[ Team LiB ]

Chapter 3. Dynamic Behavior

The goal of this chapter is to understand dynamic behavior. We begin by working with linear state
space models, often obtained by linearizing a nonlinear model, such as those developed in
Chapter 2. We then introduce Laplace transforms. The main advantage to Laplace transforms is
that they allow us to analyze behavior exhibited by linear differential equations by using simple
algebraic manipulations. Laplace transforms are used to create transfer function models, which
are the basis for many control system design techniques.

After studying this chapter, the reader should be able to:

Apply the initial and final value theorems of Laplace transforms

Understand first-order, first-order + dead time and integrating system step responses

Understand second-order under-damped behavior

Understand the effect of pole and zero values on step responses

Convert state space models to transfer functions

The major sections of this chapter are as follows:

3.1 Background

3.2 Linear State Space Models

3.3 Introduction to Laplace Transforms

3.4 Transfer Functions

3.5 First-Order Behavior

3.6 Integrating System

3.7 Second-Order Behavior

3.8 Lead-Lag Behavior

3.9 Poles and Zeros

3.10 Processes with Dead Time

3.11 Padé Approximation for Dead Time

3.12 Converting State Space Models to Transfer Functions

3.13 MATLAB and SIMULINK

3.14 Summary
[ Team LiB ]
[ Team LiB ]

3.1 Background
One of the major goals of this chapter is to obtain an understanding of process dynamics. Process
engineers tend to think of process dynamics in terms of the response of a process to a step input
change. Assume that the process is initially at steady state, then apply a step change to an input
variable. The majority of chemical processes will exhibit one of the responses shown in Figure 3-
1. In this plot, we assume that a positive step increase has been made to the input variable of
interest. The solid curves are examples of "positive gain" processes; that is, the process output
increases for an increase in the input. The dashed lines are those of negative gain processes. The
curves in Figure 3-1a show a monotonic change in the output; this behavior is generally known as
overdamped. The curves in Figure 3-1b are indicative of "integrating" processes; a prime example
is a liquid surge vessel, where the level continues to change when there is an imbalance in the
inlet and outlet flow rates. The curves in Figure 3-1c are known as underdamped or oscillatory
responses. This type of behavior may occur in exothermic chemical reactors or biochemical
reactors. It more often occurs in processes that are under feedback control, particularly if the
controller is poorly tuned. The behavior shown in Figure 3-1d is known as "inverse response" and
is seen in steam drums, distillation columns, and some adiabatic plug flow reactors.

Figure 3-1. Common responses of process outputs to step changes in

process inputs. Assuming a positive step change, the solid curves are
illustrative of "positive gain" processes, and the dashed curves are
indicative of "negative gain" processes. (a) Overdamped or first order,
(b) integrating, (c) underdamped or oscillatory, and (d) inverse
response.

In the sections that follow, we discuss the characteristics of process models that lead to each of
the behaviors shown in Figure 3-1.
[ Team LiB ]
[ Team LiB ]

3.2 Linear State Space Models

In Chapter 2 we developed fundamental models, which were normally nonlinear in nature. We
then developed state space models that were based on linearizing the fundamental models at a
steady-state solution. This led to the notion of a perturbation or deviation variable, which is
simply the perturbation of a variable from its steady-state value.

State space models have the following form, where the states (x), inputs (u), and outputs (y) are
all perturbation or deviation variables

Equation 3.1

Recall that in matrix notation, the first subscript refers to the row and the second subscript refers
to the column. When matrices multiply vectors, each row corresponds to a particular output of the
multiplication, while the column corresponds to a particular input of the operation. Consider the C
matrix, which relates the states to the outputs. Element cij relates the effect of state xj on output
yi.

The shorthand notation for Equation (3.1) is

Equation 3.2

It is important to always check for dimensional consistency in matrix operations. In a matrix-

vector operation y = Cx, the number of rows in C must be equal to the number of elements in y.
Also, the number of columns in C must be equal to the number of elements in x.

Stability
One of the first basic concepts that we need to cover is the notion of stability. Consider a process
where one or more states have been perturbed from the steady-state solution or operating point.
The process is stable if after a period of time, the variables return to the steady-state values. This
means that the state variables, since they are deviation variables, return to zero.

Numerically, we can determine the stability of a state space model by finding the eigenvalues of
the state space A matrix. Remember that the A matrix is simply the matrix of derivatives of the
dynamic modeling equations with respect to the state variables.

If all of the eigenvalues are negative, then the system is stable; if any single eigenvalue is
positive, the system is unstable. A system with all but one eigenvalues negative and with one
eigenvalue equal to zero is called an integrating system and is characteristic of processes with
liquid levels or gas drum pressures that can vary.

Examples of unstable systems are shown in Figure 3-2. If an eigenvalue is real and positive, the
system response is that shown in the top curves. If there are complex conjugate eigenvalues,
with positive real portions, the system oscillates (with ever increasing amplitude), as shown at the
bottom.

Figure 3-2. Unstable responses. (a) Monotonic and (b) oscillatory.

Mathematically, the eigenvalues of the A matrix are found from the roots of the characteristic
polynomial

Equation 3.3

where  is known as an eigenvalue, and I is the identity matrix. For a state space model with n
states, A is an n x n matrix, and there will be n solutions (eigenvalues) of Equation (3.3). There
are analytical solutions for two- and three-state systems; the two-state solution is shown below.

In two-state systems,

The determinant can be found by

and the eigenvalues are found as the two solutions (roots) to

Equation 3.4

The roots can be found using the quadratic formula

Equation 3.5

It is easy to show that if and a11 + a22 < 0 and a11a22 >a12a21, the roots (eigenvalues) are
negative and the system is stable. A more general method of qualitatively checking the stability,
known as the Routh stability criterion, is shown in Chapter 5.

Example 3.1: Exothermic CSTR

Models for an exothermic, CSTR are detailed in Module 8. For a two-state representation, the first
state is the concentration and the second state is the reactor temperature. For a particular
reactor with two different operating conditions, the A matrix is (the time unit is hours)

Operating condition 1 Operating condition 2

and the eigenvalues for operating condition 1 can be found using the following steps

with the solutions [using the quadratic formula (3.5)]

Since both eigenvalues are negative, operating condition 1 is stable.

The reader should show that the eigenvalues of A2 are

where the positive eigenvalue indicates that operating condition 2 is unstable.

MATLAB Eigenvalue Function

The M ATLAB eig command can be used to quickly find eigenvalues of a matrix. The reader should
use the MATLAB command window to verify the following results for the second operating
condition:

» a2 = [-1.8124 -0.2324;9.6837 1.4697];

» eig(a2)
ans =
-0.8366
0.4939

Again, the positive eigenvalue indicates that the second operating condition is unstable.

Generalization
Notice that a solution of a second-order polynomial was required to find the eigenvalues of the
two-state example; this resulted in two eigenvalues. For the general case of an n x n matrix,
there will be n eigenvalues. It is too complex to find these analytically for all but the simplest low-
order systems. The simplest way to find eigenvalues is by using existing numerical analysis
software; for example, in MATLAB the eig function can be used to find eigenvalues.

The values of the eigenvalues are related to the "speed of response," and the eigenvalue unit is
inverse time. If the unit of time used in the differential equations is minutes, for example, then
the eigenvalues have min-1 as the unit. For stable systems (where all eigenvalues are negative),
the larger magnitude (more negative) eigenvalues are faster.

For matrices that are 2 x 2 or larger, some eigenvalues may occur in complex conjugate pairs. In
this case, the stability is determined by the sign of the real portion of the complex number. As
long as all real portions are negative, the system is stable.

[ Team LiB ]
[ Team LiB ]

3.3 Introduction to Laplace Transforms

Most control system analysis and design techniques are based on linear systems theory. Although
we could develop these procedures using the state space models, it is generally easier to work
with transfer functions. Basically, transfer functions allow us to make algebraic manipulations
rather than working directly with linear differential equations (state space models). To create
transfer functions, we need the notion of the Laplace transform.

The Laplace transform of a time-domain function, f(t), is represented by L[f(t)] and is defined as

Equation 3.6

The Laplace transform is a linear operation, so the Laplace transform of a constant (C) multiplying
a time-domain function is just that constant times the Laplace transform of the function,

Equation 3.7

The Laplace transforms of a few common time-domain functions are shown next.

Exponential Function
Exponential functions appear often in the solution of linear differential equations. Here

The transform is defined for t > 0 (we also use the identity that ex+y = exey)

So we now have the following relationship:

Equation 3.8
Derivatives
This will be important in transforming the derivative term in a dynamic equation to the Laplace
domain (using integration by parts),

so we can write

Equation 3.9

For an nth derivative, we can derive

Equation 3.10

n initial conditions are needed: f(0),..., f (n – 1) (0)

One reason for using deviation variables is that all of the initial condition terms in Equation (3.10)
are 0, if the system is initially at steady-state.

Time Delays (Dead Time)

Time delays often occur owing to fluid transport through pipes, or measurement sample delays.
Here we use  to represent the time delay. If f(t) represents a particular function of time, then f(t
– ) represents the value of the function  time units in the past.
We can use a change of variables, t* = t – , to integrate the function. Notice that the lower limit
of integration does not change, because the function is defined as f(t) = 0 for t < 0.

Equation 3.11

So the Laplace transform of a function with a time delay () is simply e–s times the Laplace
transform of the nondelayed function.

Step Functions
Step functions are used to simulate the sudden change in an input variable (say a flow rate being
rapidly changed from one value to another). A step function is discontinuous at t = 0. A "unit"
step function is defined as

and using the definition of the Laplace transform,

so

Equation 3.12

Similarly, the Laplace transform of a constant, C, is

Equation 3.13

Pulse
Consider a pulse function, where a total integrated input of magnitude P is applied over tp time
units, as shown in Figure 3-3.

Figure 3-3. Pulse function.

The function is f(t) = P/t p for 0 < t < tp and f(t) = 0 for t > t p. The Laplace transform is

Equation 3.14

Impulse
An impulse function can viewed as a pulse function, where the pulse period is decreased while
maintaining the pulse area, as shown in Figure 3-4. In the limit, as tp approaches 0, the pulse
function becomes (using L'Hopital's rule)

Equation 3.15

Figure 3-4. Concept of an impulse function.

If we denote a unit impulse as f(t) = , then the Laplace transform is

Equation 3.16

Examples of common impulse inputs include a "bolus" (shot or injection) of a drug into a
physiological system, or dumping a bucket of fluid or bag of solids into a chemical reactor.

Other Functions
It is rare for one to derive the Laplace transform for a function; rather a table of known
transforms (and inverse transforms) can be used to solve most dynamic systems problems.

Table 3-1 presents solutions for most common functions. If you desire to transform a function
from the time domain to the Laplace domain, then look for the time-domain function in the first
column and write the corresponding Laplace domain function from the second column. Similarly, if
your goal is to "invert" a Laplace domain function to the time domain, then look for the Laplace
domain function in the second column and write the corresponding time-domain function from the
first column. This notion of the inverse Laplace transform can be written

Equation 3.17

For example
Initial- and Final-Value Theorems
The following theorems are very useful for determining limiting values in dynamics and control
studies. The long-term behavior of a time-domain function can be found by analyzing the Laplace
domain behavior in the limit as the s variable approaches zero. The initial value of a time-domain
function can be found by analyzing the Laplace domain behavior in the limit as s approaches
infinity.

Table 3-1. Laplace Transforms for Common Time-Domain Functions

Time-domain function Laplace domain function

f(t) F(s)
(t) (unit impulse) 1

C (constant)

f(t - ) e-sF(s)

tn

sF(S) - f(0)

Snf(S) - Sn-1f(0) - L - sf (n-2)(0) - f(n-1)(0)

e-at

te-at

1-e-t/

sin t
 Time-domain function Laplace domain function

cos t

e–at sin t

e–atcos t

where ,

The final-value theorem is

Equation 3.18a

The initial-value theorem is

Equation 3.18b

It should be noted that these theorems only hold for stable systems.

Example 3.2: Application of Initial- and Final-Value Theorems

Find the long-term and short-term behavior of the time-domain function, y(t), using the final- and
initial-value theorems on the Laplace domain function Y(s) (we see later that this arises from a
step input applied to a second-order process):
The long-term behavior, y(t ), is found using the final-value theorem,

The short-term behavior, y(t 0), is found using the initial-value theorem,

The reader should verify that the time-domain function, y(t), can be found by applying Table 3-1
to find

and that the values of y(t ), and y(t 0), are consistent with the final- and initial-value
theorems.

General Solution Procedure

To obtain analytical solutions for differential equation-based models, the general procedure is
composed of several steps.

1. Start with an nth order linear differential equation

and known initial conditions

2. Transform each element of the differential equation to the Laplace domain,

 3. Use algebraic manipulations to solve for the transformed variable.

where N(s) and D(s) are polynomials in s.

4. Perform a partial fraction expansion to isolate the individual elements.

where a, b, and so forth are the roots of D(s). If roots are repeated the partial fraction
expansion has the form (for an example where three roots all have a value of a).

5. Invert each element back to the time domain to find the final solution for y(t).

For the case of repeated roots (say 3 that have a value of a), the solution has the form

It should be noted that chemical process systems are rarely described by an nth order differential
equation. Usually, a set of n first-order differential equations is transformed to a single nth order
equation, as shown in Example 3.3.

Example 3.3: Second-Order Differential Equation

Consider the following state space model of an isothermal CSTR (Module 5).

Equation 3.19
where x1 and x2 represent the concentrations of two components (in deviation variable form) in
an isothermal reactor; the initial conditions are x1(0) = x2(0) = 0. Solve for the output
(concentration of component 2) response to a unit step input.

1. The reader should show (see exercise 6) that this can be arranged to a second-order
differential equation,

with the initial conditions y(0) = dy(0)/dt = 0, and the input is initially u(0)= 0.

2. Taking the Laplace transform of each element,

and substituting back into the differential equation, we find[1]

Alternatively, one could take the Laplace transform of each equation in (3.19) and use algebraic
[1]
substitution to find this second-order equation in s.

3. Solving for Y(s), we find

Here we consider a unit step input, U(s) = 1/s.

4. Performing a partial fraction expansion using the roots of the denominator polynomial,

Equation 3.20
 Solve for C1 by multiplying by s,

Setting s = 0 to find

Similarly, solve for C2 by multiplying Equation (3.20) by s + 2.4053 and setting s = –2.4053

Also, solve for C3 by multiplying Equation (3.20) by s + 2.2376 and setting s = –2.2376

which yields

5. Inverting each element back to the time domain,

Equation 3.21

Although it is nice to have an analytical solution, it is generally more pleasing to engineers to

view a plot of a variable as it changes with time, as shown in Figure 3-5. In this plot we
notice that the concentration initially decreases, before increasing to a new steady-state
value.

Figure 3-5. Response of the concentration as a function of time.

See Equation (3.21).
In Example 3.3 we solved for the coefficients (C1, C2, C3) by selecting values of s to minimize the
multiplications performed during each step. An alternative is to solve for three equations in three
unknowns by using various values of s (say s = 1, 2, and 3, for example) in Equation (3.20).

Differential equations textbooks, such as Boyce and DiPrima (1992) present many examples of
applications of Laplace transforms to solve differential equations. In practice, it is rare for process
engineers to seek analytical solutions to differential equations; it is far easier to solve these
numerically. The primary goal of this section was to provide background material to understand
the transfer functions and dynamic responses shown in the next section.

[ Team LiB ]
[ Team LiB ]

3.4 Transfer Functions

A transfer function relates inputs to outputs in the Laplace domain. In particular, the Laplace
domain relationship between a manipulated input and output is called a process transfer function.

Consider an nth order differential equation,

Equation 3.22a

Since we are assuming that the model is based on deviation variables, and that the system is
initially at steady state, the initial conditions are

Equation 3.22b

Taking the Laplace transform of each term,

Solving for Y(s), we find

The ratio of polynomials is called the transfer function. When it relates a manipulated input to an
output it is commonly called a process transfer function. In general, we will use gp(s) to represent
the process transfer function.
Equation 3.23

In this case the process transfer function is

Equation 3.24

The roots of the numerator polynomial are known as zeros, and the roots of the denominator
polynomial are call poles. The significance of poles and zeros are discussed in detail in Section
3.9.

We have used capital letters to distinguish Laplace domain variables from the time-domain
variables. In the rest of this text we generally use lowercase letters for all input and output
variables. If the argument is s, then we assume that we are referring to the Laplace domain.

Transfer functions are often used in block diagrams. For example, the relationship between an
input and output is shown as

In the rest of this chapter we study the dynamic behavior of some commonly used transfer
functions. Our focus is on Step Responses, since process engineers often apply step changes to
understand dynamic behavior.

[ Team LiB ]
[ Team LiB ]

3.5 First-Order Behavior

Many chemical processes can be modeled as first-order systems. The differential equation for a
linear first-order process is often written in the following form:

Equation 3.25

This can also be written as

where the parameters (p and k p) and variables (y and u) have the following names: p is the
process time constant (units of time), kp the process gain (units of output/input), y the output
variable, and u the input variable. Taking the Laplace transform of each term (notice that we are
now using lower-case variables to represent the Laplace domain input and output), and assuming
that the initial condition is y(0) = 0,

So the Laplace transform of Equation (3.25) can be written

or solving for y(s) we find a first-order transfer function

Equation 3.26

Step Response
Consider the case where the output is initially zero (steady state in deviation variable form), and
the input is suddenly step changed by an amount u. The Laplace transform of the input is

Equation 3.27

So Equation (3.26) can be written

Equation 3.28

Using a partial fraction expansion and inverting to the time domain, you should find (see Exercise
1)

Equation 3.29

Here the notion of a process gain is clear. After a substantial amount of time (t >>  p), we find,
from Equation (3.29),

Equation 3.30

That is,

and, since y(0) = 0, we can think of y(t ) as y, so

Equation 3.31

We can think of the process time constant as the amount of time it takes for 63.2% of the
ultimate output change to occur, since when t =  p,
Remember that this holds true only for first-order systems.

Impulse Response
Consider now an impulse input of magnitude P, which has units of the input*time; if the input is a
volumetric flow rate (volume/time), then the impulse input is a volume. The output response is

You should find that the time domain solution is

which has an immediate response of Pkp/p followed by a first-order decay with time.

Example 3.4: Stirred-Tank Heater

Recall that an energy balance on a constant-volume stirred-tank heater (Example 2.3) yielded

where the subscript s is used to indicate that a particular variable remains at its steady-state
value. Defining the following deviation variables

The equation can be written in the form

or

where the parameters of this first-order model are

The gain and time constant are clearly a function of scale. A process with a large steady-state
flow rate will have a low gain, compared to a process with a small steady-state flow rate. This
makes physical sense, since a given heat power input will have a larger effect on the small
process than the large process. Similarly, a process with a large volume-to-flow rate ratio is
expected to have a slow response compared to a process with a small volume-to-flow rate ratio.

Consider a heater with a constant liquid volume of Vs = 50 liters and a constant volumetric flow
rate of Fs = 10 liters/minute. For liquid water, the other parameters are cp = 1 kcal/ liter°C. The
process gain and time constant are then

Step Response

If a step input change of 10 kW is made, the resulting output change is [from Equation (3.29)]

A plot of the step input and the resulting output are shown in Figure 3-6.

Figure 3-6. Step response of a stirred-tank heater, characteristic of a

first-order system. Deviation variables.
Remember that the inputs and outputs in this expression are in deviation variable form. If the
steady-state values are (for an inlet temperature of 20°C)

Then, the physical temperature response is

Impulse Response

If an impulse input of 30 kJ is made, you should be able to show that the temperature changes
immediately by 0.143°C (see Exercise 15).

[ Team LiB ]
[ Team LiB ]

3.6 Integrating System

We found in the previous chapter that material balances on liquid surge vessels or gas drums
often yielded models with the following form:

Equation 3.32

In the Laplace domain, this is

Equation 3.33

Consider an integrating process initially at steady state, with y(0) = 0.

Step Response
If a step input change of u is made at t = 0,

and we find the time-domain value

Equation 3.34

That is, the output ramps with a constant slope of ku.

Impulse Response
If an impulse input of magnitude P is made at t = 0,
then the output immediately changes to a new steady-state value of

Example 3.5: Tank-Height Problem

The mathematical model for a liquid surge tank is (see Example 1.3)

where h is the liquid height, A is the constant cross-sectional area of the tank, F1 is the inlet flow
rate, and F2 is the outlet flow rate. Assume that the outlet flow rate remains constant at a steady-
state value of F2s. Defining the output and input in deviation variable form as

For a constant cross-sectional area of 10 m2, the model is

Step Response

For a step input change of 0.25 m3/min, the output response is

which is shown in Figure 3-7. If the steady-state height is 2 meters, then the height as a function
of time is

Figure 3-7. Step response of a liquid surge tank. Deviation variables.

Impulse Response

For an impulse input of 1 m3, the output response is

which makes sense, because the cross-sectional area is 10 m3.

[ Team LiB ]
[ Team LiB ]

3.7 Second-Order Behavior

Second-order models arise from systems that are modeled with two differential equations (two
states). In this section we separately consider transfer functions that do not have "numerator"
dynamics and those that do.

Pure Second-Order Systems

Consider a linear second-order ODE, with constant parameters

Equation 3.35

This is often written in the form

Equation 3.36

where (obviously a0 0)

and the parameters are as follows: k is the gain (units of output/input),  the damping factor
(dimensionless), and  the natural period (units of time).

The second-order model shown in Equation (3.35) or (3.36) generally arises by changing a set of
two first-order equations (state-space model) to a single second-order equation. For a given
second-order ODE, there are an infinite number of sets of two first-order (state-space) models
that are equivalent.

Taking the Laplace transform of Equation (3.36),

Equation 3.37

where y(s) indicates the Laplace transformed variable.

Assuming initial conditions are zero, that is , we find

Equation 3.38

The characteristic equation of the second-order transfer function is 2s2 + 2s + 1. We can find
the roots (known as the poles) by using the quadratic formula

which yields the following values for the roots:

Equation 3.39

The following analysis assumes that  >0 and  > 0. This implies that the real portions of p1 and
p2 are negative and, therefore, the system is stable. The three possible cases are shown in Table
3-2.

Step Responses

Now, we consider the dynamic response of second-order systems to step inputs (u(s) = u/s),

Equation 3.40

where u represents the magnitude of the step change.

Case 1: Overdamped (• > 1)

For  > 1, the denominator polynomial,  2s2 + 2s + 1, can be factored into the form
where the time constants are

Table 3-2. Characteristic Behavior of Second-Order Transfer Functions

Case Damping factor Pole location Characteristic behavior

1 >1 Two real, distinct roots Overdamped

2 =1 Two real, equal roots Critically damped

3 <1 Two complex conjugate roots Underdamped

We can derive the following solution for step responses of overdamped systems,

Equation 3.41

Note that, as in the case of first-order systems, we can divide by ku to develop a dimensionless
output. Also, the dimensionless time is t/ and we can plot curves for dimensionless output as a
function of . This is done in Figure 3-8, which includes the critically damped case, as discussed
next. Most chemical processes exhibit overdamped behavior.

Figure 3-8. Step response of a second-order overdamped system.

Case 2: Critically damped (• = 1)

The transition between overdamped and under damped is known as critically damped. We can
derive the following for the step response of a critically damped system

Equation 3.42

Notice that the main difference between overdamped (or critically damped) step responses and
first-order step responses is that the second-order step responses have an S shape with a
maximum slope at an inflection point, whereas the first-order responses have their maximum
slope initially.

The initial behavior for a step change is really dictated by the relative order of the system. The
relative order is the difference between the orders of the numerator and denominator polynomials
in the transfer function. If the relative order is 1, then output response has a nonzero slope at the
time of the step input; the step response of a system with a relative order greater than 1 has a
zero slope at the time of the step input.

Case 3: Underdamped (• < 1)

For  < 1, we find [from Equation (3.39)] that the poles are complex,

Equation 3.43a
where the real and imaginary contributions are

Equation 3.43b

We can derive the following step response for an underdamped system,

Equation 3.44a

where

Equation 3.44b

Again, dividing Equation (3.44a) by ku, we can produce the dimensionless plot shown in Figure
3-9.

Figure 3-9. Step response of a second-order underdamped system as a

function of the damping factor (• ). Interpolate between the curves for
the behavior of other damping factor values.
A number of insights can be obtained from Figure 3-9 and from an analysis of the step response
equations. For  < 1, the ratio of the imaginary portion to the real portion of the pole [from
Equation (3.39)] is

As the imaginary/real ratio gets larger, the response becomes more oscillatory ( becomes
smaller). We also notice that a decreasing ratio corresponds to a larger negative value for the real
portion. As the real portion becomes larger in magnitude (more negative), the response becomes
faster. We use these insights to interpret pole/zero plots in Section 3.9.

Underdamped Step Response Characteristics

The following common measures of underdamped second-order step responses are shown in
Figure 3-10, and defined below: (1) rise time, (2) time to first peak, (3) overshoot, (4) decay
ratio, and (5) period of oscillation.

Figure 3-10. Step response characteristics of underdamped second-

order processes.
Rise time is the amount of time it takes to first reach the new steady-state value. Time to first
peak is the time required to reach the first peak. Overshoot is the distance between the first peak
and the new steady state. This is usually expressed as the overshoot ratio, as shown in Figure 3-
10. Decay ratio is a measure of how rapidly the oscillations are decreasing. A b/a ratio of 1/4 is
commonly called "quarter wave damping." Period of oscillation is the time between successive
peaks.

Second-Order Systems with Numerator Dynamics

The previous discussion involved pure second-order systems, where the relative order (difference
between the denominator and numerator polynomial orders) was two.

Consider now a second-order system with numerator dynamics with the gain/time constant form

Equation 3.45

which is relative order one. The reader should show that the pole-zero form is
where the parameters are

The gain/time constant form has the following time-domain response to a step input (see Exercise
4):

Equation 3.46

The reader should show that, if  n = 2, the response is the same as a first-order process.

Example 3.6: Illustration of Numerator Dynamics

Consider the following input-output relationship:

The unit step responses are shown in Figure 3-11. Notice that negative numerator time constants
(corresponding to positive zeros) yield a step response which initially decreases before increasing
to the final steady state. This type of response is known as inverse response and causes tough
challenges for process control systems. Positive zeros are often caused by two first-order transfer
functions, with gains of opposite sign, acting in parallel (see Exercise 5).

Figure 3-11. Step responses of a second-order system with numerator

dynamics.
Notice also that a numerator time constant that is greater than the denominator time constant
causes overshoot before settling to the final steady state. Also notice that the inverse response
becomes "deeper" as the numerator zero (–1/n) approaches a value of 0 from the positive side.

[ Team LiB ]
[ Team LiB ]

3.8 Lead-Lag Behavior

Lead-lag transfer functions have the same order numerator polynomial as denominator
polynomial. This occurs when the input has a direct effect on the output variable. In terms of the
state space model (3.2), this means that D 0.

Consider a lead-lag transfer function where the numerator and denominator polynomials are first
order:

Equation 3.47

For a step input of magnitude u, the output response is

Equation 3.48

A dimensionless output can be defined by dividing Equation (3.48) by kpu, and t/p is a natural
dimensionless time. The responses to a step input at t = 0 are shown in Figure 3-12. Notice that
there is an immediate response that is equal to the n/p ratio. This can also be found by applying
the initial value theorem to Equation (3.47"/>) for a step input change (see Exercise 3).

Figure 3-12. Step responses of the lead-lag example.

It is rare for processes to exhibit lead-lag behavior, but many controllers exhibit such behavior.

[ Team LiB ]
[ Team LiB ]

3.9 Poles and Zeros

There are a number of different ways to represent process transfer functions. The "polynomial"
form is

Equation 3.49

The values of s that cause the numerator of Equation (3.49) to equal zero are known as the
"zeros" of the transfer function. The values of s that cause the denominator of Equation (3.49) to
equal zero are known as the "poles" of the transfer function.

The "pole-zero" form is

Equation 3.50

and complex poles (or zeros) must occur in complex conjugate pairs.

The "gain-time constant" form is the one that we use most often for control system design.

Equation 3.51

where ni is a numerator time constant and pi is a denominator time constant. This form is
normally used when the roots (poles) of the denominator polynomial are real.

Example 3.7: Comparison of Various Transfer Function Forms

Consider a transfer function with the following gain-time constant form
The polynomial form is

The gain-polynomial form is

and the pole-zero form is

The zero is 1/10, and the poles are –1/3 and –1/15.

Notice that the zero for Example 3.7 is positive. A positive zero is called a right-half-plane (RHP)
zero, because it appears in the right half of the complex plane (with real and imaginary axes).
RHP zeros have a characteristic inverse response, as shown in Figure 3-11 for  n = -10 (which
corresponds to a zero of +0.1).

Also notice that the poles are negative (left-half-plane), indicating a stable process. RHP (positive)
poles are unstable. Recall that complex poles will yield an oscillatory response. A pole-zero plot of
the transfer function in Example 3.7 is shown in Figure 3-13 [the pole locations are (-1/3,0) and
(-1/15,0) and the zero location is (1/10,0), with the coordinates (real,imaginary)]. For this
system, there is no imaginary component and the poles and zeros lie on the real axis.

Figure 3-13. Pole-zero location plot for Example 3.7 (x, poles; o, zero).

As poles move further to the left they yield a faster response, and increasing the magnitude of the
imaginary portion makes the response more oscillatory. This behavior is summarized in Figure 3-
14. Recall also that a process with a pole at the origin (and none in the RHP) is known as an
integrating system; that is, the system never settles to a steady state when a step input change
is made.

Figure 3-14. Effect of pole-zero location on dynamic behavior (x,

poles; o, zero). As poles become more negative, the response is faster.
As the imaginary/real ratio increases, the response becomes more
oscillatory.

Multiple RHP zeros cause multiple "changes in direction"; for example, with two RHP zeros, the
step response, initially going in one direction, switches direction, then switches back to the initial
direction.

[ Team LiB ]
[ Team LiB ]

3.10 Processes with Dead Time

Many processes have a delayed response to a process input, either due to transport lags (such as
flow through pipes) or high-order effects. In Section 3.3.4 we found that a time-delay of  time
units had a transfer function of e–s.

A first-order process combined with a time-delay has the transfer function

Consider a process with kp = 1, p = 10, and  = 5. A unit step input change at t = 0 yields the
response shown in Figure 3-15. We see that the time-delay shifts the response of the output.

Figure 3-15. Response of a first-order + dead time process.

[ Team LiB ]
[ Team LiB ]

3.11 Padé Approximation for Dead Time

As discussed in the previous section the transfer function for a pure time delay is e–s, where  is
the time delay. Some control system design techniques require a rational transfer function; the
Padé approximation for dead time is often used in this case.

A first-order Padé approximation is

A second-order Padé approximation is

Example 3.8: Application of the Padé Approximations for Dead Time

Consider the first-order + dead time transfer function, where the time-delay dominates the time
constant

The first-order Padé approximation yields the transfer function

and the second-order Padé approximation yields

A comparison of the step responses of g(s), g1(s), and g2(s) is shown in Figure 3-16. Notice that
the first-order approximation has an inverse response, while the second-order approximation has
a "double inverse response." The reader should find that there is a single positive zero for g1(s),
and there are two positive, complex-conjugate zeros of the numerator transfer function of g2(s).

Figure 3-16. Comparison of first-order + dead time response with first

and second-order Padé approximations for dead time.

Most ordinary differential equation numerical integrators require pure differential equations (with
no time delays). If you have a system of differential equations that has time delays, the Padé
approximation can be used to convert them to delay-free differential equations, which can then be
numerically integrated.

One of the many advantages to using SIMULINK is that time delays are easily handled so that no
approximation is required.

[ Team LiB ]
[ Team LiB ]

3.12 Converting State Space Models to Transfer

Functions
A general state space model can be converted to transfer function form, using the following steps.
Starting with the state space model

Take the Laplace transform of each term, assuming zero initial conditions

Solving for x(s), then y(s) (it should be noted that often D = 0)

where G(s) is a transfer function matrix. For example, the transfer function relating input j and
output i is

or in matrix form (with m inputs and r outputs)

Example 3.9: Isothermal CSTR

Consider the isothermal CSTR shown in Example 3.3 and Module 5. The state space model is
The first input (u1, manipulated) is the dilution rate (F/V), the second input (u2, disturbance) is
the feed concentration (C Af), and the output is the concentration of the intermediate component,
B. The eigenvalues of A (obtained by solving det(I – A) = 0) are –2.4048 and –2.2381 min-1.

The sequence of steps used to find the transfer function matrix is

and multiplying,

Also,

The manipulated input-output process transfer function for the reactor is

and the disturbance input-output transfer function is

The transfer function poles (–2.2381 and –2.4048) are equal to the eigenvalues of the A matrix.
Also, the positive zero (1/0.3549) in g11(s) yields the inverse response shown in Figure 3-5.

We see that it is straightforward to convert state space models to transfer function models. An n-
state system results in transfer functions that have a denominator polynomial that is nth order in
s, that is, with n poles. Sometimes the resulting transfer functions can be factored into lower
order transfer functions because of pole-zero cancellation (a value of a pole is equal to a value of
a zero). An example of pole-zero cancellation is shown in Exercise 13.

[ Team LiB ]
[ Team LiB ]

3.13 MATLAB and SIMULINK

MATLAB and SIMULINK are ideal for analyzing and simulating dynamic behavior. The Control Toolbox
uses LTI (linear, time invariant) objects to represent dynamic models. These objects can be state
space or transfer function models; details on how to use these are presented in Module 4.

Here we first present some useful MATLAB functions not covered in other modules.

(Convolution) and
Often it is necessary to multiply polynomials in s together. Although the following multiplication is
quite simple: (5s + 1)(3s + 1) = 15s2 + 8s + 1, a more complex multiplication, such as (2s2 +
1.5s + 1)(3s3 + 8s2 + 4s + 1) could be quite time consuming. The following conv (convolve)
command makes this straightforward. Also, the roots of the resulting polynomial are easily
calculated using the roots command.

» g = conv([2 1.5 1],[3 8 4 1])

g =
6.0000 20.5000 23.0000 16.0000 5.5000 1.0000
» p = roots(g)
p =
-2.1095
-0.3750 + 0.5995i
-0.3750 - 0.5995i
-0.2786 + 0.2836i
-0.2786 - 0.2836i

First, define an object, such one of the transfer functions in Example 3.6.

» num = [-10 1]; den = [45 18 1];

» sysc = tf(num,den)

Transfer function:
-10 s + 1
-----------------
45 s^2 + 18 s + 1
then plot the step response (use help step to view other options)

» step(sysc)

A number of other useful functions, such as pole and tzero are shown in Module 4.

Simulink
It is highly recommended that you interactively work through the initial sections of Module 2 to
understand how SIMULINK can be used to simulate dynamic systems.

[ Team LiB ]
[ Team LiB ]

3.14 Summary
The main objective of this chapter was to develop an understanding of dynamic process behavior.
Although Laplace transforms can be used to obtain analytical solutions to differential equations,
we will not be using them for this purpose during the rest of this textbook. The concept of a
transfer function is very useful for control system design and analysis. In future chapters we find
that transfer functions allow the use of algebra rather than differential equations when analyzing
control systems composed of a number of components (controller, valve, process, sensor, etc.).

The stability of a process is determined by the eigenvalues of the state space A matrix, or the
poles of the Laplace transfer function: if all are negative, the system is stable. Complex
eigenvalues (poles) correspond to underdamped (oscillatory) behavior, characterized by damping
factors less than 1. Dynamic responses are also affected by the values of the zeros (roots of the
numerator polynomial in s) of a transfer function. If zeros are positive (in the right-half-plane),
step responses have a characteristic inverse response.

A first-order + dead time model is the most common process representation. Time-delays are
represented by an irrational term (e–s) in the Laplace domain. First and second-order Padé
approximations are sometimes used for controller design; the approximations lead to right-half-
plane zeros and inverse response.

The initial and final value theorems of Laplace transforms can be used to find the limiting behavior
of transfer functions without having to obtain a full solution. The process gain is a measure of the
long-term change in a process output for a given input change.

State space models are easily converted to Laplace transfer function models. The use of LTI
objects (see Module 4) allows easy interconversion of models in MATLAB.

[ Team LiB ]
[ Team LiB ]

References

For more details on analytical mathematical methods (including Laplace transforms) to solve
differential equations, see: Boyce, W., and R. DiPrima, Ordinary Differential Equations and
Boundary Value Problems, 5th ed., Wiley, New York (1992).

[ Team LiB ]
[ Team LiB ]

Student Exercises

1: Solve for the time-domain output of a first-order transfer function to a step input
change.

2: A second-order process with one pole at the origin has the transfer function

Find the output as a function of time, for a unit step input change. Sketch the expected
behavior.

3: Use the initial value theorem to find the immediate response of a lead-lag transfer
function to a step input change at t = 0.

Also, use the final value theorem to find the long-term response of a lead-lag transfer
function to a step input change.

4: For the following second-order process with numerator dynamics, solve for the time-
domain output response to a step input change of magnitude u at t =0.

For k p = 1°C/Lpm, 1 = 3 min,  2 = 15 min, n = 20 min find the peak temperature

and the time that it occurs.

5: Consider an input-output transfer function that mimics two first-order processes in

parallel

where
 If the gain of g1 is positive and the gain of g2 is negative, find the conditions
(relationship between gains and time constants for the two transfer functions) that
cause a right-half-plane zero (resulting in inverse response to a step input change) in
gp(s).

6: Consider the state space model

Find the second-order differential equation in y. Hint: first solve for x 1 from the second
equation, then take the derivative and substitute into the first equation.

7: Consider the following state-space model

Which has the following input-output transfer function relationship

For a unit step change in the input, u(s) = 1/s:

i. Find the output at t=0 and as t approaches infinity, using the initial and final
value theorems.

ii. Find the time domain solution, y(t)

iii. Sketch the time domain behavior of y(t)

iv. Are your results for i, ii, and iii consistent?

8: As a process engineer with the Complex Pole Corporation, you are assigned a unit with
an exothermic chemical reactor. In order to learn more about the dynamics of the
process, you decide to make a step change in the input variable, the coolant
temperature, from 10°C to 15°C. Assume that the reactor was initially at a steady
state. You obtain the following plot for the output variable, which is reactor
temperature (notice that the reactor temperature is in °F). Use Figure 3-9 to help
answer the following questions.
 i. What is the value of the process gain? (show units)

ii. What is the value of ? (show units)

iii. What is the value of ? (show units)

iv. What is the decay ratio?

v. What is the period of oscillation? (show units)

vi. Write the second-order transfer function.

Hint: Use Figure 3-9 to assist you.

9: Match the transfer functions with the responses to a unit step input, shown in the
figure.

i.

ii.

iii.

iv.

v.
10: Consider the following state space model:

i. Find the transfer function model.

ii. Find (analytically) the time-domain output response to a unit step input change.
Sketch the expected response to a unit step input change.

11: As a process engineer, you decide to develop a first-order + time delay model of a
process using a step test. The process is initially at steady state, with an input flow
rate of 5 gpm and an output of 0.75 mol/L. You make a step increase of 0.5 gpm at
3:00 p.m. and do not observe any changes until 3:07 p.m. At 3:20 p.m., the value of
the output is 0.8 mol/L. You become distracted and do not have a chance to look at
the output variable again, until you leave for happy hour at a local watering hole at
6:30 p.m. You note that the output has ceased to change and has achieved a new
steady-state value of 0.85 mol/L. What are the values of the process parameters, with
units? Show your work.

12: Use the initial and final value theorems of Laplace transforms to determine the initial
and final values of the process output for a unit step input change to the following
transfer functions.
 i.

ii.

iii.

13: Consider the following state space model for a biochemical reactor. Since there are two
states (the A matrix is 2 x 2) we expect that the process transfer function will be
second-order. Show that pole-zero cancellation occurs, resulting in a first-order
transfer function. Find the values of the gain and time constant.

14: Match the transfer functions with the responses to a unit step input, shown in the
figure.

i.

ii.

iii.

iv.

v.
 15: Consider Example 3.4. For an impulse input of 30 kJ, find the value of the vessel
temperature immediately after the impulse input is applied.

16: Consider the following second-order transfer function

For a unit impulse input, find the output response as a function of time. What is the
peak change and when does it occur?

[ Team LiB ]
[ Team LiB ]

Chapter 4. Empirical Models

The previous two chapters have focused on (i) the development of fundamental models based on
material and energy balances (Chapter 2), and (ii) understanding dynamic behavior, with an
emphasis on linear systems (Chapter 3). In this chapter we discuss the development of empirical
models, that is, developing models based on plant tests. For continuous control-system design
(most of the focus of this text), a study of Sections 4.1–4.3 will be sufficient. The discrete models
developed in Sections 4.4 and 4.5 will be useful when model predictive control (MPC) techniques
are presented in Chapter 16, and when digital control is studied in Module 16.

After studying this chapter the reader should be able to:

Develop continuous first-order and integrator + dead time models from step tests

Estimate parameters for discrete-time autoregressive models based on input-output data

Calculate poles and zeros of discrete-time models

Develop finite step and impulse response models

The major sections of this chapter are as follows:

4.1 Introduction

4.2 First-Order + Dead Time

4.3 Integrator + Dead Time

4.4 Discrete-Time Autoregressive Models

4.5 Parameter Estimation

4.6 Discrete Step and Impulse Response Models

4.7 Summary

[ Team LiB ]
[ Team LiB ]

4.1 Introduction
For many processes there is not enough time, or it is not worth the effort, to develop fundamental
process models. Particularly if your main interest is tuning a specific control loop, it is likely that
you will develop a transfer function-based model by performing a plant test. The most common
plant test is to make a step change in the manipulated input (controller output) and observe the
measured process output response. Then a model is developed to provide the best match
between the model output and the observed plant output.

There are a number of important issues in developing an input-output model. Foremost is the
selection of the proper input and output variables. For many processes this is not trivial, because
a particular manipulated input may affect a number of measured outputs. Similarly, a measured
output may be affected by a number of manipulated inputs. In this chapter we assume that the
manipulated input and measured output have already been selected. The important discussion of
the selection of a particular input to be "paired" with a particular output is postponed until
Chapter 13.

It is common to base an input-output model on step responses. In this procedure we first bring
the process to a consistent and desirable steady-state operating point, then make a step change
in the input variable. An important decision is the magnitude of the step change to make.

If the step change in input is too small, the measured output may not change enough to
develop a reliable model. This is particularly true if the measured output is "noisy." Clearly,
the magnitude of step input must be enough so that the output "signal-to-noise" ratio is
high enough to obtain a good model.

If the step change in the input is too large, the output variable may change too much and
produce product that is "off-specification." This is not desirable because of the severe
economic penalty (i.e., the plant loses too much money while the step test is being
performed). Also, if the step input change is too large, nonlinear effects may dominate. That
is, the operating condition may become significantly different than the desired condition.

Clearly there is a trade-off here. The input must be changed enough to observe a change in the
output variable (it must increase above the noise level), yet not so much that the output variable
change is too great (causing an economic penalty).

As a process engineer conducting a plant step test, you will usually have some basic knowledge of
(or experience with) the input-output pair under consideration. Observations of the measured
output with time will provide an estimate of the standard deviation or variance of the
measurement noise. A rough estimate of the process gain will enable you to select an input
change magnitude so that the output change is "above the noise level." An estimate of the
process gain can often be obtained through steady-state material and energy balances.

In the following sections we show how to estimate parameters for some common simple models.
By far the most commonly used model, for control-system design purposes, is the first-order +
time-delay model discussed in the next section.

[ Team LiB ]
[ Team LiB ]

4.2 First-Order + Dead Time

Recall that a first-order + dead-time process, represented by the transfer function relationship

Equation 4.1

has the following output response to a step input change,

Equation 4.2

where the measured output is in deviation variable form. The three process parameters can be
estimated by performing a single step test on the process input.

The gain is found as simply the long-term change in process output divided by the change in
process input. Also the time delay is the amount of time, after the input change, before a
significant output response is observed. There are several easy ways to estimate the time
constant for this type of model.

Time for 63.2% Approach to New Steady State

In Equation (4.2), if we set t = p + , we find y(t = p + ) = kpu(1 - exp(–1)) = 0.632k pu.
Now, since k pu is simply the long-term change in the process output, then t63.2% = p +  is the
amount of time it takes for the output to make 63.2% of its ultimate change. This method of
estimating the time constant is illustrated in Figure 4-1.

Figure 4-1. Estimating the time constant from a response to a step

input.
Example 4.1: Numerical Application of 63.2% Method
Consider the response to a step input change at t = 1 minute, shown in Figure 4-2. The measured
output is shown as the noisy solid curve, and a "best fit" first-order + time-delay model is shown
as the dashed curve. Here we find the process gain from

Figure 4-2. Example of the step response with measurement noise.

The time delay is observed to be 1 minute, and the time constant is the time it takes (after the
time delay) for the output to change by 0.632(-2) = -1.3°C. That is, the time when y = 23.7°C.
In this case,

and the process transfer function is

Maximum Slope Method

We can also find (see Exercise 1) that the maximum slope of the output response to a step input
change at t = 0 occurs at t =  and is

So the time constant can be estimated from

A major disadvantage is that the slope estimate may not be accurate if there is significant
measurement noise. This approach is particularly useful for higher-order responses that are
approximated as a first-order + time-delay response, as shown in the following example.

Example 4.2: Maximum Slope Method

The response of a high order process is shown in Figure 4-3. Here, the dashed line is drawn
through the maximum slope of the output response. The slope is

Figure 4-3. Example of the slope method for estimation of time

constant.

Since the process gain is kp = y/u = 6°C/2 gpm = 3°C/gpm, and the time-delay is 8-1 = 7
min, the first-order + time-delay transfer function is estimated as
Two-Point Method for Estimating Time Constant
A technique similar to the first one, but using two points on the output response, is shown in
Figure 4-4. Here the time required for the process output to make 28.3% and 63.2% of the long-
term change is denoted by t28.3% and t63.2%, respectively. The time constant and time delay can
be estimated from (see Exercise 2)

Figure 4-4. Example of the two-point method for estimation of time

constant.

Limitation to FODT Models

The primary limitation to using step responses to identify first-order + dead-time transfer
functions is the amount of time required to assure that the process is approaching a new steady
state. That is, the major limit is the time required to determine the gain of the process. For large
time constant processes it is often desirable to use a simpler model that does not require a long
step test time. The integrator + dead-time model shown next will often be good enough for
control-system design.

[ Team LiB ]
[ Team LiB ]

4.3 Integrator + Dead Time

An integrator + dead-time process has the input-output transfer function relationship

Equation 4.3

and the output response to a step input

Equation 4.4

with the response shown in Figure 4-5. The gain, k, can be found from the slope by solving

Figure 4-5. Estimating parameters for an integrator + time-delay

model from a step input test.
and the time delay is clearly identified by the time required for a change in output.

Integrator + dead-time models are good for describing the behavior of "integrating processes,"
such as vessel liquid levels or gas drum pressures. They can provide a good short-term
approximation to the step response behavior of a first-order + dead-time process. Consider the
step response of a distillation column, which has a fairly long time constant. We see from Figure
4-6 that it takes roughly 5–6 hours to obtain good estimates for a FODT model.

Figure 4-6. Distillation column overhead composition response to a

step change of 0.01 kmol/min in reflux rate at t = 10 minutes.
The reader should be able to show that the transfer function is approximately

where the gain has units of mol%/(kmol/min) and the time unit is minutes.

The initial response of Figure 4-6 is "blown-up" in Figure 4-7. Notice that fewer than 20 minutes
(the step change is made at t = 10 minutes) is required to obtain a satisfactory integrator + time-
delay model. Clearly, plant operators (and managers) will be much more willing to tolerate a 20-
minute test than a 4- to 5-hour test.

Figure 4-7. Distillation column overhead composition response to a

step change of 0.01 kmol/min in reflux rate at t = 10 minutes. Focus is
on the short-term response.
The reader should be able to show that the estimated transfer function is approximately

[ Team LiB ]
[ Team LiB ]

4.4 Discrete-Time Autoregressive Models

Most of the models developed and used in this text are based on continuous time. In the most
general case of fundamental models, these are nonlinear ordinary differential equations. State
space models consist of sets of linear ordinary differential equations, where the states and inputs
are in deviation variable form. Transfer functions are used to denote a linear relationship between
inputs and outputs. All these previous models assume that the manipulated input and the
measured output signals are continuously available.

In practice, manipulated input changes are made at discrete time intervals and measured outputs
are available at discrete sample times. The physical inputs and outputs are continuous; only the
changes and measurements occur at discrete times. Since the focus of this textbook is on
controller design based on continuous models, the topic of discrete models may be skipped on a
first reading. Discrete models are primarily useful for model predictive control (Chapter 16) and
digital control design and analysis (Module 16).

In this section we cover discrete autoregressive models, in Section 4.5 parameter estimation of
discrete autoregressive models is covered, and in Section 4.6 we cover finite step and impulse
response models. It should be noted that the M ATLAB Control Toolbox has routines that convert
continuous-time models to discrete and vice versa. The details of the conversion are beyond the
scope of this text, but are summarized in the appendix; read Module 4 for explanations on how to
use these routines.

Introduction to Autoregressive Models

Discrete autoregressive models assume that an output at the current time step is a function of
the outputs and inputs at previous time steps. The general form is

Equation 4.5

where k represents the current time step, k - 1 the previous time step, and so forth. The notation
y(k) is used to indicate the value of the output at step k. If all of the outputs are brought to the
left hand side, Equation (4.5) can be written

Equation 4.6

This form is primarily used to lead to the notion of discrete input-output transfer functions
covered next.
Z-Transforms
Similar to the use of Laplace transforms for continuous systems, Z-transforms are used for
discrete systems.

The Z-transform of a discrete variable, y(k), is

Equation 4.7

The so-called backwards shift operator (z-1) is represented by

Equation 4.8

so Z[y(k - 2)] = z-2y(z), and so forth. Taking the Z-transform of each term in Equation (4.6), we
find

Equation 4.9

Solving for y(z) yields the discrete transfer function relationship

Equation 4.10

where the discrete transfer function is

Equation 4.11

The discrete transfer function is the discrete-time analogy to the continuous-time transfer
function based on the Laplace transform. Similar to continuous (Laplace) transfer functions,
discrete (z-transform) transfer functions can be written in several forms. Multiplying Equation
(4.11) by zn/zn, we find
Equation 4.12

For most process systems there is not an immediate effect of the input on the output, so b0 = 0.

Poles/Zeros of Discrete Models

Recall that the stability of a continuous input-output (transfer function) model is determined by
the values of the poles (the roots of the denominator polynomial in s). If the poles are negative,
then the model is stable. Similarly, the stability of a discrete input-output model is determined by
the poles of the denominator polynomial in z. For a discrete-time model, if all poles have a
magnitude less than 1 (i.e., they are "inside the unit circle") the model is stable. If any pole has a
magnitude greater than 1, the model is unstable.

The polynomial forms of Equations (4.11) and (4.12) can be factored into gain-pole-zero form

Equation 4.13

where zi and pi are the zeros and poles, respectively.

Example 4.3: Discrete Poles and Stability

Consider the simple first-order model

which has the transfer function

The pole is found by solving for the roots of the denominator polynomial. In this case

To illustrate the importance of pole values, we will study the following values for a1:
yielding the corresponding pole values

Also, realize that the modeling equation can be written

For simplicity, we will assume no input change, so u(k) = 0. Also, let the initial value of the output
be y(0) = 1. The first few values of y(k) for each value of a1 are shown in the table below.

a1 0.5 -0.5 1.5 -1.5

p -0.5 0.5 -1.5 1.5
y(1) -0.5 0.5 -1.5 1.5
y(2) 0.25 0.25 2.25 2.25
y(3) -0.125 0.125 -3.375 3.375
y(4) 0.0625 0.0625 5.0625 5.0625
Characteristic Oscillatory, Monotonic, Oscillatory, Monotonic,
behavior stable stable unstable unstable

As expected, when the magnitude of the pole is less than 1.0, the process is stable; when the
magnitude is greater than 1.0, the process is unstable. One thing that is unusual for this first-
order discrete example is that the output can oscillate even with a constant (in this case, 0) input
value. For continuous-time systems there must be at least two states (the process must be at
least second-order) for oscillation to occur. This is because the condition for oscillation in
continuous-time systems is that the eigenvalues must be complex; for this to happen, there must
be at least two eigenvalues, each complex-conjugates of the other. Discrete time systems can
oscillate, as long as a pole has a negative value; if the negative pole has a magnitude less than
one, it is a stable oscillation.

Final and Initial Values Theorems for Discrete Systems

Recall that the process gain of a continuous system could be determined by setting s = 0.
Similarly, the gain of a discrete-time system can be found by setting z = 1.

Final Value Theorem for Discrete-Time Systems

The formal statement of the final value theorem is (where t is the sample time)
For a discrete-time input-output model

subject to a unit step input

we find

so, applying the final value theorem

so we can find the long-term behavior of the process output, subject to a unit step input, simply
by setting z = 1 in gp(z).

Initial Value Theorem for Discrete-Time Systems

Similarly, the initial value theorem is

[ Team LiB ]
[ Team LiB ]

4.5 Parameter Estimation

Often when discrete linear models are developed, they are based on experimental system
responses rather than on converting a continuous model to a discrete model. The estimation of
parameters for discrete dynamic models is no different from linear regression (least squares)
analysis.

The measured inputs and outputs are the independent variables, and the dependent variables are
the outputs. For simplicity, consider the following model where two previous values of the input
and output are used to predict the next value. This model has four parameters

Equation 4.14

For the system of N data points we can write the following

Notice that k = 1 is an arbitrary starting point for calculating the output, and that two previous
data points for the output and the input are needed to start the algorithm. To make it clear that
we are performing a model-based prediction of yk, we will use the notation k to indicate a model
prediction

Equation 4.15

or, using matrix-vector notation

Equation 4.16

where
Equation 4.17

The objective is to choose a set of parameters (a1, a2, b1, b2) to minimize the square of the
residuals (the differences between the model outputs and measured outputs).

Equation 4.18

The sum of the squares of the residuals can be written in matrix notation as

Equation 4.19

where Y is the vector of measured outputs, Y = [y(1) . . . y(N)]T.

The solution to this optimization problem is (Ljung, 1996)

Equation 4.20

Example 4.4: Process Identification

Although the response to a step input has often been used for process identification, here we
apply a pseudo-random binary sequence of inputs (the input changes randomly between two
values) to estimate parameters in a discrete time model. The input and output responses are
shown in Figure 4-8; the open circles represent the measured outputs obtained with a sample
time of 0.25 minutes. The variables are in deviation form.

Figure 4-8. Output response to a pseudo-random binary input

sequence. Open circles represent output data obtained with a sample
time of 0.25 minutes.
The output and input data as a function of the discrete time index are:

k -1 0 1 2 3 4 5
y 0 -0.0889 0.0137 0.1564 0.4618 0.1771 0.3446
u 1.0000 1.0000 1.0000 -1.0000 1.0000 -1.0000 -1.0000
k 6 7 8 9 10 11 12
y 0.2171 -0.1558 0.0485 -0.1879 -0.1123 0.0463 0.2003
u 1.0000 -1.0000 1.0000 1.0000 1.0000 1.0000 -1.0000
k 13 14 15 16 17 18 19
y 0.5007 0.3846 -0.0172 0.1513 -0.1162 0.1134 0.0502
u -1.0000 1.0000 -1.0000 1.0000 -1.0000 -1.0000 -1.0000

The  (input-output data) matrix,  parameter vector, and discrete transfer functions are shown
below.
The M ATLAB and SIMULINK .m and .mdl files used to generate this example are shown in the
Appendix. It should be noted that the discrete transfer function can be converted to a continuous
transfer function using d2c to find

This is virtually identical to the continuous transfer function used to generate the original data.

This example illustrated the response of a perfectly modeled system (no measurement noise).
The approach can also be applied to a system with arbitrary inputs, and with noisy
measurements. See Exercise 8 for the same process, with output data corrupted by measurement
noise. To verify model parameter estimates it is common to use a portion of experimental data to
estimate parameters, then another portion of the data in "simulation" mode to see how well the
model predicted outputs match the data.

The data was analyzed in a batchwise fashion, that is, all of the data were collected before the
parameter estimation was performed. For on-line (real-time) estimation and control, we would
prefer to re-estimate parameters each time we obtain a new data point. This can be done by
using a "moving horizon" of past data points (each time a new data point is collected, the oldest
one is thrown out), or by using recursive identification techniques. Recursive techniques are
commonly applied in adaptive control, where the model is updated at each time step and a control
calculation is performed based on the updated model.

[ Team LiB ]
[ Team LiB ]

4.6 Discrete Step and Impulse Response Models

The common model predictive control techniques presented in Chapter 16 are based on step or
impulse responses.

Step Response Models

Finite step response (FSR) models are obtained by making a unit step input change to a process
operating at steady state. The model coefficients are simply the output values at each time step,
as shown in Figure 4-9. Here, si represents the step response coefficient for the ith sample time
after the unit step input change. If a nonunit step change is made, the output is scaled
accordingly.

Figure 4-9. Illustration of step response parameter identification.

The step response model is the vector of step response coefficients,

where the model length n is long enough so that the coefficient values are relatively constant
(i.e., the process is close to a new steady state).
Impulse Response Models
Another common form of model is known as a finite impulse response (FIR). Here, a unit pulse is
applied to the manipulated input, and the model coefficients are simply the values of the outputs
at each time step after the pulse input is applied. As shown in Figure 4-10, hi represents the ith
impulse response coefficient.

Figure 4-10. Illustration of step response parameter identification.

There is a direct relationship between step and impulse response models:

Figure 4-11 illustrates how impulse response coefficients can be obtained from step responses.
The impulse response coefficients are simply the changes in the step response coefficient at each
time step. Similarly, step response coefficients can be found from impulse responses; a step
response coefficient is the sum of the impulse response coefficients to that point.

Figure 4-11. Obtaining impulse response models from step response

data.
[ Team LiB ]
[ Team LiB ]

4.7 Summary
The first sections of this chapter illustrated the development of first-order + time-delay and
integrator + time-delay models based on step input changes. These models will be sufficient for
many simplified controller designs. Higher order models can easily be developed by "fitting"
(adjusting) parameters to obtain a best match between model predictions and the actual
measured process output responses.

Linear regression (least squares) techniques can be used to estimate parameters in discrete time
autoregressive models. Usually, parameters are estimated based on a subset of the available
experimental data. The model parameters are then verified by applying the model to a different
subset of the data. If the model predictions match the measured outputs reasonably well, then
the discrete model is usually acceptable for discrete control system design and analysis.

The stability of a discrete transfer function model is determined from the values of the poles of
the denominator polynomial. If the poles are less than 1 in magnitude, the discrete model is
stable. Module 16 covers classical discrete controller design procedures. Discrete step and impulse
response models are often used in model predictive control (Chapter 16).

[ Team LiB ]
[ Team LiB ]

References

An overview of system identification techniques is provided by Ljung: Ljung, L., System

Identification, Chapter 58 ( pp. 1033–1054) in The Control Handbook, W.S. Levine (ed.), CRC
Press (1996).

[ Team LiB ]
[ Team LiB ]

Student Exercises

1: Consider a first-order + time-delay transfer function

which has the following response to a step input change at t = 0:

Show that the maximum rate-of-change (slope) of the output occurs at t =  . Also, find this slope.

2: Consider the response of a first-order + time-delay transfer function to a step input change. Find the
value of the output (y ) at the following times, as a fraction of the long-term output change

Find the time constant and time-delay based on the values of t 1 and t 2 .

3: Consider the following step response. Estimate parameters for a first-order + time-delay model using
the three techniques shown in Section 4.3 . Include the units for each parameter.
4: Consider the following step response. Estimate the parameters for an integrator + time-delay model,
including the units for each parameter.

5: Consider the following continuous state space model:

 a. Find the continuous transfer function model (do this analytically).

b. For a sample time of 0.25, find the discrete state space and transfer function models (use M
; see Module 4 ).

c. Compare the step responses of the continuous and discrete models (use MATLAB ). What do you
observe?

6: Consider the discrete-time model

using z -transforms, find the corresponding model represented as equations (4.11 ), (4.12 ) and (4.13
). Discuss the stability of the process.

7: Consider a unit step input change made at k = 0, resulting in the output response shown in the plot
and table below.

k -1 0 1 2 3 4 5 6 7 8 9 10
y 0 0 0.1044 0.3403 0.6105 0.8494 1.0234 1.1244 1.1616 1.1531 1.1184 1.0746
u 0 1 1 1 1 1 1 1 1 1 1 1
k 11 12 13 14 15 16 17 18 19 20
y 1.0336 1.0023 0.9828 0.9744 0.9742 0.9790 0.9860 0.9929 0.9985 1.0022
u 1 1 1 1 1 1 1 1 1 1
 Estimate the parameters for a discrete linear model with the form

Compare the step response of the estimated model with the data. Use MATLAB to convert the discrete
model to a continuous model. Compare the step responses of the discrete and continuous models.

8: Consider Example 4.4 with measurement noise on the output variable, as shown below. Estimate the
discrete model parameters based on this data. How do the parameters compare with those of Example
4.4 ? Compare the step responses of the two models.

The input/output data are

k -1 0 1 2 3 4 5
y 0.0741 -0.0857 -0.0399 0.1663 0.4065 0.1521 0.3910
u 1.0000 1.0000 1.0000 -1.0000 1.0000 -1.0000 -1.0000
k 6 7 8 9 10 11 12
y 0.2284 -0.2569 0.0910 -0.1737 -0.1260 0.0668 0.1958
u 1.0000 -1.0000 1.0000 1.0000 1.0000 1.0000 -1.0000
k 13 14 15 16 17 18 19
y 0.4976 0.3724 0.0119 0.0927 -0.0528 0.1357 0.0580
u -1.0000 1.0000 -1.0000 1.0000 -1.0000 -1.0000 -1.0000
 9: Consider the continuous state space model (where the time unit is minutes)

a. Find the eigenvalues and the transfer function (use MATLAB for these calculations, if desired)

b. Using a sample time of 3 minutes, find the discrete state space model and the discrete transfer
function. Refer to the Appendix for the form of the discrete state space model, and Module 4
understand how to use M ATLAB for these computations.

[ Team LiB ]
[ Team LiB ]

Appendix 4.1: Files Used to Generate Example 4.4

Figure A4-1. Simulink File .

% use PRBS to estimate a discrete model of the vdv reactor

%
T = 0:0.25:5; % time vector (row)
T = T'; % time vector (column)
seed = [3;0.5]; % seed for noise generation
randn('state',seed); % using the same seed each time
uu = 0.5 1 0.25*randn(length(T),1); % mean of 0.5 with variance
% of 0.25
U = 2*round(uu)-1; % creates PRBS with -1 and 1 values
%
sim('est_vdv'); % runs simulation of linear van de vusse
% diagram
figure(1); % plot input-output data
subplot(2,1,1),plot(tp,yp,'k',t,y,'ko');
xlabel('time, min'), ylabel('y')
title('PRBS estimation example')
subplot(2,1,2),plot(tp,up,'k'); xlabel('time, min'),ylabel('u')
axis([0 5 -1.1 1.1])
% % generate phi matrix for estimation
for j = 4:22;
phi(j-3,:) = [y(j-2) y(j-3) u(j-2) u(j-3)];
end
%
theta = inv(phi'*phi)*phi'*y(3:21) % estimate parameters
%
num = [theta(3) theta(4)]; % numerator of discrete transfer function
den = [1 -theta(1) -theta(2)]; % denominator of discrete transfer function
sysd = tf(num,den,0.25) % create discrete tf object
tzero(sysd) % calculate zeros
pole(sysd) % calculate poles
syszpk = zpk(sysd) % zero-pole-k form

[ Team LiB ]
[ Team LiB ]

Appendix 4.2
Discrete time models can be obtained directly from continuous time models, as summarized in
this appendix. The MATLAB Control Toolbox can be used for these conversions, as shown in Module
4.

Consider a continuous state space model of the following form (where we have assumed D = 0)

Equation A4.1

A similar discrete state space model has the following form

Equation A4.2

Assuming that the sample time is t, and that the input u is held constant between time tk and
tk+1, (A4.1) can be integrated to yield

which has the solution (although the matrix exponential is not an intuitive concept, the
computation is readily performed by MATLAB)

which is usually written in the following form

where
The stability of this discrete state space model is determined by the eigenvalues of , which must
have a magnitude less that 1.0 for stability.

The discrete transfer function is, similar to the continuous time case

[ Team LiB ]