0,.1, 2, ... In a similar manner, the output equation in equation (2.9) is
. discretized by setting t = kT, then
Equation (2.?6) can be simplified by setting
t
C(kT) = D x (kT).+ E U (kT) ... (2.33)

aCt - kT) f<P(t-1:)Bd1: ... (2.27) The discrete state equation in equation (2.29) along wiht the output
kT equation in equation (2.33) are defined to be the dynamic equations of the
system in Figure (2.2).
, Thus equation (2',26) is written as
xtt) = <I> (t- kT) x (kT) + e (t -kT) U (k'I') .~. (2.28) In certain situations, the sampling operation is not done wihtrespect
th
to time, and the Iatter is not the independent variable. Under these
Notethat although the i component of U(kT) , Uifk'I') is held constant -
circumstances the discrete dynamic equations describe the discrete
only for thetimeinterval from t =k to t= (k+ 1)T, the solution in.equation
'events' and are expressed as
(2.28) is valid for the entire interval including t = (k + 1) T, since x(t) is a
continuous' function of t. For numerical iteration, it is more conveilientto x (k--L) = <I> (1) x (k) + e (1) U (k) . ... (2.34)
describe x(t) only at sampling instants. We let t = (k + 1) T, and equation C(k) = D x(k) + EU(k) ... (2.35)
(2.28} becomes .
which are obtained from equations (2.29) and (2 ..33), by setting T = 1.
x[(k _+ 1) T] = <f>(T) x(kT) + 8(1') U(kT) ... (~.29) Another way of looking at the notation is that" in equations (2.34) and
where from equation (2.16) (2.35) the sampling period T has been normalized to be one.

th(T) =. eAT = I +AT + l.-A2T2 + l.-A3T3 + ... ... (2.30) 2.1 STATE EQUATIONS OF DIGITAL SYSTEMS WITH ALL-DIGITAL ELEMENTS
'I' . 2i 3i
From equation (2.27) When a digital system is composed of all-digital, the system may be
described by the following discrete dynamic equations
(k+l)T

e (T) = f <p[(k+l)T-1:]Bd1:
KT
... (2.31)
x (k + 1) =' A x (k) + BU (k)
C(k):: D x (k) + EU (k)
... (2.36)
... (2.37)
By setting m =- ~T -'t, we get dm = dr, (K + 1) T ..;.. 't =T- ID, the last
equation is written as
 (TheState-Variable Technique 2..9 ( 2.1'0 ) Computer Control" of pr8ess~

The difference between (2.29) and (2.36) is that, in the for:mer, the Since from the properties of state transition matrix
coefficient .matrix <!>(k) is the state transition. matrix of the A matrix of <p (tl - t2) <I> (t2 - t"s) <p (tl- ts)
the continuous-data p~ocess, and 8Ck)js. a function of $(t) and B; thus 'they
:. Which implies <p(T) <p(T) - <p(2T)_ ... (2~40)
must all confirm with the .properties of the st~te transition matrix'. In
equation (2.36), however,' the matrices A.anda can be arbitrary and depend Equation (2.39) can bewritten as
entrely on the characteriatics of the digital system.
x(2T) = $(2T)x(O) + ~(T)8(T) U(O) + e(T) U (T) ... (2.41)

2.8 THE STATE TRANSITION EQUATIONS Continuing for k = N - 1

The solutions of the discrete state equations in equation (2.29), (2.34) x(NT) =<f>([(N-l) T] x [(N-l)Tl + e [(N -1)T] U [(N - 1)' T] ... (2.42)
or (2.36) are the state transition equations of the discrete - data system, Substituting the first N - 1 equations for x[(N-l)Tl; x[(N-2)TJ ... x(T),
Since these equations of all same form, their solutions will be very similar. into last equation sucessively, we' have
However, equations (2.29) and (2.34) contain statetransit.ion.matrix N-l
<peT) of the A matrix of equation (2.8), which is .nonsingular. Thus, ~he x(NT) = cj> (NT) x(0) + L<t>[(N-i-l)T]e(T)U(iT) ... (2.43)
.transit,ion of state for the systems described by equations (2.29) and (2.34) L=O
can take place in the forward or backward direction.
which is the solution of equation (2.29) given the initial state x(O)
For the discrete state equation' in' equation (2.36), generally, the and the input U(iT), for i = 0, 1, ... N-l.
matrix A has no restriction. However if the state transition is to progress
Inderiving the equation (2.43) we have use the property
in the negati~e direction. A must be non singular.
We'.shall derive the state transition equation of the state equation <p(T) <p(T) = <PCT) = <p(NT) .... (2.44)
in equation (2.29) and then extend these to equation (2.34) and (2.36). which follows directly from the property of the state transition
matrix <per), equation (2.44) implies that if <P(t) is a state transition matrix
.2.8.1 The Recursive Method satisfies all the properties of state transition matrix, the multiplying <P(T)
by itself N times is' equivalent to replacing all the T in <p(T) by NT.
The most straight, forward method of solving equation (2.29) is by
, recursion. The following 'equations, are written by substituting The time frame of equation (2.43). Equation (2.44) implies that if(T)
k = 0, 1, 2, ... in equation (2.29) sucessively, is a state transition matrix that satisfies all the ,properties of state
transition Matrix, then multiplying <p(T) by itself N times is equivalent to
k =0 replacing all the T in$(T) by NT .
x(T) ~(T) x(O) +, SCT) U (0) ... (2.38) The time frame of 'equation (2.43) c~n be shifted forward by any
positive integer M, So that solution of x[(N+M)] T is given in terms of the
k = 1
initial state x(MT) and inputs U(iT) for i=M,
x(2T) ~(T) x(T) + SCT) U(T)
M + 1 ..... N +M - 1. 'I'his is accomplished by solving x(O) from
= ~(T) [<!>(T)x(O) + H(T) UCO)]' + SeT) U(T) equation (2.43).
<p(T) <p(T) , xeD) + <pT S(T) UCo) + S(T) U(T) ... (2.39) N-l
xeD) cj>{-NT)x(NT)- 2:<1>[( -i-l)TJ8(T}U{iT) ... (2.45)
T _1'1 '
 ( 2.12 ) computer Controlof Processes )

Where we have utilized the properties of, 2.8.2 The Z transform Method
[4>CNTJ]'-l The Z transform method can be applied to solve the linear time-in
<PC-NT
varios discrete state equations in equation (2'.29)" (2.34) and (2.36). This is
$(-NT)$. ~IN "- i - DT] = <I> [(-i-t)TJ ... (2.46)' carried' out in the following for Equation (2.29).
Replacing N with N+ M in equation (2.45), we have The state equation in equation (2.29) are represented as
. N+M-l
x[(K+l)T] = q, (T) x(KT)+8(T)U(KT) ... (2.29)
x [(N+M)~J = <I>[(N +M)T]x(O)+ L <t>[(N+M-ic-l)T]8(t)U(iT)
i=O TakingZ transform on both sides of the last equation, vee have,
.., (2.46) z.~x(Z) = z x(O) = cp(T) X(z). + e (T) Ll(z) h_ (2.53)
Substituting equation (2.45) into the last equation and simplifying
where x(Z) is defined as the Z-transform of the state vector x(KT)
'. the .state transition equation for any initial state x(MT) is written as
and is given as
N-l '
00
x [(N+M)T]'=cI>[(NT)T]x(MT)+L<t>[(N -i-l)T.J.S{t)U[(M + i) TJ x(Z) Lx(KT)Z-K ... (2.54)
i=O'
K=O
... (2.46)
and the same applies to U(z). Note'that taking the Z-transformof a matrix
The solution of Equations (2.34) and (2.36) are easily obtained and is equivalent to taking Zvtransformofevery element of the matrix. Solving
are presented as follows.' forx(Z) from equation (2'.53),. we have
State equation
-xfz) = '[Zr-$(T)J-1Zx(O)+[ZI-$(T)J-1e(T)U(z) ... (2.55)
x(k+l) = <1>(1) x(k) ,+ 8(1) UCk) , ...(2.34)
The inverse Z - transform of the last equation is
State Transition Equation
1
N-l x(KT) = Z-1[(ZI-HT)f Z]x(O)+Z-l {(ZI-HT)tEl(T)U(z)}
x(N) <I> (N)'x(0) + L q> [(N- i .; 1)] e(1) U (i) ... (2.49).
......(2.56)
i=O
WOe shall show that inverse Z -transform of [Zl - <f>(T)]-l Z is the
State equatdon
discrete state transition matrix q, (KT).
x(K+l) = A x(K) + B U(K) ... (2.50)
The Z _. transform of q,(KT) is defined in the usual manner as
State Transitio~ Equation
00

N-l q,(Z) Lq,(KT}Z-K ... (2.57)

x(K+1) ANx(O) + LAN":;-l BU(i) ... (2.51) K=O
i=O .
Pre multiplying both sides of the last equation by q,(T) Z-l.and
where' , subracting the result from equation (2.57), we have
AN = A.A.A A ... (2.52) [I - q,(T) Z~l] $(Z) = I ... (2.58)
(The State-Variable Technique [ 2.13 ) ( 2.14,) Computer ControlofProcesses]

Thus,
IExample 2.1 I
... (2.59)
In this example we shall illustrate the analysis of an open-loop
Taking invers Z transform on both sides. discrete-data system by the state variable method presented' above. The
block diagram of the sytem under consideration is shown in 'Figure 2.3 .
... (2.60) The dynamic equation that describe the linear process are.

. Thus the, equation (2.60) represents the Z-transform method of

determining the state transition matrix of a discrete state equation.
The last term of equation (2.56) is evaluated by the use of the real
rdX~t(t)] . [ 0 1]
... (2:.61)
d X2(t) ~2-3
convolution theorem and equation (2.60). Therefore it can be shown that
convolution theorem. l dt

[:convolution theorem C(t) = Xl(t) ... (2.65)

If the function fl(t) and f2(t) haveZ transform Fl(Z) and F 2(Z) . r(t)
respectively and fl(t) = 0 andf2(t) = 0 for t < O. XT c(t)

then F'igure 2.3: An open-I,oop digital system-

where Xl(t), X2 (t ) are 'state variables C(t) is the scaler output, and
UCt) is the scalar input. Also since U(t) is the output~ Output of the ZoH
U(t) U(KT) =r(KT) ... (2.66)
. K~ _
for KT s t s (K+l) T
Z':l [[ZI"';cP(T)fe(T)U(Z)] = ~[cP(K-i-1)TJxe(T)U(iT). ...(2.61)
1=0 . Comparing equation (2.64) with the standard state equation form of
, equation (2.8).
The entire state transition equation is

x(KT) = cp
'x-i
(kT).X(O) + LcP[(K-i-l)T]e(T)U(iT) ... (2.62)
A = [~2' _~] B = [~] ... (2.67) .

i=O The followingmatrix is formed

Which is the same form of equation (2.43). In a similar maner, the

state equations of equation (2.34) and (~,.36) can be solved by-means of Z-
(SI-A) = [ ~ -1 ]
8+3
... (2.68)

transorm 'method illustrated above. Now taking Z-transform of AK implies

1
[SI-AJ-l 2
1 [8 + 3 8 ] ... (2.69)
8 +38+2 -2
... (2.63)
The state transition Matrix of A is obtained by taking inverse Laplace
transform of (SI-A)-l. Therefore,. from equation (2.14)
(The State-Variabl,e Technique ( 2.16 ) Coinputer Control of Proces~)

~ -t -e -2t
ze e-t- e-2t] , ,[ 2e-NT _e-2NT e -NT -e -2NT]
~(t) = L- l [SI.-Al-l = ,[ -2e-t+ 2,
e-3t ,-t '2 -2t
-e+e .. (2.70) where ~(NT) = -2e-N'r + 2e- 3NT _e- NT + 2e- 2NT '
(2.74)

Substitution of B in equation '(2.67) and <I>(t) of equation (2.70) into

equation (2.32) (l - e-T ) e-, (N'-It-l)T, -O.5(1_e)-,2T e,-2T e-2(N-K-l)Tj
<f> [(N-K-l)T] SeT) =
. T [- (1 -e-T) e -(N-K-l)T (1 -e-2T) e--2(N-K-l) T .
e(T) =' f<p(T-t)Bdt .
, 0

[ -2(T-t}] [0] Xl (NT)] 2e -NT -e

. -2NT e-NT _e-2NT

-
-
T

f
-,
0 -
'2e:-(T-t) - e-2(T-t)
'2 -(T-t)
e
2 -3(T-t)
+ e
-(T-t)
:-: ~(T-<) ::e -2(T-<) 1 dt
.. [ X2(NT) r
L
_2e~NT + 2e-2NT 2e-2NT

, e-(T-t)_ e-2(T,- t) ]
. , ,dt (1- e-T)e-(N-K-.l)T -0-.5(1-. e-2T)e -2(N-K-l)Tj
=
R
o -e
-(T-t) 2 -2(T-t)
'+ e
... (2.71)
+~
N-l[

LJ
K=O -
(1 -e-T)'e -(N-K-l)T (1 -e-2T) .e -2(N-K-l)T U(KT)

T 2T
_ [0,.5 _e- ,+ 0.'5 e- ]
_... (2.72)
-
e -T -e-2T ,2g9 METHODS OF COMPUTING STATEIRANSITION MATRIX
Now substituting equation (2.70) with t = T and equation (2.72) into It is worthwhile to emphasize the difference between the formulation
equation (2.29), the discrete state equation of the system is written; of the state transition matrix under two different conditions. For the
i.e., x[(K+1)T] = <I>(T) x K(T) + SCT) U(KT) sampled data system.
x(t) = A x (t) + B U(t) "Q(2.75)
'where
U(t) = U(KT) KT 5: t :::; (K+l) T ~ .. (2.76)

2T
o .5 - e- T +0.5 e- ] U(kT)
The discretize state equations are express . as
+ .... (2.73)
[ e- T _e-2T x[(K+l) T] = <P(T) X (KT) + aCT) U(KT) ... (2.77)
where <peT) is a state transition matrix of A with t replaced by T, that is
Using equation (2.43), the solution of equation (2.73) is, .
Equation (2.43) implies

N-l
The state transition matrix <p(T) 'is given by,
, x(NT) = ~(NT)x(O)+L~[(N -K-1)T]8(T)U(KT)
k=O
... (2.78)
[The State-Variable Technique (2.18 )

Since q.(t) is the state transition matrix of A, $(NT) is determined 2.9.-2 . TheZ-tra:n'sform' method
from $(t) by substituting NT for t. For the sampled -data system of equation (2.77), the state transition
The state equations of digital control system are expressed as matrix of <j>{T) is expressed in terms of the Z - transform in equation (2.60) .
as,
x (K+1) = Ax(K) + B U(K) (2.79)
~(NT) =. Z-l {(ZI - <j>(T)-1 Z} '... (2.85)
The state transition matrix is given as
For the digital control system of equation (2.79), the state transition
matrix of A is expressed as,
... (2.80)
~(N) = Z-1 [eZI - A)-l z] ... (2.86)
In this case <peN) is obtained by multiplying A by itself N times.
In. the last two equation the evalution of STM involves the matrix
We shall first present. two methods of computing ~(NT) giveser'I') in inverse and then inverse Z-transform. For second-order systems, these
the sample data case, or <j>(NT) given A in. the digital. case. analytical steps can be carried out with ease.
However for higher order systems, the amount of work involved can
2.9.1 The Cayley-Hamilton T~~oremMethod be quite tedious if carried out by hands.
Given Matrix $(T) or A, as the case maybe equation (2.78) and (2.80) The task of finding {ZI-A)-l Z in equation (2.86),- given A, can be
can be computed using the Cayley-Hamilton theorem. simplified by the following method. Let
The theorem states that every square matrix must satisfy its own
F =' (ZI - Ar' Z ... (2.87)
characteristic equation. For example, consider that the characteristic
equation. of A (n x n) is written as, Pre multiplying both sides of the last equation by (ZI _. A) and
. rearranging we get
... (2.81)
ZF = ZI +AF "~ (2.88)
then,
Premultiplying both sides of last equation by ZI+A and rearranging
A D +a n-1.An-1 +a n -:-2 A n - Z + +a1 A + a.-o I ... (2.82) we g e t Z2F = A2 F + ZA.+ Z2 I ..,.(2.89)
By repeatedly multiplying by ZI + A,
... (2.83)
we' have
Similarly, if the stateequation is given in the formof equatidll (2.77),
A 3, F + Z A 2 + Z2 A + 'Z3 I ... (2.90)
the state transition matrixotn'I') must satisfy the characteristic equation
2
of $(T) -(nxn) that is, A 4F + ZA3 + Z2 A +'Z3 A + Z4I ... (2.91)
. r

AnF + ZAn- 1 -+- Z2 An-2 + +Zn- 1 A +z ... (2.92)

~(nT) = {an-l ~[(n -l)TJ + a n -2 ~[ (n - 2)TJ + + a1~(T) +aDI}
Let the characteristic equation of A be as an equation (2.81). We form
...(2.84)
the following equation by using coefficient of the characteristic equation
where and .equation (2.87) through (2.92).
ai ; i = 0, 1,.2...... n-1 denotes the coefficient of the characteristic aoF . = ao F
equation of$(T).
a1ZF =. a1 AF + al ZI
(TheSt'lte-Variable Technique ( 2.20 ) ComputerControlof Processes )

a2Z2F aaA2F + 8.2Z A + a 3Z2I ~le2.2 1

agZ F 3 = a3A~'F + a3 ZA2+a3Z2A+agZ31 ... (2.93) Consider that the. digital control system described. by the equation
(2.79) has the A matrix.

~ an.-:l zn-l F= a n- 1 An - 1 + a n- z Z An - 2 + a n=l zn-2. A+ an-l zn-l I 0 , 1 011

ZnF = ANF + Z An-l + Z2 An-2, + zn-l A + ZnI
A = 0 0
[
~6 -11 -6 J
The equation in equation (2.93) are summed on both sides to give,
The characteristic equation of A is
n n n
~a.AiF+ ~a.Ai-lZ+ ~a.Ai-2
I ZI-A I = Z3 + 6Z2 + lIZ + 6
L..J L..J L..J
i=O
1
i=l
1
i=2
1
rz-n (Z+2) (Z+3) = 0
+ I' a, A n
n
i - +1 zn-l +ZnI', The coefficient of the charateristic equal are as
and ao = 6.
= 1, a2 = 6, al = 11
i=n-'-l
Using equation (2.95), the matrix F is written as
where an = 1. Due to Cayley-Hamilton Theorem, the term on the right-
hand side of the last equation is a null matrix. 3 3
~ zj .~ a' Ai-j
'.i'hus equation (2.94) .leads to, F = (ZI - A)~l. Z =~
)=1
~
I=J

n 'n
IZI-A!
Izj Iai Ai-j
j=i i=j = a3 Z
31
+ (a2 I+a3 A)Z2 + (al 1+ a2 A+ a3 A 2)z
F = =
IZI-AI
... (2.95)
jZI-A!
Substituting the coefficient of a3, 82, a1 and A into the last equation, '
we have
The numerator of the last equation is also known as the matrix.
[adj (ZI - A)] Z
Once F is evaluated from equation (2.95), <P(N) is given by,
<p(N) = Z-l (F) ... (2.96) F ... (2.96)
(Z-1)(Z-2) (Z-3)
For the case.in equation (2.85), simply replace A by <p(T) in equation (2.95).
Performing partial -:- fraction expansion Equation (2.96) gives
 ( 2.22 ) Computer Centrolof Processes]
(TheState-Variable Technique . [ 2.21 )
The Laplace Transform Method

r2(ZZ~ 1)[ ~: :]
7 8
7 11 J z [27 16
---'-6
Z-2
Repeating equation (2.14)
f(t) = L-1 I(SI-A)-l] ... (2.99)
-17 1 . -12 -28

:]
F = ... (2.97) The matrix inverse of SI-A can be conducted using the same procedure

l +2(Z~3)-6
[38
-18

Now taking the inverse Z-transform on both sides of the last

9
27
~39
in equation 2.~6 through 2.94.
Thus
n
LsjLajAi-j
n

equations we get
j=l i=j
(SI-A)-l = ... (2.100)

~ ]-[~ }~O.693K
7 8
! [ -6
18
7 16 : where n is, the dimension of A, and
2 -6 -17 1 -12 -28 lSI - A I = sn+ a n- 1 sn-l + + ajS + ao ... (2.101)
ep(K) = 9 -1
... (2.98)
Direct Power Series Expansion Method
+~
[ 38
-6 27 3 ]e-
1 1K
.
. The power-series representation of 4>(T) is
-18 -39 9
A2T 2
We can check this by verifying that q>(O) = I, the identity matrix. $(t) = eAT. = I + AT + 2! + ... (2.102)

This expression, can be programmed recursively. For' example, the

'-2.9.3 Computing State Transition Matrix. ,(T) AK TK . . A K +1 T K +1
The discussion conducted thus far have been concentrated on the k th
term of the series is K! and (K + 'l)st term is (K + I)!
computation of state transition Matrix <p(NT) or <p(N) of the sample data
or digital control system. When the sample - data system is' encountred, These we can w~ite (K+l)stterm = AT x Kthterm ... (2.103)
the starting point in equation (2.75) in which the. matrix A and B are given. . K+l
K = 0, 1, .2 Normally, when computing the series, a cheek is
In order to find <p(NT) using -equation (2.78) we must find'
. made on the convergence.rand the recursive iteration can be stopped after
. <p(T) which is given by' equation N terms.
4>(T) = eA"r. =$(t) It=T Truncated Power .' Series Expansion Method
Earlier in the chapter the state transition matrix of A,. <p(t) is given (Cayley Hamilton Theorem)
by. equation (2.14). and (2.16). Equation (2.14) is the Laplace Transform
The power series of (T) 'can be truncated. after N jerms, where N is -
solution, whereas equation (2.16) is the power series representation. le
some positive integer, Then,
shall investigate va~ious.methods of computing <p(t), given the matrixA.
 ( 2.24 ) Computer Control of Processes )
(The.State-Variable Technique , [2.23 )

Applying the Cayley - Hamilton theorem to the last equation, we have, .' IExample 2.4
To illustrate the eigenvalue method described above, consider the
... (2.105)
. A . A= [3 2] .
matrix z gIven In as
2 .3
Where n is the dimension of A and cq are constants that depend on
the coefficients of the characteristic equation of A and can be computed The eigen values of A are 81 = 5 and 8 2 = 1 which are distinct from
recursively-in general N andn are not related, so N 'can be greater than n. equation (2.108).

The Eigenvalue, Method (Sylverter's Expansion Theorem : A Has ... (2.110)

.Distinct Eigenvalue only). ... (2.111) .
, For system with distinct eigenvalues sylverter's .expansion theorem' Equation (2.106) gives
states that if,
0.5 o.5l
... (2.112)
-r(A) =L
00

Ck A k ... {2~104)
[ 0.5 0.5 J
k=l

n F(S2) = A-S11 = [0.5 -0.5] ... (2.113)

82 - 81 -0.5 0.5
then, f(A) Lf(Si)F(Si) ... (2.105)
i=l Thus 4>CT) = fCS!) F(8 1 ) + f (82) F(S2)
where,' Si, i 1; 2, .....n, are the eighen values (all distinct) of A

! A-SjI
... (2.114)

.F(St") -- J" =1 Si -Sj , 1, 2 , ... n ....

1 (2 106)
i:t:j 2.10 .STATE DIAGRAM OF DIGITAL SYTEMS
For the state transition matrix ,problem we have When a digital system is described by difference equations or discrete
state equations, a state diagram may be drawn to represent the
f(A) = <p(T) 'V'.oo.
= ..J
A K. T,K = ~AT ... (2.107) relationships between the discrete state variables. In contrast to the
k=9 K! similarity between an analog computer diagram and a continuous - data
Thus f(Si) = e Si T ... (2.108) state diagram, the digital state diagram, the digital state diagram portrays
the operations on adigitalcomputer.
and from equation (2.105),
Some of the basic linear operations of a digital computer are
n
~ultiplication by a constant, additicn of variables and or storage -the
'<p(T) = L~SiTF(Si) ... (2.109)
i=l
mathematical descriptions of there basic digital operation and their,
corresponding 2-transform expressions are given below.
where, F(Si) is given by Equation (2~106)
( TheState-VariableTechnique ( 2.25]. Computer Contr-ol ofProcesses )

1. Multiplication by a constant I Example 2.4. I

x~ (KT) = .aXI (KT) . ... (2.115) Consider that a digital system is described by the difference equation,
X2 (z) ax! (Z) C (K+2) + 2 C(K) + 3 C(K) =r(K) ... 2~121
2. Summing The state diagram of the system Is .drawn as show in Figure 2.4 by
... (2.116) first equating the highest - order term to the rest of the term in
X3 (KT) = Xi (KT) +x2(KT)
equation (2.121). .
X2 (z) = Xl (Z) + X2 (Z) ... (2.117)
The state variable of the systems are designed as the outputs of the
3. Time delay or storage time-delay units of the state diagrams, the state equation of the system-
X2 (KT) = Xl K+l)T) .u (2.118) are written as,

(K). t rO] U(K)

X2(z) Z XI(Z) - Z Xl(O) ... (2.119)
or [~30 -21] [Xl (K)J
X2 Ll
... (2.122)

x, (z) Z-l X 2 (z) + x, (0) .. (2.120)

a
x1(kT) 4 ~ x2(kT) x1(kT) 0
(a)
x1 (kT)
x1 (kT) x 2(K)
x2(K+1) x1(K+1)Z~1 x1(K) C(K)

y(K) X2(Z) X1(Z) C(K)

x2(kT)
x1(O)
x 2(z)
(b)

-3
Fig. 2.4
. z-
'x2(kT) 4 Storage ~ x1 (kT} - x2(kT) 0----.----0 x1(z) Figure 2.4 A state diagram of the digital' system described by
(equation 2.122).
(e)
The output equation is,
Figure 2.3 Basic Elements of a digital state diagram and. the
corresponding digital computer operations.
... (2.123) .

The state' transition equation of the system, which are the solutions
of the state equation in Equation (2.212) are obtained from the state
(The State-Variable Technique Computer Controlof Processes )
( 2.27 )

diagram using the gain form when ,Xl(Z) and' X2(Z) are regarded as output 2.11.1' Direct decomposition
nodes and R(z), Xl(O), X2(O) are input nodes, we have, Consider that the transfer function of a digital controller or process
is,

... (2.124) C{z) b m + b m - 1 z' + + b 1 Z-m+l + b o Z-m

Dtz) = R(z) = Z-l' , Z-2 . Z-n+l Z ... (2.127)
an + an-l + an - 2 + ..... + al + ao
~ = 1+2 z' + 3Z-2
~ere nand m are positive integers.Fo~ physical realizability, an "* O. -'
Equation (2.124) is the state transition equation in the Z-transform domain.
The general form of equation isgivenhy Equation (1.125). To find the state diagram realization of Equation (2.127) by direct
decomposition we first multiply the numerator and denominator of the
x (Z) = [ZI - A]-l 'z x (0) + [ZI - A]-l B U (z) ... (2.125)' right - hand side of equation (2.127) by a variable xfz), we have,
One advantage of using state diagram is that equation -(2.125) is
obtained' simply by 'using the gain formula. This saves the effort of C(z) _ .(bm+ bm_1.z-1 + bm_z,Z~2 + + b1Z- m + 1 +b o z-m )x(z)
performing .the matrixinverse of (ZI.. . .A}wh~n~qu~tion '(2~ 125) is used. ~ (z )
- (an +,a - l z' + a - 2 Z-2 + ..... + al Z-n+l + ao z-n) x ()z
... (2.128)
n n

The state transition equation in thetime domain can be obtained by

Next we equate the numerators on both sides of the last equation,
performing the in~erse Z-transform on both sides of equation (2.12~).
and then do the, some with the denomintors.
The transfer function between the output and input is determined
The resulting equations are,
from the state diagram by applying the gain formula betweenRtz) and
C(z) with zero initial states. We have, C(Z) = (b~ + bm- 1 Z-l + b m- 2 Z-2 + b1Z-m +1 + b o z-m) x(Z) (2.129)
R(Z) = (an + an-l Z-l +an~2 Z-2 + al z-n+l + ao z-n) x(Z) (2.130)
.. C(z) Xl(Z) Z-2 1
... (2.~26)
_R{z) R(z) 1+2Z,...1 +3Z- 1 = z2 +2z+3 Toconstruct a state diagram, equation (2.130) must first bewritten
in a cause - and - effect relation. 'I'h.erefor e , dividing both side of
equation (2.130) byan, and then equating X(z) in terms of the other terms,
2i11' OfCOMPO,S ITI;DNOf DISCR'ETE - DATATRANS,FER FUNCTION we get,
Before embarking on the state diagram ofdiscrete - data system with
sample and hold devices, it is appropr-iate first to' discuss the state ~ R(z) .:- an-l Z-1 x(z)~n-2 Z-2,x(z)",
representation ofZvtransfer functions. an B: n an
X(z) = . ... (2.131)
In general, the transfer function of a digital controller or system ~ Z-n-l x(z) - ao Z-n x(z)
a
Dtz) may be realized by digital processor, digital circuit, microprocessor an an.
ora' general-purpose digital computer. The steps involved in realizing
The state diagram portraying equation (2.129) and (2.131) is drawn
the state diagram and dynamic equations of a' Z-transfer function are
as shown in Figure (2.5) for n = m = 3. For simplicity, the initial states are
termed decompositions.,
excluded from the diagram.
There are three basic types of decomposition: direct decomposition, . A digital computer program can be derived from the state diagram
cascade decomposition and parallel decomposition. These three methods
of Figure (2.5); the branches with the gain Z-l are realized by a time delay
of decomposition are described in the following,
of Ts which is the sampling period,
[he State-Variable Technique,' ( 2.29 J. ( 2.30 ) Computer Control ofProcesses]

The state diagram of Figure (2.5) can also be used for analytical ?11.2 Cascade decomposition
purposes., If the numerator and the denominator of a transfer function D(Z)
By defining the variables at the nodes of all the time-delay units are expressed as products of first- order factors, each of these factors
state variables, the dyna~ic equations and the state transition equations can be realized by simple digital program on state diagram. The entire
can all be determined from the state diagram by applying gain formula. D(Z) is then represented by the cascade connection of the individual
programs.
Consider that the transfer functionof a digital process is written as.

*0_ (2.133)

C(z)
R(z)n---~rq.---""""--4;C}--""""'---I----""~--------I"'-~
Where K, Zi, i = 1, 2, ..... m and Pj, j = 1, 2, .... n are real constant.'
.D(Z) to be physically realizable n ~' m
Writting n(Z) as a product of the. constant Kand.n first order transfer"
funtion yields.
DeZ) = KD l (Z) D2 (Z).~ .... Dn(Z) ... (2.134)
where,
. Figure 2.5state' diagram representation of the transfer function of
Z+Zj 1+ Zi Z~~
equation (2.127) with n = m = 3'by direct decomposition. Di(Z) = L = J,2,.....m ... (2.135)
(Z+Pi ) l+Pi Z-l
The discrete state equations are obtained from Figure 2.5'by applying
the gain formula with Xl (Ks-L), X2 (Ks-L) Xg (Ks-L) as outputs. After deleting 1 Z-l
the time - delay branches with gain Z-l, the inputs are r'(K), xi(K), x2(K) DiZ) = - - j =m+l n (2.136)
Z+P j l+P Z-l
j
and Xg (K) ~
The state diagram representation of equation (2.135) is ,constructed
The state equations' are,
in Figure 2.6 (a) and that of equation (2.136) .is in Figure 2.6 (b), all by
Xl (Ks-L) = X2 (K) direct decomposition.
'X2 (K+l) = Xg (K) The overall state diagram for D(Z) is obtained by connecting the m
diagrams of Figure 2.6 (a) and n - m diagram of Figure 2.6 (b) with gain
X:l (K+1) = -a2 Xl (K) - al X2 (K)- (K)+~r(K)
ao Xs ... (2.182) branches of K, all in series.
ag' ag ag' ag .

From ,these state. equations, we can see clearly that direct The advantages with the cascade decomposition is that gain K and
decomposition will always yield a state variable model in the phase. - the constant Zi and P j are all preserved in the state diagram. In design
variable canonical form. This is apparently one advantage of using direct situation these parameters may be varied individually, and the effects
can be. observed independently.
decomposition.
(TheState-Variable Techniq ue [ 2.31] (2.32 ) computer Control of Processes] _

The parallel decomposition of a transfer function will lead to ~ set

of state equations that is in canonical form for distinct eigenvalue or Jordan
Canonical form for systems with multiple - order eigenvalues.

2.12 STATE DIAGRAM O.FDISCRETE DATA SYSTEM SYSTEM WITH

ZERO - ORDER HOLDS , ,
-Pi, A discrete -data control system' uaually contains digital as well as .
(a) analog elements and the two types of elements are coupled by sample
and hold devices. A typical discrete - data control system is illustrated by
the block diagram in figure 2.6.
The system consists of a digital controller, a zero -order hole
(or D/A converter) and a continuous - data rprocess which is to be
-Pj
controlled. We show in this section how state diagram and the state -
variable analysis of this type of system are conducted.
(a)

Fig. 2~6: State ~iagramrepresentation of Equations (2.135) & (2.136)

2.12.1 State diagram of the Zero - order Hold
Before the state diagram for the discrete - data system inFigure 2.6
2.11.3- Parallel Decomposition,
can be drawn, we must establish the state diagram representation of the
The transfer function Dtz) with real pole can be realized by parallel zero - order hold.
decomposition, in which case only the denominator Dtz) must be faelored.
Let the transfer function of a digital-system be written as Let the input and the output ofthe ZOH be denoted bye*(t) and h(t)
respectively.Then, for the time interval
I
. D(Z) = C(z) = K Zm + b m-I 2m- .... + bIz + b o KT ~ t < (Kt1) T
, R(z) Zn Zn':-.-l .
... (2.137)
~ , + an-l . + a1z + a o
h(t) =e CKT) ... (2.139)
where nc-m, Assume for illustration purpose there is no pole ~ero
cancellation in Dtz) and, that, among the real eigenvalue i are distinct
and rest are of multiplicity n-i, then D(z) is expanded by partial fraction
expansion_ as . e(KT)
C(z) i=3 K n K
R(S) = -S- ... (2.140)
D(Z) = - - = 2: -. _k_ + L .' k k-' (2 138)
R(Z)K=l 2 + Pk K=i+l(2 +P )
k l . . For KT :5; t < (Ks-I) T. Therefore, the state diagram of the Zero - order
If Dfz) has at least one zero at z = 0, then the partial - fraction e*(t)
expansion of rule of Z - transfer functions should be fo Ilowed
_ _.....-1 Digital~
J, -
.. cantrolle!!
equation (2'.138) implies that the, individual state diagram. T

.CoD,1llon~nt.ofD(Z)~re all the form of figure 2.6 (b) and. the over all
state diagram is 'obtained by connecting thesein parallel. Fig. 2.6': A sampled - data control system
(The State-Variable Technique [ 2.33,J ( 2.34 ]

5-1 Therefore, with reference to Figure 2.8, observability usually deals

o 0 with the relationship between the state x(K) and the outputs C(K),wher~as
,e(Kt) H(s) controllability deals with the relations between the control U(K) and the
Fig. 2.7: State diagram of the zero - order hold for KT t< (K+l) T states ro.
Hold consist of a single branch with gain of 8-1, connected between, Among the many forms of performance specification used in the the
the nodes e(KT) and.H(S) as shown in Figure 2.7.' It should be pointed out design of control systems, the most important requirement is that the
that while the gain 8-1 of the ZOH is analytically anintegrator, it is not system be stable at all times. Relative to the conditions of controllability
counted into the order of the system. ' .and observability stability is a end condition that the system must have,
since an unstable system is virtually useless.
2.13CONTROLLABILITY OBSERUABILITY ,AND STABILITY OF LINEAR TIME Even linear systems have man:)' different types of stability, hut to
INVARIANT DISCRETE DATA SYSIEM introduce the concept, we can state that stability is the condition where
in the system output must be bounded for 'any bounded input.
2.13.1 Introduction Before .embarking, on the formal discussions of controllability,
The concepts of controllability, observability and stability play observability and stability, let us first provide simple illustrations of the
important role in the analysis and design of 'control system. The ideas. Figure 2.9 shows, state diagram of a.Iinear discrete - data system
significance of the control' ability of a system can be stated with reference with two state variables. Since the input U(K) affects only the state
to the block diagram of Figure 2.8. "The controlled processG is said to be variable X1(K), we say that X2(K} is uncontrollable and the system is not
controllable, of every state of G can be affected or controlled in finite time completely or simple uncontrollable. .
by some uriconstarined .control .signal U(K)".
~
~ x(K+1)=Ax(K)+BU(K) ~
I~~D'I.C~(K}
Intuitively we understand that if anyone of the state variables of
the system is not accessible fromthe control U(K), then there is no way of
driving this particular state to a desired state in finite time by means of Fig. 2.8: Block diagram of a linear, discrete ... data system
some control effort. This particular state variable is said to be
U(K) C(K)
uncorrtrollable and the system is said to be un~'o.ntrollable. Thus,
controllability is an important prerequisite before attempting to design
the control of any process.
The concept of ,observability is the complement of controllability.
Essentially, the process G is said to be observable if every state variable
of the process eventually effects some of the outputs of the process.
It is often desirable to obtain' information on the state variablesfrom Fig. 2.. 9: State diagram of a precess that is not completly controllable
measurements of the~ outputs and inputs. For example, we may want to
Figure' 2.10 shows a linear discrete -.data' system in which the state
estimate those state variables that are not accessible for feedback control
variable x2(K) is not connected to the output C(l() in any way. Therefore,
from measurements of the outputs and inputs. If anyone of the states
we can observe the state variable x1(K) from 'measurements of the output
cannot be observed from, the measurements of the outputs, then the state
C(K), since C(K) = Xl (K).
is said to be unobserable, and the system is unobservable.
 Computer Control of Processes ]
(the State~Var,~ableTechnique, (2.35 )
Where x CK) is an n-vectorvUt.Klis an r-vector, C(K) is a P-yector; A,
However, the state variable x2(K) cannot be observed from the
B" D and E are. coefficient matrices with appropriate dimensions.
information on C(K). Thus the system is not completly observable, or
, simply unobservable. In general, the conditions of controlability and Definition 2.1: Comple State Controllability
observability usually cannot. ,be decteated by mere inspection of the 'The system described by e.quation (2.141) is said to be completely
system's statediagam or block diagram. state controllable if for any initial time (stage) k = 0, there exists a set of
We .shall show that stability of linear discrete-data systems depends unconstrained controls UCk), K = 0, 1, 2 N - 1 which transfers each
purely on the roots of the characteristic equation. That 'is for stability, initial state xeD) to any final state x(N) for some finite N.
the roots of the characteristic equation must allIie inside the unit circle Definition 2.2: Comple Output Controllability
I z] = 1 in the Z-pl~ne.
The system by equation 2.139 and 2.140 is said to be completely outp:ut
From earlier discussion on signal flow graphs ~nd transfer functions, controllable if for any initial time (stage) K =0, there exists a set of .
it should be clear that characteristic equation of a linear system depends unconstrained controls U(K), K =0, 1, 2 .... N - 1, such that any final 'output
only on the loops of the system. C(N) can bereached arbitary initial states in finite time (stage) N.
Note also that controlollability, observabi'li ty-artd stability are
mutually exclusive conditions..Thatis, any sytem can be controllable and 2.14. 2 Theorems on controllability
observable 'but 'unstable, it can be stable and corrtr-ol l able but un The follwingtheorein are given on state and output controlability of
observable, etc. linear time-invariant system. Some of the theorem are stated without
proofs.
Theorem 2.1: State Controllabity
The system describe by equation 2.141 is completly state c-ontrollable
if and only if the n x Nr matrix. '
S = [B A B A 2B ~ .. ~ An-l B] 0 (2.143)
is of rank n. The matrix S is called state controllability matrix, This test
is also called Kalman's test.
Fig. 2.10: State diagram' of a process that is not completely observable
If the system is completly statecontroallable, then we also say that
the pair [A, B] is controllable.
~.14 CD.NTRDLABllITY OF LIN~AR TIME-INUARIENI DISCRETE-DATA
Proof
SYSTEM.
The state transition equation of equation 2.141 is [from equation (2.51)]
2.14.1 Definition' of controlability
Consider that a linear discrete-data or digital control system is x(N) ANX(O) + L AN-i-1B U(i)
N-l

... (2.144)
described by the following dynamic equations: 'i=O

x (K+l) = A x (K) + B lr(K) ... (2.141)

C (K) = D x:(K) + E U(K) ... (2.142)
 ( 2.38, ) computer Control of Processes)
(The State-Variable Technique [ 2.37 )
The matrix T is refered to as output controlability matrix.
The left side of equation (2.144) can be represented by an n x 1 vector
Proof:
x(N); then the equation is written as .
Substituting the state transition equation of equation. (2.144) into
x(N) .= S U ... (2.145)
equation (2.142) with K:= N and rearranging, wehave,
Where S is given by equation (2.143) and U is the n x 1 vector.
N-l

U(O) C(N) - D AN - 1 o: = I DA N - i - 1 B U(i) + EU(N) ... (2.150)

i=O
U(l)
u= ... (2.146) Since the last equation is similar to equation (2.145), the proof of
this theorem parallel to that of theorem 2.1. The following examples, are
U(N~l)
provided to illustrate the concepts and application of controlability,

For complete state' controlability every initial state X(O) must be

transfer by unconstrained controls Uti), i = 0,1, 2 n to any final state
IExample 20
x(N) for finite N. Thus, the problem is that given S and every vector x(N) Consider a linear discrete-data control system whose input-output
in n dimensional state space, solve for the controls from equation (2.145). relation is described, by the differenc.e equation.
Since equation (2.145) represents n simultaneous linear equations, from C(K+2) + 2 C (K+l) + C(!\) = u CK+l) + UCK) ... (2.151)
the theory of linear equations, these equation's must be linearly
The difference equation is decomposed by direct decomposition into
independent for solutions to exists,,' and the necessary and sufficient
the following dynamic equations.
condition for this to ,be true is that the matrix S has a, rank of n. In
otherwords, S must have at least. n independent columns. x (K~l) A X (K) + B U(K) ... (2.-152)
If the system in eqution (~.141) has only one input, r = 1 then S is an C (K) = D X (K) ... (2.153) ,
n x n square .matrix; the condition of complete state controllability is that
S must be non singular. where A= [~1 ~2] B [~]
If the state equations of-the linear systems are represented as
D = [1 1]
X [(K+1) T] = $(T)x(kT)+S(T)U(KT) .... (2.147)
The state controlability matrix is
then the state controllability matrix is given by
S = [B AB] = [~ ~2] ... (2.154)
S = [S(T) ~(T)S(T)~2(T) 8(T).....~n-l(T) S(T)] ... (2.148)
Which is nonsingular. Thus the system is completly state controllable
Theorem 2.2:"Output Controlability for the state variable xl(K) and x2(K).
The system described by equation (2.141) and (2.142) is completely The output controlability is tested by forming the matrix
output controlable if and only if the following, Px (N + 1) r matrix is of
rank P, where p is the dimension of the output vector C(K). e
T = [DAB DB] = [-1 1] ... (2.155)
T= [D A N - 1 B . DA N - 2 B DAB DB E] ... (2.149) Which has rank of one. Since the dimension of C(K) is one, the system
is completly output controllable'.
(The State-Variable Technique [ 2.39 )
( 2.40) Computer Control of Processes J
Whenj assumes values from 1 through N.-l, equation (2.158) together
2.15'OBSERVABllITYOF, ll:NEAR TIME-INVARIANT DISCRETE - 'DATA with equation (2.142) represents pN linear algebraic equations that can
SYSTEMS be put in a matrix form as follows:

2.15. 1 Definition of observability C(K) D'

Definition 2.3 C(K+) DA
C(K + 2) DA 2 xCK)
The digital system' described in equation (2.141) and (2.142) is said
to be completly observable if for any initial time (stage) K = 0, any state
=
x (0) can be ,determined from knowledge of the 'output C(K) and input C(K+N -1) DA N - 1
U(K) forO:::; k :::; N, for N is some finite time (stage).

E 0 0 0 0 U(K)
2.15.2 Theorem on Observability
DB E 0 0 0 U(K.+,1)
The linear digital system described by equation (2.141) and (2.142)
+ DAB DB E 0 0 U(K+2)
is completly observable if and only if the following n x PN .matrix is .of ... (2.159)
rank n.
D A N - 2B D A N- 3 R' D A N - 4B DB E U(K +N -1)
... (2.156)
Where n is 'the dimension of x(K) and p is the dimension of 9(K). To determine x(K) from the last equation, given C(K + j) and U(ks-j)
The matrxL is known as observability matrix. for j = 0, 1, 2... :...N- 1, the pN x n
matrix.
Proof:
Following e9-uation '(5.64), the state transition equation of equation '
(2.141) is written as,
... (2.160)
j-l
Aj x(K) +'LAj-i-l BU(K + i]: ... (2.157)
i=O

for J == 1, 2 N- 1 must have n independent rows. (or assuming that pN ~n). The condition
Substituting Equation (2. 157) into output equation, equation (2.142), in equation (2.160) is equivalent to the matrix L in equation (2.15) must
we get have a rank n.

. J-l For a discrete - data system that is described by the state equation
C(K+j) = DAjx(K)+ LDAj-i-1BU(K+i)+EU(K+j) ... (2.158) in equation (2.147) and the out equations of equation (2.142) the condition
i::::O of complete .observabifity is that the matrix.

for J = 1, 2 N- 1 L = [D T ~(T)T nT ~(2T)T n" .... ~[(N - 1) T]T D T .~. (2.161)

must be of rank n.
 ( 2.42 ] Cort!pu:ter,Control'.ofP,rocesses )
( The State-Variable Technique ( 2.41 )

I Example 2.5 I 2.16.1 Definitions onStabiUty

Definition 2.4: Zero - State Response
Consider alinear discrete - data control system whose input-output The output response of a 'discrete - data system that is due to the
relation is described by the difference equation input only is called the Zero '- state resopnse; all the initial conditions of
C (K+2) + 2C (K+1) + C(K) = U, (K+l) + U(K) ... (2.162) the system are set to zero.
,This difference equation is decomposed by direct decomposition into Definition 2.5: Zero - input Response
the following dynamic equations. The output response of a discrete - data system is due to the initial
x (K+1) A x (K) -+ B U(K) ... (2.163) condition only is called the zero - input response; all the inputs of the
system are set zero.
C (K) = D x (K) ... (2.164)
From-the principle of superposition, when a system is subject to both
A = [~l ~2] B [~] ... (2.165) inputs' and initial conditions, the total output response is
, Total response = Zero - state response +
D = [1 1] Zero' - input response ... (2.i68)

In this chapter we shallgive definition of' stability and, investgate Definition 2.6: Bounded - Input - Bounded - StateSt'ability
the methods of testing the 'stability of linear discrete. - data and digital Consider a linear time - Invariant discrete - data system, that is
control systems. described by the following dynamic equations.
The state controlability matrix is x (K+1) =A x,(K) + B U(K) - ... (2.169)
... (2.166) C (K) = D x (K) + E U(K) ... (2.170)
The system is said to be bounded Input - bounded - state (BIBS) stable
if for any bounded input; the, state x (K) is also bounded.

The norm of an-vector x(K) is denoted II x (K)II a~d is defined" as

... (2,,167)
i

[~x~ (K)J
Since L is.singlular, the system is unobservable.
Ilx (K)II = - ... (2.171)

'2.1 & STABILITY O'F LINEARDIGITALCONT.ROL SYSTEMS DEFINITIONS AND,

THEOREMS The bounded state implies that the state vector x(K) sastisfies,
One. of the most important requirements in the performance of ~I x,(K)11 s M for all 'K ... (2.172)
control system is stabilrty. This is true for continuous ~ data systems as
well as discrete - data and digital control systems, where, M is a finite number
In this chapter we shall give definition of stability and 'investigate
the method of testing the stability of linear discrete - data and digital
control system.
 ( 2.43) ComputerControl of Processes )

Definition2.7,:Bounded Input Bounded - .output stability

iii 2.17.1 The Bilinear Transformation Extension of the Routh - Hurwitz
The system described by the equations (2.169) arid (2.170) is bound - Criterion
input bounded - output stable if for, any bounded input, the output C (K) The Z - transformation Z = exp "(Ts) transforms the imaginary axis of
is also' bounded. S - plane onto the unit circle I z I = 1 in the Z plan and thus renders
. the Routh - Hurtwitz criterion ineffective for discrete data systems. We
Definition 2.8: Zero - Input stability
can attempt to transform the unit circle in the Z plane back to the
. The system described by the dynamic equations (2.169) and (2.170) imaginary axis of another complex plane, but we cannot use the inverse
is said to be zero -'input stable or simply stable if the zero - input response Z transform, since it would transform an algebraic character equation in
C (K) subject" to the finite initial conditions, reaches zero as K approaches Z into a transcendental equation' of S, which still cannot be tested with
infinity, otherwise, thesystem is unstable, Mathematically, zero - input Routh - Hurwitz criterion. Transformation that are algebraic and
stability require that, first transforms circles in the Z plane onto vertical. lines in a complex variables
plane say the r-plane, are of following form:
II C (K)II s M < 00 ... (2.171)
ar + b
Z = Cr +d ... (2.173)
. and second lim
K~oo .Iei (K)I = 0 ... (2.172)
Where a; b; c and d are real constants. One such transformation that
Definition: Asymptotic stability
transforms the interior of the unit circle I Z I = 1 onto the left half of r-
The condition given in equation (2.171) and (2.172) are also the plane is,
requirements for the asymptotic stability. Thus Zero- input stability
r+l
implies asympt?tic stability. Z = r-1
... (2.174)

Which is referred to the r - transformation..Another posibility is to

2.17S1AII:LITY JESISOF:':DISCR:ETE DATA'SYSTEMS
let a = b == d = 1 and c =-1 so that
A linear time - invariant" discrete - data system is asymptotically
r+l
stable if roots of the characteristic equation lie inside the unit circle in Z =- -
-r+l
... (2.175)
Z - plance.
As it turn out; amount of algebraic work involved in applying
The Routh .. Hurwitz criterion which is useful for the stability test equation (2.175) in stability,studies when constants a, b, cand d vary is
of. linear continuous -data with respect to thecharacteristic equation" relatively same. The computer program RWT and DCSP software package
roots in Z- plane, cannot be directly to the Z ~ plane, due to the- difference used to carry out r-transformation.
in the stability boundary. Figure (2.11) illustrate the rnapping of unit circle in Z - plane on to
In this- section we shall introduce a bilinear transformation that the imaginary axis of r - plane.. Notice that interior of unit circle
transform the unit circle in-the Z - plane. on to the imaginary axis in corresponds to the left half of r - plane. Thus once the characteristic
another complex plane so the Routh - Hurwitz criterion can applied to equation in Z domain is transformed into the r - domain using equation
discrete .. data system.. (2.174), the Routh - Hurwitzcriterion can again be applied to equation in
r in the normal fashion.
(The State-Variable Technique ( '2.45 ) computer ControlofProcesses )

We observe that all the coefficients are of the same sign, and none of
coefficients is zero, which are the necessary and sufficient conditions for
all the roots of the second - order equation in equation (2.179) to be in the
left .half r - plane. Thus, all the roots of equation (2.178) are insidethe
unit circle.
I Example,2.7 I
Consider that the characteristic equation of a third - order discrete
- data control system is given as
@ F(Z) = Z3- 1.25 Z2- 1.375 Z - 0.25 = 0 ... (2.180)
Z-Plane y .. Plane, Which is known to. have roots at Z = 0.25, 0.5 and -2.0

Figure 2.11 Mapping of the unit circle I Z "=1 onto the imaginary axis . Transforming F(Z) into the r - domain using Equa~ion.(2~180), we get
of the r - plane through, the r - transformation Z = (r + 1) '/, r-1 '-1.875 r + 3.87 r
3 2
+ 4.875 r +1.125 = 0 ... (2.181)

I Example 2.6 I Since the coefficient of the last equation do not all have same sign,
we know that the system has at least one zero outside the unit circle in
Consider that the matrix A of the state equations in equation (2.169) is the- Z - plane. Nevertheless applying the Routh - Hurwitz Criterion to
equation (2.181), we have the following Routh tabulation.
'[-0.5 0]
A = ,0 0.5 ... (2.176) r3 - 1.875 4.875
Sign change
- The characteristic equation of the system is
r2 3.875 .' 1.125
IZI - A I = (Z - 0.5) (Z + 0.5) =0 ... (2.177) r 1
5.419 o
F(Z)= Z2-0.25=O ... (2.178) 1.125
Transforming F (Z) into' a polynomial in r using Since there is one sign change in first column of the Routh tabulation,
equation" (2.181) has one root' in the right-half of r - plane or 'equation
r+1 (2.180) has one root outside the unit circle in the Z - plane, which in this
Z = --'
r-1
we get
case is known to be at Z' = -2.
, Singular Cases
(r+lr -0.25
r-1
= 0 For the first singular case in ther -plane, the same remedy is applied
by replacing the zero element in the first row by ail arbitarily small
(r + 1)2'_ 0.25 (r - 1)2 = 0 number E.
0.75z-2-2.5r + 0.75 = 0 ... (2.179) When a row of zeros occur in the Routh tabulation of the equation'
.looking from the transformation relationship between Sand ,Zand r we
conclude that
.
the, following conditions may exist.
'
(TheState-Variable Technique ( 2.48 ) ComputerControl of Process!!]
1. Pairs of real roots in the Z - plane that are the inverse of each Taking the derivative of ~(r) with respect to r we get,
other; eg Zl = 2 and Z2 = 1/Z l = 0.5.
dA(r)
2. Pairs of roots on the unit circle 1Z,I= 1, including Z = 1 and --'-
dr
= I.Sr ... (2.185)
Z = -1 simultaneously.
The second condition corresponds to roots on the imaginary in the r plane. dA(r)
.
. . . The coefficient of the r 1 row are now filled with the coefficient .~
The' remedy to the singular case is simila~ tothat for the S -.domain.
and the remaining Routh tabulation is,
1. Form the auxilary equation A (r) = 0 using the coefficient of the r1 1.8 0
row in the Routh tabulation just above the row of Zeros.
rO - 0.1
.. ,2. Performd A (r)/dr = 0
Since one sign change occurs in the first column of Routh tabulation
3. Replace the row of zeros in the tabulation with the coefficien~s of the characteristic equation has one root in the right half of the r - plane,
d A(r) I dr. or Ffz) has one root outside the unit circle in the Z - plane. Solving forthe
4. Continue with the Routh tabulation in theusual manner. roots of the auxiliary equation A(r) = 0, we have,
r = 0.333 and - 0.333 ... (2.186)
IExample 2.8 ~.
Which are two of the roots of the characteristic equation. These two
Consider the equation roots correspond to, the root at Z =- 0.. 2 and - 0.5 respectively.
F(Z) = Z3 + 3.3 Z2 + 3Z + 0.8=.0 . ... (2.182)

Which has roots at Z =- 0.5, - 0.8 and - 2.0. The roots at ., 0.5 and
IExample 2.9 I
- 2.0 form an inverse symmetry about Z = 1, and should create a singular Consider adiscrete -data system that has the following equations.
case for the Routh test. F(Z) = Z3 + Z2 + Z + 1 = 0 .~- (2.187)

. . . r+1.. The roots of F(Z) = 0 are at Z =1, Z =j

and Z :.; ~ i, which are all on
Tran~for~ing F(Z-) into the r-plane using Z = r -1 '. we
get the unit circle. Thus the system is unstable. Applying r ~ transformation
of equation (2.188) we get
8.1 r 3 + 0.9 r 2 - 0.9 r - 0.1 =0 ... (2.183)
r3 + r = 0 ... (2.. 188)
The Routh tabulation of the last equation is
Since the left-hand side of the last equation is an odd polynomial,
r3 8.1. -0.9 the r 2 row of its Routh t~bulation would contain all Zeros. In this case,
r2 0.9 -0.1 equation (2.189) is also its own auxiliary equation.
r 1
0 0 ~ A (r) = 0, the complete tabulation is now
/I'hus the r1
row contains all Zeros, and the' Routh test could not be r3 1 I
continued in the usual manner. The auxilliary equation is formed by using
r2 3 1
the coefficients in, the r 2 . row: '
r1 2/3 0
A(r) =O.9r 2 - O.1 ... (2.184)
rO 1
 ( 2.50 )
(TheState-Variable Technique [ 2.49 J
Since the coefficients in the first column of the tabulation are all of Note that the elements of the (2k + 2)th row (k = 0,1,2, ... )con-sist of
the same sign, equation(2.188) does not have any roots in the right half of the coefficients of the (2k + l)th row written in the reverse order, The
the r - plane. or equation (2.187) does not have any roots outside the unit elements in the table are defined as
circle in the Z - plane. In this case, we know that all three zeros" of F(Z)
are on the unit circle. bk = [ ao
an
an-k
ak
1
2.17.2 Jury's Stability Test Ck= [ b
o bn - 1- k ]
bn - 1 bk "
One of the first direct methods devised for testing the location of
the roots of a polynomial in Z with respect .to the unit circle in the
Z - plane is the Schur - Cohn Criterion. The criterion given the necessary dk= [ Co
Cn - 2
Cn - 2- k
Cit
1
and sufficient conditions for the roots to lie inside the unit circle in terms'
of the signs of the Schur - Cohn determinants. The Schur - <?ohn Criterion
will not be coverved in this text simply because it is very cumbersome for
ql.= [po
. Ps
ps]
po
q2 = [po
. Ps
P1]
p~
... (2.190)
equations higher than secod order. Instead -Jury's stability test is . The necessary and sufficient condition for the polynomial F(z) to
described.:
have no roots on or outside the unit circle in the Z-plane are
,Given an nth - order" equation in Z, F(1) > 0
Zri~l + ~
F (Z) = an zn + an-I a2 Z2 + Rl Z + RO= 8

where ao, al ... an are real coefficients. Assume that an is positive or

... (2.189)
F(l){: ~
n even
n odd
that n can" be made positive "by changing the signs of all coefficients of
Ftz). The following table is made using the coefficients of Ftz). lao] < an ]
Row Zo Zl Z2 ... zn-k ... zn-l' ... zn Ihol > Ibn-II
1 ao al a2 ... an-k ... an-l' ... an
Icol > ICn - 21l (n - 1)constraint s
H~ (2~191)
2 ... ... ...
Idol> Idn-sl
an an-l an-l ak ftl aD
3 ho hI b2 ... bn-k ... b~-l

4 bn- l bn- 2 bn- s ... bk ... bo

5 Co Cl C2 ... Cn-2 For a second - order systemn = 2, the Jury tabulation contains only
6 Cn-2 Cn-3 Cn-4 ... Co one row. Therefore, the requirements listed in Equation (2.191) are reduced
to
....
2n-5 Po PI P2 Ps
F(l) ->> D'0"}
F(-I)
2n-4 Ps P2 P1 Po ... (2.192)
2n-3 qo ql q2
laoi < a2

I.
 (The State-Variable technique ( 2.51 ) ,(2.52 ) Computer Control of Process:!!J

One advantage of the Jury stability test is that, for systems 'of any Since the equation. is of the second order, under Jury's test the
order the eonditions on F(1), F(-l) and between ao and iriin conditions in Equation (2.192) are necessary and sufficient for the system
equation (2.191) from necessary conditions of-stability that are very simple to be stable.
to check without carrying out Jury tabulation. For second order .systems, Thus F(l) > 0 F(1) = 1- 0.25= 0.75> 0
the conditions in Equation (2.192)-are necessary and sufficient.
F(-I 0 F(-I)= 1- 0.25 = 0.75 > 0
For equations that pass the necessary conditions and would require
carrying out Jury's tabulation, the inequality tests of the calculated laol> (a2) laol = 0.25 a2 = 1
coefficients in Equation (2.191) could be quite tedius if the equation Thus, laol < a2 The system is stable expected.
contains variable parameters.
Singular Cases IExample 2.11 I
When some or all of the elements of a row in the Jury tabulation are Consider the characteristic equation, in Equation (2.180)
zero', the tabulation ends prematurely. We refer to this situation as a F(z) = Z3 -1.25Z2 -1.375Z -''0".25= 0 ... (2.195)
singular case. A singular case can be remedised by expanding or
contracting the unit circle I z I = 1 infinitesimally, which is equivalent to ag = 1 a2 = -1.25 al;: -1.375 8.0 = -0.25
moving t'hezeros of Ffz) of the unit- circle. The transformation for this Checking the first three conditions in Equation (2~ 191)
purpose is
F(1) > 0
Z = (1 + s) Z . ... (2.193)
.Ft L) 1 - 1.25 - 1.375 - 0.25 = ~1.875 ... (2.196)
where e is a very small real number. When is a positive number in
Equation (2.193), the radius of the unit circle is reduce to (1 + E), and F(-l)< 0
when. e isnegative, the radius of the unit circle is reduced to 1 + e. The F(-1)= -1- 1.25 + 1.375 - 0.25 = -1.125 ... (2.197)
difference. between the number of zeros found inside (or outside) the unit
circle when the circle is expanded or contracted by E is the number of laol < a3
zeros on the circle. I ao I 0.25 < 1 ... (2~198)
The transformation of Equation (2.193) is actually quite'simple to Since F(l) is negative, not all the roots of Equation (2.195) are not
apply, since inside the unit circle and system is instable. There is no need to carry out
(1 + s)" zn == (1 -i- ns) zn ... (2.194) the Jury tabulation.
For positive or negative E. This means that the coefficient of the zn
terms i~ multiplied by (1 + fie). IExample 2. 12 1
Consider the following characteristic equation of a discrete-data
1 Example 2.10 I system.
Consider the characteristic equation given in Equation (2.178) Z3 + 3.3 Z2 + 4Z+ 0.8=0 ... (2.199)
F(z) = Z2 ~ 0.. 29 = (2 - 0.5) (2 + 0.5) = 0 The equation has roots at Z = u.2463, Z = -1.5268 + j 0.9574 and
With reference to Equation (2.189),'the coefficient of the equations are' Z =-1.5268 - jO.9574. Thus the system is unstable.

. RZ = 1, at = 0, . ao = -0.25
 ( 2~54 ) ComputerControlof Processes )
(The State-Variable Technique
The Liapunov Function
Applying the necessa~y condit.ions of the Jury test in Equation'
(2.191), we have Any function V = V[x(k)] of definite sign (positive definite "or
negative definite is called a Liapunov function. If V( 0) = O. and x(k) is the
F(l) > 0 F(l) = 9.1 ... (2.200) solution of the state equation.
F(-l)< 0 F(-l)= -0.9 ....(2.201) x(k +1) = f(x(k .... (2.206)
laol < a3 .1 ao I =0.8 < 1 ... (2'.202) We define a difference opeartion 8 so that 8V [x(k)] is the solution of
Thus, we must carry out the Jury tabulation as follows to determine the state equation.
the stability: of the system x(k + 1) = f[x(k)] ... (2.207)
Row ZO Zl Z2 Z3
We define difference operation ~ so that ~v [x(k)] is defined. as
1 0.8 4.0 3.3 1.0
LiV [x(k)] =. V[x(k +1)] - V[x(k)]
2 1.0 3.3 4.0 0.8
Stability Theorem of Liapunov
3 .bo bI b2 , Consider a discrete-data system described by

[ao .aa] =ao2- a32 X(k + 1) = f[X(k)] ... (2.209)

bo = ---0.36 ... (2.203)
a3 ao where X(k) is the n x 1 state vector f[X(k)] is an n x 1 function vector
with the property that
bi = [aa3o a2
a1
] =aOa1 - a2a 3= -o.1 ... (2.204) f[x(k)] = 0] o for all k
Suppose that there exists _a scalar function V[x(k)] continuous.in x(k)

h2 = [ao a1=aoa2-a1a3=
] . -1.36 ... (2.205)
such that
a3 a2' .
' . 1. V[x(k)]= O} = V(O) == 0
Since I bo I is not greater than Ib 2J Equation (2.199) has at least one 2. V[x(k)] = Vex) > 0 for. x "* 0 ... (2.210)
root out side the unit circle.
3. vex) approaches infinity as II x II 4 00

- 2.17.3 The Second Method of Liapunov 4. LlVex) < 0 for x "* 0

The powerful method of determining the stability of linear and Then the equilibrium state x(k) =o (for all k) is asymptotically stable
nonlinear system. The- second method of Liapunov is based on the in the large andV(x) is a Liapunov function.
determination of a V function called the Liapunov function. From the
Instability theorem of Liapunov
properties of the V function, we can show the stability or instability of
the system. However, the main disadvantage of Liapunov's stability For the system described by Equation (2.206) if there exists a scalar
criteria- is that it .gives only the sufficient conditions for stability, not the function Vix) continuous in x(k) such that
necessary condition. Furthermore, there are no unique methods of ~V(x) < 0
determining the V function for a wide class ofsystems,
(TheState-Variable Technique ( 2.55 )
(2.56 ] ComputerControl of Processes')
Let the Liapunov function be
then, .
. 1. The system is unstable in the finite 'region for which V is not
Vex) = XI 2(k ) + X22 (k ) , ... (2.218)

positive semidefinite' (~O). Then AV(x) = V[x(k +1)] .; V[x(k)]

2
2. Theresponse is unbounded as k approaches infinity of V is not 1.25 XI (k ) - O.75x22(k) ... (2.218)
positive semidefinite for allx(k). ' Since AVex) is inndefine in sign, the test using Liapunov function of
Equation (2.218) fails, and no conclusion on the stability condition of the
IExample 2. ill system can be reached.
Consider the discrete-data system described by Now let us turn to the instability theorem of Liapunov. Let the
x1(k + 1) - 0.5 xICk) ....(2.211) , Liapunov function can be defined as

X2 (k +,1) = - 0.5 x2(k) ... ;2~212) Vex) = al X1

2
(k ) + 2a2 Xl (k)X2(k) + ag ~22(k) (2.219)
Let. us assign the Liapunov function to be and LiVex) .be of form

Vex) = X1 2 (k) + X22 (k ) ... (2.213)- L\V(x) = - XI 2 (k) - X22(k ) ... (2.20)

Which is positive for all values of XI(k) and A2(k) not equal to zero. So that it is negative definite for all Xl (k) and x2(k) *- 0
Thenthe function ~V(x)is evaluated, using Equation (2.208) FormingAv'(x) according to Equation (2.208)'
L\Vex) V[x(k + 1)] - V[x(k)] We have
= V[x(k + 1)] ~ V[xCk)] LlVex) = V[x(k + 1)] - Vlxtk)] ... (2.221)
= x~2 (k + 1) + x2 2 (k + 1) -'X12 (k ) - X2 2 (k) .. .{2~214) = 1.25al X1 2
(k) - 0.5 a2 xICk) x2(k) - 0.75 ag ,x2
2
(k ) ... (2.222)
Bubstitutefor x1(k +1) and x2(k + 1) in equation (2.214) Comparing Equations (2~222) and (2.220)
= 0.25 X1
2 (k ),+ 0.25 2 2 k) - X22(k )
X2 (k ) - XI C ag = 1.333 ... (2.2'23)
= -0.75 X1 2 (k) - O. 75 X22(k)~ ... (2.215) - Thus from Equation (2.219) .
Since ~V(x) is negative for all x(k) *- 0, the. system is asymptotically Vex) = -O~8' X1 2 (k) + 1.33 X2
2
(k ) (2.223)
stable. ','
Since Vex) is indefinite there is again no conclusion of the stability
I Example 2. 14 I condition.

Consider the digital system described by the state equations Liapunov Stability Theorem for Linear Digital Systems

xI(k + 1) = -1.5 xI(k) ... (2.16) The stability and instability theorem of Liapunov are valid for both
linear and non linear systems. The last two examples iflustrate that a
X2 (k + 1) = -0.5 x2(k) ... (2.17)
successful execution of the Iiapunov tests depends upon 'the correct guess
It is simple to show that the eigen values of the A matrix are at -1.5 or assignment of Vex) or LiVex), whichis always a' difficult task. However,
and -0.5 and that system is unstable. However without prior knowledge for linear digital systems a simple procedure is available.
of the stability conditions of the system, the stability theorem of Liapunov
is applied.
(TheState-Variable Technique' Computer Control of Processes J
(2.57 )

Consider that a linear time - invariant digital system is described IExample 2.15 f
by the difference equation
Consider the discrete-data system discribed by
x(k + 1) = Ax(k) ... (2.224) xICk + 1) = -0.5 xI(k) ... (2.230)
where x(k) is n x 1 and A is an n x n matrix. The equilibrium state x2(k+ 1) = -0.5 x2(k) ... (2.231)
Xc = 0 is asymptotically stable, if and only if, given any positive definite
The coefficient matrix is .
real symmetric matrix Q, there exists a positive definite real symmetric
matrix P such that
... (2.225) .
A= [-0.50 -0.50]' .~. (2.232)

Then The equillibrium state X, = O. Let Q be the identity matrix

[~~]
... (2.226)
Q =. ... (2.333)
is a liapunov function for the system,and further
and let P be of the form
dV(X) = -xT(k) Qx(k) (2.227)

where ~V(x) is as defined in Equation (2.208) p == [PI1 P12]. ... (2.334)

P21 P22
Proof: Then the Equation (2.225) becomes
The' proof of this theorem is based on Sylvester's theorem, 'which ATpA-P = -Q
states that ifP is a positive - definite matrix, then Vex) = x T PX is positive
definite. Using Equation (2.226) as the Liapunov function
= [~ ~] ... (2.335)

AV(x) = V[x(k + 1)] .: V [x(k)]

= XT(k +.1) P x(k ~ 1) - XT(k) Px(k) ... (2.228) -Solving the elements of the Pmatrix from the last equation yields

Now substituting the state equation of Equation (2.224) into 1.33 0 ].

p = [ 0 '1.33 ... (2.336)
Equation (2.228)

We have which. is positive definite. Therefore

'dvex) = xT(k) [ATpA - P] x (k) Vex) = xT(k)P x(k) ... (2.337)

is the Liapunov function and is positive definite. The function ~Vex) .given
d Vex) .- -xT(kJ Q xlk) .. (2.229)
by Equation (2.229), which is negative definite, and the equillibrium state
Thus, from Sylvester's theorem, if dV(x) is to be negative definite, Q is asymtotically stable.
has to. be positive definite,
 (TheState-Variable technique ( 2~59') . ( 2.60 ) ComputerControlof processeiJ
In many applications, all. the state variables cannot be measured
2.18 POLEPLACEMENT DESIGN AND SIATEDBSE'RUERS
because of cost consideration or because of the lack of suitable
transd ucers.
2~ 18.1 Introduction
In these cases, those state variables that cannot be measured must
The goal of. this chaptter is to solve the identical problems using be estimated form the ones that are measured .. Fortunately, we can
different techniques based on state variable formulation. We prefer to seperate the design into two phases. During the first phase, we' design
refer to the two approaches to design as. state variable methods and the system as though all states of the system will be measured. The second
transform -methods. ' phase is concerned with the design of the state-estimator.
The transform methods of design are powerful methods of practical
design. Most 'control systems are designed using variations of these y
methods. An important property of these method is robustness. The
resultant closed - loop system' characteristics tend-to be insensitIve to
small inaccuracies in the system model. This property is very important
because of the difficulty in finding a~ accurate linear model of a physical
system and also, because many systems have significant nonlinear 1- - - - - - - - - - - - - - - - ------,
I
.operations. I
Compensation I
The state variable methods appear to be much-mere dependent on I
I I
having an ,accurate system model for the design process. An -advantage of . I I
these methods 'is that the system representationprovides a complete : I
I Statevectorestimate:
(in ternal) description of the system including possible internal' oscillations :L Constantgain matrix I
or instabilites that might be hidden by inappropriate cancellations in the
~_~

transfer function (input/output) description.

Fig. 2.12: Schematic diagram of a state-feedback control system
The power of state variable techniques is especially apparent when
we design controllers for system with more than one control input or' Figure 2.12 shows how the state-feedback control law and the -state
sensed output. However" in this chapter, we will .ilfustrate the state estimator fit together, and how the combination takes the place of what
variable design methods ~sing single-input, single-output (SISO) systems. we refer as 'dynamic compensation'.

In t hi s chapter, we present a 'design method known as pole

placement or pole assignment. This' method is similar to the root-locus 2.18..2 Stability Improvement by State Feedback
design ill; that, the closed - loop poles may be placed in desired locations. An important aspect of feedback system design is the stability of the
control system, Whatever we want to achieve with the .control system, its
However- 'pcle-placement design" allows all' closed-loop poles to be
stability must be assured. Sometimes the main goal ef feedback design is
placed in desirable Iocationsvwhere as the root-locusdesign procedure
.actually to stabilize a system ifit is initially unstable, or to improve its
- allows only thhe two dominant poles to be placed. There is a cost associated
stability if transient phenomena do not die out sufficiently fast.
with placing all closed-loop 'poles, however, because placing all closed
loop poles requires measurement and feedback of all the state variables The purpose of this section is to investigate how the stability
of the system. properties of linear system can be improved by state feedback.
(The State-Variable Technique ( 2.61J ( 2.62 ] Computer Control of Processes )

Consider the singel-input linear time-invariant system with nth order

state differential equation.

x(t) ~ Axtt) + Bu(t) ... (2.337)

If we suppose that all the n-state variables XI, X2, Xn can be

accurately measured at all times, it is possible toJmplement a linear
control of. the form.
, 'u(t), = -k1Xl(t) - k2X2(t) ... -knxn(t) = -kx(t) ... (2.338)
wherek == [kj, k 2 , .. knl is
a constant state-feedback gain matrix,
With this state-feedback control law, theh closed . . loop system is described
by the state differential equation.

x(t) = (A, ~- Bk) xtt) ... (2.33-9)

.and the characteristics equation of the closed - loop system is Fig. 2.13: Control configuratiori' for a state regulator
ISI-(A-Bk)] = 0 ... (2.340) In this structure, these is no command input (r = 0). Control systems
When evaluated, this yields' an nth - order polynomial in s containing in which the output mustfollow the command signals (called-servo system)'
the n gains k}, k 2 , ... k n The control . . law design then consists' of picking. will be considered later.
the gains. So that the roots of Equation (2.340) are in desirable location. Selection of desirable locations for the closed-loop poles requires
In the 'next section we find that a mildly restrictive condition some iteration by the designer. For now, we will assume that the desired
(namely, the system (2.337) must "Qe completely controlable),all the eigen locations are known, say,
values of (A - B]{) can be arbitrarily-located in the complex plane, by S = AI, A2' ... An
choosing k suitably (with the restriction that complexeigen values, occur
Then the desired characteristics equation is
in complex . . conjugate pairs).
... (2.341)
If all the eigen values of (A - Bk) are placed in the left-half plane,
the closed- loop system is, ofcourse, asymtotically stable; x.(t) 'will decay <The required elements of k are obtained by matching coefficients in
zero irrespective of the value ofx(O) the initial perturbation in the state'. Equations (2.340) and (2.341) then forcing the system characteristic
The system. state is thus maintained at zero value in spite of distrubance equation to be identical with the desired equation.
that act upon the system. The canonical form used in control-law design is the controllable
Systems with this property are called 'regulator systems'. The origin canonical form. Consider, a system represented by the transfer function
of state space is the 'equilibrium state' of the system.
Y (8) ~I Sn-l + f32sn~2 + ... + f3 n
Control configuration, for, a state regulator is shown in Figure 2.13. U(S) Sn+XlSri-l+ .. '+Un
(The State-Variable Technique [ 2.63,) ( 2.64 ) Computer Control of Processes )

A companion-form realization of this. transfer function is given below 2.18.3 Necessary and Sufficient conditions for Arbitrary Pole-
x = Ax. + Bu Placement
y = ex ..(2-.342) Consider the linear time-invariant system (2..3 37) with, state-feedback
control law (2.338); the resulting closed-loop system is given by
Equation (2.339)~ In the following, we shall prove that necessary and

~1
0 1 0 0
sufficient condition for arbitrary placement of closed-loop eigen values
0 0 1 0
in the complex plane (with the. restriction that complex eigen values occur
.where A b
in complex-conjugate pairs) is that the system (2.337) is completely
,0 1
~J
0 0 controllable. We shall first prove the sufficient condition, i.e., if the system
:-a.n -U n-1 -cx,n-2 -0.1 (2.337) is completely controllable, all the eign values of (A - Bk) in Equation
(2.339) can be arbitrarily placed.
C = [~n Bn - 1 ... B2 , B l ] In proving the sufficient condition on arbitray pole-placement, it is
The matrix A in Equation (2.342) has a very special structure. The convenient to transform the state equation (2.337) in to. be controllable
coefficients of the denominator of the transfer function ,preceded by minus canonical form (2.342).
signs form a string along the bottom row of the ~atrix.The rest of the Let us assume that such a transformation exists and is given by,
matrix is zero except for the superdiagonal terms, which are all unity. It
can easily be proved that the pair (A, B) is completely controllable forall x = Px
values of aiS. For this reason, the companion form realization given by
equation (2.342) is referred to as the controllable canonical form.
PI 1 P12 Pin 1

:~j
Pi1 P22
One of the advantages of the controllable canonical formis that the
controller gains can be" obtained from it, just by inspection. The closed- Pn l Pn2
loop system matrix.

0 1 0 0
Pi = [~i1. Pi2 ... Pin] := 1~ 2, .... n.
(2.337)
0 0 1 0
A-Bk = Under the transformation (2.343), system (2.337) is transformed to
the. following controllable canocical model:
0 0 0 1
-an -k1 . ~a1, -kn
-cx,n-l -k2 -a n - 2 - k a x = Ax+BU ... (2.344)

has the characteristic equation

~I
0 1 0
s" + (a~ +kn ) sn-1 + ... +(an~2-+ kg) '8 2 + (cx,n-l + k 2) s + Un + k~ - 0. 0 0 -I
and the controller gains can be found by comparing the coefficients where, K= PAP-l =

-~J
of this characteristic equation with equation (2~341). 0 0 0
-an -Un-l -U n - 2
(The State-Variable Technique, [ 2.65 ) ( 2.66 ) Computer Control ofProces$es,J

From Equation (2.344), we have

o

B= PB =
o 0
0
r 'PlAB
PIB

=
lPIA~_2B
'0 PB
1 0
1 PlAn-lB
I~I - AI = s" + n l
U1S - + ... + an-l s + an = 'I SI - A I
Or
. (characteristic polynomial is invariant under equivaIerice
Pl~B n 2B' An-lbB]
transformation), l. AB ... A - [ 0 ... 01]
=0
The first equation in the. set (2.343) is given by
This gives
Xl = PII Xl -f:'P12 X2 + ....+ Pin Xn = :PIX
Taking the derivative on' both sides of this equation, we get where
=. PI X .=
,~

xl PlAx +.PIBu is the controllability matr-ix, which is non s ingular because of the
But xl (= x2) is a function of X only as per the canonical model (2.344) , assumption of controllability of the system (2.337).

Therefore, Therefore, the controllable state model (2.337) can be transformed

to the canonical form (2.344) by the transformation.
PIB =0 and "x2 = PlAx .
Taking derivative on both sides once again, we get x = Px (2.345)

PI AB =, 0 and Kg =PI A2x

Continuing the process, we obtain
p
P IAn- 2 B = 0 and xn = PI An- 1 x
Taking derivative onceagain, we obtain
PlAn - l B = i Under the equivalence transformation (2.345)., the state-feedback
Thu~ 'control law (2.338) becomes

r PI u = -kx kx .. ~ (2.346)

x= Px ,= l .
PIA X where k-- k p~l= [k1 k2 .. k n ]

PI~n-1
...
~

'where PI must satisfy the conditions

~IB =. PlAB - ... - PIAn-2B = 0, 'PIAn-lB~ 1
 ( 2.68) , Computer Control of Processes )
C!!!Oe State-Varlable,Technique [ 2'.67 )
We now derive the necessary condition by proving that if the system
.With this control law, system (2.344) becomes
(2.337) is not completely controlable, then there are eigenvalues of
(A - Bk) that cannot be controlled by state feedback.
x = (X-Bk)x
It was shown that an uncontrollable system can be transformed into
controllability canonical form
o o o
o
~12] 1:1]+ [Be] u
1 o
o o o
[ ~1 ] = [AC
X2 0 A 22 LX2 0
Ax+Bu

where th.e pair (Xc, be) is completely controllable.

... (2.347)
The set of eigen values of A is the union of the sets of eigen values
of Ac and A22 . In view of the form of 13 , it is obvious that the matrix 1\22
lSI - (A-Bk}1 =sn+(U1 +kn}sn-1 + (U2 +kn_1}sn-2 +...+(an-1 +k 2}s+(un +k1) is not affected by the intorduction of any state feedback of the form
... (2.348) u = -kx. Therefore, the eigen values of A22 cannot be controlled. This
Since the .coefficients ki are arbitrarily chosen real numbers, the proves the necessary condition.

coefficients of the characteristics polynomial of (A - B k) can he given any

, desired values. Hence, ,the .closed-loop poles can be placed at any desired
2.18.4 State Regulator Design
locations in the complex plane '(subject to conjugate pairing: coefficients Consider the nth-order~ single-input linear time - invariant system,
of a characteristic polynomial will be read' only if the complex poles ~re
,present in conjugate 'paIrs).
x(t) = Ax(t) + butt) .... (2.350)

With state-feedback control law

Assume that the desired characteristic 'polynomial of' (A - Bk) and
hence, (X - 13k), is u(t) = -kx(t) .... (2.351)-

The resulting closed-loop system is

From Equation (2.348), it is obvious that this requirement is met if

x(t) = CA - 13 k) x(t) ... (2.352)

It is chosen, an The eigen values of'{A - Bk) can be arbitrarily placed in the complex
plane (with the restriction that complex eigen ~alues' occur in complex-
conjugate pairs) by choosingk suitably if, and only if, the system (2.350)
Transforming the feedback controller (2.346) to the original is completely controllable. The important result on pole placement was
proved in the previous sectiou. The following design steps' for pole
coordinates, w'e obtain
placement emerge from the proof, '
.... (2.349)

This proves that -if (2.337) is controllable, the closed loop poles can
be orbitrarily assigned (sufficient condition). .
(TheState-Variable Technique ( 2.69] ( 2.70 ) Computer Control of Processes)

Step 1: From equation (2.356) (2.354), we get

From the charactersti~ polynomial of matrix A
,lSI-AI = sn + alS+
n-l
... + cx.n- l .
S + an ... (2.353)
k
Determine the values of 0.1, 0.2, .. an-ban

,Step 2:
Determine the transformation .matrix P that transforms the system 1
= [0 0 .. , 0 1]uc- [(a l -u1)A
n-1 +(a2 -u n-2 + ... +(a -un)]
2)A n
(2.350) into controllable canonical form.
(2.357)
The characteristic polynomial of matrix A is equation (2.353)
p I. SI - A I = s
n
+, alS rr-L + ans n-2 + .... + an-l S + Ctn

Since the Cayley - Hamilton .theorem states that a matrix satisfies

its own characteristic equation, we have
~~~,An + Ct1An- 1 + Ct2 A n- 2 + ... + Ctn-l A +Ctn I = 0
P1 = [0 0 .. ~ 0 1]u-1
.. An = - alAn-l - 0,2 An - 2 - .... - an-l A- an I ... (2.358) .
Uc = [b All ... An-1b] ... (2.354), From Equations (2.356) (~.357), we get

Step 3: k = [0 0 ... o 1]UC-1 ~ (A) '.. (2.359a)

.Using the desired eigenvalues (desired closed-loop poles) A.I, A.2, An where ... ~2.359b)
write the desired characteristic polynomial.
(s - A.I) (s - A2) ... (8 - An) = Sn.+ alS n- 1 + ... + an-l S + an ... {2'.355) ... (2.359c)

and determine the vaues of ai, a2, ... , an-I, an. Equations (2.359 (a describe the Ackermann's formula for the
Step 4: determination of the state-feedback gain matrix.

'I'hevreqtrir'ed state feedback gain matrix is determined from the

following equation: ' '2.18.5 Design of State Observers
The pole-placement. design' procedure introduce in the preciding
... (2.356) ,
sections results in controlIaw of the form.
There are other approaches also, for the determination of the state- utt) = -kx(t) ... (2.360)
feedback gain matrix k. In what follows, we shall present a well-known
which requires the ability to directly measure the entire state vector
formula, known as the 'Ackermann's formula', which is convenient. for
xtt). Full state-feedback control fo.: many second-order system, requires
,computer solution..
feedback of position and rate' variables which can easily be measured.
However) for must of the higher-order system, full state measurements
are not practical.
(The State-Variable Technique ( 2.71 ) {:2072 )" Computer Control of Process~
Thus, either a 'new approach that directly accounts for the non-
availability of the entire state vector is to be devised, or a suitable
where X is the estimate of the actual state X. We know A, Band
utt), and hence this estimate is satisfactory if we can obtain the correct
, approximation of the state vector must be determined. The later approach
is much simpler in ninny situation. initial condition x(O) and set X equal to' it.
The purpose of this section is to demonstrate the estimation of all
Plant
the state variables of a system, from the measurements that can be made u y
on thesystem, If the estimate of the state vector is 'denoted byx, it would
be nice if the true state in the control law giveriby Equation (2.360), could
be replaced byits estimate

utt) = -kx(t) ... (2.361) Mode 1

This indeed is possible, as we shall see in the next section. x =Ax+ Bu

A device (or a computer program) that estimates the state variables
Fig. 2.14: Open-loop estimator
is called a state observer, orjsimply an observer. If the state.observer
estimates all 'the state variables of the system, regardless ofwhether some Figure 2.14 depicts this 'Open-loop' ,estimator. However, it is
state variables are available for direct measurement, ~t is called a full- precisely the lock of information on x(9) that requires the construction of
order state observer.' However, if accurately measurements of certain
an estimator. If i(O):;(: x(O), the estimated state i(t) obtained from the
states are possible, we may estimate only the remaining states,' and the
open-loop scheme of Figure 2.14 would have a continually growing 'error
accurately, measured signals are then used directly. for feedback. The
or an error that goes to .zero too slowly, to be of any use. Furthermore,
resulting observer is called a reduced-order' state observer.
small errors in our knowledge of the system (A, B) and the.. disturbances
that enter the system, but not the model, would also cause ~he estimate
2.18.6 Full-Order State Observer to slowly diverge from the true state.
Consider ,a process described by the state equation Inorder to speed up the estimation process and provide a useful state
estimate, we feed back the difference between the measured and
x(t) = Ax(t) + Bu(t) ... (2.,362a)
estimated output 'and correct the model corrtirruousl.y with this error
where A and bare, respectively, n x nand n x 1 real constant signal. 'This scheme, commonly known as 'Luenberger state observer', is
matrices. The measurement yet) is related to the state by equation shown in Figure 2.6 and the Equation for it is
yet) = ex (t) .~. (2.362b)
i(t) = M(t) + Bu(.t) + m(y(t)-y(t)) ... (2.364)
where C is 1 xn real constant matrix, without loss of generality, the
where m is an n x 1 real constant gain matrix. The state' error vector
direct. transmission' part' has assumed to be zero.
One method of estimating all the state variable that we may consider, X(t) = x(t) - x(t) ... (2.365)
is to construct a model of the plant dynamics

i(t) = Ai(t) + bu(t) , ... (2 ..363)

 ( 2.74 ) computer Control of Processes ]
(The State-Variable Technique ( 2.73 )
If we do not have a very accurate model of the plant (A, B, (I) the
Differentially dynamics of the error are no longer governed by Equation (2.366). However,
x(t) i ( t) - i (t) "m' can typically be chosen so that the error system is stable and the error
is acceptably small, even with small modeling errors and disturbance
Plant inputs.
y
X = Ax + Bu The selection of 'm' can be approached in exactly the same fashion
as the selection of k in the control law design. If we specify the desired
<Iocation of the observer-error roots as
.r-----------~----~-~--------~~---------~
1
I"
1
I s = Ai, A,2, .. , Am
1 I
I i The desired observer characteristic equation is
I I
I I
1 I ~ .. (2.367b)
I I
1 I
I n- parallel 1
and one can solve for m by comparing coefficient in Equations (2.367
: integrators : a) and" (2.367 b). However, as we shall see shortly, this can be done only if
1
I
1
I
the system (2.362) is completely observable.
I 1
1 1 The calculation of the gains using. this simple technique becomes
L
1 Observer ~J
I
rather tedious when the order of the system is larger than three. As in
Fig. 2.1S~ Luenberger state observer the controller design, there is an observable canonical form for which the
observer design equation are particularly simple. Consider' a system
Substituting for x(t) and i(t) from equation 2.362. and 2.364 represented by the transfer function.
respectively, we get
y(s) [31 Sn-1 + J32sn- 2 +.;. + J3n
x(t) = Ax(t) + Bu tt] -Ai(t) ,- Bu(t) -,mc(x(t) -x(t)) U(s) = 's" + a1 Sn-l + ... + an

= (A-mc)x(t) ... (2.366) A companion-form realization of this transfer function is given below

The characteristic equation of the error is given by ,

... (2.368)
1,81 - (A - me) I= 0 ... (2.367a)-
If 'm' can be chosen so that (A - me) has stable and reasonably fast
.roots, x(t).$,will decay to zero irrespective of x(O). 'I'his .means that x(t)
will converge to xtt) regardless of the value of x(O), and furthermore, the J3n
dynamics of the error can be chosen to be faster "thari the open-loop J3n-1
dynamics. Note that Equation (2.366) is independent of applied control. B =J3n-2 C = [0 0 .... 0 1]
This is a consequence of ansuming ,A, Band c to be identical in the plant
and the observer. Therefore; the estimation error x converger 'to zero
and remains there, independor of any known forcing function urt) on the
plant and its effect on the state x(t).
(TbeState-Variable Technique ( 2.76 ) Computer Control of Processes}

It can easily be proved that the pair (A, C) is completely observable Since
for all values ofcq's. For this reason, the conipanion form realization given
det W
by Equation (~.a68) is referred to as observable canonical form.
One obtains
One of the advantages of hte observable canonical form is that the.
observer gainsmcan be obtained from it, just by inspection. The observer- , det lSI - (AT - CT mT)] = det lSI - (A - mC)]
error matrix is (i.e.,) the eigen values of (AT - C T mT) are same as the eigen values of
(A- me)
0 0 0 -an - f i l
Table 2.1 Duality between and control and Estimation
1 0 0 -an-l -m2

(A - me) = 0 1 0 -U n - 2 - rn 3 Control Estimation

A AT
0 0 1 ~al ~mn
B CT
which has the characteristic' equation
k mT
sn + (al + mn)sn-l + ... + (an- 2+ ms)s2 + (Un-l + m2) s + an + mj = 0 and
the observer gains can be found by comparing t.he coefficients of this By comparing the characteristic equation of the closed-loop system
equation' with equation (2.367 (b. (2.352) and that of the auxilIary system (2.370), we obtain the duality
A procedure for observer design, therefore, consists of transforming relations given in Table 2.1 between the control and estimation problem. '
the given state variable. model to observable canonical form, solving for The Ackermann's control-design formula given by Equations (2.359)
the observer gains, and transforming the gains back to the original state. becomes the observer-design formula if the substitutions of Table 2.1 are,
made.
V!e can, howevervdirectly use' the equations of the control-law design
for computing the observer gain matrix ID, if we examine the resemblance A necessary and sufficient condition for determie of the observer
between the estimation and control problems. In fact, the two problems gain matrix ill for the desired eigen values of (A - me), is that the auxillary
are mathematically equivalent. This property is called duality. The design system (2.369) be completley controllable. The controllability condition
of a full.. order observer requires the determination of the gain matrix 'm' T T 1 T
for this system is that the rank of [C A TC ... (A Tr- C ] is n. This
such that (A _. me) has desired eigen values Ai, i = 1, 2, ... n. This is
mathematicaly equivalent to .designing a full jstate.. feedback controller is the condition for complete observability of the original system defined
for the 'transposed auxillary system'. by Equation (2.362). This means that a ne~essary and sufficient condition
for estimation of the state of the system defined by Equations (2.362) is
~(t) = AT l;(t) +CTll(t) ...(2.369a)
that the system be completely observable.
With feedback
Again by duality, we can say that for the can of single-output system,
l1(t) = -inT ~(t) ... (2~369b) the gain matrix m, which places the observer poles at desired locations,
So that the closed-loop. auxiliary system. is unique.. In the multi-output case, the same pole configuration can be
achieved by various feedback gain matrix.
. ~(t)= (AT _CTmT)~(t) ... (2.370)
has eigen values A.i; i = 1, 2, ... n
 [ 2.78 ) Computer Control of Processes )
(Th,e State-Variable 1echnique ( 2.77 )
where the two rightmost terms are known and can be considered as
2.18.7 Reduced .. Order StateObserver an input into the Xe dynamics. Since Xl = y, measured dynamics are given
The observer developed' in the previous. subsection reconstruct the entire by the scalar equation
state vector. However, the measurements usually available are some of the states
... (2.373)
of the plant. For example, for the satellite-attitude control problem considered
in the previous subsection, the measurement is orientation of the satellite, which If we collect the known terms of Equation (2.373) on one side, we get
is Xl(t). The measurement of a state, in general will be more accurate than any
y' -- -allY - hI U = aie x, ... (2.374),
estimate of the 'state based on the measurement. Hence, it is not logicalin most -v--'
Known Measurement,
cases to estimate states that we are measurement.' One possible exception is
Note that Equations (2.372) (2.374) have the same relationship to
the case in which a measurement is very noisy. The state observer for this case
the state X e that the original Equation (2.371) had to the entire state x,
may furnish some beneficial noise fillter,
Following this line of reasoning, we can establish the following'
Since we will not usually want to estimate any state that we are measuring, substitutions in the original observer design equations, to obtain an
we prefer' to design an observer that estimates only those states that are not (reduced order) observer of X e :
measured. This type of observer is called a "reduced-order state observer".
x -(- x,
We develop design equations for such an observer in this subsection. A +- A ee
We consider only the case of one measurement. It is assumed that the
bu~ aai Y + beu
state variables are always chosen such that the state measured is ~l(t);
we can, do this without loss of generality. The output equation is given by Y <f- Y- allY - b l u
yCt) = Xl(t) = Cxtt}. C ~ aie ... (2.375)

where C = [1 0 0 ... 0] Making these substitutions into the equation for full-order observer
(Equation 2.364) we obtain the equation of the reduced -order observer.
To derive the reduced - order observer, we partition the state vector
into two parts; one part is Xl which is directly measured and the other
part is Xe, representing the state variables that need to be estimated. i e= Aeex e + "aeiY + bell + m[y-allY-blU-aiexe)
-------v-----' ~--'
... (2.376)
input . measurement

x(t) == [Xlx, (t)]

(t)
, If we define the estimation error as
... (2.377)
If we partition the system matrices accordingly, the complete
the dynamics of error are given by subtracting Equation (2.376) from
description of the system is' given by ,
Equation (2.372).

[X~le J = [al~
a
aie J[X~J + [blJu
A ee x,
el be .
...(2.371a) ... (2.378)

Its characteristic equation is given by

y = [1 0][::] . . (2.371b) I 81 - (Aee + maie)1 = 0 ... (2.379)
We design the dynamics of this observer by selecting 'm' so that
The dynamics of the unmeasured statevariables are given by
equation (2.379) matches a desired reduced-order characteristic equation.
= A ei x, + aeixl + beu
~
... (2.372) To carry 'out the design using, state regulator results, we form a
Known input 'transposed auxilIary system'.
 ( 2.80 ) ComputerControl of Processes )
[The State-Variable Technique [ 2.79]
2.18.8 Digital Control System With State Feedback
~(t) .- AeeT ~(t) +aie
T
11(t)
This section covers the key results on the pole-placement design,
11(t) - -IDT l;(t) ... (2.380} and state observers for discrete-time systems. Consider the discretized
Use of Ackermann's formula given by equations (2.359) for this x(k + 1) = Fxfk) + gu(k)
auxiliary system gives the gains m of the reduced-order observer. We
should point out that the conditions for the existence of the reduced-order
y(k) = ex (k) ... (2.383) ,

observer are same as for the full-order observer - namely observability of Where x is the nx 1 state vector, u is the scalar input, y is the scalar
the pair (A, C). output; F g and c are, respectively, n x n, n x 1 and 1 x nreal constant
matrices and k = 0, 1, 2_ ~ ....
Let us now look at theimplementational aspects of the reduced-order
observer given by Equation (2.376). This equationcan be rewritten as We will carry out the design of digital control system for-the plant
(2.383) in two steps. One step assumes that we have all the elements of
xe = (A ee -male)xe~(ael -mall)y+(b e -mb1) u + mY .. (2.381)
the state vector at our disposal for feedback purposes. The next step is to
The fact that the reduced-order observer requires the derivative of design a state observer which estimates the entire state vector, when
yet) as aninput, appears topresenta practical difficulty. It is known that provided with the measurement of the 'system indicated by the output
diffenentiation amplifier noise. So if y is noisy the use of y is unacceptable. equation in (2.383).
To get around this difficulty, we define the new state as, The final step will consist of combining the control law and the
... (2.382a) observer, where the control law calculations are based on the estimated
state variables rather than the actual state.
Then, in terms of this new state, the impelementation of the reduced
order observer is given by
2.18.9 State Regulator Design
... (~.382b) - Consider the nth - order, single-input, linear time- invariant system
and if no longer appears directly. A block diagram representation of the (2.383) with state feedback control law .
reduced order observer is shown in Figure 2.16. u(k) = -kx(k) ... (.2~384)

y------.--,..---------.. .
The resulting closed-loop system is
x(k + 1) = (F- gk) x (k) ... (2~385)
If all the eigen values of (F - gk} are placed inside the unit circle in
the complex plane, the state x(k) will decay to the equilibrium state x = 0
u
irrespective of the value of x(O) the initial perturbation. in the state.
n-1 parallel
integrators A necessary and sufficient condition f~arbitray placement of closed-
loop eigen values (with the restriction thatcomplex eigen values occur in
conjugate pairs), is that the system (2.383) is completely controllablet

Fig., 2.16: Reduced - Order observer structure

(The State-Variable Technique [ 2.81 ) ( 2.82 ) Computer Control of Processes )

The characteristic equation of the closed-loop system is Step 3:

I.ZI - (F -gk) I = 0 ... (2.386a) Using the desired eigen values (desired closed-loop poles) At, ).,2, ... ,
An write the desired characteristic polynomial.
Assuming that the desired characteristic; equation is
(Z - A1) (Z - A2) ... (Z - An) = zn + a1 zn-l "+ ... + an-1 Z + an (Z - I~l) (Z - ""2) ... (Z - An) = zn + al zn-l + ... + an-l Z + an ... (2.389)
i
and determine the values of ar, a2 ... , an-l, an.
o ...{2.386b)
Step 4:.
The required elements of k are obtained by matching coefficients-in .
Equations(2.386a) and (2.386b) . The required state - feedback gain matrix is determined from the
following equation.
The calculation of the gains using this method becomes rather tedious
when the order of the system is greater than three. The algebra for finding ... (2.390)
the gains becomes especially simple when the state variable equations
are in controllable canonical form. A design procedure .b a s ed on t h e use The Ackermann's formula given below is more convenient for
of controllable canonical state variable model, is given model. computer solution,

Step 1: k [0 0 ... 0 1]U-1 ~(F) ... (2.391)

From the characterstic polynomial of matrix F; where"
IZI - F.I = Z" + a1 zn-l + ... + Un-'-l Z + an ... (2.387)
Dc [g Fg ... Fn-lgJ
determine the values of a1,. u'2, ~ .. an' IExample 2. 15]
Step 2: Consider the problem of attitude control of a rigid satellite. A state
Determine the transformation matrix P that transforms the system variable model of the plant is
(2.383). into controllable canonical- form:
x = Ax + bu= [~~]x + [~]u
where Xl = e is the attitude angle and uis the system input.
p
The discrete-time description of the plant (assuming that the input
u is applied through a zero - order hold (ZOH) is .given below.
x(k'+ 1) = Fx(k) + gu (k) ... (2.392)
.P1 = [0 0 ... 0 1]

Dc = [g Fg ... Fn~lg ]
where F = eAT = [~ ~]
... (2.388)

g
 ( 2.84) Computer Control of Processes J
(The,State..v ariable Technique "f 2.83 )
2.18~11 Prediction Observer
The characteristic equation of the open - loop system is
An estimation, scheme employing a full-code observer is shown in

JZI-FI
r Z - 1 _T l
Z- 1 J I=(Z _1)2' = 0
,
Figure, and the equation for it is
! 0 x(k+l) = Fx(k) +gu(k) + m(y(k)-Cx(k))'
'L ... (2.394)
with, the control law where m is an n x 1 real constant gain matrix. We will call this a
u(k) = -kx(k) = -[k1 k 2]x(k) predication observer because the estimate x(k+l) is one sampling period
ahead of the measuremet y(k).
the clos sd loop system becomes
x(k + 1; = .F(- g k) x(k) u(K) y(K)

The characteristic equation .of the closed-loop system is

IZI - (F - gk) I
r--------------------------------------~
... (2.393a) I I
i I
I I
We assume that T = 0.1 sec, and 'the desired characteristic roots of i I
the closed-loop system are Z12 = 0.875 < 17.9. " I
I
I
Note that these roots .corr-v-pond to t
I
+ I
S 0.5 and ron = 3.6 I
I
I
I
I
I
L ~ ~ ~
The desired characteristic Equation is then I Observer i

... (2.893b)
Fig. 2.17: Prediction Observer
Matching coefficients in Equations (2.393 (a) and 2.334 Eb) we obtain
A difference equation describing the behaviour of the error is
k 1 = 10; k 2 = 3.5 obtained by subtracting Equation (2.394) from Equation (2.383).

2.18.1'0 Design of State Observers

i(k+}~) = (F-.mc)x{k) ... (2.395)

The control law designed in the last subsection assumed, that all where x = x - X
states were available for feedback since, typically, not all states are
The characteristic 'equation of the error is given by
measured; the purpose of. this subsection is to show how to determine
algorithms which will reconstruct all the states, given measurements or IZI - (F - mel] == 0 ... (2.396a)
a portion of them. If the' state is x, then the estimate is x and the idea is Assuming that the desired- characteristic equation is
to let u':-= --k X; replacing the true states by their estimates in the control ... (2.396b)
law.
 ( 2.86 ) Computer Control' of Processes]
(The State-Variable Technique ( 2.85 )

The required elements ofm are obtained by matching coefficients in The improved x(k+l) is
i(k~+l). 'I'he state observer based on this
Equation (2.396 a) and Equation (2.396 b). A necessary and ~ufficient formulation is calledthe current observer. The current observer equation
condition for the arbitrary assignment of eigen values of (F - me), is that are given by
the system (2.383) is completely observable. x(k+l) = Fi(k) + gu{k] ... (2.398a)
The problem of designing a full - order, observer is mathematically
equivalent to designing a full state - feedback controller for the 'transposed x(k+1) = 'x(k+l) + m[y(k + 1)-cx(k + l)J ... (2.398b)
auxiliary 'system'.
In practice, the current observer cannot be implemented exactly
t;(k +,1) = FTt;Ck) + C T llCk) ... (2.397a) because it is impossible to sample, perform calculations, and output with
With feedback absolutely no time elapse. However, theerro.cs introduced due to
computational delays will be negligible if the computation time is quite
ll(k) = _mT l;(k) ... (2.397b)
short-compared to the sample period.
So that the closed-loop auxiliary system
The error Equation for the current observer is similar to the error
t;(k + 1) '=' (F T - C T m T) s(k) ... (2.397c) eqution for the prediction observer given in (2.395). The current-estimate
error equation is obtained by subtractly Equation (2.399) from (2~383).
has eigen values Ai, i = 1, 2 ... n.
This duality provides may be used to design full-order state ob-servers .x(k+l) = x(k+l) - x(k+1)
by Ackermann's formula (2.387) (2.390).
= Fx(k) + gu(k) - Fi(k) --, gu(k) - mc[x(k+l)-x(k)]
,2.18.12 Current Observer
FX(k) - mcFx(k)
The prediction observer given by Equation (2.394) arrives at the state,
estimatex(k) after receiving measurements up through yCk-1). Hence (F-mcF)x(k) ... (2.399)
the control u(k) = -lci.(k) doesnot utilize the information on the' current
Therefore, the gain matrix m is obtained exactly as before, except
output y(k). For higher- order system controlled with a slow computer, that C is replaced by CF.
or any time the sample rates are fast compared to the computation' time,
this delay between .making a .measurement and using it.in control law
2.18.13 Reduced - Order Observer
may be a blessing. In many system, however, the computation time
required to evalute Equation (2.394) is quite short-compared to the sample The observers discussed so far, are .designed to reconstruct the entire
period- and the control based on prediction observer may not be as state vector, given measurement of some of the states. To pursue an
accurate as it could be. observer fOl only the unmeasured states, we partition the state vector
into two parts. One part is Xl which is directly measured, and the other
An alternative formulation of the state observer is to use y(k) to
part is Xe , representing the state, variables that need to be estimated. If
obtain the state estimate x(k) . This can be alone by separating the
we partrtion the system matrices accordingly, the complex description of
estimation process into two steps, In the first step we determine x(k+l),
the system (50)i8 given by
an approximation of x(k+ 1) based on x(k) and ufk),using the model of the
plant. In the second step, we usey(k+1) to improve x(k+l). ,
rXl (k+l)ll
... (2.400a)
Lx, (k+l)-J
 (The State-Variable Technique (2.88 ) Computer Control of Processes )
( 2.87 )

IExample 2.16 I
y(k) = [1 0] [Xl (k)] ... (2.400b)
xe{k) As an example of complete design, we will add a state observer to
the satellite-attitude control, The system equations of motion are (refer
The portion describing the dynamics of unmeasured states is to Equation (2.392).

xe(k + i) = Feexe (k) + Fel Xl (k) + geu (k),

.. \
... (2.401) , The formula for conversion of discrete time state variable model to
Know~ input the transfer function model is given by

The measured dynamics are given by the scalar equation y(z) _ a(Z) = C(ZI _ F)-l g+ d ... (2.408)
u(z) -
y(k + 1) - f 11y (k ) - gju'(k) = FIe x, (k) ... (2.4~2)
Applying this result to the model (2.407)~ Wa obtain
Equations (2.401) and (2.402) have the same relationship to the state
x, thatthe original equation (2.38---) had to the entire state x. Following . u(z) _ D(Z) = k(ZI - F + gk + me)"! m ... (2.408)
this reasoning, we arrive at the desired observer by making the following -y{z) -
, . substitutions into the observer equation.
x(k + 1) = Fx(k) + gu(k)

F ~Fee [~ ~]X(k) + [T~2 ]U(k)

gu(k)~ f e l y(k) + geeu(k)
y(k) ~ y(k+1)- f 1I y (k ) - glu(k) We assume that the position state Xl is measured and the velocity'
state X2 is to be estimated, the measurement equation is therefore,
C ~ FIe' ... (2.403) ,
y(k.) = Cx(k) = [1 0 ]x(k)
Thus, the reduced - order observer equations are
.Xe (k+l) = We will design a first-order observer for the state x2(k)

Fee xe(k)+ FeIy(k)+ geu(k) + m(y(k + 1)-fI1y (k)- glU(k)- fIe x(k)) The partitioned matrices are
\- J \ -I

. input . . measurement

... (2.404)
SubtractingEquation (2.404) from (2.401) yields the error equation
From Equation (2.406), we find the characteristic equation interms
... (2.405)
of'm:
IZ - (1 - mT) I= 0
The characteristic equation is given' by For the observer to be about four times faster than the control; we
... (2.406) place the observer pole at

We design the dynamics of this observer by selecting ill so that Z = 0.5 (= 0.835)4
Equation (2.406) matches a desired reduced-order characteristic equation. ' .. I-roT = 0.5
'rhe design may be carried out directly or by using duality principle.
 ( 2.90 ) Computer Control of Processes)
(The State-Variable Technique [ 2.89 )
TWO MARKS QUESTIONS AND ANSWERS
For T = 0.1 sec,'
1. What is state space analysis?
m=',5.
The state space analysis is a modern approach ;or powerful
The observer equation 'is.
technique for the analysis and design of control systems, and also
easier for analysis using digital computer.
2. Write the advantages 'or state space analysis?
.. This analysis can be carried with initial conditions and can be
= O.5X 2 (k ) + 5(y(k+l)-y(k)}+O.075u(k) carried on multiple input and multiple output systems.
.Substituting for u(k) from the control law This approach is easily amenable to solution through digital
computer.
u(k) = ~10y(k)-3.5X2(k) ... (2.409a)
The modern control are specially apparent when we design
We 'obtain controllers for systems with more than one control input or
semed output.
,X2 (k + l ) = O.2375X 2 (k) + 5y(k+l)-5.75y(k) ... (2.409b)
3. Define state.
To relate the observer based state-feedback design to a classical The state of a dynamic system is a minimal set of variables such
design, one needs to compute the Z-transform of Equations(2.409a) and "that the knowledge of the inputs for t 2:: to, comp.letely determines
(2.409b) obtaining the behaviour of the system for t > to.

u(z) 27.5(Z -'0.818) 4. Define state variable.

=
-y(z) Z-O.2375 The smallest set of variables that determine the setof the dynamic
_system at any time instant are called state variables.
1
It looks similar to lead compensation that could for 2"" plant. 5. What is state vector?
. s
The state variable are needed to completely describe the behaviour .-
of a given system, then these n state variables can be considered
the n components of a vector x. Such a vector is called a state
vector.
6. Define state space.
The n - dimensional. space whose co-ordinate axes consist of the
Xl axis, X2 axis x n axis, is called a state space. Any state can
be represented by a point in the state space.
8. Define state diagram.
The pictorial representation' of the state model of the system is
called state diagram.
[The State-variableTechnique .[ 2.91 ) ( 2.92 ) computer Control ofProces~~

9. How the state dtagram of the system. is represented? 14. What are the different methods available for computing A k?

Signal flow graph form. Z - transform

Block diagram form. Canonical transformation.
10. Write the fundamental elements are used to construct the Cayley - Hamilton theorem.
state diagram. 15. What is the stability criterian for sample of data control
Scalar" system?

Adder It is states that the system is static if all the poles of the Z - transfer
function of the system lies inside the unit circle in Z -. plane.
Unit delay
16. Define characteristic equation of sampled data system,
11. Define state transition matrix.
The characteristic equation is the denominator polynomial or the
The solution of the 'homogeneous state equation.
Z - .transfer function of a sampled data control system.
x(k + 1) = Ax (k) ... (1)
17~ Write the types of stability.analysis of sampled data control
as x(k) =~ (k) x (0) ... (2) systems,
where, 4> (k) is a unique n x n matrix satisfying the condition. . Jury's stability test.
<I> {k -+- 1) =A <I> (k) $ (0) = I ... (3) Bilinear transformation.
k ... (4)
<1>-. (k)= A Root locus techniques.
From equation ,(2) the solution of equation is simply a 18. What are the necessary and sufficient conditions' to be
transformation of the initial state. Therefore, the unique matrix satisfied for the stability of sampled data control system
<I> (k) is' called the state transistion matrix. It is also called the using Bilinear transformation and Jury's stability test"
fundamental matrix..
Jury's stability test
12. Write the pr-oper-tf es of "the state transistion matrix of
Necessary condition
discrete time system?
Let F(Z) be the characteristic equation. of the system F(l) >0,
<I> (0) I and (_l)n F (-1) > 0
$-1 (k) <I> (- k)
or
ep (k, lto) = <I> (k - k o) A(k...,. kO), where k > k o
F(-I 0 Even power
13. What is state transition matrix of discrete time systemf
F(-1)< 0 Odd power :
The.matrixq> (k) (or A k ) is called state transition matrix of discrete
time system.
It is given by, A k = Z-l {(ZI - A)-l Z} It is used to find the state
of the system, at any discrete time instant k.
(The.State.;VariableTechnique [2.93 ). Computer Control of Processes )

Sufficient condition 22. Define complete output controllability.

'I'he system described by equations.
I ao I < an
x (k + 1) Ax (k) + Bu (k)
Ibo I > Ibn- 1 I
I Co I ,> Icn -21 y(k) = cx (k) + Du (k)
is said to be completely output controllable for any initial time
Iro I > Ir21 (stages) k= 0 there exist a set of unconstrained controls, u (k),
Bilinear Transformation k = 0, 1, .... N-l. Such that any final output C(N) can be reached
Necessary condition from arbitary initial states in finite time (stage) N.
F(l) > O. 23. Write the features of controllability?
(_l)n F(-I) > 0 The controllability verifies the usefulness of a state 'variable.
Sufficientco~dition ~ In the controllability test we can find whether the state
variable can be controlled to achieve the desired output,
There 'should not be any sign change in the elements of first
column of routh array.' Controllability is an important pre requisite before attempting
to design" the control of any process.
19. What is Bilinear transformation.
24. What, are the two methods are used to find controllability
It is a tranformation used to map the interior of unit circle in
and observability?
the Z-plane into the left half of r - plane. The transformation is
achieved' by choosing.' 1. Gillbert's test.
2. Kalman's test.
1 +r
Z = 1- r 25. Siate the condition for controllability by Kalmarr's ~ethod?_
For a nth order system described by state equation
20. State the concept of controllability and observability test.
x(k.+l)=Ax (k) + Bu (k), we can form a composite matrix, Qcwhere
The concept of controllability and observability were introduced by
R.E Kalman. They play an important role in the optimal control of Qc = [B AB A 2B An~l B]
multivariable .systems, Controlability and observability are properties In this case the system is completely state controllable if the
which structural features of dynamic system. Properties play an rank of Qcis n. '
important 'role in modern ,~ontrol system design theory.
26. What is 'the disadvantage in Kalman's test for controllability?
21. Define complete state controllability.
In Kalman's test we can't find the state variable which is
A system' described by equation, x (k + 1) = Ax (k) + Bu (k) is said uncontrollable.
to be completely state controllable, if for any initial time (stage)
k =. 0 there exists a set of .unconstrained controls, utk), k =0, 1,
2.....N-l, which transfer each initial state xeD) to any find state
. x(N) for some information.
(Th.e State-VariableTechnique ( 2.95 J (2.96 ) Computer Control of Processes )

27. Write the concepts. of, obser-vabtlfty? 320, State pole - placement design.

The concept of observability is useful in solving the problem of In pole. - placement. method, the closed loop poles may be placed
reconstructing unmeasurable state variables. at any desired locations by means of state feedback through an
appropriate sta~e feedback gain matrix 'k'. The necessary and
It is the complement of controllability. sufficient condition to satisfied by the system for .arbitrary pole
It is necessary to estimate the unmeasurable state variables in placement in that system be completely state controllable.
order to contruct the feedback control signals. 33. Write the expression .for state feedback gain matrix?
28. Definition ofobservability?
K = [k 1 k 2 knl
Complete observability: The digital system describe in equation
x (k + 1) =Ax (k) + Bu (k), y(k)' ~ ex (k) +Du (k) is said to he
completely observable if for .any initial time (stage) k = 0, any
state 'x (0) can be determined from knowledge of the output C(k)
and input u (k), for 0 ~ k :::; N, where N is some finite time.
29. Write the features of observability?
It can be used to find whether the state variable is observable
or measurable.
It is very useful in solving the problem of reconstructing
unmeasurable state variables from measurable ones in the
minimum possible length of time.
In state feedback control the estimation of unmeasurable state
variables is estimated in order to construct the control signals.
30. What are the methods used to find observabifityf
Gilbert's' test.
Kalman's test.
31. State the Kalman's test for observabffity? Consider a system "
with stata model.
x (k + 1) = Ax (k). + Bu (k) and

y(k) = Ex (k) + Du (t)

For the system a composite matrix, Qo: can be formed such that
Qo' = [C T AT CT {AT)2 C~ (AT)n-l C T]

where n=order of the system (or) number of state variables. The

rank of matrix is n, if the determinant of composite matrix Qo is
non - ,zero.
 3.2 Computer Control of Processes )
CHAPTER

~
3
h
INTRODUCTION TO Volume V
Concentration C
~ FlowF
SYSTEM IDENTIFICATION Concentration C

INTRODUCTION Fig. 3.2: A Stirred Tank

A model is an image of a dynamic system. A dynamic system is shown Suppose we, want to design a controller which acts on the flows Fl(t)
in .Fig.' 3.1. The dynamic system is driven by input variables tr(t) and and F 2(t) using the measurements of F(t) and c(t).,The purpose of the
disturbances vet). The user can control u(t) to obtain desired output yet). controller is to ensure that F(t) or C(t) remain conta.nt even if the
For a dynamic system the control action at time t will influence the output concentrations Cl(t) and C 2( t ) 'vary considerably.
at time instants s > t. For. such a controller design we need some form of mathematical
model which' describes how the input, output and disturbing variables
Disturbance
v(t) are related.
Many industrial pr,?cess, for example for production of Iron, Sugar,
Glass, Paper must be controlled in order to run safely and efficiently..
Input Output, To design controllers, some type of model of the process is needed.
u(t) y(t)
The models can be of following various types.

Fig. 3.1: ADynamic System .. Distributed and lumped model

.. Static and dynamic model
'3.1 NEED FOR MODELING DYNAMI,C SYSTEMS Discrete and continuous model
The following example of dynamic system illustrate. the need for .. Deterministic and stochastic model
mathematical model.
Explicit and implicit model
Consider a stirred tank as shown in Fig. 3.2, where two flows are
mixed. The inputs variables are input flows Fl(t) and F2(t). The output Linear and nonlinear model
variables are outflow F(t) and output concentration .ctt). The disturbing.
variables are input concentration C1(t) and C2(t ). The flows F 1 and F 2 ,ar e
controlled with control valves.
 3.4 Computer Control of Processes )
(Introduction to System ldentification [3.3 )
For example the model to be fitted is
The main uses of model are,
1. To improve understanding of the process Y(s) K -td
- -.- - e s
R(s) - 't s+ 1 .
2. To optimize process design and hence operating conditions.
Wh.ere,
3. To design a control strategy for the process. K - Process steady state gain
4. To train operating personal , 'r - Time constant (in seconds)
5. To design model based controller (Example: MPC,Model td - Delay time (in seconds)
Refer-ence 1\dativeController) which helps in achieving. .From experimental data, the process parameters such as K, 't and td
uniformity, disturbance rejection and setpoint tracking, all-of are estimated using system identification technique.
which translate into better process economics,
6. To plan and sche-dule production. .3.3 SYSTEM IDENTIFICATION STEPS
Since the system identification is an experimental approach, the
3.2 MATHEMATICAL MODELING AND SYSTEM IDENTIFICATIO.N steps listed in Fig. 3.3 are used to estimate model of a process.
IIi order to analysis the behaviour of a process and to design its
controller, we need a mathematical representation of the physical arid
chemical phenomena taking placein it. Such a mathematical representation
constitutes the model of a process. The activities leading to the constrution Construct an . Priori knowledge
experimental setup of the model
of the mathematicalmodel will. be referred to as mathematical" modeling.
There are two ways of constructing mathematical models: Conduct an experiment
and collect data .
1. Mathematical Modeling
2. System Indentification Determine I choose
model structure
1. Mathematical Modeling
This is an analytical approach. Basic laws from' physics (suchas
Newton's laws and balance equations) are used, to describe the dynamic
behaviour. of a process. Model validation
I
2. System Identification I--
I
System identification is the field modeling dynamic system or process
from experimental' data. 'So it is anexpert~ental approach. Some
No
~ew dataset
experiments are performed on the process' to collect input and output
data. Then a model is filled to the experimental data by assigningsuitable
values to' its parameterer.
Fig. 3.3: Schematic Flowchart of System Identification
(Introduction to System Identification
Computer'Controlof Processes )
Step 1: Construct an experimental setup for the given process or plant.
3. Wh,at is mathematical modeling?
Step 2: Conduct an experiment by exciting the, process (using input
Models obtained from fundamental laws. (such as Newton laws,
signal as step, sinusoidal or random. signal) and observing
material and energy balance equations".... ) are known as
its input and output over a time interval' and record it for
mathematical modelsralso known as first principles models) which
preprocessing.
a~e used to describe the dynamic behaviour of a process. It is an
Step BeDetermine /choose appropriate model structure (typically analytical approad}t.
"linear differential equation of a certain order).
4. What ~s empirical model?
Step 4: Choose a suitable satistically based method to estimate the
It is an alternative approch to a mat.hen.at.ical unodel. Some
'unknown parameters of the model (such as coeficientsin the,
experiments are performed on the process to collect input and output
differential equation).
data. Then a model is fitted to the-experimental data by 'assigning
Note: 'In practice, the, estimationsof structure and parameters are suitable values to its parameters. So it is an experimental approach.
often done iteratively. 5. What are the different approaches to modeling?
Step 5: The model obtained is validated- to test whether it is an exact a) Mathematical model
model of the process. If the model is correct the identification
process will be stoped. Otherwise the, identification process b} Emprical model
will be repeated by selecting new exparimental data, new 6. What is system identification?
model structure (complex structure if re~uired), new
System identification is the field of. modeling dynamic process from
estimation methods etc.
experimental data. The model obtained using system identification
technique is known as emprical model. Unmeasured dynamics and
TWO MARKS QUESTIONS AND. ANSWER~ uncertainties are difficult EO handle using mathematical model. They
1. What is model? have to be handled using emprical models.

A model servers as a good mathematical substitute for the process,

A model usually consists of a set of .differential or. difference and
algebric equations. Simulations using models are very economical,
safe and powerful substitutes for experiment using model. It is safe
to study response of a process un~er faulty conditions.
2. What are the advantages of model?
To improve understanding of the process.
To optimize process design and 'hence operating conditions.
To design a' controller. for the process.
To train oprating personal.
To plan and schedule production.
To obtain inferential estimate of a physical quantity.

. i
 CHAPTER 4.2 Computer Control of Processes )

2. Frequency Analysis,
The input is sinusoidal. For a linear system the output will also be
sinusoidal at steady state. The change in amplitude and' phase will give
the frequency response (Example: bode plot). From the frequency response
the model will be estimated.

NON PARAMETRIC, METHODS OF Sinusoidal input Sinusoidal output

SYSTEM IDENTIFICATION a siruet) b sin(rot + tP)

Fig. 4.2: Process with sinusoidal Input and Output

4.1 INTRODUCTION 3. Correlation Analysis
This chapter describes four different non parametric methods for The weighting function h(K) is used in this method to model the
system identification. Non-parametric identification methods are process.
characterised by the property that the resulting models arc curves or
functions whichare not necessarly paramateri~zed by a finite dimensional 00 ,

parameter vector. In this, method estimated parameters will have some yet) Lh(K),u(t-'K) -+- vet) ... (4.1)
, physical insight of the process. K=O

, Thefour different nonparametric methods for system identification are Where utt) Input '

1. Transient Analysis yet) Output

In transient analysis, input is applied as a. step or impulse and . h(K) = Weighting sequence
recorded output helps to idendify model. Fig. 4.1 shows the transient
Vet) =' disturbance
response of a first order system with delay
.
(~lJto
"Cs +
a unit step input.
The input is a white noise. A normalized cross-covariance \function
From Fig. 4.1, we can easily' estimate model parameters such as steady between output and input will provide an estimate of the weighting
state gain K, time constant t and delay time td. : function h(K).
y
Spectral Analysis:

KI-----~--~~-
Equation (4.1) is used as a model in spectral analysis. The model can
be estimated for.arbitrary inputs by dividing the cross-spectrum (between
output and Input) to the input .spectrum.

4.2 TRANSIENT ANALYSIS

,~ In this method, the model is estimated from step response or impulse
Fig. 4.1: Responseof a first ordersystem with delay (G (s) = _K_
rs +1
e- ld S
Jto a unitstep response of the process. In this section, how a model can be estimated
from step response using transient analysis technique is explained.
 4.4 Computer 'Control,ofPro,cesse~
( Non Parametric Methods of System Identification 4.3
Identifying Second Order Model
Iedntifying First-Order-Dead'!'Time (FODT) model
Let us consider a'second order 'system which is described as a
Let us consider a FODT system which is described as transfer transfer function model as shown' below
function model ' as shown below.
Y(s)
Y(s) K ... (4.2)
U(s)
= 9"(s) = --e
Ts + 1
-St
... (4.2) U(s)

Where,Y(s) Laplace Transform of the ~utput Signal Yet) Where, yes) = Laplace Transform of the Output Signal yet)
Ll(s) = Laplace Transform 'of the, Input signal urt) Uts) = Laplace Transform of the Input signal uit)
G(s) Transfer function of FODT system G(s) Transfer function of second order. system
T Time constant K = Steady state gain
r' = Dead time =
0)0 Undamped natural' frequency
K = Steady state gain o= Damping ratio
Apply a, unit step as a input to FODT system as shown in Figure 4.'3 Apply a unit step asa input to second o~der system as shown in
and obtain a step response which is shown in Figure 4.4. Figure 4.5 and obtain a step response which is shown in Figure 4.6.
FOOT system
'2
K Output
Unit step input Kmo Output
Unit step input ""'"tS
---e 2 2
T8'+ 1 u(t} .8 + 2omoS + 0)0 Y(t) .
u(t) Y(t) ,

Fig. 4.3: Unit step Inpu.t to FOOT system Fig. 4.5: Un.it step Il1put to second order system

Figure 4.4 demonstrates a ,grap.hical method for determining the

Y(t)
FODT paramters K, T, and t from step response.
First Maxima
The gain K is given by the final value of the response. By drawing
Second Maxima'
the steepest .tangent, T and t can be obtained as shown in Figure 4.4. The
tangent crosses the t-axis at t = 'to
I
Y(t)
KI----I----I----\------I----I-----lr---
I
I
I
I
KI---------~=-""-- I
I
: First minima :

Flg. 4.6: Step response of a second order system to a' unit s~ep input'

Fig. 4.4: Step response of FODTsystem to a unit step ~nput

(Non <Parametri:cMetbods of So/stem tdentification (4.5 ) 4.6.J Computer Control of Proces~!!J

The gain K is given by final value (after'convergence) as shown in Apply a following sinusoidal Input utt) to a system (described in (4.9)
-Figure 4.6. "I'he maxima and Minima of the step response occur- at times. as shown in Figure 4.7.
ri(t) a sin (rot) ... (4.10)
... (4.4)
Where a -amplitude of the sinusoidal inpututt)
co = Frequency of the sinusoidal input utt) in rad/sec
n = 1, 2, ... ~
Sinusoidal input .Sinusoidal output
and that ... (4.5)
u(t) Y(t)
where the overshoot M is given by
Fig. 4.-7: SinusoidallnputtoG{s)
~S1t ] If the system G(s) is asymtotically stable, then the output y (t) is
M = exp [ ~1- 32 ... (4.6)
'also a sinusoidal signal.
yet) b sin (cot + <p) ... (4.11)
Note: First maxima occurs at n = 1, First minima' occurs -at n = 2,
second maxima occurs at n= 3 and so on.. Where, b = amplitude of sinudsoidal Output yet)

From the step response as shown in Figure 4.6, the first peak time <P Phase diff~rence between Input u(t} and output
tl and first peak overshootM can be determined from (4.5). yet)' (Shown in Figure 4.8)

-logM Y(t)
s = 1
... (4.7)
/" . . , '.r: Y(t)
[1[2 + (log M)2 J2 I 'f .
\
\
Then roo can be- determined from (4.4) at tl \
\

7t I
(since n =. 1) ... (4.8) J
t 1 ~1-o2 I
I
u(t) I
I
From (4.7) and (4.8) the parameters S and roo can be determined. \ I
\ I
' ......... /

4.3 FREQUENCY ANALYSIS

Fig. 4.8: Input and Output waveforms of G(s)
For a frequency analysis, it is .convenient to use following system
with transfer function model as From(4~11), we can write
yes) = G(s) U(s) ., (4.9) b = a l Gtjwl] .. (4.12a)
Where, yes) = Laplace Transformof Output signal Yet) <p arg I G(jw) I ... (4.12b)
Uis) = Laplace Transform of Input Signal utt)
Gts) = Transform function of a system
 { 4.8 ComputerControl of'Processes )
(Non Parametric MethodsefSystem Identification [ 4.7]
Since we can represent
This can be proved as follows. Assume the system is initially at rest.
G(iro) re ie
Then the system Gts) can be represented using 'a weighting function htt)
as' follows. ~here, .r = magnitude of G(iro) IG(iro) I
t
O
e = argument of G(iro) eiargG(ico)

Y (t) = fh(t)u(t <'I:)d'l: - ... (4.13)

o
where h(t) is the function whose laplace transform equals G(s).
t
(.~. IG(iro) I = I G(-iro) I = I G(iro) I) .
:. G(s) = fh('I:)e-sTd'l: ... (4.14)
o
(e irot _ e-i rot)
Since sinrot = 2i
... (4.15)
;i.IG(iOJ)lsin(rot + argG{iro)) ~
Equations (4.10) (4.13) (4.14) and (4.15)
t
yet} fh('I:)u(t - 'I:)d'l:
o ' " ~sin[rot ~]
b '
+
~

t
= b sin (rot + ~) ... (4.16)
fh( '1:) a sin (ro(t-'l:))d'l:
o From above equation (4.11) and (4.12) are proved.
t (eiro(t-T) _ e-iro(t-T))
By measuring the amplitudes a and b as well as the phase difference ~,
= fh('I:)a 2i dt one can draw a bode plot (or Nyquist 'or equivalent plot) for, different co
o values. From the bode plot, one can easily estimate the transfer function

;i t
fh( '1:)( eiro(t":'-r) - e-iro(t-T))d'l: .
model Gts) of a system.

o 4.4 CORRELATION ANALYSIS

= ;i t
f h ( '1:)(eirot e-iroT - e-irot e+iroT ) dr The form of model used in correlation analysis is
o
OJ

yet) Lh(K)u(t-K)+ v,(-t) ... (4.17)

K=O

= ;i t t
e irot Jh(1:)e(-iroT)d1: - e-irot Jh( '1:) e-(-irot)d'l: Where . Yet) = Output, Signal
ult) Input Signal
o
'---y-------I
0
"-------y-----
G(iro) G(-iro) ,h(K) = Weighting sequence
vet) = Disturbance term
 ( 4.10] Computer Control of Processes )
( Non ParametricMethods of System Identification 4.9
't = 0,1,2, ...
V(t)

Where
r--~-------------- --~ N = Number of experimental data
I I
i I 00 I
nput :.-.... L h(K)u (t-K) + + ~.""'11_ ...Output r ('t) =
Yu Estimated cross covariance function between Y(t)
u{t) I K=O I Y(t)
I i
and utt) from input and Output data
~--------------------~ ru ('t) = Estimated covariance function of urt) from Input
System

Fig. 4.9: System used in cerrelefien analysis data

Assume that in putut t) isa stationary stochastic process (white noise)

which is independent of the disturbance vet). Then the following relation
Then an estimate {h(K) L~9f the weighting function (h(K)} can be
holds for the cross covariance function determined by solving followingequation
00

ryu('t)= Lh(K)ru(~-K) ... (4.18) ~--';.: (4.20)

K=O

Where The above equation represent a linear system of infinite dimension.

ryu('t) = Cross covariance function between output yet) a~d input u(t)
/\

= EY (t + 't) u (t) ru (0)

~~O)""J'
/\

ryu (0) /\

ru('t) = .Covariance function of input u(t) ru (1) ou (4.21)

= Eu (t + t)u (t) /\
h(oo)
ryu (co) /\

Note: E is called expected value = meanvalue r u (00)

Conduct an experiment and collect Input u(t) and Output Yet) dat~.
Solving above infinite dimensional equation is very difficult. This
The covariance functions in (4.18) can be estimated' from the Input and
problem can be simplified by using white noise (mean value =0, variance
Output data as
.0 = 1) as input.
2

1 N-max(t,o)
= N . L.
Y(tH)U(t)
t:=l-min(~,o)
r =0, 1, 2, ... ... (4.19)
For white noise,
rue 't) = 0 for t =J, 2, ...
r u (r) = Constant value for t = 9 (i.e., non zero value)
1 N-t
= N LU(tH)U(t) r u (0)= Constant value
t=1.
(Non, Parametric Methodsof System Identification ( 4.11 ) ( 4.12 J ComputerControl; of Processes)

.Si~ce input is awhite noise,the equation (4.20) can be written as <!>Yu(CO) ... (4.24)
Where
h(K) = ryu (K) K=O, 1, ... ... (4.22)
ru (0) Cross spectral density between Input u'(t) and
Output yet)

ryu (0)
(0)
. ru
'rh(O) ryn.(l)
h(1) <Pu (00) Spectral density of input u (t)
ru'(O)
,lh(~) ... (4.25)

H (e-iro) = Transfer funciton

, ,

From the above. equation, the weighting function {h(K)} (i.e, h(O), hf l), ... ,
L h (K)e
co

-iKro
h( 00) can be easily estimated.
K=O

4.5 SPECTRAL ANALYS'IS Now the transfer function H(e-iro) can be estimated from (4.24) as
'Theform of model used in spectral' analysis is
co

yet) Lh(K)u(t'-K) + v('t) ... (4.23)

K=O
Conduct an experiment andcollect'N' number of Input u(t) and
Where,
Output yet) data. Then cross spectral density <PYu (co) and spectral density
Yet) Output Signal <Pu (00) can be estimated from experimental data as
u(t) Input Signal.
. h(t) = Weighting sequence ... (4.27)
vet) Disturbance term
1
The system described in (4.23) is shown in Figure 4.9
Taking discrete fourier transforms for the system described in (4.23)~
/\

<Pu (co) = -
21t-r=_N
L ru{'t)e-
N /\ .
r rm
_
... (4.28)

the following relation for the spectral densities can bederi~ed'from cross
Using (4.19), equation (4.27) can be written as
covariance function described in (4.18).
, N N-max(t,o)-
$yU (00) = _1_.
21tN
L L N t=l-IDln
' . (-r,O )
Y(t+ -r)u(t)e-itro
,
... (4.29)
t="-
( Non ParametricMethods of SystemIdentification,
( 4.14 ) Computer Control of Proc~
[ 4.13 )
Similarly, .
Next make the substitution s= t + 'to Figure 4.10 illustrates how to
derive the limits for the new summation index.
1 N N . . "
Since --~" u(s)U(t)e-lSIDe-ltco
21tN LJL...J
s=.i t=l

We get,
... (4.31)

By computing UN(m) by using FFT algo..: ithm, we can estimate

spectral density ~u ( ill).

Then we can estimate transfer function

-N
Fig. 4.10: Change of summation variables
(summetien is over the 'shaded area)

... (4.30)
... (4.32)
Where,
N
YN(ro) = Ly(s)e- isro = 'Discrete fourier transformofylt) This transfer function is called empirical transfer function estimate.
s=1

N .
UN(ro) = L u(s)e- isro
= Discrete fourier transform of net)
5=1

21t 41t ';

For ro = 0, - , - ... , 1, YN(ro) and UN(O) can be computed efficiently
N N . -
using Fast Fourier Transform (FFT) algorithms. Then using (4.30), we
. /\
can estimate cross spectral density $y~ ( ro).
( Non Parametric Methods of System Identification ( 4.15] ComputerControlof Processes )

TWO MARKS QUESTIONS AND ANSWERS 4. What is correlation analysis method for system identification?

1. What is .non parametr-ic method for system identification? The model used in correlation analysis is

Nonparametric identification methods ate characterised by the 00

. property that the resulting models are curves or functions,which yet) Lh(K)u(t-K) + v(t)
K=O
are not necessarily parametrised by a finite dimensional parameter
vector. The four different non parametric methods for system Where, Y(t) = Output Signal
identification are net) = Input Signal
1. Transient Analysis h(K) = Weighting sequence
2. Frequency 'Analysis vet) = Disturbance term
3. Correlational Analysis The input u(t) is a white noise signal, a normalised.crossconvariance
function between 'output and input will provide an estimate of the
4. Spectral Analysis
weighting function h(K).
2. What is Transient analysis merhodfor system identification?
5. What is spectral analysis' method for system Identdficatdon?
, In this method the input is taken as a step or impulse' and the
The model used in spectral analysis is
recorded output (example transient response ora FODT system to a
unit step input) helps to Identify a model. C()

yet) Lh(K)u(t-K) + v(t)

y K=O

K I - - - - - - - -........- -Where, yet) = Output Signal

u(t) Input Signal

'. o
<>
Ir
I' "
Jransient
response to
.~'J a.unit step}nput
h(t)
vet)
Weighting sequence
Disturbance term
J .' t.
The model can be estimated for arbitarary inputs by dividing the
j~ ~~I~ td ~I
0
cross spectrum (between output and input) to the input spectrum.
.' 'The above transient response ~(~Qorded datal-helps to find out FOl::>T
system parameter~ s~c~ asK" tJand td. e, (,

s. What Is Frequency analysis metil0d fpr ~ystem identification?

In this method Input.rto a syst'eni tt~ b~ :i,dcen~ified) i~ 'sinusoidal.
For a lin~ar system the output wil,l also be sinusoidal at-steadyst~te.
The change in amplitude and phase will give the .frequency response
(example bode plot)." From the frequency .response,. the model will
. be estimated. . 1,'
 CHAPTER
5.2 J Computer Control of Processes J
IExample 5.1 I
5 .Consider a following first order linear discrete model
Y{ t) + aye t - 1) = b u (t - 1) ... (5.2)

The model represented in (5.2) can be written in linear regression

model as follows
PARAMETRIC METHODS OF yet) - ay (t - 1) + bu (t - 1)

,SYSTEM IDENTIFICATION [-y(t - 1) u(t - 1)J[:]

<l>T (t) 0 ... (5.3)

5.1 INTRODUCTION Where', <I>(t) [-y(t -1) .u(t - l)J'

A parametric method can be characterised as a maping from the
recorded data to the estimated ,parameter vector. T~~ estimated
parameters donot have any physical insight of,the process. The various
parametric methods of system identification are
The elements' of $(t) are often called regression v ar i ables or .:
1. Least squares (LS) estimate regressors while y( t) is called the regressed variable. The e is called" ,
. 2. redictionerror method (PEM) parameter vector. The variable t takes integer values..
3.. Instrumental variable (IV) method
IExample 5.2 I '<"~~.:,; " v
5.2 LEAST "SQUARES ESTI,MATION Consider a truncated weighting fuii~~Oft~"'in~deL, t

The method of least squares is about estimating parameters by Yet) == h o u(t) + hI u (t - 1) + .. ~~+Jlrti~l u(t - M + 1)
minimizing the squared error between observed data .a~d their expected values. The input. signal ut t ) ... u(t ~ '.M + .1) are recor~ed during thee
The linear regression is' the simple type of, parametric model. experiment. Hence. regression variables
This model structure can be written as
~ (t) '=, (urt) u(t - ,1) ... u (t'- M + 1) is a )\1 - vector of
Yet) = ~T (t) e ... (5.1)
.known quahtitie~. and, para~eter vector.
Where,
'0 ~ 'eho hI' .... '- 'hni_1)T is a lVI-Ve~tor of unknown
Yet) = Measurable Quantity " parameters to be estin.1at-ed.
4>(t) = n - Vector of known quantities
[~Y(t-l), -yet - 2), ... ,.-Y(t - n a) utt - 1) ... -u(t- nb)]T
The .problem: is to. find an tistiJl}.ate e of the parameter vector e as
shown in (5.1) from experimental measurements Yi I), ~(1), Y(2), $(2), ...
e n . Vector . of unknonwn parameters Y (N) ' ~(N). Here 'N' represents number of .experimenta~ data and 'n'
The following two examples shows how a model can be represented represents number' of known quantities in <p(t) or nu.mberof unknown
using linear regression" model" form; parameters in O.
(Parametric Methods of System'Identification 5.4 Computer Control of Processes"]
5.3
and stack these in a vector E defined as
Given experimental measurements, a system of linear equations is
obtained as
Y(l) = <l>T (1) o
Y(2) = <f>T(2) 8

In satisticalliterature the equation errors are often called residuals.

YeN) = <f>T (N) e The least square estimate of e is defined as the vector ethat minimizes
This can be written in matrix notation as the loss function
Y = <I>8 ... (5.4)

Where~ .... (5."7)

Note:
an (N x 1) vector ." (5.5a)
The other form of loss functions are

... (5.8a)

an (N x n) matrix ... (5.5b) "

... (5.8b)

One way to find e from (5.4) would of course be to choose the number where IHI denotes the Euclidean vector norm.
of measurements, N to be equal to n. Then <I> become's a square matrix. If
this matrix is non singular the linear system of equations (5..4) could easily Thee will be estimated from experimental measurementsYt I), p(l);,
be solved for O. In practice, however noise, disturbances and model misfit Y(2), ~(2) ... , Y (N), ~(n) by minimizing the loss function V(S) in OJ.7) and
are good reasons for using a number of experimental data 'N' greater than (5.6). The solution to this .optimized problem is "
en'. With the possible to get an improved estimate. When N> n,exact
solution for linear system of equations (5.4) in general not exist.. ... (5.9)

To find an estimate e, introduce the equation of errors as For this solution, the minimum value of Vee) is "

e(t) = Y (t) <l>T (t) e ... (5.6)

-J, t
observed expected
... (5.10)
value value
(Parametr~c MethodsofSystem Identification [ 5.5' 5.6 Computer'Control of Processes J

Note: Disturbances
1. The matrix $Tep is positive definite.
2. The form (5.9) of the least squares' estimate can be rewritten in
the equivalent form

... (5.11)

-In many casesott) is known as a function oft. Then (5.11) might be

Estimated
easier to implement than (5.9)' since the matrix cD of large dimension is parameters
not needed in (5.11). Also the form (5.f1) is the starting point in deriving
several recursive estimates. Recursive'
identifier

5~3 RECURSIVE IDENTIFICATION METHOD Fig. 5.1: A general scheme for adaptive control
In recursive (also called on-line) identification methods, the Most adaptive systems, for example adaptive control systems as '
parameterestimates are computed' recursively in time. This means that shown in Figure 5.1 are based (explicity or implicity) on recursive
,,'.'" _:';"'. C'-'.;~<'",' ' , ' 1\

if' thereis an e-stimate"6(t-l) based on data upto time (t - 1), then '9(t). is identification.
, 1 \ '
computed by sOll1e':(simp!~ modification' of_,9(~-1). Then a current estimated model of the process is available at all
times. This time varying model, is used to determine the parameters' of
The counter parts to one-line methods are the so-called off-line or the (also time-varying) regulator (also called controller).
batch 'methods, in which' all the recorded data are 'u~ed simultaneously to
find the parameter estimates. In this way the regulator will be dependent on the previous' behaviour
of the process (through the information flow : Process ~ model -7'
Recursive identification methods have the following .general features:
regulator).
They are central part of adaptive systems (used, for example,. for If an appripricate principle is used to design the regulator, then,
control or signal processing) wherethe action is based on ,the most
the regulator should adopt to the changing characteristics of the
recent model.
.process.
Their requirement on primary memory is quite less compare to offline
The various identification methods are
identification methods' which, require large memory.tostore entire data,
a) .Recursivc least squares method
They c~n be easily 'modified into real-time algorithms" aimed at
tracking time-varying parameters. b) Real time identification method

, They can be first step in a fault detection algorithm, which is used c) Recursive instrumental variable method
to find out whether the system has changed significantly. ' d) Recursive prediction error method
( Parametric Methods of System.Identification 5.7 5.8 ) Computer Control of Processes )

Then using (5.15), (5.14) can be written as

5.4 RECURSIVE LEAST SOUARES'ESTIMATIDN
T'he linear time-varient system can be represented as
... (5.17)
'A (q-l) yet) = B (q-l) u (t) + (t) ... (5.12)
Notes
Where, A (q-l) = 1 + alq-l + . '." + ana q-na
If we replace t by (t - 1) in (5.17), we get
B (q-I) = hI q-l- + ... + bnb q-nb
E(t) = Equation error S(t-l) = . P(t-l)[~Hs)y(s)J
This model can be expressed in 'regression model form as
yet) = <l>T (t) e .+' e (t) ... (5.13) t- l ]
e(t-l)p-1(t-l) = [
f;Hs)Y(s) ... (5.18)
Where, <I>(t) [-yet -1), ... , -yet ~ na)u(t -1) ... ~u(t - nb)J
The equation (5-.17) can be rewritten as

_Then the least squares parameter estimate is given by S(t) = P(t)[~HS)Y(S) + Ht)Y(t)] ... (5.19)

By substituting (5.18) in (5.19), we get

The argument t has been used to stress the dependence ofe on time.
The (5.14) can be computed in a recursive fashion. By substituting (5.16)

Introduce the notation

e(t) = P(t>[{p-l (t) - ~(t)+T (t)}e(t-l) + Ht)y(t)]

P(t) = [t~(SHT(S)r ... (5.1'5)

- P(t)[p- 1(t)6(t-l) - ~(tHT (t)6(t-l) + Ht)y(t)]
/\ ' / \

9(t-l) -P(t)q,(t)<t>T-(t}8(t-l} + P(t}q,(t)y(t)

p-l(t) [t~(SHT(S)]
= 9(t-l) + P(t)+(t)[y(t) - <l>T (t)e(t-l)] ...(5.20)

p-l (t) [~~(SHT(S) + HtHT(t)] Thus (5.20) can be written as

/\ /\

9(t) = 9(t-l) + K(t)e(t) ~ .. (5.21a)

p-l (t) = p-l (t - 1) + <p(t) <pT(t) K(t) = p. (t) 4 (t) ... (5.21b)
:. p-l (t ~ 1) = p-l (t). - <pCt) <l>TCt) ... (5.16) E (t) = y(t) ~ <t>T (t)e(t-l) ... (5.21c)
 ~ Compute,r Control of Proc-ess~
(Parametric Methods ofSystem, Identification ( 5.9 )

Here the term E(t) should be interpreted as 'a prediction error. It is

TWO MARKSOUESTIDNS AND ANSWERS
the difference between the measured output yet) and the one-step-ahead L Define parametric method of system tdenrifieatdon.

=~T (t)-8(t-1) ofy(t) made at time t -1 based

A parametric method can be characterised as a m aping from the
prediction ;(t I t-1;8(t-1)) recorded data to the estimated parameter vector. The estimated
on, the model corresponding to the estimate- e(t - 1). If E(
t) is small, the
parameters do not have any physical insight of the process. The
various parametric 'methods of system identification are
/\
estimate. 9 (t -1) .is 'good' and should not be modified very much. The vector
1. Least squres (LS) estimate
,K(t) in (5.21b) should be interprected as a weighting or gain factor showing
how much the value of E(t) will modify the different, elements of the 2. Predication error method (PEM)
parameter vector. 3. Instrumental variable (IV) method
To complete the algorithm, (5.16) must be used to compute P(t) which 2.. Define Least squares estimati..>nmethod.
is needed in (5.21b). However, the use of (5.16) needs a matrix inversion
at each time step, This would be time consuming procedure. .Using matrix The LS estimate is to find 8 of the parameter vector e in a regression
inversion lemma, however (5.16) can be rewritten in updating equation model yet) = T (t") 8 from experimental measurements by minimizing
form as the loss function.
T(t)P(t-1)
P(t-l) _ P(t-1H(tH
... (5.22)
P(t) = [ 1 + ~T (t) P(t -1H (t)]

Note that in (5.22) there is now a scalar 'division (scalar inversion) The LS estimate of parameter vector q is
instead of matrix inversion. From (5.21b) and (5~22),

K(t) = .P(t-:1H(t) _> (5.23)

[1 + ~T (t)P.(t-1)Jl(t) J
3. What are the advantages of LS method?
The recursive least squares algorithm (RLS) consists of
Primary 'tool for process modeling because of its effectiveness and
A A
completeness.
Set) =9(t ~ 1) +K(t) +E(t)
SinceLS method uses data very effectively, good results can be,
get) = y(t}- <j)T(t)8(t-l) obtained with reletively small data sets.
Easy to constant statiscal techniques for prediction, calibration
T and optimizations.
P(t-1)- P(t-1H(tH (t)P(t-l)
P(t)
, [1.+ ~T (t)'P(t -l)~(t)] .. Simple to use.

= P(t-l)~(t)
K(t)
[1 + <j)T(t)P(t-1)~(t)J
(Parametric Methods of System Identification' (5.11 ) Computer Control of Processes )

4. Where the limitations of LS method? The various recursive identification methods are
The LS method gives consistent .parameter estimates only under 1. Recursive least squares method
following restriye conditions,
2. Real time identification method
If E<I>(t) <l>T(t) is non singular
3. Recursive instrumental variable. method
If E <!>(t) v(t) =0 4.' Recursive prediction error method
If .input is persistently exciting of order nb, ,7. What are the advantages of recursive identification method? .
, .

v(t) should be, white noise.

They are central part of adaptive system
5. Differentiate par-ametr-ic and non parametric method of system Requires less primarymemory
identification. ' -:
They can be easily modified into real-time algorithms, aimed at
'Parametric Method Non-parametric Method tracking time-varying parameters.
a
1. 'Characterised as', maping Characterised by the property '. Used in fault detection algor-ithm. '
from recorded data to the that the resulting mo.dels are
8. Write down' recursive least square (RLS) algorithm.
estimated parameter vector. curves or functions which are
not necessarly parametrized. I
The RLS algorithm consists of
A A
2. Parameters does not have Parameters have physical 9(t -1) + K(t) E(t)
e(t)
any physical insight of the' insight of the process.
process. Example: 't - time constant
.-
'f;{t) = y(t)~ <pT (t)e(t-1)
3. Different methods are Different methods are

Least square estimate Treansient Analysis P(t-'-l) _ P(t-lH(tHT(t)P(t-l)

P(t)
[1+ +'r (t) P (t -1) <P (t)]
Prediction error method Fraquency Analysis

Instrument variable method

1.oo--_'" -
Correlation Analysis
K(t)
P(t-1)<J>(t) ,
[ 1 + <I>T (t) P ( t -1) <p ( t)]
6. Define, recursive identification method.
Inrecrusive (also called on-line) identification' methods';
the parameter estimates are computed r ec ur s ively in time .
. This means tha~ if there is an estimate 6(t -1) based on data upto
time (t ~ 1) then 8(t) is computed by some 'Simple modification'of e(t -1).
This method is a central part of adaptive systems.
 CHAPTER Q.2'J Computer Control of Processes)

Now take t.he laplace transform of equation (6.5).

f (f*(t)} = {f(O) so + i!{f(T) 8(t-T) + i!{fC~T) a(t -2T)} + ...
rco) f {St} + f(T) e{(t - T)} + f(2T) e{St - 2T} + ... ... (6.6)
e{f*(t)} = f*(s) = reo) + f(rr) e-ST e {8(t)} + f(2T)e-S2T e {oCt)} + ...(6.7)
Since
'R {Set)} = 1; Equation (6.7) becomes
DIGITAL CONTROLLER, DESIGN
L f (nt) e
00

e{f*( t)} = f*( S) = -nST

n=O

&.1 ,REVIEW OF Z-TRANSFORMATION If we introduce a new variable Z = eST in equation (6. 7) we get
co
"We found laplace tr ansforrns to be helpful in' the analysis of
conventional - control systems. In subsequent chapters we will see that f*(s)lz~eST = Lf(nT)e- nST .. (6.8)
n=O . '
theanalysis of sampled-data control problems is conveniently handled in
terms ofZ-transforms. This result is defined to 'the Zvtransform of f(t) and is denoted as
F(z). Thus
In this chapter we shall develop some of the important concepts of
Z-transforms. In this .chapter we shall develop some of the important 00

concepts of Z-transforms. To begin, recall that the laplace transform of 'a F(Z) = z{f(t)}= f~(s)lz=eST = Lf(nT)z-n ... (6.9)
function f(t) is defined as . n=O

a:>

. L{f(t)} = F(s) = ff(t)e-stdt ... (6.1)

6.2 la TRANSFORM OF VARIOUS FUNCilDN
' 0 Unit step Function
Also recall that the sampled function f* (t) and the continuous 0 t-cO]
function f(t).are related by f(t)
{ U(t) t~oJ ... (6.10)

0Cl
By definition
f*(t) = f(t) L8(t-nT) ... (6.2) .

L U(nT)z-n
C1J
o
Z{U(t)} = F(Z) = ... (6.11)
00 n=O
f*(t) = Lf(nT)8(t.... nT) 00

n=O ),Z-n = 1+Z-1+Z-~+ ... _, 1 ... (6.12)

k-J l-z-1
= reO) 8(t~ + f(T) B (t - T) + f(2T) () (t - .2T) + ... ... (6.S) n==O

where f(T) o(t - T) represents an impulse at time IT whose area is 1 Z

Z{U(t)} = l-Z-1
Z-l
... (6~13)
f(t).
(Digital Controller Design 6.4 Computer Control of Proces~es J
Exponential Function Sin.e Function

t<Ol
... (6.14)
f(t)
{Si:at t z OJ
... 0 (6.20)

f(t) sin at, f(nT) = sin anT

00

Z{e- at} =. Le-anT Z-n ... (6.15) e jnat _ e- jnaT

and sin ant =
n=O 2j
CXI

= L(e-aTZ-1)n, Z{sin at}

O'J

Lsin anT Z-n

n=O n=O

.. ~ (6.16)

l'
IZ-l! < (6.17)
~.
J n=O
f( e JDaT _e-janT)z-n .0. (6.20)
F(Z} = 1- e- Z-l 'at eaT

Ramp Function 1(
2j 1- e j aT z'
1
l-e-~T z' J
f(t) ... (6.18)
jaT
[ 1- e Z-l .,.. 1+ e jaT Z-l ]
=:
(e
1 + Z-2 - Z-l j aT +e -jaT)'
F(Z) = Z{KT} :::: 'IKnTz-,n
n=O .
z' (e jaT _. ~jaT)
_ KT {Z-l +2 Z-2 + ... }
2j
KT Z-l {l + 2Z- 1 + 3Z- 2 + 4Z- 3 +...} ..' (6.18)
1+ Z-2_ 2Z- 1 (jaT
e ~e -jaT)
KTZ-1
=
(l-Z~l t Z-l sin aT
= 1 + Z-2 - 2~-lcos aT

Z-l Biua~
Z{sin at} .00 (6.21)
1 - 2ZcosaT + Z-i
F(Z) =KTZ ... (6.19)
(Z_1)2 ZsinaT
Zlsin at} ... (6.2,2)
Z'-2 - 2ZcosaT + 1
(Digital Controller Design 6.. 5 6.6 ) ComputerControlof Processes ]

6.3 THE INVERSE z.. TRANSFORM 0.5(0.52 + 0.5 +1)

A=
Three methods are outlined here for inverting the z-transform: (0.5-1)(0.5-0.5- jO.74)(O.5-0.5+ jO.74)
1. Partial fraction Expansion
2
A = 0.5(0.25 +0.5+1) _ 0.5{1.75)
2. Long Division
(-0.5)(0.5476) - 0.2738
3. Tables of Z-Transform
The operation of inverting the Z - transform is denoted as A = 3.19
f*(t) = z-l {F(Z)} ... (6.23)
. Z2(Z2 + Z + 1) , I
Note that inverse yields the sampled function f*(t) and not the B = jO.74)(Z-(O.5~ at Z,=J.
continuous function f(t). The f*(t) so obtained is unique, but it is conceived I(Z-0.5)(Z-(O.5+ jO.74)
that sample functionf*(t) could be derived from two different continuous B 7.5
function f*(t) could. be .derived from two different continuous functions
fl(t) and f 2 ( t ) . It follows, therefore that the .inverae of F(Z) does not
necessarely yield aunique continuous function F(t).
Thus we, should not expect any additional information about f(t) from
the inverse other than the values at sampling instants. c = -1.67 + j 0.52

IExample 6.1 1 Similarly

Find the inverse Z transform of D= -1.67-jO.52

F(Z). = 7.5 3.18-1.67 + 0.52j -1.67 ~ 0.52j

FeZ) = Z2 (Z2 +Z+l) ------- + .
Z Z - l Z - O.5Z ~ (0.5 + j 0.74) Z - (0.5 - j 0.74)
(Z-O.5){Z-1}(Z2 ~Z+O.8)
... (6.25)

F(Z) _ Z(Z2 +Z+1) Now, recognize that the presence of areal number other than one in
... (6.24) the denomination of the Z-transform expension willgive rise to an
Z - (Z-O.5)(Z-1)(Z-(O.5+ jO.74))(Z-(O.5- jO.74))
exponential function, in its inverse.

F(Z) _ _A_+~+ . C + D Therefore, we must express the number 0.5 in exponential form.
Z - Z-O.5 Z-l .Z-(O~5+.jO.74) Z-(O.5-jO.74) Similarly, complexnumber give rise to exponential and sine/cosine terms.
Thus, by converting from rectangular to polar coordinates.

A = Z(Z2+Z+1) I e-O.693 = 0.5

(Z-1)(Z-(O.5+jO.74))(Z-(O.5~ jO.74)}. at Z = 0.5 e-O.ll 0.98 j 0.5 O.74j
 ( Digital Controller Design 6.8 Computer Control of Processes )

6.1.1 MODIFIED Z-TRANSFORM

z Z
7.5 Z -1 - 3.18 _ e-O.693 The Z transform method enables us to determine the transient'
Z
response of sample data control systems only at sampling instants. To
, Z
:. F(z) = +(-1.67-0.52) Z _ e-O.ll+0.98j ... (6.26) obtain the value of response between sampling instants, modified
Z-transforms are used. Modified z-transform mainly useful in .analysing
+( '-1.67 -O.52j) Z _ e-:'11~O.98j sample- data' control systems, containing transportationlag (dead time).

f*(t) = 7.5 - 3.18 e-O.693t1T

6.4.1 Evalution of Modified Z-Transform
+ (-1.67 + 0.52 j) e- CO 11- O.98j) Suppose that the transfer function of a process with dead time is
-(1.67 + O.52j) e-(O11 + j O.~8)tlT ... (6.27) respresented by the following expression
Which reduces to Gp(S)= G(s) e-6dS 40. (6.31)
f*(t) = 7.5 - 3.18 e"':O.693t/T - 2 e-:-o1 ltlT where g(s) contains no dead time, and
{1.67 cos (O.98t/T) + 0.52 sin (O.98t/T)} ... (6.28)
8d= dead time
Several values of the' function f(t) at the' sampling instants are
ad = NT + a ... (6.32)
t o T 2T ' 3T 4T 5T' 6T \ _N - Largest integer number of sampling intervals in 8d
T - Sampling period
f(t) 1 3.5 ' 7 9.3 9.3 7.6 6.1
Equation 6.31 becomes,
Another method of inverting F(z) is by long division. By this method Gp(s) = G(s) e-(NT + 9) s ... (6.33)
F(z) is expanded into a power series of z-l by long, division. This can be
seen easily if we write the equation that defines the Z-transform' of a For example if ad = 1.5 and T ~ ~.5 secN ::: 2 and 8 value-is
function as ... (6.34)
00 1.5 2 x 0.5 + 8
F(Z) = Z{f(t)} = Lf(nT)z-n ... (6.,29)
e 0.5
n=O .
Now, let us take the z-transform of equation 6.33
F(Z) =' f( 0) + f(T) Z-l + f(2T)Z-2 + ... ... (6.30)
_Gp(s)= Z{G(s) e(-NT + e)S} ... (6.35)
If fez) is expanded into a power series, we. have F(z) = ao +al Z-l
'.+ a2 Z~2 + a3 Z-3' + ....' Z{G(s} e-N'fS + G(s) e-aS}

'By comparison ao = f(O), al = f(T) and an = f(nT) which is the value of We .know that Z{e-~Ts}= Z-N
f(t) at the nth sampling instant. ,. Gp(Z) = Z-N Z{G(s) e-9S} ... (6.35)

, T4e quantity Z{G(s) e-6S} is defined as the modified Z-transforms of

G(s)and is denoted by Zm{G(s) or G(Z, m) thus,
G{Z, m} = Zm {G(s)} = Z{G(s) e-6S} (6.36)
(Digital ControllerDesign 6.9 ( 6.10 ) Computer Control of Process~~

6.4.2 Modified Z-Transforms of some Simple Function m = 1- -

e ... (6.40)
T'
1. Unit step function
Zm{e-at } = e~amT Z-l + e-amT e-aT Z-2 + e-amT e-2aT Z-3 + ...

Z-1 e-amT { 1 +Z-l e-at + ... + e-2aT Z-2 + ...

Z-l e- amT
Zm{e-at } = 1 - Z-l e-a
T ... (6.41)

Zm{F(s)} Zm{l/s} = {-as}

Z. ~
&.5 PULSE TRANSFER FUNCTIONS
Z{U (t --8)
In conventional control 'system the laplace transform 01 :n input
00 function Xes) is related to the laplace transform of the output, function
= IU(nT-e)z-n ... (6.37) yes) by the transfer function of the system G(s) according to the equation
n=O

Zm{F(s)} = 8 + U(T -8)Z-1 + U(2T - 8)Z-2 + U(3'r - 8) Z-3 + ... (6.37)

Y~s) = G(s) ... (6.42)
X~s)
= Z~~ {1+Z-1 + Z-2 + ... } Schematically the transfer function is represented as

Z-l 1 Input III ~(s) ~ Output

z, rue = Z-1 ... (6~38) ~~ . ~~~
1- Z:"'l
In sampled data system, we must relate the pulsed input to the pulsed
output.
By analog with the transfer function representation of the
1~ conventional control systems, it is tempting to postulate that the
Gts) = s+a Z transform of the pulsed input can, be related to the Z transform of the
pulsed output according to the expression .

{ I .} {e-as}
Z'-=z-
m s+a s+a
... (6.39)
y (Z) =
X(Z)
G(z) ... (6.43)

The term G(z) is called the pulse transfer function or Z transfer

function of the systems.
00

IU(nT-8)e-a (nT- a)
6.5..1 Complex Series Representation of the Sampler
n=O
Recallthat the sampling operation gives us the values of the output
=e-a (T - 9) Z-l + e-a (2T - 9) Z-2 + ... ... (6.39) at each sampling instant and can be represented as

mT = T-8 Continuous input .---1 Samples ~ Sampled output

([)igital coaitl'oIler tJesign [ 6.11 ) ( 6.12 ) ComputerControlof Processes )

We have said that the sampled output is related to continuous input

by the .expression
X* (t) = OT (t) X (t) ... (6.44) ... (6.49)
where,
00
Now, since
OT(t) = L o(t-nT) ... (6~45)

Since BT(t) is a periodic funetion, it ,can be expanded into a complex

1 ... (6.50)
Fourier 'series'

LC
00

OT( t) = n ejnwst ... (6.46)

. It follow that
n=-oo- .

Ws = Sampling frequency, in radians per unit time 1

... (6.51)
Cn = T
21t
Ws = T Substituting this result in Equation (6.46) given

4- f
, and the Fourier Coefficients are given by
OT(t) = ejnwsl ... (6.52)
n=-oo

... (6.47)
and X*(t)= 4f n=-'oo
X(t)~jnWsl ... (6.53)

Substituting for OT(t) from equation (6.45) in Equation (6.47) give

Taking Laplace transform of this equation we have
<,

~+! e-jnwstLtoo o(t-nT)ot

1
= -T L
00

en = X*(s)= l{X*(t)} X(s+jnws ) ... (6.54)

n=-co
.
2"
6.5.2 Development of Pulse Transfer Function
We wish to relate the sampled input X*(t) to the sampled output
'Y*(t) as shown in the 'following Figure.
1-------+
./'~*(t).
X(t) I Vet) v*(s)
. g(5) I ~
Xes) X*(s) ......J Yes)
(Digital Controller Design ( 6.13 ) ( 6.14 ) Computer COntr()I~ProtessesJ .

From this Figure By the same reason in which led to Equation (6.54), this can be written
as
yes) = x*(s) G(s) ... (6.55)
Y*(s) =, X* (s) G*(s) , ~ .. (6.63)
By analog to Equation (6.54)
Taking Z transform .of this Equation gives

... (6.56) . Y(Z) =X(Z) G(Z) , ... (6.64)

GCZ) = Y(Z) ... (6.65)

In the view of Equation (6.55) X(Z)
yes + jnw.) =. g(s + jnws)'X* (s + jnw.) ... (6.57) Note that similarly of Equation (6.65) to a familer transfer function
G(s) given in Equation, which would be the transfer function of the system
Substitution in Equation (6.56) yields
,if the sampler removed. A step-by-step procedure for developing pulse
. 1 transfer function is given below.
L Y(s+ jnws)
00 .

Y*(s)=' T .... (6.58)

1. Derive the tranfer function G(s) of the system by conventional
n=--oo
techniques.

~ I
n=-oo
X*(s+jnws)g(s+jnws) ... (6.59)
2. For X (t) = S(t) ( i.e., the impulse function)

G(s) =' yes) = yes) ... (6.66)

But because Xes)
Since L{S(t)} = 1
X*(s + jnw.) = f {X*(t) einwst }
30 Invert G(s) to get G(t) which is numerically equal to the impulse
= e{X* (t) ein 21ttlT} response y(t) .according to equation.
But t= n'I' gives. 4. Note thatfor X(t) = oCt), X(Z) = Z{o(t)}. Therefore
X*(s +)nw s ) = . .e {X*(t) ei 21tn 2 } .: (6.60)

Since G(z) = ~~~j =y(z) .. (6.67)

5. Thus g(Z) can be obtained fromy(t) [or get)] substituting nT for'

t and evaluating the sum.of series as
X" (s+ jnw.) = .e {X* (t)} = .X* (s) ... (6.61)

Equation (6.59) becomes, 00

GCZ) = Ly(nT)z-n ... (6.68)

Y*(s)= ~ i
n=-<X)
G(s+jnws)X*(s) . ,
n=O

For physically realizable systems, convergence of this series assured.

Frequently Z transfer functions, with G(s) available as the starting point .
4.
1 . are desired, in which step Sand 4 suffice. When G(s) is complicated
= TX*(s) L G(s+ jnws)
co

n=-co
.
... (6.62) function, it may be split into partial fractions then steps 2 and 5 can be
applied to the individual partial fractions.
 Computer Control of Processes) -
(Digital Controller Design [ 6.15 )

Xes)
1 1
Y(s) fig. 6,,'1: Block diagram of a closed loop system
s+1 s+2

Solution: R(s) ---.+

1 \..._--~-------------.-/

- G(s) = (8+1)(8+_2) ... (6~69) g(5)

G(t) e~t _e-2t = Y(t) Fi'g. 6.2: Equivalent block dia9ram

G(nT) = e-nT _e-2nt = Y(nT) Showing the transfer function"

1- e Ts
G(Z) =
00

Ly(nT)z-n L(
co
e-nT'_e-2nT) Let Gho(s)=
s
... (6.70)
n=O . n=O
From Figure
1 1 C(s) = G(s) Drs) E*(s) ... (6.71)
GCZ) = 1 _ e-aT Z-l 1- e-2aT Z-l
Or
After ,simplification we get C*(s)= G*(s) D*(8) E*(s) .~. (6.72)

C(Z) = G(Z) rxz: E(Z) ... (6.73)

In terms of Zi,,]cu.nsform notation
E(Z) =R(Z) - C(Z) ... (6~47)

6.5.3 Deve!opement of Closed loop Transfer Function We have

The block diagram of a single loop, computer based, closed-loop C(Z) = D(Z) G(Z) (R(Z) - C(Z
control system is shown in Figure 6.1. This ,is the most general description
of the -feed back control system in that the dynamics of all the elements C(Z) D(Z)G(Z)
... (6'.75)
are" assumed to be significant and therefore are included in. the block R(Z) 1 + D(Z)G(Z)
diagram.
Equal (6.75) is the closed -loop pulse transfer function of the sampled,
.The derivation of the closed-loop pulse transfer function' proceeds data system to set-point change. Note the similarity of this equation with
in manner analogous to that of a continuous control system. As an its counterpart in conventional control system.
illustration, letus derive the closed -loop pulse transfer function of the
block diagram of Figure 6.2.'
 Computer Control of Processes )
l DigitalConlroUerD'esigl1 ( 6.t7 )
The proportional controller output CO at-the nth sampling instant is
l.Example 6.3 I related to error by the expression.
crz). (OO)n = K, en + (CO) steady state
a ) Determine the pulse transfer function R(Z) for the sampled data
Similarly, at (n_l)th .samply instant.
control system of.Figure 6.3 with L(s)-i= O.
(CO)n-l = ICc en-l + (CQ) steady state.
b) Evaluate the transient response to a step change in set point.
Therefore
The closed lo~p transfer function is given by
(CO)n - (CO)n-l = :ICc (en -en-l)
C(Z) , D(Z)G(Z)
... (6.78) , Taking the Z transform of this equation gives
R(Z) 1 + D(Z)G{Z)
CO(Z) _Z-l COeZ) x, EeZ) - Z-l x, E(Z)
Load L(s) = 0 C*(s)
CO (Z) [1 - Z-l] = x, "E(Z) [1 - Z-l].
.... :11:......
~

.: T Therefore, D(Z) = CO(Z) =

E(Z)
C*(s) Substitution of the expressions for G no Gp(Z) and ncZ) into the pulse-
T' transfer function will give
Fig. 6.3:Pr~porlion(l;J -entrel of sampled system

G(Z) = Gno(Z) Gp(Z) = Z{Gno (s) Gp(s)}

C(Z)
= z{!x s
1-
s'
eST} R(Z)

zt~} -z[ e:~]

= zt~} _Z-l zt~ } b) Evaluation of transient closed-loop response for a unit step change
in set point,"

( ~Z-1)Z(s~ J

TZ-1
=
(l-Z-l)
 computer Control of processe;],.
[Digital ControllerDesign ( 6.19 J
The inverse of which is e-at
Substitution ofR(Z) in the closed-loop pulse transfer function given

where
C(Z) = (Z-l) [Z+(K c T-l}J = ZC 1( z-)

1 [
(l-I~c T)J
1 l
a ln
T
Now we expand C 1(Z) in partial fraction., Thus,

, Kc T A B ., C(t)
(Z-l) [Z+(Kc T-l)J= (Z-l) + Z-(l-K c T)
u(t) _ e1n(1 - KcT)tlT

Multiply by Z-l and set Z ='1

urt) = (1 - KcT)t/T
Then,
A= 1 6.594 Design of Sample Data Control System
Multiply byZ-(l - KcT) and set Z = 1 -ReT Conventional controllers came into existence because the necessary
hardware was available. Hardware could be built to produce the desired
Then, B = -1
proportional integral, and derivative relationships..
Therefore,
The availability of digital computers for control applications sparred
much research to produce better designs without'regard to hardware.
Early designs w e re primar ly concerned with response
characterisitcs. Recent research has shown that there are many other
1 important considerations..The following is a list of desirable theoretical
Z-l properties of digital control algorithms.
.and C(Z) 1.. Open-Loop Characteristics

Z z An open loop, stable or unstable control algorithm can be used to

Z-l control an open-loop stable or unstable process when the control structure
employed is the usual sampled-data control system.
1 1 However, all other things being equal; An open loop stable control
1..... Z-1 - _In_1_,_ algorithm would probably be preferred When the control system is
l-e (I-KeT) z' expressed in Il\1C (for Internal Model .Control) from the controller and
the process must b\~ open-loop stable as we will, see.
1
The .second term is. ,of the form 1 _.e-at Z-I
 TDigital Controller Design ,( 6.21 ) ( 6.22 ] Computer Control of Processes)

1) Stationary constraints are always preset.

Load dynamics
Load L(s) An example might be a reflux valve of a distillation column that
must be morethan X% closed to avoid weeping on the trays 'or
more thanY% open to avoid floodingthe column.
Zero-order Process ;cC*(t)
hold dynamics 2) Moving constraints come on and disapear during different phases
I C(z)
of procers operations constraints are generally encountered on
C(s)
inputs, but output constraints may be present in specific
applications.' A digital control algorithm must be designed such
that it does not violate operational constraint's.
T

Fig. 6.4:. Closed loop sampled data system 6. Dead time compensations
We examined the detrimental effects of dead time on the response'
2. ResetProblem of first order system. The presence of dead time necessitates lowering of
The algorithm must guarantee offset-free performance in the controller gain to. maintain stability.
presence of modeling errors .. Modeling errors arise when the nominal W e observed that w hen the apparen t dead tim e ed exceeded the
model upon which the algorithm is based is different from the Teal process. dominant time constant of the system t the peak offsets following a step
3. Controller Tuning change in load could, approach those of the uncontrolled situation even
with best PID controller tuning. Under these conditions, the settling time
It is desirable that the algorithm. contain a small number of
approaches g8d.
independent (adjustable) tuning parameter. The larger the number, the
more difficult will be the task of keeping the controller tuned in the field. Since many chemical engineering process exhibit apparent dead time
characteristics and since dead time is detrimental to control, there is
4. Robustness considerable incentive to develop control algorithm that can compensate
A digital control algorithm is based on a nominal plant model that 'is for such time delays.
developed from first principle or 'obtained from experimental data.
The concept of dead-time compensation involves the use of a
Modeling error willfnvariably be present. Further more, new modeling
mathematical model of the process to in effect remove the dead time from
, errors will creep in due to equipment' fouling changing production
the feedback signal so that the performance of the system is improved.
strategies, etc:
The algor-ithm must maintain stable proceas operation in the 7. Working with RHP zeros
presence cf such modeling errors. Robustners issues having to do with When the roots of the numerator polynomial of the laplace' domain
the ability 'of the control system to maintain.stability in the presence of'a transfer function contain one that lies to the right of the imaginary axis
plant model mismatch, have received consideratble _attention in recent in the s-plane, the system is said to have a RHP(right half plane) zero.
years. Such a system containing all RHP zero. begins in a direction that is
,:';. ConstraintHandling opposite to the direction in which it eventually settle out. Inverse response
occurs in boiler level systems and in some distillation base level systems.
Industrial process must often be operated in the neighborhood of
operating costraints. There are two types of constraints. There are two
types of constraints:
 (6.24] ComputerControl.ofProcesses)
{Digital Controller Design ( 6.23 )
To develope the algorithm for set point chages recall from study
It should be clear that such systems will pose'difficulties for the that the closed loop pulse transfer function of the system of Figure 6.4 for
PIn type controller in that the controller will take wrong action intially set point changes is
with digital control. A facility to accommodate inverse response can be
built into the algorithm and this should be considered wherever applicable.' C(Z) D(Z)G(Z)
... (6.79)
8. -Manipulated variable -movements R(Z) 1 + D(Z)G{Z)
Excessive movement of the manipulated variable should be avoided where, G(Z)' = zro., (s)Gp(s)}
to reduce actuator wear.
The solution of Equation (1) for n(Z) is
9. Shaping of closed leep dynamics
In the absence of modeling errors the algorithm should yield a closed C(Z)
loop response having desired dynamics. 1 R(Z)
D(Z) = G(Z) 1 _ C(Z) ... (6~80)
10.' Reset windup R(Z)
If the algorithm contains- integral mode the phenomen.in of reset
windup may occur. Anti-reset windup features should be provided' The design, procedure is to specify the desired response characteristic
wherever 'appropriate. C(Z) , ..... - .
R(Z) , for example the controlled variable shall ,reach the new .set point
11. Bumpless Transfer
in one sampling period. Then Equation (6.80) may be solved for D(Z),
The facility for bumpless transfer from manual to automatic should
provided the process transfer function Gp(s) is known.
be provided.

12. Noisy Process 6.5.5 Dead Beat Control Algorithm

Industrial process often contain noise. We consider the treatment An algorithm that requires theclosed loop response to have finite
noisy process signals later in the chapter. Suffice it to say at this point settling time, minimum rise time, and zero steady state error is referred
that the algorithm must function in the presence of noise recall that to as dead control algorithm.
derivative action tends ,to amplify noise and therefore it should not be
According to this method, we 'require that the response of a process
used whenever noise a problem.
to a unit step change in the setpoint exhibit no error at all sampling
'I'o began the developement of digital control algorithms by instants after the first. Thus for unit step change in setpoint we have,
Z-transforms. Consider the block diagram of typicalsampled data control
system shown in Figure. The process transfer function Gp(S) is assumed
1
R(Z) = 1- Z-l ... (6.81)
to include the dynamics of the sensor and the. final control element. The
objective is to eyntheeize- a algorithm D(Z). If the response .is to have zero error 'at all sampling instants after
Such that the desired loop performance is achieved. Once the the first, its discrete time behaviour resembles that of a unit step delayed
expression for n(Z) has been developed, it may be inverted intothe time by one sampling instant.
domain to given an equation suitable for computer programming. It is Z-l
C(Z) = 1- Z-l ... (6.82)
possible to develope the algorithm for set point changes or load changes.
(Digital.Control:ler.Design ["6.25] [6.26 ) Computer C~ntrol of Processes )

3. Neverthless it can be shown that the steady-state error is zero.

tJ)
tJ)
c:
o Thus
a.
(/)
(I)
0::: E(Z) = R(Z) - C(Z) ... (6.85)
Using Equation (6.83)
E(Z) = R(Z) - Z-l R(Z)
E(Z) = (1 - Z-l) R(Z) ... (6.86)
Use finalvalue theorem

T 2T 3T 4T Time

Fig. 6.5: Response requirements for the design. of a deadbeat controller

= Lim [(l-Z-ltR(Z)]
z~l
= 0
Therefore
4. The dead beat algorithm is physically realizable if the time delay
C(Z)
Z-l ... (6.83) in the a(Z) is not larger than one sampling period.
R(Z) =
For example, let
and we know that
... (6.89)
C(Z)
(i.e., dead time is equal to K sampling period then from equation
1 R(Z).
(6.84)
D(Z) = G(Z) x 1 _ G(Z)
R(Z)
... (6.90)
1 Z-l
D(Z) = G(Z) x 1 _ Z-l' ... (6.84) \
And the controller is not physically realizable if K > 1, because
K-l > 0 and the numeration of the design equation has z to a
A few 'remarks on the characteristics of a deadbeat algorithm are
positive power. This implies that we need future error values
now in order
which are not available.
1. The req-uriment that the error is zero at all. sampling instant's
5. If the dead time is .larger than one sampling period, we need. to
doesnot preclude large overshoots or .highly oscilator behaviour
modify 'the response specifications requiring that the respo~se
for the process response.
zero error at all sampling instants after the first two, three etc.,
2. The fact that response reaches its desired value (error =0) in one (Figure 6.6). In such a case we have the following characterizations.
samplinginstants indicates that the rise time is minimum, which
in turn implies that the controller exhibits very strong control
action, not necessrily a .desirable feature (cause large overshoots.)
[Digital Controller Design ( 6.27 ) ( 6.28 ] ComputerContrelof'Processes )

Q)
IExample 6.4 ,
C/)
c: Design a dead beat control algorithm for a process whose transfer
8.
CJ)
(\)
0:::,
function is given by Gp(s)' = 1 . , with T = 1 sec.
0.4.8 + 1
1.0
Solution:

1
Gp(s)=
0.48 +1

C(Z)
n(Z)
1 R(Z)
= - - x _--:.--:.
G(Z) 1- C(Z)
R(Z)

1 Z-l
... (1)
G(Z) 1 -- Z-l
T 2T 3T 4T 5T Time We know that,
(b)

~.6:
Fig. Response specifications for dead beat design
with process dead time (a) i d =T (b) td =21 =
(l_e-
zJ---x--.
TS 1}
l s 0.48+1
C(Z)
= Z-2
R(Z) =Z 1 _Z_e_ -TS_ }"
{ s (0.48 + 1) } { s (0.48 + 1)

E(Z)
or . R(Z) = Z-3 etc.
z{ 1
8(0.48 +1)
}-z-lzf, 1
l8 (0.48 +1)
}
Yielding following controller design equations.
I
,1 Z-2,
n(Z) = ---- .t, (6.91) l-Z-1)Z 1. I
J ... (2)
G(Z) 1- Z~2 ( { (O.4s + 1)
'S

1 .Z-3 2.25 A B
or ... (6.92)
. D(Z) = G(Z) 1 -' Z-3 8(8+2.25). = s + 8+2.25
( Digital.ControUerDesi,gn ', (' 6.30) ComputerControl of Processes )

2.25, A(s + 2.25) +Bs We know that,

s = 0 s = --2.25 D(Z) = M(Z)

E(Z)
IA ,= 1 I
(2) becomes M(Z) 1 - 0.0820Z-1
E(Z) 0.9179(1-Z-1 )
= (1-.Z-1 ) Z { ! _ -1_'}
'
s s + 2.5 M(Z) = M(Z) Z-l + 1.09 E(Z) - 0.089 E(Z) Z-l
(1- Z-l) Z{U(~) _ e-2 .5t } M; = M n- 1 + 1.09 En - 0.089 E n- 1
where M; - Controller o/p at nth sampling instant
= (I-Z~l){1_~-1 l_e-Z~5Tz-l} Mn- 1 - Controller olp at (n_l)th 'sampling instant
En - Error at nth sampling instant

~7-1 \[, 1- e-2.5T Z-l, -l+Z-l ]

E n- 1 - Error at (n_1)th sampling instant

\1-..z.z M(1-e- Z.5T Z- 1 )

IExample 6.5 1
Design a dead beat control algorithm the transfer function is given
= Z-l (1,- e- Z .5T ) 10
(1- e-~5T Z-l) by G p = 0.58 + 1 where T d = T.

Solution:
Now (1) =>
10
1 z' where t = Is Gp'(S)=
0.58 + 1.
D(Z) = G(Z) 1 _ Z-l

e(z)
= (1_e-Z.5T Z- 1) x _ ~ 1 R(Z)
~(1_e-2.5T) 1- Z-l
D(Z) = G(Z) =
1 + C(Z)
R(Z)
(1- e- Z-l )
2 .5

1 Z-2
(1-Z-1)(1- e-Z.5 ) = --x
G(Z) 1- Z-2
... (1)

1 - O.0820Z-1 G(Z),= Z{ Gp(s) Gho

1
0.9179 (1-Z- )
(Digital Controller 'Design ( 6.31 J ( 6.32 ) Computer 'Cont-rol of Processes )

TS 1 Z-2
orz: = z{ l-se O.:SO+l} (1) =>
neZ):::: g(Z) 1- Z-2

(1"- e-2T Z-l) Z-2

G(Z) = (1- Z-l \z{ 10 } ... (2)
~--_--:...-x---

10Z-1 (1- e-2T ) 1 - Z-2

} s(0.58+1)

Z-1{1_e-2T Z - 1)
10
10
~=
20 10(I-Z-1 )(1 ~ e-2T )
8(s+2) s(s+2)
Substitute T =1

20 A
-+-
B (1- e--2 Z-1)
. Z-1
s(s+2) - 8 8+2
10(1-Z-2 )(1- e-2 )

20 ~ A(8 + 2)+ Bs
Z-1 (1- 0.1353 Z-1 )
Put s= 0 Put =-2 10 (1--Z-2) (0.8646)
20= A2
M(Z) Z-l (1 -0.1353Z-1 )
I A = 10 I I B = -10 I . E(Z) 8.646(1- Z-2)
(2) => {l_.Z-l)z{lO _~} M(Z) 8.646 - M(Z) 8.646 Z-2 = E(Z) - E(Z) 0.01353 ~2
- .. s s+~
M(Z) = M(Z) Z-2 + 0.1156 E(Z) Z-1_ 0.01353 E(Z) Z-2
= (1 _Z-1) Z{lO U(t) - 10 e-2t } Mn = Mn - 2 + 0.1156 E1:<-1 - 0.01353 E n - 2

6.5.6 Dahlin Algorithm _

Dahlin algorithm specifies that the closed loop sampled data control
system behave as though it were a first - order process with dead. time,
Thus the closed loop specification is given by

C(Z) = FeZ) aez) = FeZ)z-<n+ll ... (6.83)

R(Z)

10Z.-1(1- e-2T ) F(Z) represents first order lag and N is the number of sampling
periods in the process dead time 0d. To derive the expression for F(Z)
(1- e-2T Z-l) consider the differential equation for a first - order process.
fDigit~f ControllerDesign ( 6.34 J'-' Computer Control of Processes )
[6.33 )
.1 EXample 6.6
. 'tf
.
dy +
dt.
yet) = X(t). . (6.85) e-1. 4 s
Design a Dahlin's controller for Gp(s) = T =1 sec.
3.34s+1
Note that in Equation (6:85) the steadygain have gain of one. The
Solution:'
numerical solution of Equation (6.85) is
Dahlin's Control algorithm,
Yn (l~af)Xn+.afYn-l ... (6.89)
where ar = e- Thr (O < af< 1)' D(Z) =
1
--x----'--------
1- aZ-(N+l)

Taking Z transform of terms in Equation (6.89) and rearrangement

G(Z) 1 - u.Z-1 - (1 - a )Z~(N+l)
gives.
where, G(Z) = Z{Gho(S) Gp(s)} = G no .Gp(Z)
Y(Z) 1- elf
" = F(Z) =. 1 - . Z 1 .... {6.90)
GCZ):::: z{l-e-~S
. X(Z) af - x e-1.4s }"
. s A3.34s+1
Substituting F{z) from equation (6.90) into Equation (6.83) gives
C(Z).
(l-Z~l )z{ S(3~;::s+1)}
1 - elf Z-(N+l}
... (6.91)
R(Z) - 1- afZ-1

C(Z) this examples: 8d = NT

nrz: 1
---x
R(Z) ... (6.9"2) where,
G(Z) 1-' C(Z)
N;:: Larges t integer number of sampling
R(Z)
interval- in ed
Substitute Equation (6.91) in Equation (6.92)
T = Sampling period
. 1 1 - a Z-(N+l} .
DCZ) == _-x .' . f . ... (6.94) 8d = 1.4 sec
. G(Z) 1 -' <Xf Z-l - (1- elf )Z-(N+l).
T = 1 sec
The time constant of the closed loop response Zr on equivalently, the
.. N= 1
parameter <Xf is an adjustable parameter that is selected by trial and error
method. e = 0.4

As we will see, Dahlin's algorithm is essentially the same as the -NTs -es 1
recently developed internal model control for first order plus dead time . :. GCZ) = (1-Z-1IZ e e
types of proceses. It turns out that elf can be used to enhance robustness
\ J { 3.348+1 f
of the loop in the 'presence of modeling errors. We know that, Z{e- NTs } = Z-N
A high value of elr close to 1 gives a high degree of robustness, but
( -6s }
the response become quite sluggish. Smaller values of ar close to zero G(Z) = (l_Z-1)Z-l zJ__e _
improve dynamic response, but the system becomes more sensitive to .\ . . l8(3.34s+ 1) "
modeling error.
[Digital Controller Design ( 6.36 ) Computer Control of Processes)
( 6.35]

(l~Z-l)Z-lZ {_ _
m s(3.348+1)
l_} D(Z)
1
G"(Z) 1 ~ aZ-
1
(1 -
= - - x----C.---=---------'-
-
a)Z-(N+l)

(1 - a)Z-(N+l)

a e-Tlt '1' is not given' so 1; =T

(l-Z-i )Z-l Z { 0.3 }

We know.that,
m s(1+0.3)
f -0.741z-1 l[0.6321Z-
2
]
[ O.165Z- 2,+O.094Z-1 'I -O.3678Z-1 - O.6321Z- 2 xz

0.6321Z-1 -'o.468Z-2
0.094+0.1305Z-1 - O.12 Z- 2 - 0.1043Z-2
Where a = 0.3, 8 = 0.4
m= 1 - ~ =. 1 - 0.4 = 0 n(z) = M(Z)
T l' E(Z)
Substitute the above values into above equation

G(Z) = (l_'Z-l)Z-lZ-l
, , {I e
-(O.3xO.6 xl)
_
}
M(Z)
E(Z)
0.6321Z-1 - 0.468Z-2
0.094 + O,.1305Z-.1 - 0.12Z-2 - 0.1043Z-3
. ,..' . 1- Z":'1 1- e-(O.3xl) Z-1
00094 M(Z) + 0.1305 Z-l M(Z)

(
1 _ Z-1)Z-2{_ 1 _ - . 0.835
; 1-Z-1 1-0.741Z-1 J
1 - 0.12 Z-2 M(Z) - 0.1403 Z-3 0.031 Z-1 E(Z) - 0.468 Z-2 E(Z)

TakingInverse Z transform

6.722 E n- 1 -4.97En-2-1.~88Mn-1 + 1.276 ~n-2 + 1.1095 Mn~3

-2 [ O.835(1., Z-l ),]
,
Z 1-----
1 - 0.741Z-1 Controller output at the nth sampling instant'

1 1
'Mn - 1 = Controller output at the (n_l)th sampling instant
G(Z) = Z"'2[1-O.741Z- - 0.835 + O.835Z- ]
1 -O.741Z-1
, .
Mn - 3 = Controller output at the (n_3)th sampling instant

Z~2 [ 0
1 65 + 0.094'Z-1 1
1 - O.741Z-1 -J
En

E n- 1
= Error at the nth sampling instant
Error at the (n_l)th sampling instant

2 3
G(Z) = O.165Z- + 0.094Z-
1- 0.741Z-1
 (Digital Controller Design [ 6.37 ) 'tomputerC9ntrol of Processes J

~ample 6.7 1
We know that,

. . e-O8s
Z{e-:-NTs } = Z-N
Design a Dahlin's controller algorithmn for Gp(s) = -0.6s+1
- - ; T = 0.4 'sec

. Solution: .
acz) = (1- Z-l )Z-2
.
z{ 8(0.6s+1)
1 .}.
Dahlin's Control Algorithmn
= (1-Z-1)Z-2 Z{ 1.67.. }
1 (1 - a) Z-(N+l) . s(8+1.67)
n(Z) = --x--~--..:...---
. G(Z) 1 - aZ""! ~ (1 - a)Z-(N+l) By using partial fraction method,'

where,
z{ 1.67 }
8(s+1.67)
_ A +_B_
- s s+1.67

G(Z) =
Z{l-e..s-TS 0.6s+1
e-O.. 8~.}
x
1.67. = A(s + 1.67) + Bfs)

=> A= 1; B = -1
-O.8s }
_ 1-Z-1 Z e
- ( ).{ s(0.6s+1)
:. GCZ) = (l~Z-l
.
)Z-2 Z{-!_ _
S 8+1.67
l_}
In this example Gp(s) contains dead time Od.
ed= NT
(1 - Z-1)Z-2{ Z-l
Z - Z}
In this. example: Z_e-l.67xT.

8d = 0.8 sec T. = 0.4 sec

.T = 0.4 sec
0.8 = N x 0'.4 (lZ-l )Z-2 [Z . .Z ]
- - Z -1 - Z_e-1.67xO.4
N = 2

. (1-Z-1)Z-2 [~_ . Z ]
Z-l Z-O.51

1 - Z~lZ -:NTs
e. ]
. = .Z2[1_ Z-0.51
Z-l ]
(- ) [ s(0.6s+1)

= Z-2[Z-0.51- Z+l]
Z - 0.51.
(Digital Controller Design (6~40 ] ComputerControlef Processes J
Let D(Z) is the ratio of two polynomials in z-', as demonstrates
Z-2[ 0.49 J C(Z) f30 + f3IZ- 1 + (32 z-2 + + r\. Z-k
'Z - 0.51
D(Z) = -()
E Z
= . 12m
1+ al Z- + a2 Z- + + am Z-
... (6.95)

O.49,Z-2
The roots, of the polynomial in the' denominator are the poles of the
Z - 0.51
controller and for the stable controller, these poles should be inside or
1, (1 - a )Z-(N+l) on the unit circle inthe complex plane. Let us examine what the effect is
DCZ) = G(z) x 1 _ aZ-1 _ (1 _ a)Z.,.CN+l) of a pole's location on the controller outputCrt).
Consider a controller with the following "discrete transfer function \
where, a = e~Tlt ''t' is not given.so.r ~T e"':' l = 0.3678
C(Z) 1
D Z = Z-O.51 [ " (1 ~O.3678)Z-3 , ] DCZ) = E(Z) 1-PZ-l
... (6.96)
( ) . 0.49Z-2 1-0.3678Z-1-(1-0.3678)Z-3
An er~9r equal to' a .unit impulse enters the controller at t = o.
O.6321Z-2 -O.322Z-3 Then from the above equation, C n - P Cn- l = En
=
0.49Z-2 -O.18Z-3 -O.31Z-5
where En::;: Lfor ri e G and Ej e D forn=1,2,3 ...
Let us now invert this algorithm into thetime domain. To accomplish
LetP = -0.9, then Cn + 0.9 C n-1 = En
this we note that
Controller output, Cn
n(z) = M(Z)
E(Z) n En ForP = -0.9 For P = ~10.9

M{Z) 0.6321 -0.322Z-1 , 0 1 1 1

E(Z) = ().49-0.18Z-1-O.31Z-s
1 0 -0.9 0.9
0.49 M(Z)- 0.18 Z-l M(Z) - 0.31 Z~3 M(Z) = O.6321E(Z) - 0.322 Z-l E(Z)
2 0 +O~81 0.81
This equation can be inverted to give the algorithm for computing,
the controller output. This 3 0 -0.729 0.729
:rvin 0.49 - 0.18 Z~l Mn- 1 - 0.31 Mn- 2 = 0.6321 En - 0.32? E n-1
. , 4 0 +0.6561 0.6561
Mn = 1.29 En - 0.657 E n-2 + 0.3673 M n- 1 ,+ 0.6326 M n-:- 3
5 0 -0.59049 0.59049
6.5.7 Ringing and the Placement of Poles
There are successive changes in the sign of the controller output.
While designing a digital controller by the res pone sepcification
This 'phenomenon is known. as ringing or controller ringing and is caused
meth~ds (Dead beat, 'Dahlin's) we should not only concerned with closed by negative poles. The closer a negative, pole is at the origin, the smaller
loopz-esponse, but pay "attention to the resulting movement of the
the ringing is C as shown in Figure 6.7 (b)
controller output. Excessive control valve movement is unacceptable in
industrial practice.' Let us examine the movement of a controller's output.
 rDigi:tal'ControJ:ler ~.eslgn [6.41 ) computer Control of Processes]

Imaginary
p=~().9
+j
Unit circle

-1 +1

(a) (b) Ringing' Desired

poles poles
-j
Cn Cn
p =+ 0.9 p =+ 0.3
1.0 1.0
Fig. 6.8: Desired location of controller poles

There exist several design methodologies which eliminate the 'ringing

a Time a Time problem by placing the controller poles on the right hand side of the real
T 2T 3T 4T 5T
axis. These methodologies are quite complex and a complete discussion
is beyound the scope.

(e) (d)
IExample 6.8 I
Fig~ 6.7: EHect of pele.Ieeefien on liehaviour of controller output
Eliminating the ringing of a digital controller. Suppose that for a"
For positive poles, the controller outp.ut exhibits no ringing (as shown given process" Dahlin's method leads to a eorrteoller with the transfer
in Figure 6.7 (c) and (d). Thus from above discussion: . function.
'Any negative pole of the controller's discrete transfer function
will cause ringing. .... (6.97)

The closer a ringing pole is to the' unit circle, the higher the ringing
. of the controller output will .be. The term (1 + 0.6 Z-l) in the denomimator implies the presence of a
pole at -0.6, which will cause controller ringing.
. Hence, we should design a digital controller, so that all its' poles are'
poaitive, preferably with medium - size absolute values as shown in Dahlin has suggested that the ringing pole can be eliminated by
, .Figure 6.8. multiplying D(Z) with (1 + 0.6 Z-l), thus canceling th~ ringing term from
the denominator. But inorder to keep the static gain K the same.we should
Unfortunately, the deadbeat and Dahlin's algorithms usually contain
poles that causesevere ringing of the controller output. This may be the' divide DCZ) by the term lim(1+0.6Z-1 ) .
Z~l
= 1.6.
case of any' response, - .specification design technique 'and is not unique to
Thus the ringing - free controller has the transfer function'
the foregoing two algorithms.
,,(Digital Controller Design ( 6.43 ) [6.44 ) Computer Control of Proces'ses.] .-

Hence, the Z-transform of equation (6.102) is

1
DCZ) = K 1-0.5Z-
~(1 +_~.HZ-l) (1+O.6Z-C )( 1-Z~1-)(1~O)Z-1)
1
P'CZ) = -l +~:
1 (z-n+l + z~n+2 + +Z-l + 1 )E(Z)+ :~ (1-Z-1)E(Z)]

K 1-0.5Z-1 ... (6.104)

= 1.60 (l-Z-l )(1-0.7Z-,-1) ... (6.98)
, In the above equation, the summation (integral) term' for large values
. 1
.of n approach a limit Of(l~Z-l) .'Pl1,e;refore,EqnationC6.104) can be
6.5.8 DigitaJ PID Controller
simplified to
Consider the dealcontinous (analog) PID controller,

pet) = P+K c [e(t)~1- fe(t)dt+'t D de(t)] ... (6.99)

P'(Z) = Kc [l+Llt(~)+ tn (1-Z-1)]',E(Z)
tI 1-Z.~t
'.. (6.105)

tI 0 " dt
The digital controller transfer function is,
To convert this expression to its digital equivalent, we use the
following finite difference approximations
t 'n DCZ) = ~g}= Kc[1+~:(1_~_1)+ :~ (1-Z-1)] ... (6.106)

fe(t)dt ""Iek At ... (6.100)

o k=l, Equation (6.106) is referred to as the position form of the control
law because it yields the value of the controller output directly.
de, en -'en-I ... (6.101)
. dt ~ 'At 6.5.8.2 Velocity Algorithm
, " The digital PID controller equation can be written in' two ways, the ,The velocity form of the PIn controller is an attractive alternative
position form and velocity form substituting '(6. 100) and (6.101) into (6.. 99); to the position form because it avoids computing the summation in (6.102).
we obtain the position form of the digital PID control algorithm. Also, the velocity form doesnot require specification of the bias term 'p
and is less prone to reset windup, as discusse'd below, Since the nominal
6.5.8.1 Position Algorithm, steady-state value (or bias) Ii is a constant, the change in the controller
output 8Pn is given by
Pn = p+K c [e n+
8
,t I eK,+ 'tD{en-e
tI K=I dt '
n _1)]
,
.... (6.102)
LlPn = Pn-Pn- 1 = P'n-P'n-l ... (6.107)

Equation (6.102) can also be written aes a Z- transform ,expression. The velocity form of the ,controller' can be found by shifting (6.102) ,to
obtain Pn-l and subtracting it from (6.102).
Let pIn = Pn -P be defined as'a deviation variable.. Recall that the
Z-transform translation theorem yields
~Pn = x, [(en - en-l)+ AttI en + t~tn (en - 2e n. _1 + e n-2)]
'
... (6.109),
Z(en ) = E(Z)
,Z(en-l) = Z,...l E(Z) ... (6.103) Taking the Z-transform of (6.109) given

AP(-r)= Kef (1- Z-l )E (Z)+ ~: E(Z)+ ~ {1,..;2Z-1 +Z-2)E(Z)] ... (6.111) .
(DigitalController Design .1 645 ) ( 6.46 ] Computer Control ofProcesses )

'Note that Equation .(6.11) can be obtained by multiplying both 'sides For the position form algorithm, there are several modifications that
of (6.105) by (1 - Z-l). Another variation of the digital PID controller is can be made .to reduce reset windup.
based, on the more accurate trapezoidal approximation for the integral a) Place an upper limit on the. value of the summation. When the
(in Eqution 6.99) . - controller saturates, suspend the summation unt.il the controller
output moves away from the limit.
!e(t)dt "' ~(ek +2ek- }t 1
b) Back - calculate the value of en that just causes the controller to
saturate. If the saturation occurs, and this actual as error term
After ,this expression is substituted into(6.99), the velocity form of en-l in the next controller calculation. Field testing has' indicated
the control law becomes, that approach (b) is supperior to (a), although it is somewhat more
complicated.
APn"' ~ x, .[(en - en-l) + ~t.
'tI'
(en + en-I) + 'tn (en -
2 8t,
2 en-l + .en- 2 )]. Note that in the velocity' form of the algorithm, Equation (6.109) no
summation appears avoiding the windup. problem. The controller remains-
or as a Z-tranform (Equati~1?>_Q .ll_3J'
saturated until the error (en) decreases to a point where the control action

8P(Z) = K c [(.1.,- Z-1)E(Z)+~(1

,2'tI
+ z' )E(Z)+ 'tn (1-2Z-1 + Z-2)E(Z)]
8t
returns to normal levels. However, .the algorithm must be programmed
to disregard~Pn if P n is at the limit implying that P n should be monitored
at all times. In most situations, where the integral mode is present the
... (6.114)
velocity algorithm is preferred' over the position algorithm.
Compared to Equation (6.111), this moreaccurate ap~roximation'of
the integral may not actually achieve a significant improvement in control 2. Elimination of derivative kick
loop performance. When each' controller is tuned, slightly different values When a 'set-point change is made control algorithms (6.162 and 6.109)
of the controller settings (K e , 11, ':n) may be obtained 'in tuning (6.113) will' produce a large immediate change in the output (an impulse) due to _
versus (6.109). Note that both eqauation can be arranged to the same the derivative term. There are several methods available for eliminating
general form ' derivative kick.
a) Most c.ommercial controllers apply derivative action (or the
discrete equivalent) to the. measured variable 'rather than to the
6.5.8.3 Features of Digital PID controller
error signal. Thus en = r n -bn where is the measured value of the
1. Eliminatfon of Reset Win~up controlled variable, is replaced by -bn in the derivative term, .giving
In equation (6.102) reset windup ca.n occur when the entire summation the position form
gro~s to very large value. Suppse the controller output is at an upper limit
or lower Iimit, as the result of a large sustained error signal.
... (6.109)
Even though the measured variable reaches its setpoint (en = 0), the
controller maybe wound up because of summation errors. Until the error'
Equati?n (6.109) for the velocity form can be chang~d in an analogous
changes its sign for a period' of time, thereby reducing the value of
~~~ .
summation, the controller will remain at its limit start-up situations
performed. Under atomatic control and other set point changes are b) Instead of a step change in the setpoint, ramp the set point :0 a
particularly prone to this problem, leading to excessive overshoot by the new value, limiting the rate of change of r n
"controlled variable and saturation of the manupulated variable.
 ( Digital Controller Design [ 6.47 ) , ( 6.48 J Computer Control' ofProcesses' J
If measurement noise combined: with a large ratio of derivative time However, where the input and/or output. stgnals are converted
internally into engineering units, the value of K, will not be dimensionless.
to sampling period ('t
~t
D
\) -is causing an overactive derivative mode, fitter
. . ' ' 6. Time-delay compensarion
the error signal before calculating the derivative action. Many commercial distributed,controI system offer the option of time
3.Effect of saturation on controller performance delay compensation with a digital PI or PIn controller. The controller is .
implemented in difference equation form using.a smith predictor.
One of the difficulties that aries. in equation (6.102) is that a small
change in the error can cause the controller output to saturate for certain
values of the controller settings. Suppose thatKc'tnl~t = 100 because of a
6.5 . 9 Smith Predictor Algorithm
small sampling period, and that en, andP; are both scaled form 0 to 100%. The smith predictor algorithm: was developed by O.J. M Smith in
A small change in ~en = en-en~l of 1% willyield .a 100% change in P n , 1957. This technique is model-based approach to better control of systems'
exceeding itsupperliniit. Therefore, one must be careful to select with long dead times and has come to be known as the Smith Predictor.
controller .settings and a value of L\.t that do not cause scall ing Let us develop the Smith predictor algorithm' for a process that can"
problems. ' be .represented by a first order lag plus dead time model. The block
4. Comparison of position and velocity 'algorithms. diagram of a conventional control system for this process is shown in
Figure 6.9.
The position form of the PID algorithm (6.102) requires a value of Ii
whereas the velocity form of the algorrthm (6.109) does not explicitly As shown in Figure 6.9, the process is conceptually split into a pure
require ,a steady state value for the controller output.However, lag and pure dead time. If the fictions variable B, could be measured
initialization of either. algorithm isequally simple since manual operation somehow,we could connectit into the controller, as shown in Figure 6.10.
of the control system. Usually preceds the transfer to automatic control. This would move the dead time outside the loop.
_ Hence Ii (orPn - 1 for the velocity algorithm) is simply taken to be equal to The controlled variable C would repeat whatever B did after a delay
the signal to the final control element at the time of transfer. As noted of ed- Since there is no delay in the feedback signal B, the response of the
previously, the ve~ocity form is less prone to reset windup problems. system would be greatly improved. "I'his scheme, ofcourse, cannot be
To implement"a velocity algorithm directly,a pulse up/down ebunter implemented, because B is an 'unmeasurable (fictitous) signal.
or steping motor can be used in series with the controller to convert Now, suppose we develop a model of the process and apply the
incremental changes to 'position' (example: flowrate), Integral action is manipulated variable M t.o the model as shown in Figure 6.1~.Ifthe model
always recommended with the velocity algorithm to prevent drift of the were perfect and L = 0 (i.e., no load disturbances are prevent), then C
process out put from the set point. In equation, because en = r n - b n , note will equal em and Em': C - Cm ::: o.
that rndrops- out when 'tI ~ 00 (no dependence on set point).
The arrangement-shown in Figure 6.11 reveals that, although the
5.. Use of dimensionless controller gain fictitious process variable B is unavailable, we can get at B m in the model.
In commercial digital control system then the controller gain K, is The Bmwill be equal to B <unless modellig errors or load upsets are present.
usually expressed as a dimensionless number, if the measured variable W~ uee Bm,as the feedback, signal as shown in Figure 6.12. the difference

bn and the signal, to the final control element P n are computed in terms C - em is 'the error Em, which arises because of modeling errors or load
of percentage full range K, willbe dimensionless. This will also be true if upsets.
the computer input .and output have the same. units (Example rnA or V).
(DigitalController Design ( 6.49 )
[6.50. J Computer Control of Processes )

L
L

Fig.6.10: Desired c~nfiguration of the Feedback loop

Fig. 6.9: Conventional feedback loop .having dead time
L
The arrangement shown in Figure 6.12 will control the model well
but perhaps not the process, if load upsets occur or if our model is
inaccurate. To compensator for these errors, a second feedback loop is
c
implemented using Em, as shown in Figure 6.13.
This 'is the smith predictor control strategy. The Ge(s), is a c
conventional PI or PID controller, which can betuned much more tightly
because of the elimination of dead time from the loop. A block diagram of
the smith predictor drawn for computer control applicaitons is shown in
Figure 6.14.
The closed-Ioop transfer function of the system shown in Figure 6.13
for L = 0 is fig.6.11: Feedback arrangement incorporating a process model

C(s) . G c (s)G~ (s)e-eds To be assured of success, the model parametersmustbe known to a:

R(s) = l+Gc (s)G~ (s)~Gc (s)G~ (s)e-ems +G c (s)G~ (s)e-eds
... (6.116) high degree of accuracy.

rr L

G'm(s) = G'p(s)
c
and 8 m = ad; equation (6.116) reduces to.

C(8) = Gc (s)G~ (8) eeds .... (6~117)

R(s) l+Gc (s)G~ (s)
Equation (6.117) is the closed-loop transfer function of the desired
configuration, based on the fictions signal B, which was shown earlier in
Figure. 6.10.

Fig. 6.12: Preliminary Smith Predictor Sch~me

rDigital Controller Design 0 -_ (0 6 .51 ) ( 6.52 J Computer Control of Processes )

Fig. 6.13: Smith predictor control system

Fig- 6.14: 'Smith predictor block diagram for t-omputer control application
A comparison of equation (6.117) with the closed-loop transfer
From Figure 6.14, observe that the model output CM 0 is related to
function of the system of Figure 6.9 will show that the smith predictor
the input U by
strategy has "removed the dead time from the characteristic equation. This,
stability is' improved and the controller can be tuned more tightly.
... (6.119)-
As an example, let us consider the design of a smith predictor
algorithm for computer control of a first-order prcess .having dead time
Where,
-whose transfer function is given by
Z transform of model output
... (6.118) Z transform of input

1 - e- st
Where, Transfer function of zero-order hold - - -
s
Kp = Process gain
- o K e-9 dS
ad = dead time Transfer function of process, -..:.P_-
rs-i-I
t = Process time constant
Substituting the transfer functions in Equation (6.119) we get
The smith predictor scheme for this computer control application is
shown in Figure 6.14 .. The implementation of this scheme would require
that we develops equations for BM , eM and UK. ' ... (6.120)
(Digital Controller Design " [ 6.53.) ( 6.54. ) Computer.Control of Processes]

If we denote the integral .number of sampling periods in dead time Cross multiplying and inverting gives'
, as N, then the following equality holds
... (6.124)
ad = (N + (3) T ... (6.121)
From Figure (6.14) the error is computed from the equation
Where, f3 = a fraction
between 0 and 1. With this expression .for ed,
equation (6.120) can be written as ... (6.125)

Now that if the model is perfect and L is zero.

... (6.122) ... (6.126)

and the input to the controller will be based on

Taking Z transform of equation (6.122) gives ... (6.127)

[:_,_1_~
Thus, dead time has been compensated. The digital controller must .
C,M(Z)=
U(Z)
K p
Z-N(l_Z-l)Z-l
1- Z-l
e-(l-fl)T ]
!
... (6.123)
bea conventional control algorithmCi.e., P, PI or PID). If the controller is
based on Z transforms, it incorporates the effect of dead time, and the
. .' 1-etZ~
. . smith predictor algorithm is not applicable.
Cross multiplying and inverting gives an equation for the model
output containing true delay. 6.5.10 Internal Model Control
Thus, Internal model control (IMC) is a comprehensive model based des~gn
method developed byMorari and coworkers. The I~,1C method is based on
CM,K = A ZCM,(K-1)+A zk p (;3-l)Uk-(N-Z) +kp (1-1:)U
k-(N+li an assumed process model and leads to analytical expressions for the
controller settings. Both IMC method and direct synthesis method
.. ~ (6~124)
produce identical' controllers if the design parameters are specified in a
Where consistent manner. However, the IMC approach has the advantage that it
allows .model uncert aining and. trade offsbetween performance and
A2 = e-TIt robustness to be considered irra niore systematic fashion.
A3 = e- J3Th
The IMC method is based on the following simplified block diagram
Now, we need to get the model output containing no dead time. For as shown in figure 6.15.
this' case, .
A process model G and the controller output P are used. to calculate
1 . .:. e- sT K p ] the model response, Y. The model response is subtracted from. the actual
Z [ s. r s j I response. Y and the difference, Y - Y is used as the input signal to the
IMC controller, G*c. I~ general, due to modeling errors (G;tG) and
kp(1-Z-
.'
1 ) [_1_ --
l-Z-1
' 1. J
l~e-T/'t. Z-l
... (6.123) unknown disturbances (D:;t:O) that are not accounted for in the model.
(Digital Controller Design (6.55 ) ( '6.56 J Computer Control of Processes )

D The IMC controller is designed in two steps:

Step 1:

y The process model is factored as, G = G+ G_ ... (6.131)

Where G+ contains any time delays and right half plane zeros.
In addition, G is required to have a steady state gain equal to one inorder
(a) Classical feedback control to ensure that the two factors in (6.131) are unique'.
o
Step 2:
The controller is specified as
I------,r------+ y

... (6.132)
y

where f is a lowpass filter with a steady state gain of one.

1
f ... (6.133)

(b) Internal model control In analog with the direct synthesis method,
Fig. 6.15: Feedback centrel strategies t c is the desired closed loop time constant
From Figure (6.15) it can be shown that the two block diagrams-are r is positive integer (usually r = 1)
dentiealifcontrollers Gc andG*c' satisfy the relation
The IMC controller in (6.132) is based on the invertible part of the
G* process model; G_ rather than the entire proces model, G. If G had been
Gc = ' c ... (6.128) used, the controller could contain a prediction term e 9s (if G+ caintains a '
1- G:G
time delay 8), or an unstable pole (if' ,0-+ contains a right half plane zero).
Thus any IMC 'controller Gc* is equivalent to a standard feedback Thus, by employing the factorization 'of (6.131) and using a filter of
controller G c and vice' versa; From Figure (6.15 (b)) , the closed loop the form of (6.133), the resulting controller o', is guaranteed to be
relation for TMC is physically realizable and s,table. -

In general, the noninvertible part 'of the model, 0-+, places limitations
... (6..129)
on the performance that can be achieved by any control system. Because
the standard IMC design method is based 'on pole-zero cancellation, the
For a special case -of perfect model, G-= G ' IMe approach' must be modified for process that are open loop unstable.

:. Equation (6.129) reducesto y = G:GYsp + (1-G:G)D ...{6.130)

(Digital Controller Design, (' 6.57 ) ( 6.58 ) Computer ControloiProcesses )

b) First order Taylor series approximation

For ideal situation where the process model is perfect (G=G),
substituting (6.131) gives the closed loop expression: e-es == 1- as ... (6.139)
a) Substitute (6.138) in (6.137)
y =. G+ f . Ysp + (1- f G+ )D ... (6.134)

Thus the closed loop transfer function for setpoint changes is:
G(s) = ... (6.140)
y -
-'= G f ... (6.135) .
Ysp +

The IMC' and direct synthesis (DS) methods can produce equivalent Factor, this. model as
controllers and identical closed loop responses, .even when modeling .. .. .. _. ' e
errors are present. This equivalence occurs if the desired transfer function G = G+ G - where G+ = 1 - -s
2 .. 0 (6.141)

YJ. in (6.136) is set equal to y.

(y. Y in, (6.135). and G_ = K
... (6.142)
sp d. sp
(1 + ~s)(~S+l)
In DB.method,
Substitute (6.142) and (6.133) in (6.132) and set r =.1

Gc = ... (6.136)
(1 + ~S)(~S+l) ... (6.143)

,
l-(lJ Ysp d
K(t cs+l)

The equivalent controller Gc can be obtained from (6.128)

IExample'.' I
Use IMC design method to design two controller for the FOPTD ... (6.144)
(First order plus time delay) model given by ,

_ . Ke-es
G(s) = - , - .. ~ (6.137) Rearranged into the PID controller of
rs-tI

Assume thatf is specified by (6.133) with: r=l and consider two

approximation for the time' delay term:
G = x, (1 + ~
c
'tr
+ tDs)
S
... (6.145)

a) 1/1 pade approximation

e .' with
1- ~s .
e....es _ 2 .... '(6.138)
1+
e
-s
.')
(Digital Controller Design ( 6.60 ) Computer Control~fProcesses J
Digitally based feedforward control involves the extension of theory
't1 = -2e+ ' t of continuous transfer functions. Here we return to the block diagram of
sampled data, feedback control system and add the necessary components
. for feed forward control. It is assumed that the measured value of the
'tD .= ... (6.146) disturbance is available as a ,sampled signal.
As with continuous systems, we select Gr so that any disturbances
are cancelled (i.e., C = R = 0). This implies that C(n8t)= C at the sampling.
b) Repeating this derivation for the Taylor series approximation instants, but not necessary in between. Perfect control, such as is achieved
gives ~ standard PI controller: in principle with continuous feed forward control,may not be attainable.

Ke= .l[_
K
. 't c
't
+8
), ... (6.147)
Considering only the feedforward path (and ignoring the feedback
loop), the appropriate equation for perfect control (with R = 0, C = 0 and
L *- 0) is
LGL (Z) + LGt (Z) GF (Z)HGr Gp(Z) = 0 .: (6.148)
A comparison of (6.14.6) and (6.147) indicates that the type of
controller that is designed depends on the time delay approximation. The Solving for Gf to obtain the FF controller gives
IMC controller in {(6.~47) is identical to DB controller for the" FOPTD
model. ... (6.149)
Gt<Z) = LGt (Z)HGv Gp (Z)

6.5.11 ~igital Feed Forward Control 'I'his result is net as attractive as for continuous control, because
L . 'LGL{Z) GL{Z)
the load variable cannot be factored area cancelled (i.e~,). LGt{Z) of. G (Z)
Computer t

_. HG L LG L (Z)
in gen.eral. However, the ratio HG .may be used in place of LG (Z)
t . .t
where L is a step input or a sequence of steps (piecewise constant input).
I
I
This assumption removes the requirement that Equation (6.149) be
I
I revaluated for every different input L(Z). The resulting expression will
. I -

RJX~+
be reasonably accurate except when the load change is not well
:.~~- I
I c approximated by the piecewise constant function.
.1L _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ JI

If Gt(s) = Kt e-8t s and at =T~t then HGt(Z) = Kt Z-T

Assuming a piecewise load change, Equation (2) bec-omes

B
Gt<Z) = -HGL (Z) ... (6.150)
Kt Z-T HGv G p (Z)
Fig. 6.16: Block diagram for digital feedforward controller
[Digital Controller Design [ 6.61 ) ( 6.62 ) ComputerControlof Processes)

The feedforward controllerwill not be realized unlessBj, > (St + 8p ) , 'fWD MARKS QUESTIONS AND ANSWERS
where Sp is' the process time delay. An unrealizable controller is indicated 1. Define Z -transform.
bya positive power of z premultiplying Equation (6.150).
The Z - transform plays an important role 'in the communication
For effective feedforward control, the manipulated variable must engineering .and control engineering. It is used in the analysis and
be able to act on the process output before' the disturbance affects the representation of discrete. time linera shift invariant system. The Z
response through the load transfer function, hence the time delay - transform. provide, a method for the analysis of discrete time
restriction. systems in the frequency domain.

Another way to handle the approximation of Gr is to tune Gf in the 2. What ,is region of 'convergence (ROC)?
field for typical disturbances using adjustble parameters, such as 'done The Z - transform of a discrete sequence is an infinite power series,
for continuous control. In this way, the mathematical intractability of hence the Z - transform exists only for these values of Z forwhich
Equation (6.149) can be avoided. A lead-lag digitalmodelwith time delay the series convergen. If theF (Z) is 'a Z.- t.ransform of f(k) then the
is given. by region of convergence of F (Z) is the set of all values of Z, for which
F(Z) attains a finite values.

... (6.151) 3 State the Inrtfalvalue theorem.

If. f(k) is a causal signal and F (Z) exist then the initial value of the
signal is given by,
where K f, af, br, andNj can be tuned to give the desired compensation,
This digital controller is roughly equivalent to the lead - lag compensator f( 0) = lim F(Z); [F(Z) = Z{f(k)}]
z~oo
used in continuous feedforward control. '

We can easily implement a pure time delay in Gr. This is an advantage 4. State final value theorem.
for digital control because a time. delay term can' only be approximated If f(k) is causal and .stable signal and F(Z) exists with Z = 1 included
using lead-lag components for continuous feed forward control. in the ROC then the final value theorem is given by

Suppose Gr can be obtained from laplace' transform' models of the

transmitter and process. Once Gris computed, we can then discretize G:t<s)
to obtain Gt<z). Again 'some on line adjustment need to be performed, since The final value theorem can be applied only if FfZ) is analytic' for
discretization errors will be present.
IZI > I
However qr designedin this way should be a close approximation to 5. What is the relation between Sand Z domafnfs.
the more rigorous version present in Equation (6.149).
1
The transformation S = T In < maps the S - 'plane into the Z - plane.
Every section of jw axis of length NWs of S - plane maps into the
unit circle in the anticlockwise direction, where N is on integer and
Ws in sampling frequency. Every strip in S - plane of width Ws maps
into the interior ofthe unit circle. The mapping of S - plane is to
Z - plane is 'shown in figure.
(Digital Controller Design (6.04 ] Computer Control o~ Processes)

A strip L with width

jw
lmaqlnary 'z'
+j1 '
9. Major Disadvarrtages of digital feedforward controller.
Requires identification of all possible disturbance.and their direct
measurement.
Cannot used to unmeasured disturbance.
Semitive to, process parameter variation.'
-jw s
~~~"""""~~
Requires good knowledge of the process', model.
, -j1
10. State closed loop specification for Dahlin's eontz-olatgor-itbm.
S - Plan Z - Plan The time constant is a design parameter whose value is to be
6. Define one sided and two sided Z - transform. selected by the designer according to the desired closed-loop
response. Thus we choose low' value of time constant to get faster
The one sided Z - transform of f(k) is defined as the powre series.
co
response,' while larger values of time constant yield more sluggish
F (Z) ., Z [f (k)] = L f(k)Z-11 closed loop response.
k=O 11. Ringing.
The two sided Z. - transform of discrete sequence f(k) is defined as
the' power, series. The successive changes in the .sign .of a controller output is called
00 ringing and is caused by negative poles. The closer a ringing pole is
F (Z) = Z [f (k)J = 'L f(k)Z-ll
k =-co
to unit 'circle, the higher the ringing of the controller output will be
sluggish in nature.
where, Zis a complex variable.
Method to eliminate ringing
7. Write the, inverse Z - transform typer?
We should design a digital controller so that all its poles ,are positive.
Partial fraction expansion,
Suppose that for a given process Dahlin's method leads to a
Power series' expansion'. controller, with transfer function.
Direct evaluation by control integration (or) complex inversion integral. 1
'K 1.05Z-:-
3. What is the equivalent representation pube sampler with Z04? nrz: = (1 + O.6Z ~ 1) (1-Z-1) (1-"-0.7 Z-l)
The. pulse sampler with Z04 shown in figure (a) can be replaced by The term (1 + 0.6 Z-l) -in the denominator implies the presence, of a
an equivalent system consisting of an impulse sampler and a block pole at -0.6 cause ringing. Dahlin's suggested Ringing pole can be
with transfer' function. eliminated by multiplying ncz) with (1 + 0.6 Z-l), thus cancelling the
,, '1 - e-st ringing term from denominator. -To keep static gain k the same we
Go(s)= ,8 as shown in figure' (b) should divide n(Z) by the term lim ( 1 + 0.6 Z-l) ~ 0.6. Thus ringing
PT(t) STet) for controlles has the transfer function.
~~ 1~se-ST~
~~, Pulse
ZOH' ~,Frs,(t),
, ImpUIS~~
sampler
~
sampler
(a) (b)
 7.2 .] computer Control of Processes )
CHAPTER
Multiloop Control

7 Each manipulated variable depends on only a single controlled

variable, i.e., each conventional feedback controller will control each
output variable.

MULTI-LOOP REGULATORY -------:.-y_1 ....

CONTROL Proce-:s

7.1 INTRODUCTION fig. 7.3: Multiloop process control system (2 x 2 system)

.Process that have only: one. controlled variable and one manipulated
Multivariable Control
variable iscalled.Single-input, Single-Output (8180) process. This iscalled
Single-loop process control system. , Each manipulated 'variable can depend on. two or more of'the
controlled variables. A single controller will control all the output variables.
Examples:
Decoupling control, Model predictive control
Process .

-----y1
Fig. 7.1: 5150 Controi system 1-----.----....
In industry, many process may have more number of output variables Process
which can be controlled and may have more number of input variables
which can be manipulated. Such a system is called Multi-Input,
Multi-Output (MIMO) process.

Disturbances Fig. 7 . 4:2 X 2 Multivariable pr.ocess control system

1.2 MULTILDOP CONTROL

The multiloop approach, using n.ultiple single-loop controllers, was
the "first approach used' for MIMO control in process industries. One
advantage of multiloop control is the use of simpel algorithm (easy to
Fig. 7.2: MIMO process
implement). A second advantage is the ease of undetstanding by plant
The MIMO process control is further divided into Multiloop process
operator, which results from the simplicity, . of the control structure.
'control system and Multivariable process control system. \
.(Multi-Loop Regulatory Control ( 7.3 ) 7.4 Computer Control of Processes )

Since each controller uses only one measured controled varibale and
adjustes only one manipulated variable, the actions of the controllers are
relatively easy to monitor. A third advantage is that hardware and Xsp
software are readily available for multiloop controllers (using multiple SISO
controllers).
A MIMO process is said to have interaction when process input
,<manipulated) variables affect more than one process output
(controlled) 'variable. Fig. 7.. 5: Blending process
In most industrial MIMO process, each manipulated variable (input
signal) may affect severalcontrolled variables (output signals) causing Loop 1
interaction between the input/output loops. For this reason, control of 1- - - ~ - - _. - - - - - - - - - - - - - - -

MIMO process is typically much more difficult compared to the 8180

-----_ .... __ ..
Loop 2
I '
process. It is therefore of great importance to quantity the degree of
, interaction so that proper input/output pairs that minimize the impact of Xsp
the interaction can be formed. For this, dedicated interaction measures
canbe used.
Compared to SISQ. process, the control design for MIMO process is
more elaborate. One reason for this is, as mentioned above, that different
parts of a MIMO process 'may, interact and cause couplings in the process.
Although most industrial processes are 'MIMO,SISO, control loops are (a) Y1- u 1 and Y2- u 2 pairing
formed in MIMO process by selecting a measured output that is most
strongly affected by a particularmanipulated input.
Loop 1
~----------------

7.3 PROCESS INTERACTION .-- - - -l-Loop"2- - - - - ~

A specific example of 2 x 2 MIMO process is shown in Figure 7.5. I
I u, Xsp
Here, two different streams are blended together. The objective is to ,
I
manipulate individual stream' flow rates to meet specifications on the

I
output flow rate and output composition. Let the outputsbe represented
by Y2 (total flow) 'and Yl (composition), and the inputs by U1 (flow rate of
streamFg) and U2 (flow rate of stream FB).
How should outputs and in~ts be paired? If _we assumethat the
output-input ,pairings are Yl"U1' and Y2-U2, the instrumentation diagram is (b) Y1-U2 and
.
Y2 -u 1 pairing
, ,

shown in Figure 7.6(a)., If we assume that the output-input pairings are Fig. 7 . 6: Blending system alternatively closed loop pairings
Yl"u2and Y2-U h the instrumentation diagram is shown in Figure7.6(b).
(Multi-L{)op Reguiatory Control' 7.5 Computer Control of Processes)

We can see the potential problem in the operation of the control

Diluted ethanol concentration Y1 Total output'flow rate Y2
strategy in either Figure 7.6 a or Figure 7.6 b. For discussion purposes,
consider Figure 7.6 a. A setpoint change in Yl (concentration) will cause 0.32
4.6
the concentration controller to change Ul (FA flow). ,-
0.31 1-,.....---------.. .
'I ' 4.4
The change in FA will aslo affect Y2 (total flow). Once Y2 is disturbed,
>: 0.3 .;: 4.2
the flow controller will cause a change in u2,(FB flow),
0.29 4
The U2 will disturb the concentration. causing the concentration
controller to change Ul (FA flow), and we
are back to where we started. 0.28 3.8 ........._...r....._...1.-_~_.....L..._...J
The effect of one control loop on the other is referred as control loop o 2 4 6 8 10 o 2 4 6 8 10
Time,min ---+- Time,min~
interaction.
80 % concentrated ethanol flow rate u1 Pure water flow rate u2
IExample 7.1 I 1.5 r-----r----,.--.......---------
Consider now a specific blending problem to see whether physical 3.6

intution can be used to determine the proper pairing. 3.4

N

Assume that a control engineer has decided to blend a concentrated :r 3.2 ::r

ethanol stream of 80% concentration with pure water to produce a diluted

3
ethanol stream with 60% concentration. -
2.8 0.5 _ _ ""--_~_-J- _ _L__....J
Assume that the control objective is to meet both qual ity and 0 2 4 6 8 10 o 2 4 6 8 10 -
production.rate requirements. Assume that the 80% con~entrated 'ethanol Time,min -----.. Time,min - - .
flow rate (FA) is the first manipulated input (U1) and pure water flow rate
(FB) is the second manipulated input (U2). Fig. 7.7: Response to diluted ~f-honal concentration set point change:
Loop 1 has diluted ethanol eeneentretlen as the measured out~ul '11
The blended product (diluted ethanol) flow rate (F) is the output Y2 and 80% concentrated ethanol flow rate
and blend concentration (x) is output Yl. ' as the menlpuleted input u 1 ,Loop2 is open.
Your intuition is' probably to pair output Yl (diluted ethanol
Assume that it is desirable to maintain the total output flow rate Y2
concentration) with input ui (aO% ethanol flow rate). This control strategy
at a constant value. This requires that loop2 be closed. Loop2 has total
is shown in FIgure 7.6 a. Intr-oduce a setpoint change in ethanol
output flow rate as the measured process output (Y2) and pure water flow
concentration andobserve the response.
rate as the' manipulated input (U2).
In this case Loopl is closed and Loop 2 is open. From the response
Assume' that it has been tuned independently from loopl, i.e, loop!
as shown in Figure 7.7, it is observed -that the closed loop behaviour of
is open. The response to a setpoint change in total output flow rate' is
ethanol concentration is quite good, but that the total output flow rate Y2
shown in figure 7.8. Notice that the total output flow rate response is
derivates, because loop2 is open.
quite good, but the diluted ehanol concentration deviates from desired
value, because loopl is open.
( Multi-Loo~RegulatoryControl I 7.7 7.8 Computer Control of Processes]

Diluted ethanol concentration Y1 Total output flow rate Y2

Diluted ethanol.concentration Y1 Toteloutput flow rate y 2
0.4--.---------.;...----... 10

4.6 0.35 8

4.4 0.3 C'\l6

>: >.
N
>4.2 0.25 4

4 0.2 2

3.8 0.15-+--...,...--....,....--------1 0
4 6 8 '10 0 2 4 6. 8 10
o 2 ,4 6 8 10 0 2 4 6 8 10
Time,min --+ Time.min ---Ito-
Time,min ---+ Time,min ---+

Pure water flow rate u 2

80 % concentrated ethanol flow rate u1 Pure waterflow rate u2
80 % concentrated ethanol flow rate u 1
1 0 - r - " " - - - - - - - - -.... 3-.-----...-...------....
I

3.6 ,8 2.5
2
3.4 ~6
:::s ~ 1.5
::r 3.2 4
1
, 3 I--------------~ 2 0.5
o -+-_~----....,---.l-J'_t_-.u. __-J...I_I O-+---,-I-L-.a.,--..1...,-01...........1,---1
2.8 L.-_oI...-_oI...-_'---_L..-..~
o 2 4' 6 8 10 2 4 6 '8 10 o 2 4 6 8 .10 o 2 4. 6 8 10
Time,min - + Time,min --+
Time,min----ll> Time,min ----+
Fig. 7.. 9: Response to diluted ethanorc~ncentrafion setpelnt change
Fig. 7.8: Resoponse to total output flow rate setpoint change: with both loops closed. Lo.op2 pairs Y2: Uland leep 1.pairs Y1- u2
Loop2 has lotal output flow rate as the measured oulpul Y~
and pure water flow _rate as the manipulated input u 2 Loop 1 is open. If you select variable pairing as Yl - U2 and Y2 - ,lib would this system
be stable?
We have seen that the individual control loops have quite good
performance. Since We desire to control both the diluted" ethanol The response .to step increase in 80%- concentrated ethanol flow rate
concentration (Yl) and total output flow' rate (Y2), we need to have both Ul under two conditions:
loops closed simultaneously. The response to a setpoint change in diluted 1. With the flow controller (loop2) open (solid)
ehanol concentration with both loops closed is shown in Figure 7.9.
2,. With the flow controller closedIdashed) are shown in Figure 7.10__
Notice that the closed loop system is unstable. What characteristics
From the response we can understand that the closed loop system is
of this ,control strategy caused the instabilityto occur when both loops stable. -
,are closedr. W.9uld this behaviour have occured ifvatiable pairing of
Yl-Ul and Y2-U2 have been used?
From the responsesshown in Figure 7.9 and Figure 7.10, we can
understand that selection variable pairing YI-U2 and Y2.;.Ul will make the
blending system to be stable whereas selection of variable pairing YI-Ul
and Y2-U2 will make the blending. system to become unstable.
[Multi:.toopRe,gulatory'Control 7.9 ( 7.10 ) ComputerControl of Processes, )

[~xample 7.2
Diluted ethanol concentration Y1 Total output flow rate Y2
Interaction of control Ioops in Stirred T.ank Reactor
0.33
4.6 In the CSTR of Figure 7.11, the temperature is controller by the
/ ' - - Flow loop open
0.32 I - - ~ Flow loop closed flow of coolant in the jacket (Loop 2) while the efficient concentration is ..
I
4.4
I controlled by the inlet feed flowrate (Loop l), Assume that initially both
~ 0.31
~ 4.2
1~---------f effluent concentration and temperature are at their desired values.
0.3 r,
4 ,-------------- FtGAi'
Feed
,0.29
3.8 t---~-....,--__-__..-____J
o 2 4 6 8 10 o 2 4' 6 8 10
Time,min ---+ Time,min~

80% concentrated 'ethanol flow rate u 1 Pure Waterflow rate u2

1.5
3.6 1-- Flow loop open
.; - - Flow loop closed
II
3.4

(/_l..---....
::r

~AI",,----,"---'
3.2 "
Coolant
,
\
3.0 --------_.-._---- Effluent
..
2.8 1-----,.----.-----...--_-~ 0.5 I
o 2 4 6 8 10 o 2 4 6 8 10 Fig. 7.11: Control loops in CSTR
Tirne,min ---...: Time.min ~
1. Consider a change in inlet concentration (load change) or the
desired effluent conceritration (setpoint change). Loop I will
Fig. 7.10: Response of diluted ethanol" concentration to step change:
compensated for these changes by manupulating the feed flow
Loop2 open (solid curve) and loop 2 closed (dashed curve).
rate. However, this change in the feed flow rate also disturbsthe
Hence choice of pairing is crucial since, a bad choice" will make the reactor temperature away from the desired value. Then Loop '2
closed loop system to become unstable, This problem could arise due to attempts to compensate for the change in temperature by varying
interaction' between the different loops. ,the coolant flow rate, which inturn affects the effluent concentration.
Eventually there is a need for a measure that can both give some 2. On the otherhand, an attempt to compensate for changes in feed
advise when solving th~ pairing problem and the also quantifies the le~el temperature (load change) or the desired setpoint of "reactor
of interaction occuring in the .system. , temperature (setpoint change), it 'also causes the effluent
concentration to vary. The loopl attempts to compensate for the
change in effluent concentration by varying the feed flow rate,
which in turn disturbs the reactor temperature.
It is clear form the above that, loopl interacts with loop 2 and also
loop 2 interacts with loopI (i.e., in both directions).
(Multi-Loop Regulatory Control .[ 7.11') ( 7.12 ) Computer Control of Processes )

The input output "relationships are given by

1.4 INTERACTION Of CONTROL LODPS
Yl(S) gIl (8) UI (8) + g12 (s) U2 (s) (7.1)
Consider a process with two controlled outputs and two
manipulated inputs as shown in Figure 7.12 (a). Y2(S) g21 (8) lil (8) + g22 (s) U2 (8) (7.2)
Where,
gIl (8), g12 (8), g2l(S), g22(S)- four transfer fu.ions relating two
outputs [Y1(s) &Y2(S)] and two inputs (Ul(S) and U2(S.
Equation (7.1) and (7:.2) indicate that a change in Ul or Uz will affect
both the controlled outputs.
. Let us from two controlled loops by coupling mj with Yl and m2 with
yz as shown in Figure 7.12 (b). For simplicity, we assume that the transfer
functions of the measuring devices and final control elements in both loops
are equatl to 1. IfG C1(s ) and GC2 (8) are the transfer functions of the two
controllers, the values of the amnipulated variables are given by,
Process Ul(S) = Ge1(s) [Y1sP(s) - Y1(s)] (7.3)
(aJ. Open loop li2(S) = G e2(s) [y sp (8) -
2 y /8)] (7.4)
To understand the nature of interaction between two control loops and
Loop 1 how it arises, we will study the effects of input changes on the outputs when
1. One loop is closed and the other loop is open.
2. Both loops are closed.
1. One loop is closed
Assume that loop 1 is closed and loop 2 is open as shown inFigure 7.13 (a).
Also assume that u 2 = constant (i.e u 2(s) = 0), and make a change in the set
point Y1SP (s), After substituting equation (7.3) into equations (7.1) and (7.2),
I" we take "

gIl Gel
1---_ _...1...-. . . . . . Y2 (5) Y1 = Yisp ... (7.5)
1+ gIl Gel

Process
g21 Gel
Yz = 1 + gil Gel
Yls p ... (7.6)
"Loop 2
(b) Closed loop From the above equation, it is clear that any change in the set point
y 1 sp will affect not only controlled outpu y 1 but also the uncontrolled
.fig. 7.1.2: Block diagram of 2 x 2 process output Y2.
rMulti~LoopRegulatory Control ('7.13 ) ( 7.14 ) Computer Control of Processes )

Loop 1

Y1 sp

fa) One loop Closed

U1 Direct effect
:: - - -+- - - - ,,-+----+-+ Y1
.... ..... -' "
.... .".
..... ./

/./'"
"',"
./ "
..... -1
I
I
I Indirecteffect :
--------------+-------------
(b) Both loops Closed

Fig. 7.13: interaction among control 'loops

The dashed lines in Fi~re 7.13 (a) indicate schematically the routes
through which Ylsp affects the two outputs. Simillar conclusions are drawn if
we consider loop 1 open and loop 2 closed. The situation become more complex
when both loops are. closed.
..~ (7.7)
2. Both Loops Closed
This case is shown-in Figure 7.13 (b), Initially the process is at steady
state with both outputs at their desired values. Consider a change in the Yis) = g 21 (s) Gel [Ylsp - Y1 (s)] + g 22 (s) G e2 [ Y2sp (S) - Y2 (s)] '
setpoint Y1 sp only, and keep the setpoint of loop2 the same (i.e Y2sp = 0).
Y2 (s) +[1 + g22 (s)G c2 ] +g 21 (s) Gel Y1 (s)
g21 (8) Gel Ylsp + gZ2 (S)Ge2 Y2sp (8) ...(7.8)
Solve equation (7.7) and (7.8) with respect to' controlled outputs y 1 and
Y2 and take the following closed loop input-output relationships:
(Multi-Loop RegulatoryControl . ( 7.15)
( 7.16 ) Computer ControlofProc~sses J
1~5 STEADY STATE EFFECTIVE GAIN-llf TWO INPQI AN,D TW,D OUTPUT
,y1 =' Pi l (s) Y1 sp + Pl 2 (s) Y2 sp (7.9)
!
PROCESSES
. Y2 = P21 (S)YI sp +P22 (S)Y2sp ... (7.10)
Let us consider the relationship between Y1 - u l and Y2 - u 2 under a
Where;
number of conditions for a 2 x 2 (two input - two output) process as shown in
= gil Gel + Gel Gc2 (gil g2~ - gl2 g2l) Figure 7.12 (a).' The' open loop input-output relationships are
'Q(s)

... (7.13)

which is written in matrix form as

g22 G c2. + Gel Gc2 (gIl g22 - g12 g21) ... (7.14)
Q(s)
"'. (7.11) The corresponding feedback control system, if pairing are' lil - Yland
liz - Y2 is shown in Figure 7.12 (b).
Remarks:
, 1. WIlen Hi.2 = H 2l = 0 (Refer diagram 7.1) in equation (7.7) and (7.8), The open-loop steady state relationship for ,a 2 ?C 2 process is
there is no interaction between the two loops. The closed loop outputs
are given by, the following equations.
H' (7.15)

... (7.12)
a) Steady state effective gain between u 1 and Yl with loop2 closed
Tuning, two control loops without interaction is' very simple, and we that is Y2 = 0 (Refer Figure 7.14).
can tune each loop separately.
2. Equation (7.9) and (7.10) describe the response of outputs Y1 and Y2
when both loops are closed (i.e they have accounted for the interaction
between two loops). ,Tuning two control loops with interaction is very u2 = ...-(7.16)
difficult and we cannot able to tune each loop seperatly.
3. Interaction between control loops is a. significant factor which affects and we can then find the following relationship from equation (7.15)
the "goodness" of a control system. By a suitable pairings 'of
manipulatd variables (u, and u 2) and controlled outputs (Y1 and Y2)'
'we can minimize interaction between two control loops 'and then we
can easily tune the controllers. If the interaction between the two
control loops is more, we have to design special control system which
eliminates the interaction (i.e decoupling the loops).
(MultiLoop Regulatory Control
.@s] Computer Control-of Processes ]
[ 7.17,)
From equation (7.15),

Y1, ~ (KI I - Kl}C22

K JUl
2 2I Y1 = K ll u l +'~2 u 2
x, u l + ~2(O)
Y1 = ~l eff U l (7.17)
YI -- u
..L""11 u I ... (7.18)
This is steady state effective gain relation betweenu, and YI with the,
loop2 closed. This is the steady state gain relation between u l and Y1 with loop2 is
open,
~--~-~-~-~--~-~--~-~----------, c) Steady state effective gain between u 2 and Y2 with loop! closed.
"

I
Let us ass~e perfect control of YI (loop is closed) that i~ Y1 '=' O.
J
u l + ~2 u 2 ~ 0
I
I Y1 = Kn
I
I
I
I u1 = , ... (7.19)
I
I
f From equation (7.15), we can write
I
I
I y 2 ,= ~l
+ ~2 u 2 UI , ... (7.20)
I-
I Substitute equation (7.19) in equation (7.20)
J

= x, (~ KK12 U2 J+ K 22 U2
I
~~-------------~---------~----3 Y2
11
911 eff

Fig. 7.14: Block diagram of Yl - u 1 relationship with loop2 closed. Y

, .2
= (K 22 K12K21 JU 2'
. K 11
.. (7.21)

b) Steady state gain between u 1 and 'Y1 with Ioop,2 open

This is the steady state effective: gain relation between u 2 an~ Y2 with
'lfthe'second inputu, is c~nstant{loop2 is open), then 'output Yl depend loop 1 closed.
only on the first input u l as
shown in Figure 7.15. '
r-~-----------------------------
I
1-

1------.,.. Y2 ~ ~
............. oIIIIII!II ..... _,
"'""'"'"----.
~ _ -.-,;. ~ - _ - - -.. -'- - - -
I
-.--..

Fig. 7.1-5: Block diagram of loop 1 and loop 2 open with u2 =0 ,Fig. 7.16: Block diagram of Y2 -u 2 relcltionship:'with loop 1 closed
 ( 7.20 ) ,Computer Control of Processes )
[Multi-Loop Regulatory Control ( 7.19)
From equation (7.15), we can write
d) Steady state gain between u 2 and Y2 with loop! open
If the first input u l is constant (~~opJis',open), then output Y2 depend
Y1 = K n u 1 + K12 u 2
only on the second input u 2 as shown inFigure 7.17. K n (0) + K 12 U
2

-. K 12 U2 ... (7.23)
This is the steady state gain relation between u 2 and Y1 with loop 1
is open. -
f) Relative gain between u 2 and Y1 with loop 2 closed ,

Let us assume 'perfect control of Y2 (loop 2 is closed) that is Y2 =0

u =0 - - . - - - - - - - . . - - -.....
I
I
Fig. 7.17: Block diagram of loop1 and loop2 open with u 1 =0 I
I
!
From equation (7.15)' I
I
i
Y2 = ~1 u 1 + ~2 u 2 - i
!
K 2l (0) + K 22 u 2 I
I
I
~2 u 2 (7~22) i-
r
This is the steady state gain relation between u 2 and Y2 with loopl I
I
open.
I
e) Relative gain between u 2 and Y1,with loop 1 open 912eff

If the first inputu, is constant (loop 1 is open) then output Y1 depend Fig. 7.19: Block diagram of Yl- u2 relationship with I~op 2 closed
only on th~ second input u 2 a~ shown in Figure 7.18.
From equation (7~15), we can write
y2 = K21 U l + K22 U2 =0

From equation (7.15), we can write

y1 = Kn Ul + K l2 U2
1 - - - - - - + Y2
Y1 = K n (- K
22
K 21
U2 J+-K12 U2
Fi~. 7.] 8: Bloc~ diagrarno' Yt-U2 relationship with loop 1 _open
 ( 7.22 ) Computer ControlofProc~sses ]
(Multi-Loop Regu!atory Control ( 7.21]
h) Relative gain between u l and Y2 with loop! closed

Y1 =(
~
KIt K 22 + K 12
K 2I
J
U2
Let us assume perfect control of Y1 (loop 1 is closed) that is Y1 = 0

~-------------------------~----~
I I
: I
1-
: ~

Y1 =- K 12 eff u 2 ..:(7.24)
This is the steady state effective gain relation between u 2 and- y 1
with loop 2 closed.
g) Relative gain between u l and Y2 with loop 2 open -----~--~----~-~----------~-~~~
. I
I
If the second input u 2 is constant f loop 2 is open) then output Y2 921 eft
depend only on the ,first input u l as shown in Figure 7.20.
Fig. 7.21: Blockdiagram of Y2- u t relationship with loop 1 elesed

From equation (7 .1~), we can write

Y1 = Ku Ul + K 12 U2 = 0

u 2--

From equation (7.15), we can write

1------. Yz y 2 = ~1 UI + ~2 U2

Fig. 7.20: Brock diagram of Y2- u t relationship with loop, 1 open

From equation (7.15) we can write

Y2 = ~1 u l + ~2 u 2
~lUI + ~2 (0)
Y2 = ~l u l
This is. the steady state gain relation betweenu1and Y2 with the
loop. 1 open.
 (. 7.24 ) Computer Control ofProcess~s] ,
(Multi-loop Regulatory Control. ( 7.23 )
b) Relative gain between input 1 and output 1 for a 2 x.2 process
The relative gain between input land output 1 is
Gain between HI and Y1 with U2 constant (ie loop 2 'is open)
Y2 = .....~l eff .u l Gain between lil and YI with Y2 constant (ie loop 2 is closed)
This is the steady state effective gain relation between u, and Y2
with loop 1 closed. - ( ~ JU2 = constant
7.& RELATIVE GAIN ARRAY-AND SELECTIONOF VARIABLE PAIR lJVITH (
OJI JY2 = constant"
aUI .
MINIMUM INTERACTION
(Refer equation 7.18)
For a 2 x 2process (2 controlled outputs and 2 manipulated inputs) = (Refer equation 7.17)
there are 2ldifferent input-output variable pairs are possible. The four
different input-output variable pairs are u 1 - Yl ' u 2 - Y2 , u 2 - y1and U 2 - Y2: %
Which .one is best? one way to answe.r this question is to consider the
interactions between input - output variable pair of ali 4 combinations
and select one where the interactions are minimaL
1
The Relative Gain Array (RGA) is a heuristic, technique to predict K I I K 22 - K I 2 K 21
possible interactions between input-output variable pairswhen multiple K I I K 22
8180 loops (each loop is constructed using one input and one output
"variable) are used in MI1VIO process. It was first proposed by Ed -Bristol
... (7.28)
and today is a popular tool to determine optimuminput-output pairs of a K I I K 22 - K l 2 K 21
MIMO process. c) Relative gain between input 2 and output 1 for a 2 x'2 process"
a) Definition of the Relative Gain The relative gain between input 2 and output 1 is
The relative gain (Ai) between input j and output i is defined in the Gain betv/ee,n U2 and YI with-u'l constant (ie loop 1 is open)
following fashion: ' Gain between U2 and y 1 with y 2 constant (ie loop 2 is 'closed)

(8ylJU1= constant
Gain between inputj and output i with aU"other loops open
A..
I;
Gain between input j and output! with all 'other loops closed OU2

(~)UK =constantK;tj
A..
IJ
... (7.27) K 12 (Refer equation 7.23)
= (:: )YK = constant'K;t i =
. -[ KllK22~2~12K 21}Refer equation 7.24).
-K12 K 2I
... (7.29)
(Multi.loop Regulatory Control ( 7.26) Computer Control of Processes )

d) Relative gai:il between input, 1 and output 2 for a 2 x 2 process f) 'Relative Gain Array (RGA)

The relative gain between input. 1 and output 2 is Arrange the four relative .gains "'12' A21 an~ "'u' "'22 (for a 2 x 2 process)
into a matrix form, which is known as RGA.
Gain between Ul and y 2,' with u2 co;nstant (ie loop 2 is open)
Gain between Ul and Y2 withy, constant (ie loop 1 is closed)

( Oy2
BUI
J
U2 = constant
'

= (Oy2 JYl == constant g) Propertfes of RGA

Oul'
The sum of the relative gains -in any row or column of the array is
, K 21 (Refer, equation 7.25) equal to 1. Thus for a 2x~ process.
A 11 + A 12 1 A 11 + tv 21 1
.and
A 21 + A 22 1 A 12 + A 22 1
For a 2 x 2 proces, we need to know. only one of the four relative
... (7.30). gains, while other three can be easily computed.
- [K I I K 22 - K 12 .K 21 ]
e) Relative gain between input 2 and output 2 for a2 x 2 process
IExample 7.3 I
The relative gain between input 2 and output 2 is
For a 2 x 2 process, if A 11 = 0.75, find other three elements of RGA.
Gain between U2and y 2 with UI constant Cie loop 1 is open) Since the sum of the relative gains in anyrow or column or the' array
Gainb'etween U2 ,and Y2 with Yi constant (ie loop 1 is closed) is equal to 1.

(~ J Ul = constant
All + 1...12 = 1
0.75 + A,12 =1
A12
0.25 +
+ A22
"'22
=1
=1
1...21 + =1
A22
"'21 +0.75 =1

A12 = 0.25 .. A22 = 0.75 .. A21 = 0.25

(:: JYl =constant
Ul U2
K 22
K 11K 22 - K 12 K 2l
(Refer equation 7.22)
]
:. The RGA x=
,
'[75 0.25] Yl
0

0.25 0.75 Y2
[ (Refer equation7.21)
KI I
Note:
... (7.31) For a 3 x 3 process, only four of the nine elements would need to be
calculated.
[Multi.Lo()p Regulatory Control [ 7.27 ) ( 7.28 ) Computer Control of Processes)

b) Interpretation of the .relatfve gain IA

U = 0.51 Then the relative gain array is
Depending on the values of All' we can interprete the following information,
til u2
I Au I = 1 Then the relative gain array is 0.5 0..5] Yl
[ 0.5 0.5 Y2
Ul u2

1 O]Yl Since all the elements in' RGA are equal" the amount of interaction
[0 1 Y2. between the two loops is the same in both configurations of Figure 7.22
and Figure 7.23. We can use two possible different combination of variable
The 1's in the diagonal elements indicate that we can form two non- , pairing.
interacting control loops by coupling u l with YI and u 2 with Y2 as shown in Combination 1 : Yl - u l and Y2 - u 2
figure 7.22.
Combination 2 : Y1 - u 2 and Y2 - ul
The degree of interaction remains the same in both combination.

2x2
I O<An < 0.51 (say Au = 0.25). Then the RGA is
Process Ul u2.

I\, =
[0.25 O.75]Yl
0.75 0.25 Y2
The two larger numbers (i.e 0.75) indicate the recommended variable
Fig. 7.22: 2 X 2 proces with Y1 - u 1 and Y2 - u 2 paring pairing with the smaller amount of interacti?n. Thus we can couple u l
with Y2 and u 2 with Y1 as shown in Figure 7.23.
I
Au = 0 Then the relative gain array is
I O.5<A.n < 1 I(say An = 0.8). Then the RGA is
Ul U2

A = [o.?
0.2]Yl
0.2 O.S Y2 ,
The 1's 'in the off-diagonal elements .indieate that we can form two non- In this case also the. two larger numbers '(ie.O.S) indicate the
interactingloops by coupling u l with Y2 and u 2 with YI as shown in figure 7.23. recommended variable pairing with the smaller amount of interaction.
Thus we can couple ill with Y1 and u 2 with Y2 as shown in figure 7.22. This
is opposite to previous case.

Y1 SP
2x2
AU >1 I
Then A22 = Au > 1 and A12 = A21 =1 - An < 0 (say An = 5).
Process Then the RGA is,
Y2SP

(Multi-Loop Regulatory Control (' 7.29 ) ( 7.30 ) Computer Control of Processes )

In this case we can couple u 1 with Y1and u 2 with Y2as shown in Hence we can pair Yl with .0 1 and Y2 with u 2i n this case because their
figure 7.22. The larger the values of the relative gains above unity, the corresponding gains (0.95, 0.95) are positive and closer to unity.
"larger the "holding back" effect will be. Thus we need larger values for Note:
the controller gains. .
U1 U2
Note:
However, if A 0 .05 ' O.95]Yl
Never forms variable pairing by coupling inputs to outputs with [ 0.95 0.05 Y2
negative relative gains.
In this case, if you couple y1with U 2 and Y2 withu., the corresponding Then we would pair Y1 with u 2 and Y2 with u 1
gains, A. 12 and A21 are negative. Then the interaction will take the controlled
outputs in the opposite direction from that desired by the controlled effort IExample 7.51
and control will be lost altogether. - Consider a blending process with steady state gain matrix as shown
We can summarize all the above observations with the following rule below.
for selecting' the input-output variable pairs.

A...
IJ
Possible pairing

A..
IJ
= 0 Avoid 'pairing yi,with u j How would you choose input - output pairings, for this process.
pr~cess gain matrix isgiven as
A...
IJ
=- I Pair s. and u j
The steady state'

A... < 0 Avoid pairing Yi withuj " K = fK.'11 K12 J = [0.025 -0.0751
IJ
LK 21 K 22 . 1 1 J
A...::;
IJ
0.5- Avoid pairing Yi with u j
The RGA can be calculated from K matrix as
A...
IJ
>-1 Pair v. with u j

IExample 7.4 I
Determine optimum- input-output pair for a 2x2 process whoseRGA
K 11"K22 - K 12 K 21 K 11 K 22 - K 12 K 21
is,
. -K 21 K 12 K 11K22
Ul U2

A = [0.95 ,O.05]Y1'
U1 u2
0.05 0.95 Y2
0.25 0.75]Y1
We can select an optimum input-output variable pairs in such a way [ 0.75 0.25, Y2
that the relative "gains are positive and as close as possible to' unity. \ .-

The RGA indicating that the input-output pairing should be y 1 - U2

and Y - u -- '\ ' -
2 t
 [ 7.31 ) ( 7.32.) Computer Control df Processes )

Substitute' uz(s)value in equation (A)

1 1 0.2 8 +1
Y1(8) = -.-ul8+.
8+1
- ---
0.1 s-r I 0.8 0.58+1
Ul(8)

... (A)
1
Substitute uI(s)= - and find the final steady state value of Y1(s) .as,
s .
, -0.2. ( ) . 0.8 ()
Yz (s).= ---Ul s +--U2 s ... (B)
0.5s+1 s+1. Y1ss(S) ;::: lim(s Yl (s))
s~o

How would you choose. input-output pairings for this process?

Let us compute relative gains:.
Step 1:
Make a unit .step change in u l (L e, u~( s) = 1/ s), w hile keeping
(
Byl J . 1.25
1 --
'1.25
U
z = constant (ie.,
U z == 0). Then from equation (A), we can get au! Y2 = constant

'Step 8:
yes) = -.1_(!J+O
1 8 +1 8 Determine All
By using final value theorem, we can find new steady state value of

~
Y1 as,
( 12=constant _1__ R
;::: limi_1_.~ = 1.25 - O.v
040 (s + 1) (t) 1 .(fJyl J
.au 1 Y2 =constant

Step 4:
Determine A12 > A21 andA22 from A. n
Step'2: Since
Keep yz = constant (ie., yz ;::: 0).. Introduce .a unit step change in ul
'All + A12 1 Au + ~2l 1
(ie., U1 (s) =.; J . The.equation (B) can be written as 0.8 + Al 2 1 0.8 + A21 1
.. "'12 = 0.2 .. A,21 0.2
-0.2 ()0.8 () Similarly
o = -O-.5-5';"'-
'+-lUl 8 +-8-+-1 U2 s
A.12 + 11..22 =1
() 0.2. s + 1 ( ) 0.2 + A.22 =1
Uz 8 = 0.8 0~55+1 Ul ,8
.. A,22 = O~8
 ( 7.34 ) computer Control of Processes )
[ 7.33 )
The steady state mass b~lance equations of the mixer process are,
step 5:
F =F I + F2 ... (A)
Ul U2
Fx= F 1 Xl + F 2 X 2 ... (B)
The RGA is A= [0.8 0.2]Yl Given data:
0.2 0.8 Y2
F = 200mole/hr
Step 6:
x 0.6
We can draw a conclusion fromRGA as
We can pairy, with u l and Y2 with u 2 to form two loops withminimum Xl = 0.8
interaction. ' x 2 ,= 0.2
Note: with given data and equations (A) & (B)
If you couple differently ie YI with Uz and Y2 with ul' then'the 200 = FI + F2 ... (C)

interaction of the loops would have been four times ~arger (ie., ~:~ = 4)' (200)(0.6) = F (0.08) + F (0.2)
1 2

120 = 0.8 F I + 0.2 F 2 ... (D)

then tuningcontroller will' be much difficult.
From equation (C),'
lExCU11ple7.t .1 F1 =,200-F2 ... (E)
'two streams with flow rates 14\
and F 2 and compositions (mole
Substitue equation (E) in equation (0), we get
percent) Xl = 80%ancl. Xz :: 20%, ina chemical A 'arc mixed in a' vessel
(Fig. 7.24). How do you pair control variables to regulate theproduct 120 (0.8) (200- F 2 ) + 0.2 ,F2
composition x and flow rate F. Let F = 'YI and x = Yz' be the two controlled
160 - 0.8Fz + 0.2 'F2
variables, while F I = u l andF, = uzare the two manipulated variables,
The desired steady state for operational purpose is F = 200 mole/hr, , 0.6 F 2 - 160 -120 =. 40
x = 60% (by moles).
40
=
0.6
= 66.6 mole/hr ... (F)

Substitute equation (F) into equation (A), we get

200 = Fl + 66.6
Mixer F 1 = 133.4 mole I hr

Fig. 7.24: 2 X 2 Mixer Process'

(Multi.Loop Regulatory Control Computer Control. of Ilrocesses)

From equation (C), we can get F 2 = 66.6 IExample 7.8 I

To compute the relative between F and F l' do 'the following: Consider the following RGA for a process with three inputs and three
outputs. .
1. Change F l by one unit (i.e., F 1 = 133.4 + 1 = 134.4) while holding
'F 2 = 66.6 (the same). Solving equations (C) and (D) for F and x and
find .the following new steady states:

F = 201 arid

x 0..6012

(:~ 12 =
(201- 200) 1
.. eonstan t
=
(134.4 - 133.4) "1= 1

2. ChangeP'l by one unit (i.e. Fl = 133.4 + 1= ~34.4) while holding Second, we donot wantto pair with a relative gain of 0 because that
, x = 60 % constant. Solve equations (C) and (D) and find means that the particular input does not have an effect on theparticular
F = 201.67 and output, when all of the other loops are open.
Look at Row 3 in relative gain array, this corresponds to output 3,
F 2 = 67.27
we would not pair Ys with Us because of the 0 term. We also would not

(~FJ .
(201.67 - 200) 1.67
-1-= 1.67
pair Yg with u1because of -3 term. This means that Yg must be paired
with. u 2 " Let us indicate this- choice in our RGA.
8 Fl x = constanf = . (134~4 - 133.1) =
Therfore relative gain between F and F 1 is

(L\F / AFl )F2 =constan t 1.

. - 0.6
All
r-
=. (.6.F
. . '
IL\F)
.
.
1 x = constan t
1.67

:. Relative gain array is

Now we have 'eliminated the third output (row) and second input
(column) from our selection process. From the first row (output), we see
that we would not pair Y~ with Us because of the -1 term. We also cannot
pair, Y1 with u 2 because we have already paired u 2 with Y30 Our only choice
is to pair Y1 with u l ' Let us indicate this choice in our RGA,
F! F2

A= . [0.0.4.6 O
..4]F
0.6 x r{l} 1
A = 3 4
l-3
I
From relative gain array,. the optimum input - output pairing is
{4}
F -F1 and x - F 2
 (Multi-Loo.p Regulatory Control [ 7.37,J ( 7.38 ] Computer Control ofProcesses )

NOV-Twe have eliminated outputs Land 3 (row -1 and row 3 as well as Note: RGA for Non-square systems
input's 1 and 2 (column 1 and column 2) we have no choice but to pair Y2 The relative gain array is a square matrix, which implies that the
with Us and we make _that choice below: number of manipulated va~ables,is equal to number of controlled outputs.
Now, suppose that we have a process with two outputs (Y1 and Y2) and
three manipulated inputs (up u 2 and us) How to pair three inputs (up u 2
and us) with two outputs (Y1 and Y2) there are three possible pairs of
manipulated variables (up u 2 ) (u 2 , us) and (us' u.). Therefore, we can from
three .different relative gain arrays:

U1 u2

IExample 7.9 I A1 = [All Al2]Yl

A21 A,22 Y2
Consider the followingRGA for a process with three inputs and three
outputs. u2 U3

U1 U2 U3 U4 Iu2 = [Al2 Al3]Yl

A22 A23 Y2
3.. 0 -2.83 -0.04 0.86 Y1
-0.08 0.05 1.55 -0~52 Y2 U3 U1
"A,
Iu = [A l3 All]Yl
-5.03 4.67 ,-0.04 1.39 Ya
3
3.11 -0.90 -0.48 -0.73 Y4 /...23 /...21 Y2

How would you choose input-output 'pairings for this process? and we need to examine all of the above RGAs arid we can select set of
two variable pairs with minimum interaction.
Looking at columri B, we see only one relative gain that is not
negative, ~herefore Y'e must pair Y2 withu.. IExample 7.10 I
Looking at row 4, we notice that there is only one relativegain that Non - Square Process
, is not negative,therefore, we must pair Y4 with u r
Consider process with 2 measurements and 3 inputs as shown below:
Looking at column 2, there are 2 non negativerelative gains, but we
have already used output ~, so we must pair Y3 with u 2

ry~] =
, Y 0.5 0.07 0.04 U1]
This leaves Y1 and u 4 which fortunately, has a favourable relative
[0.004 -0.003 -0.001] [ ::
gain (0.86).
'Note: How to pair input-output variable pairs for this process.
{ If the Y1 - u 4 relative gain had not been favourable, we would have , Step 1:
been forced to drop output 1 and input 4 and simply have three control
loops for this pr.ocess. Forma three possible pairs of manipulated variables (Up-.u2 ) , (u 2 , u g)
and (U
3,U1).
( Multi-~oop Regulatory Control'
( 7.40)

Step 2:
1.1 MULTI LOOP'PIO CONT.ROLLER
In multiloop PID controller scheme, each controlJoop is controlled
Form a three possible 2' x 2 process as shown below:
by a single PID controller. After selection of input-output variable pairings

[Y2Yl] = [0.5 .0.07] [Ul] with minimum interactions (using RGA) one can design PIDcontrollers
... (A) for individual loops. This multiloopPID controller scheme was the .first
0.004 ....-Q.003 U2
approach used for MIMO control in process' industries.

Y1
[y
,].=. [0.07 ~"O.04] [U 2] ... (B)
Advantages of Multiloop PID controller scheme are,
2 ,-0.003 -0.001 _u3 i) Use of simple Pl Di.control algorithm.
Ii) Ease of understanding by plant operator.
[Yl] . [0.'5 '0.04 _] [U1] ... (C) iii) The actions of the controllers are relatively easy to monitor.
Y2 =0.004 -0.001 u 3
The matrix notation for multiloop PID controller scheme is explained
Step 8:' as shown below.
Determine 'three RGAs such as RGA 1, RGA 2 and RGA 3 for the For a single loop PID control scheme as shown in Figure 7.25.
three 2 x 2 process (A), (B) and-(C) respectively.
D
Ul U2

RGA! = [0:84 0.16Yl Ysp

. 0.16 0.84 Y2

u2. Ug

RGA2 = [-104 2.4 yl Fig.,7.25: Single loop PID control scheme

2.4 -1.4 Y2
Gp Gc
ul Us The closed loop transfer function y = t:t-G G, Ysp ... (7.32)

RGA3 = [0.76 0.24Yl and characteristic equation is 1 +Gp G c

p

=0 ... (7.33)
0.24 0.76 Y2
For a multiloop PID control scheme for a nxn process, the closed
Step 4: loop transfer function Y = (I + Gp Gc)-l G p Gc Ysp (7.34)

Determine best combination where

Fromtheabove RGAs, we can select RGA 1 as the best combination Y (n x 1) vector of control variables
because of the following- reasons.
Ysp = (n x 1) vector, cf set. - points
"i) Pairs the input-output variables with relative gains closer to 1~
G p = (n x n) matrix of process transfer functions.
Ii) Do not pair the Input-output variables with negative relative gain. G c = (n x n) diagonal matrix ofPID controller transfer functions.
Hence we can pair u 1 with Y1 and u 2 with Y2 we should keep u 3 a~ and the characteristi~ equation is det (I + G p G c) =0 .... (7 ~35)
constant.
 . [7.~~.] ComputerControlof Processes )

For a 2 '5 ~. precess.vas shown in Figure 7.1~. Steps involved in tuning multiloop PID controller scheme using BLT
method are~.
P'],
Y Sp = .,[. Y1S.
Y2sp . Y 2
Y = [Yl.]" G=
p
Step 1:
By ignoring process interactions, each PID controllers are first.
where designed using Ziegler - Nichols tuning method (as shown in Table below).

Gel = PID controller transfer function in loop 1 .Ziegler - Nichols tuning

G e2 PIDcontrollertransfer function in loop 2 s, T1 Td

The methods used to tune multi loop PIDcontroller scheme are, . P 0.5 K cu - -
i) Detune method
PI 0.45 Keu Pu/1.2 -
Dueto the. presence of interaction and retaliatory effects.from other
PID 0.6 K cu P u/2 P u/8
loops, the performance of PID controller in each loop will degrade. In
order to improve the PID controller performance, each PID 'controller
Note:
should be detuned. The various detuningrnethods are,
For a n x n proc~ss with multi loop PI controller scheme, theclosed
.a)BLT Detuning method
loop transfer function is Y(s) = {[I + Gp(s) Gc(S)]-l G/s) Ge(s)}Ysp (sland the
b) Sequential loop tuning closed loop characteristic equation isdet[I+Gp(s) Ge(s)] = 0
c) Independent loop tuning If Nyquist plot of detfI+GpUw) Gc(jw)] encircles origin, the closed
ii) Decouple method loop system is unstable.
If we define a new function W(s) as W(s) =. -1 + det[I+Gp(s) '9"e(s)]
In this method, Decoupler 'is used to cempens ate for process
, then encirclement of (-1, 0) by W(jw) would indicate instability.
interactions and .thus reduce/eliminate the proces interactions. In this
'method, set point change in one loop does not affect the loop. Thevarious Step 2:
decoupling methods are, Assume a factor F : 'Typical values between 2 and 5.
,a) Partial decoupling method Step-S:
b) Static decoupling method Calculate new values of controller parameters by
c) Non linear decoupling method _ Kci,ZN
Kei -' F ,Tn = FT Ii ZN i = 1,2 ..... n

where n is the dimension of multi loop process. For 2 x2 process, n =2. In

step 3 PI controller is detuned.
BLT is a popular detuning method used in multiloop PID controller
scheme, Due to presence of interaction and retaliatory effects from other Step 4:
loops, the performance of PID controller in each will degrade. This method Compute W = .:...l+det(I+GpGc) for os w < 00
is used, to detune the pin corrtr ol l er in each loop for acceptable
for example, 2 x' 2 system
performance. This method was. suggested by- Luyben in 1986.
det (I + GpGc) = 1+ Gel Gpll + Gc2 Gp22 + Gc1GC2 (GPll GP22 - GP12 GP21)
( Multi~Loop Regulatory C.ontrol [ 7.43) Computer Control of Processes )

Step 5: The RGA indicates how the inputs should be paired with outputs to
form loops with the smaller amount of interactions. In some process, even
Determine multi loop .closed loop log modulus
t,hough the interactions between the loops are small, the two control loops
still affect each other's operation very seriously and the overall control
LIDaX =
C
max{2010gl~I}'
W 'l+W system is characterised as unacceptable.
When the designed process is given with two mildly/strongly
Layben suggested that Lrgax_ = max.
w
L, = (2n)
.
dB
interacting loops, he or she introduces in the control system special new
where n is the dimension of the multiloop process.' elements called decouplers. The purpose of d-ecouple is to cancel the
interaction effects between the two loops and fhus rreduce to two
Step 6:
noninteracting. control loops.
If L~ax =2n, stop the iteration and take Kci andT'Ii as calculated in
Let us now study how we, can 'design the decoupler for a 2 ,>.c 2 process
step 3as detuned PI controller parameters. If L~ax* (2n) dB, select a
with two strongly interacting loops as shown in Figure 7.26. In this figure'
newv~.lue of F and return to step 2 until L c max = (2n) dB.
the decoupling matrix is restricted to the form,

7.9 DECOUPLER
D(s) = [d ll (s()=) 1 d 1(2 ()s) .] ... (7.36)
We have seen the problems with multi loop interaction. These occur d 21 S d 22 S = 1
because a manipulated input affects more than one controlled output. One
Since the purpose of the -decoupler D(s),.ls to remove the interaction
approach toha~dling this problem is known as decoupling. The idea is to
in the process as shown in Fig. 7.26. The, interaction between loop 1 and
develop "Synthetic" manipulated inputs that, affect only one process output
loop 2 is zero means g12 = g21 = 0 in the 2 x 2 process. So we can write
.eac1l,.This approach is shown in Fig~ 7.26.
Loop 1

"iIIIIIIIIII __

u*1 I
~_"""'''''' __ ''''
- - - - - - - - - - -- - - - -.
I

... (7.37)

From the above equation (7.37), we can write

I
1 _
00. (7.39)
Decoupler Process

Loop 2 FrOID, equations (7.38) and (7.39) the decoupling elements

. ' d 12 'and d 21
Fig. 7.26: A decoupling centre] system can be calculated.
Cu l , u2 are synthetic manipul~tedinputs)
(Multi~LOOPR.~glllat()ryControl [ 7.45) ( 7.46 ) Computer Control of Processes)

, By fhe addition of decoupler Dts) with interacting 2 x 2 process as IExample 7.11 I

shown in Fig 7.26, the process reduces to two noninteraction control loops Consider a wood-Berry, distillation column process
as shown in Fig 7.27. 0

12.8e-s -18.ge-35
XD(S)] = 16.7s + 1 218+1 [R(S)]
[ XB(s) 6.6e-78 -19.4e-8
0 0
S(s)
10.9s + 1 14.4s+1

where

Y1 = Xn Distillate composition

Y2 = XB Bottom composition
Fig. 7.27: Equivalent block diagram of Fig 7.~6 with complete decoupiing
(i.e., two non interacting loops) u1 =R Reflux flow rate

Remarks:
u2 =S Steam flow rate in Reboiler
Design a deeoupler for this process
i) Two interacting loops are perfectly decoupled only when the
o process transfer function is perfectly known ie gll" g12' g21 and g22 0
Step 1:
are known correctly. Identify steady state gain' matrix for this process
i i) For a non linear and non stationary process, adaptive decouplers
are required. '
x, XB

iii) Perfect decoupler allows independent tuning of each controller

K = [12.8 -18.9]R
06.6 -19.4 S
without risking the stability of the overall process.
Step 2:
i v) A close examination of Fig 7.26 reveals that the decouplers are Identify the RGA
essentially feedforwardcontrol elements.
XB
Static decoupler is designed using steady state gains or static gains
of the process. In some process, dynamic decouplers (as shown inequations -K12 K I 2 R
(7.~8), and (7.39)) are not realizable. In such process,static decouples are 1\ K I 1 K 22 - K12 K 21 K I 1 K 22 - K 12 K 2I
used. The static decoupling elements are, -K l 2 K 21 Fell }(22 S
Kil K 22 - K l 2 K 21
....:.K12
d 12
KI I Xn XB

~K21
/\ = [2.01 -l.OI]R
d 21 -1.01' aoi S
K.22
where KIt' K 12 ," K 21 and ~2 are steady state gains of the process .

..
(Multi.Loop Regulatory Control [ 7.47 )

Step 8: .IExomple:i.'12,
Identify the .input-output pairing using RGA, fromRfkA, the best Consider a following 2 x 2 process
.input - output variable pair is ~ ~ Rand 'XB - s.
Step 4: 1-18 .9 e-3s -12.8 e-s
Steady state decoupler is = l.-19.4e-
21s +J
s
16..78 +.1. [So (s)]
6.6e-7s R(s)
-K12 . 14.48 + 1 10.98 + 1
D 12 =
K 11
Design a decoupler for this process
-(-18.9)
= = 1.477
12.8 Step 1:
-K21 Identify steady state, gain matrix for this process
D 21 =
K22
x, XB
= -(6.6) K = [-is.9 12.8] S
= O~34
-19.4 -19~4 -6.6 R
Step 5:
. Step 2:
The dynamic decoupler is
Identify the RGA matrix
-(-18.9 e~3S) XD XB
21s+1 A _= [~1.01 2.01 ],8
12.8e-s 2.01 -1.01- R
16.7s+1
Step 8:
= (1.477) 16..7-s+ 1 e-2s Identifying input - output, pairing using RGA .
12.8s+1
From RGA, the best input - output variable pair isX, - R and ~ - S
-6.6 e-7s Step 4:
10.9s+1 Steady state decoupler is
, = , -19.4'e-s
.14.48+1
du = -K12 = -0. 677.
K 11
= (0.34) 14.48 + le-6s
10.9s+1
(Multi.L~~p Regulatory Control ( 7.49 )
( 7.50 ) Computer Control of Processes )
Step 5:.
TWO MARKS OUESTIDNS AND ANSWERS
The dynamic decoupler is
1. What are the advantages of RGA?
_ . -g12 (s) i) 'Provides two types .ofuseful information
- -gll (s) a) Measure of process interaction
b) Recommendation about best pairing of manipulated and
-12.8e-s controlled variables.
l-6.7s+1 ii) Requires knowledge of steady state gains but not dynamics of a
=
-18.ge-38 process.
21s+1. 2. What are the problems arising from interactions?
i) Closed loop system may become destabilized.
= (0.677) 21s + 1 e2s ii) Controller tuning becomes more difficult.
16.78+1
3. How do you pair input u with output y for a process?
.. Select a control . loop by pairing the controlled outputs Yi with
-g2l (8)
g22(S) manipulate variables mj in such a way that the relativ~.gains Iij are
positive, and as close as 'possilble to unity.
-19.4 e-8 4. What are properties' of. RGA?
. (-)14.4s+1 The properties of RGA are,
= 6.6e-78 i) Summation of elements in any row = 1
10.9s + 1 i i) Summation of elements in any column =1
iii) RGA is independent of scaling used for input' and output variables.
10.9 s+ 1 68
= (2.94) ----e 5.. Comment about interaction exisiting a system in the followipg .
14.48+1
RGAval~es. .
Thedecoupler elements Dj, and D 21 are physically unrealizable
and 'e6s respectively. This is due to current decoupler
because the term s e 2s ii) A. = [0.5 0.5J
0.5 0.5
output to depend on the' future' inputs.
iii) At = [o.s 0.2] Iv)
0.2 0.8 .
A. ~ [01 11
OJ
i) No interaction (pairing should be Yl - u l and Y2 - u 2) .
ii) Strong interactio existing. in the system (pairing can be either
(Yl - u l an Y2 - u 2 ) or (Y1 - u 2 and Y2 - u l )
iii) Mild interaction exists in the system (pairing should be Yl~ - u l
and Y2 - u 2 ) .
iv) No interaction (pairing should be Y2 - ~1 and Yl -u2 )
 8.2 computer Control of Processes]
CHAPTER
Distrubances

Manipulated 1----..,-.--+ Controlled

variables variables

MULTIVARIABLE ..
REGULATORY CONTROL
Centralized
1 .. .- ..... oW 11IIo""
control . . . . . _ ... t

8.1 INTRODUCTION
The chemical plant usually consist of Multi-Inputs and Multi-Output Set point
(MIMO) variables,
Fig. 8.2: Centralised control scheme
Distrubances
, \

Decentralized f+
control,
Manipulated 1---"-'" Controlled
variables Process variables

Fig. 8.1: MIMO Process

The MIMO controlproblem is handled by two approaches:

i) Centralised (Multivariable) control

ii) Decentralised (multi-loop) control

, i) Centralised (Multivariable) control
Decentralized
control
......
'Ituses all process inputs and outputs m~asurement simultaneously
to determine all manipulated variables. Fig.8.3: Decentralised control scheme
H) Decentralised (multi-loop) control
It uses multiple single loop controllers, i.e., each single outputIs
connected to a single input forming a network of
multiple loops.
(Multlvariable:RegulatoryControl [8.3 ) 8.4 Computer Control of Processes )
Decentralised (multi-loop) .control
The diagonal controller (decentralised control) is shown in
S.N Advantages Disadvantages Figure 8.4 can be expressed mathematically as

[~~~:~] = [el~s) e2~S)][gCri(S) gc~(S)][:~~ ~:~ :~: ~:~][::~ ~:~ ::~:~].

1. Usually use a simple Does not handle cross-loop
algorithm.. interaction.

2. Easy to understand by' Does not optimize the use of the error diagonal decoupler process
plant personal. manipulated variables. controller

3. Standard control designs There are many. possible control ... (8.1)
Loop 1
have been developed for configuration.
common' unit operation.
4. Maintaining. a failed loop Does not handle input constraint
doesnot affect the. other efficiently.
loops.

Centralised (multivariable) control

, .

S.N Adoantage Disdvantages

1. It can handle cross-loop Control algorithm and

interaction. calculation may be complex.
2. It can incorporate input May be difficult.to understand.
constraints directly. by plant operations. Loop 2

Fig. 8.4: 2 x 2 Centralised control combining a decoupler

3. It can provide. optimal usage Failure of the central control and a decentralised controller
of the manipulated variables. system may affect all loops.
In diagonal controller, .the diagonal elements are controllers and off
diagonal elements are zero. In this controller decoupler is used to
minimise/remove interaction.
8.2 MULTIVARIABLE PIO CONTROLLER
When interactions are significant, a full matrix 'controller
Generally most industrial processes 'are multivariable systems. Two-
(cnetralised .control) is used. This is shown in Figure 8.5 and is expressed
input Two output (TITO) system is one of the "most prevalent ~ategories
mathematically as .
of multivariable systems, because there are real process of this nature or
because a complex process has been decomposed in 2 x 2 blocks with non
negligible interactions between its inputs! and outputs. .
Yds)l = [e~" .(s) 0 l[gcl(S) gC2(s)][gll (8) g12{s)j '.. (8.2)
r~Y2 (S)J .0 e2 (s)J "gc3 (S)gC4(s) g21 (s) g22 (8)
When the. interactions in different loops of the process are modest, error Full matrix . Pr ocess G p (s)
a diagonal controller (decentralised control) is often .adequate. . controller G ~ ('s )
" (MultivariableRegulatory Control 8.6 Computer Control of Processes )

The closed loopTiF is" 'I'hestability of the closed Ioop system is determined by, the poles of
Y(s) = Gp(s) U(s) ... (8.3) the characteristics equaiton. '

U(s} Gc(s}.~r(s)-yes~ ... (8.4.) II + G(s) Ge(s) I = 0 ... (8.6)

Frosn the' above equations ", 'Usually process with equal number of i nputs and outputs are
controlled by methods as shown in Figure 8.4 and Figure 8.5. Such a system
Y(s) = (1+ des) Gc(s~i ,G(s)'G~(s) ~(s) ... (8.5)
is called square system process with unequal number of inputs and outputs
Where, (called norisquar systems) often arise in the process industry. Such
nonsquare systems may have either more outputs than 'inputs or more
Y,' 1(8)] inputs than outputs. Examples of nonsquare systems are mixing tank
yes) [
,Y2 (s) process: 2 x 3 system, shell standard control problem: 5 x 7 system, crude
distillation system: 4 x 5 system etc.
The two simple methods to control non square system are
G(s)
1. Davison's method
2. Tanttu and Lieslehto method-
Davinson's methods of designing centralised PID controller
Davision has proposed a centralised multivariable PID controller
Res)
'[r (s)]'
1 tuning method for square systems. Here proportional Ingegral and
r2 (s) derivative gain matrixare given by

"Loop 1 'S[G (s = 0)]-1

KI E[G(S = 0)]-1

Kn [G(s= 0)]-1

where [G(s = 0)] is called the rough tuning matrix' and 8 and E are
the fine tuning parameters, which generally range from 0 to 1. This method
can also be extended to non-square systems. As inverse not exist for non
square system, Moore-pe.nrose pseudo inverse is used. For matr-ix A,
Moore-penrose pseudo inverse is
A+= A H (A x AH)-l
where A H is the Hermitian matrix of A.

Full matrix Process So for a non-square system, PIn controller gains are
controller Kc = 8[G (s= 0)]+
Loop 2 K1 = c;[G (8 = 0)]+
Fig. 8.5: 2 X 2,Centralised controller coti1prising ,'ullma'trix controller Kn = [G(s) = 0]+ ... (8.8)
(MultivariableRegulatory Control 8.7 8.8 computer Control of Processes]

8.3 MODEL PREDICTIVE CONTROL (MPCJ However, the fundamental framework of MPC algorrthma is in
common for any kinds of MPC schemes. The basic elements of MPC are
MPC .is a multi'variable. controller but also a wide class of optimal
illustrated in Figure 8.. 6 and can be defined as follows: ' e

control based algorithms that use an explicit process model to predict

.the behaviour of a plant. There is a wide variety of MPC algorithms that 1. An appropriate model is 'used to predict the output behaviour of
have been used in' industries. such as a plant' over a future time. interval or normally known as the
prediction horizon (P); For a discrete time model this means it
Model Predictive Control(MPC)
predicts the plant output from. ~(k+1) to y(k+P) based on all
Dynamic Matrix 'Control_,,~DMC) actual past control inputs u(k), u(k-l), .... u(k-j).and theavailable
Generalized Predictive Control (GPC) current. information y(k).
.Internal Model Control (IMC) 2~ A sequence of control actions adjustments (Aufk) .... Au (k+M) to
The main diff~rence for all these MPC algorithms are the types of be implemented over a 'specified future time Interval, which is
models used to represent the plant dynamic. and the cost function to be known as the control horizon (M) is calculated by minimizing some
.minimized, MPCis a multivariable control algorithm that uses. specified objectives such as' the deviation of predicted output from
setpoint over the prediction horizon and the size of control action
An internal dynamic' model of the process adjustments in driving the process output to target plus some
A history of past control moves and operating constraints. However, only the first move of computed
An optimization cost function J over the receding prediction control a~nion sequence is implemented ~hile the other moves
horizon to calculate the optimum control moves, are discarded. The entire process step is' repeated at the
subsequent sampling time. This theory is known as .the receding
Past Future
horizon theory.
Controller
~ w_ ""._ _ ...... ~.~

:
y

Model prediction

1
p
~:... Prediction Set point
current horizon
step
.................. max

I
.

u
: y(k+i)
... "' .IiI_Willllllll.fIII_

~
I

. Past control moves

Control horizon .'
Fig~8.7: The MPC block diagram
Fig. 8.6: MPCStreategy
[Multivariable Regulatory Control 8.9 ) Computer Ccatrel of Processes J
3. A.nominal MPC is impossible, or in other words that no model
can constitute a perfect representation of the real plant. Thus, 0.8 C
the prediction error, E(k) between the plant measurement Ym(k) o 0.6-
C)
andthemodel prediction y(k) will always occur. The E(k) obtained ~ 0.4 h1
is normally used to update the future prediction. The Figure'S. 7 ci

illustrated the error feedback of MPC.

~ 0.2-
o 0
h21 h31 0
0 0 0 0 0 J
0'"
Cost functions -0.2 -I.----...,.----..;.-..,-----.,----"""-r-------,r------I
-2 0 2 4 6 8 10
The most popular cost function used in MPC algorithm is least 'square Discrete -time step
objective function. The least square objective function forP predictive
horizon and M control. moves is written as
P M-I 0.8
J ;:: Lkl
wI (rk+i~h+;)2 + W2 L !m~+i
~o
... (8.9) 0>
'00
0,6
c.
0.4'
where y represents the model predicted output, r is the setpointv.Au .is
0.2
the change in manipulated input from one sample time to f.he next, WI is
o
the weight for the error, W2 is the weight for the change in the manipulated
input and k is the current sample time. -2 o 246 8 10
Discrete -time step
Models
. Fig. 8.9: Impulse response parameter identification
2 0 0 O' 0
The models used In MPC algorithm are finite step response model
1.5
U or finite impulse response model; The step response model is the .vector
0),
CD 84 55 .Etc
"'0 83 of s~ep response coefficients. -
52
1
ci
E 0.5 ... (8.10)
$ 51!
0 0 0
The step. response coefficients can be obtained as shown in
-2 0 2 4 6 8 10 Figure 8.8. Similarly the .impulse response coefficients canoe obtained as
Discrete -time step shown in Figure 8.9. The impulse response model is the vector of impulse
response coefficients.
H = [hI, h 2 , . , hN]~ ... (8.11)
0.8
0), 0.6
MPC Advantages
'fi Streaight forward formulation, based on well understood concepts
0.4
0.2 Explicity handles constraints
o
"'----,~----..,.---__r_---_.,...---__r_--- Explicity use of a model
-2 o 2. 4 6 8 10
Discrete -time step
Fig. 8.8: Step response parameter identific~tion
 [.s.11 J (8.12 J Computer Control of Processes )

N-l
Well understood tuning parameters .
~Uk-i+l + SN Uk--N+1
"C
Yk+1 LSi + dk +1 ... (8.16)
- Prediction horizon i=l .

Optimization problem setup' N-1

Development time much shorter than for competing advanced Y~+l = Sl d U k + L S i dUk:"'i+1 + SN Uk-N+l + ak+~.
i=2 ." ~
control methods.
Easier to maintain: changing model or specs does not require So, for the jth step into the future, we find
complete redesign, sometimes can be done on the fly. "C "
Yk+j = Yk+j +dk +j
8.4 DYNMIC MATRIX CONTROLLER IOMC) N-j N-l .

DMC is amultivariable controller which .was developed by' Shell Oil L S i L\Uk_i+j + L s, L\Uk_i+j + s~ Uk-N+j+ dk +i
Company in the 1960s and 1.970s. It is based on a step response model, ~ ~=j+1 . 'Co~rc;;;term
effect of future effect of pasteontrol moves
which has the form control moves .

'Yk = 8 1 ~Uk-1 + 82 AU k- 2 +.... + SN-1 LiUk-N+l+ 8N Uk-N ... (8.12) ... (8.17)
and we can separate the effects of past and future control moves
which is written in the form
"C . ' . ( effect of current and
N-l
Yk+j = 8 1 L\uk+j-l + 8 2 L\u k +j - 2 + ... + SJo L\uk ~.
Yk = LSi~Uk-i + SNuk-N ... (8.13) 1future moves
i=l

where. Ykis the model prediction at time step k, and Uk-N is the
+'SNUk-N+j + Sj+l tiuk_1 + Sj+2. LlU k _2.
. + +S
...
AU
N-l Ll k-N+j+1
I
.
J effect of past moves
manipulated input N steps in the past.
+dk +j {correction term
Note that the model-predicted output is unlikely to .be equal to the
actual ..measured'output at .time step k. The difference between the ... (8.18)
~ measured output (Yk) and the model prediction is called the additive
The most common assumption is that the correction term is constant
disturbance. in the future (this is the "constant additive disturbance assumption"):
... (8.14)
... (8.19)
The "corrected prediction" is then equal to the actual measured
output at step k, Also, realize that there are no control 'moves beyound the control
horizon of M steps, so '
"C
Yk =Yk + d k .. (8.15) L\Uk+p-1 =. 0
Similarly, the corrected predicted output .at the first time step' in
the future. can be found from,

"C
Yk+l = "
Yk+1 + d" k+l
(Multivariable Regulatory Control [ 8.13 J computer Control of Processes' )

In matrix-vector form, a prediction horizon ofP steps and a control In equation. (8~21) the correct-predicted output response is naturally
horizon of M.steps, yields composed ora "forced response" (contributions of the current andfuture
control moves) and a "free response" (the output changes that are;
predicted if there are no future control moves). The difference between
the setpoint trajectory, r, and the future predictions is
.. c 81 0 0 o o
Yk+l
.. c
Yk+2'
82 81 O o 0' r _ y C= ... (8.22)
"---v--'
corrected
= S predicted
.. c J Sj-l Sj-2 error.E?
Yk+j

Sp Sp_l Sp_2 which can be written

.. c
LYk+P
'---v---' PxMdynanrlc;matrix,Sc' E C = E - 8fAur ... (8.23)
Pxlcorrected output
predictions, yc S2 S3 84 SN-2 SN-l
AUk_l where the future predicted errors are composed of "free response"
:83 84 85 SN-l 0
AUk~2 (E) and ."forced response" (-SfAUf) contributions.
0 0
.+ The least-squares objective function is
Sj+l Sj+2 SN_l 0 0

SP+l SP+2 o o ... (8.24)

Px(N-2)dynaxclc matrix, Spast

Notice that the quadratic terms can written in matrix-vector form as

Uk-N+P d k +P
~ '---v---' ... (8.25)
Pxl past inputs, up Pxl predicted
disturbances.d

... (8.20) .
and
which we write using matrix-vector notation

. yc. = 'Sf'L\Uf + + ~Up + ~ .!. (8.21) M-l

'-r-'
corrected
predicted
outputs
'---v---'
effectof current
and future moves
Spast AUpast
effect of p~stmoves ar:::.~ces W L (.~Uk+i)2
i=O
AUk+M-l
 (MultivariableRegulatory Control ( 8.15 ) ( 8.16 ) Computer Control of Processes )

.3. Specify the weighting on the control action (w.=O if no weighting)

w 0 0 0 AUk
4. All calculations assume deviation variable form, so remember to
0 w 0 0 AUk+1
. - [AUk AUk+1 ... L\U k +M- 1] ... (8.26) convert to/from physical units.
0 0 0
Expansion of DMC to multivariable pr-oblem.
0 0 0 W AUk+M_l
Multivariable DMC can easily behandled by expanding the Srmatrix'
= AUTfWAUf in equation, (8:20) to include dynamic dat~ for all input -out.put
Therefore the objective function can be written in the form' relationships, - For example, a control .problem with two controlled
variables and three manipulated variables would have the following Sf
....(8.27) matrix.
subject.t~the modeling equation equality constraint (8.23)
E C= E - SfL\Uf . (8.28) Sr. = [S1.1.. Sfl2. Sf13.] ... (8.32)
. _ ,Sf?! 8 f 22, Sf23
Substituting (8.23) into (8.27), the objective function .can be written
Each individual 'Sfij .m atrix-Is s imil ar to the form shown in
<I> = (E '-' Sf ~Uf)T (E - Sf AUf + (AUf)T W AUf ... (8.29) Equation (8.20). The calculabion procedure is 'same as for the SISO
The solution for the minimization' of this objective function is p~oblem.

AUf = (SfTSf + W,)-1 SfT ~ ... (8.30)

~ unforced errors
k

Notice that the current and future control .move vector (AUf) is
proportional to the unforced error vector (E). That is, a controller gain
matrix,K, multiplies the unforced error vector (the future errors that
would occur if there were ~o control move changes implemented).
Because only the current control move is actually _implemented, we
use the first row of the K matrix, and
AUk = K1E ... (8.31)
where K 1 represents the first row of the K matrix,
where K =(STf Sf + W)-l STf
Perhaps it is worth summarizing the steps involved in implementing
DMC on a process.
1. Develop a discrete ste:presponse model with length NC8.13), based
on a. sample time, ~t.
2. Specify the prediction CP) and control (M) horizons, N ~ P ~ M.

I'