You are on page 1of 48

An Experimental Validation of Data-Based PID

Controller Design for Linear Time-Invariant Systems

by

Timothy Fitzgerald Stratton

Capstone Design Project Report

Submitted to the Faculty

of the

College of Engineering and Technology

in

Partial Fulfillment of the Requirements

for the Degree of

Bachelor of Science

in

Mechanical and Manufacturing Engineering

(May, 2011)

College of Engineering, Technology and Computer Science


Tennesse State University
Nashville, Tennessee
An Experimental Validation of Data-Based PID Controller Design
for Linear Time-Invariant Systems
Capstone Design Project Report

Approval Recommended:

Project Advisor

Date

Course Instructor

Date

Department Head

Date

Approved:

Dean, College of Engineering, Technology and Computer Science

Date

ii
Abstract
An Experimental Validation of Data-Based PID Controller Design
for Linear Time-Invariant Systems
T. F. Stratton, Senior Mechanical and Manufacturing Engineering

The main focus of this project is to design controllers directly from the data of
real-world, linear time-invariant (LTI) systems. This project aims to experimentally
validate the recently introduced data-based PID controller design techniques of [5].
The results of this project will show that the complete parameter set of stabilizing
proportional-integral-derivative (PID) controllers can be calculated solely from the
frequency response data of the plant to be controlled. It will be shown that con-
troller gains chosen outside the stabilizing parameter set will cause instability when
implemented upon the plants. Using the LTI system, the ECP Industrial Emulator
Apparatus, it will be shown that frequency response data is sufficient for controller
design, and that data-based controller design for real-world systems can provide all
the features that model-based techniques can provide.

iii
Dedication
To my Family and especially my Mother, who patiently bared with me through my
college education, supporting me in every way possible and serving as an inspiration
from day to day; I dedicate this project to you all.

iv
Acknowledgment
This project was written under the direction and supervision of Dr. Lee-Hyun
Keel. I would like to sincerely thank him for inspiring me to push my limits and excel
beyond mediocre. I express my deepest appreciation to him for the patience, interest
and assistance given to me.

v
Table of Contents

Abstract ii

Dedication iii

Acknowledgment iv

Table of Contents vi

List of Tables viii

List of Figures ix

1 Introduction 1

2 Data-Based PID Controller Design Approaches 3


2.1 Classical Data-Based PID Controller Design . . . . . . . . . . . . . . 3
2.2 Modern Data-Based PID Controller Design . . . . . . . . . . . . . . . 8

3 Theoretical Overview 11
3.1 Mathematical Preliminaries . . . . . . . . . . . . . . . . . . . . . . . 11
3.2 Phase, Signature, Poles, Zeros, and Bode Plots . . . . . . . . . . . . . 13
3.3 Synthesis and Design of PID Controllers . . . . . . . . . . . . . . . . 15

4 The Experimental Test-Bed 19


4.1 The Electromechanical Plant . . . . . . . . . . . . . . . . . . . . . . . 20

5 Procedure for Data-Based Controller Design 23


5.1 Collecting Frequency Response Data . . . . . . . . . . . . . . . . . . 23

vi
5.2 Data-Based PID Controller Design . . . . . . . . . . . . . . . . . . . 25

6 Analysis of the Results 31

7 Conclusion 37

References 38

vii
List of Tables
2.1 Cohen-Coon Recommended Tuning Parameters . . . . . . . . . . . . 8

viii
List of Figures
2.1 Ziegler-Nichols Step Response Method . . . . . . . . . . . . . . . . . 5
2.2 Cohen-Coon Reaction Curve Method . . . . . . . . . . . . . . . . . . 7

4.1 The ECP Systems Industrial Emulator Apparatus . . . . . . . . . . . 19


4.2 Industrial Emulator Apparatus Layout. [4] . . . . . . . . . . . . . . . 21

5.1 Industrial Emulator 50.079 rad/s Sinusoidal Response in Steady-State. 23


5.2 Graphical Representation of Magnitude and Phase Parameters. . . . . 24
5.3 Bode Diagram for Open-Loop Plant. . . . . . . . . . . . . . . . . . . 26
5.4 Graph of the function, g(ω) . . . . . . . . . . . . . . . . . . . . . . . 28
5.5 Stability Set at Kp = 0.25 . . . . . . . . . . . . . . . . . . . . . . . . 29

6.1 Step Response with stable PID gains . . . . . . . . . . . . . . . . . . 31


6.2 Step Response with unstable PID gains . . . . . . . . . . . . . . . . . 32
6.3 Step Response testing sensitivity of Kd . . . . . . . . . . . . . . . . . 33
6.4 Step Response testing sensitivity of Ki . . . . . . . . . . . . . . . . . 34
6.5 Step Response testing sensitivity of Hardware Limitations . . . . . . 35
6.6 Step Response with various PID parameters . . . . . . . . . . . . . . 36

ix
1 Introduction
Controller design can take place using model-based or data-based methods. Partic-
ularly, data-based design relies solely on data taken from the system to be controlled.
Various methods for data-based design exist. One particular method for data-based
design was introduced in [5], introducing methods for designing first-order- and PID-
type controllers for continuous, linear time-invariant systems possibly cascaded with a
time-delay. This method for data-based controller design determines entire stabilizing
controller parameter sets and, analogous to model-based controller design, can ad-
dress performance goal problems also. These techniques were mathematically proven
and illustrated via academic examples in [5]. Thus, there is a need to verify these
techniques on real-world systems. Data from real-world systems will contain noise
that is generally not present in academic examples. This report will detail the proce-
dure taken to verify the results of [5] on the Educational Control Products Industrial
Emulator Apparatus, which is currently located in the Control Systems Laboratory.
The results of this project will demonstrate that data-based PID controller design,
theoretically proven in [5], can determine entire stabilizing controller parameters sets
for real-world systems, thus showing that data-based controller design is an attractive
complement to model-based controller design for real-world systems.
The organization of this report is as follows. Chapter 2 gives a review of data-
based design methods. Chapter 3 gives the theoretical overview of the data-based
design technique, introduced in [5], used to determine the complete set of stabilizing
PID controller parameters. This project will validate this technique determine the
entire stability set for the Industrial Emulator. Chapter 4 describes in detail the
experimental test-bed, i.e. the Industrial Emulator, which was used to validate the

1
techniques introduced in Chapter 3. Chapter 5 will outline the procedure taken for
data-based PID controller design for the Industrial Emulator. Chapter 6 will analyze
the results obtained through the procedure in Chapter 5, verifying the design param-
eters obtained on the Industrial Emulator. Finally, in Chapter 7, some concluding
remarks are made.

2
2 Data-Based PID Controller Design

Approaches
Data-based controller design by H. Nyquist in [7] and H. W. Bode in [2] constituted
the “model free” part of classical control theory. Their approaches are represented by
the frequency response of open-loop systems. In 1932, Nyquist introduced his criterion
as a means of predicting stability of a closed-loop system based on the frequency
response measurements made on the open-loop system. This also allowed designers
to determine the gain and phase margins amongst other performance measures for
closed-loop stability. Bode later revised Nyquists techniques into a graphical design
approach that ultimately achieves closed-loop stability margins.
From the frequency response, various properties of the system can be determined.
Closed-loop stability can be determined using the Nyquist criterion. Gain margins
and phase margins can also be determined.
2.1 Classical Data-Based PID Controller Design
Early feedback control devices implicitly or explicitly used the ideas of propor-
tional, integral, and derivative action in their structures. However, it wasn’t until
Minorsky’s “Direction Stability of Automatically Steered Bodies”, published in 1922,
that rigorous theoretical consideration was given to PID control.
Manual PID tuning is a common method for determining PID gains. In this
method, the designer must observe the sinusoidal response of the system, setting
kd = 0 and ki = 0. Taking several values of kp , the designer must increase kp until
the output of the system to a sinusoidal, y(t) oscillates. At that point, the designer
reduces the kp value obtained by half. Setting the value kp∗ = kp , and with ki = 0,
the designer then constructs the root locus of the system for 0 ≤ kd ≤ ∞. From this

3
root locus, he selects the kd value that produces the smallest overshoot and sets that
value as kd∗ . With kp = kp∗ and kd = 0, the designer constructs the root locus for
0 ≤ ki ≤ ∞ and select his ki value, setting it to ki∗ . Finally, he collects the closed-
loop step response using the gains previously determined, i.e. kp = kp∗ , kd = kd∗ ,
and ki = ki∗ , and verifies that performance goals are met. If not, he makes minor
adjustments until they are met.
The empirical method above requires that the system be operated near instability.
This could prove harmful to the system and costly to controller designers. For this
reason, many other classical data-based approaches have be developed. Through ex-
tensive empirical work, John G. Ziegler and Nathaniel B. Nichols innovated a method
for determining PID gains for stable plants. Using the Ziegler-Nichols Step Response
Method, control design engineers merely need to collect the stable plant’s open-loop
step response, determine the maximum slope of the output, which is called the curve
of reaction, and where it intersects the horizontal and vertical axes, and measure
the corresponding lengths of the horizontal axis intercept, L, and the vertical axis
intercept, A. Using these values, PID gains can be determined through the following
equations:
1.2 0.6 O.6L
kp = , ki = , kd = .
A AL A

Inserting the gains into the equation

ki
C(s) = kp + + kd s,
s

yields the model for the controller. To circumvent the problems of pure differentiation
kd s
when the error signals are contaminated with noise, substitute kd s = 1+Td s
. Thus,

4
the controller model becomes

ki kd s
C(s) = kp + + (2.1)
s 1 + Td s

where
kp is the proportional gain,
ki is the integral gain,
kd is the derivative gain, and
Td is a small positive value that is usually fixed. [1]
The formulas given above for controller parameters were selected to obtain an ampli-
1
tude decay of 0.25, meaning after the first overshoot, the amplitude will reduce by 4

its initial value after a single oscillation. [1]

Figure 2.1: Graphical determination of parameters, A and L. [1]

The Ziegler-Nichols Frequency Response Method is similar to that of its sister


method. In this method, the frequency response is used to determine the ultimate

5
gain and ultimate period for use in calculating PID gains. One simply needs to
switch off the integral and differential actions and adjust the proportional gain until
the response of the system to a sinusoid is a pure periodic oscillation. The kp value
at which oscillation occurs is the ultimate gain, ku . One could also simply determine
the gain margin for the plant and choose the upper bound. This value can then be
used to produce the oscillation from which the ultimate gain and ultimate period, Tu ,
can be derived. Using the ultimate gain and ultimate period, the designer can now
use the following equation to determine PID gains:

1.2ku
kp = 0.6ku , ki = , kd = 0.075ku Tu (2.2)
Tu

One pitfall of this method is that it requires that the closed-loop system operate near
instability. Careful execution is necessary to avoid damaging the physical system. [1]
An alternative to the Ziegler-Nichols Frequency Response Method was proposed
by Åström and Hägglund. They proposed the use of a relay to induce a self sustaining
oscillation in the loop. The relay feedback can be used to determine the ultimate gain
and ultimate period, which can then be inserted into (2.2) above to determine PID
parameters. [1]
The tuning methods above require only a limited amount of knowledge of the
plants and simple formulas are given to calculate the controller parameters. These
methods were design to achieve a 41 th amplitude decay ratio for the load disturbance
response. Some criticisms of these methods are that they give little emphasis to
measurement noise, sensitivity to process variations, and setpoint response. While
these methods achieve good load disturbance rejection, the product could be a closed-
loop system that is poorly damped and sometimes has poor stability margins. [1]
A fourth classical method for data-based PID controller design is the Cohen-

6
Cool Reaction Curve Method. Cohen and Coon observed that for non-selfregulating
systems, the Ziegler-Nichols Step Response Method rendered undesirable results, and
thus they suggested that the deadtime of the system be considered in PID parameter
calculations. Approximating such systems with a first-order-deadtime-model, i.e.,

Ke−θs
y(s) = u(s) (2.3)
1 + Ts

and fitting the model parameters of (2.3) to the observed step response, seen in Figure
2.2, it is relatively straight forward to calculate the PID controller gains.

Figure 2.2: Estimating Model Parameters from the Process Reaction Curve.

The model parameters are found as follows, using the response to a step of am-
plitude, B:

1. Locate the inflection point, i.e., the point where the change in slope equals zero.

2. Draw a straight line through the inflection point, with the same gradient of the
reaction curve at that point.

7
3. The point where this line crosses the initial value of the output, assuming the
initial value to be zero in Figure 2.2, gives the apparent time delay, θ.

4. The straight line reaches the steady state value of the output at the time T + θ.

5. The gain K is given by A/B.

Using the procedure above, PID controller parameters may be calculated using
the equations given in Table 2.1.

Controller Type Proportional Gain, kp Integral Gain, ki Derivative Gain, kd


T θ
P Kθ
(1 + 3T )
T θ 30+3θ/T
PI Kθ
(0.9 + 12T ) θ( 9+20θ/T )
T 4 θ 32+6θ/T 4
PID ( + 4T
Kθ 3
) θ( 13+8θ/T ) θ 11+2θ/T

Table 2.1: Recommendation of Tuning Parameters for Cohen-Coon Reaction Curve


Method.

Note that the Cohen-Coon modifications to the Ziegler-Nichols Tuning rules are quite
insignificant when the deadtime, θ, is small relative to the time constant, T , but can
be important for large θ. Also, note that all the methods described above do not
determine whole stabilizing PID parameter sets. They only determine PID gains
with a prescribed level of performance.
2.2 Modern Data-Based PID Controller Design
Some modern methods for data-based PID controller design utilize the computa-
tional power of computers to design PID parameters. These methods often compute
PID gains and store them within a database for later use. One particular method was
proposed by Yamamoto, Takao, and Yamada in [9], in which they stated that “since
most processes have nonlinearities, controller design schemes to deal with such system
are required”. They introduce a new design scheme for PID controllers based on a
data-driven technique for nonlinear systems. This technique automatically generates

8
a suitable set of PID parameters based on input/output pairs of the controlled object
stored in a database. Their scheme can adjust PID parameters in an online manner
even if the system has nonlinear properties and/or time-variant system parameters.
The main feature of this new design scheme is that the PID parameters are “up-
dated corresponding to the control error in an online manner and they are stored in a
database.” They propose an algorithm to avoid excessive stored data accumulation.
Finally they present the scheme through some simulated and empirical examples.
Another modern data-based PID controller design technique was introduced for
discrete time systems by Mitra, Keel, and Bhattacharyya in [6], in which they intro-
duce a procedure for designing a digital PID controller for an arbitrary stable plant
based on the input/output data of the plant. They first collect the impulse response
for the plant, then truncate the data to a certain number of samples and approximate
the z-transform of the output. Utilizing the Tchebyshev representation for discrete-
time systems, the z-transform of the output is then used to determine the set of linear
inequalities in two parameters while one is kept constant. These inequalities are used
to calculate the stability region for the two parameters on each level of the third
parameter which is swept, thus rendering the entire stabilizing set. They show that
by increasing the number of sample data points, the stabilizing set resembles more
closely the actual stability set. Finally they address various performance specifica-
tions such as gain margin, phase margin, and overshoot, illustrating the results with
examples.
A technique proposed by Chen, Wang, and Wu in [3] introduces a new method for
obtaining the ultimate gain, ku and ultimate frequency, Tu , for the use of obtaining
PID parameters, as opposed to the Ziegler-Nichols Frequency Response Method and
Åström and Hägglund Relay Method. They show that the frequency characteristic,
ku and Tu , can be obtained by first using a high-pass filter to do some preprocessing

9
on the step response data, then doing a fast Fourier transform to obtain the frequency
characteristics. This method can obtain ku and Tu in only one step with considerable
accuracy. Finally they verify their results on a first-order inertia link with a time-
delay.
Balaguer, Wahab, Katebi, and Vilanova in [8] stated that “proper tuning of multi-
variable PID controllers is not a straightforward task” and that “this is especially true
when data-driven models and [newer] controller design methodologies are adopted to
increase the industry acceptance of controller implementation and commissioning”.
They propose a tuning methodology that selects PID parameters by matching the
PID controlled closed-loop system response to a Figure of Merit. Thus, PID tuning
is simplified to minimizing certain error metrics when compared to an optimal PID
controller. The metric to be minimize is obtain directly from the frequency response
data. The goal is to tune and retune parameters in such a way that the discrepancy
(e.g the residual) between the PID controller obtain from the frequency data and the
optimal PID controller is minimized.

10
3 Theoretical Overview
In this chapter, the theoretical background of the project will be given. In the first
section, some mathematical preliminaries are discussed, and in the following section,
the theory for PID synthesis and design will be detailed.
3.1 Mathematical Preliminaries
In this section, some notation and technical results will be developed which will
be used later. First, consider a real rational function

A(s)
R(s) = . (3.1)
B(s)

A(s) and B(s) are polynomials with real coefficients and degrees of m and n, respec-

tively. Assuming A(s) and B(s) have no zeros on the ω axis, denoting zR+ , p+
R , zR ,

and p−
R as the numbers of open right half plane (RHP) and open left half plane (LHP)

zeros and poles of R(s), and denoting ∆∞


0 ∠R(ω) as the net change in phase of R(ω)

as ω goes from 0 to +∞, we have

π −
∆∞
0 ∠R(ω) = [z − zR+ − (p− +
R − pR )]. (3.2)
2 R

(3.2) shows that each LHP zero and each RHP pole contributes +π/2 to the net
phase change, and that each RHP zero and each LHP pole contributes −π/2 to the
net phase change.
The Hurwitz signature of R(s) is defined by

σ (R) := zR− − zR+ − (p− +


R − pR ). (3.3)

11
Decomposing R(ω) into its real and imaginary parts yields

R(ω) = Rr (ω) + Ri (ω), (3.4)

where Rr (ω) and Ri (ω) are real, rational functions in ω. From (3.4), it is clear that
Rr (ω) and Ri (ω) have no real poles for ω ∈ (−∞, ∞) since R(s) has no imaginary
axis poles. To compute the net change in phase, i.e. the left hand side of (3.2), it
proves convenient to develop formulas in terms of Rr (ω) and Ri (ω). Because R(s) is
real, it is clear that ω0 = 0 is always a zero of Ri (ω). Therefore, let

0 = ω0 < ω1 < ω2 < · · · < ωl−1 (3.5)

denote the real, finite non-negative zeros of Ri (ω) of odd multiplicities, let


+1, if x > 0





sgn[x] = 0, if x = 0 , (3.6)





−1, if x < 0

and define ωl = ∞− , noting that for a real function f (t)

f (t−
0 ) := lim f (t), f (t+
0 ) := lim f (t) (3.7)
t→t0 ,t<t0 t→t0 ,t>t0

Lemma 1 : (Real Hurwitz Signature Lemma) [5]


For an even relative degree, i.e. n − m is even,

l−1
!
X
(−1)j sgn[Rr (ωj )] + (−1)l sgn[Rr (ωl )] (−1)l−1 sgn Ri (∞− ) .
 
σ(R) = sgn[Rr (ω0 )] + 2
j=1

12
For n − m is odd

l−1
!
X
(−1)j sgn[Rr (ωj )] (−1)l−1 sgn Ri (∞− ) .
 
σ(R) = sgn[Rr (ω0 )] + 2
j=1

Proof is these formulas is given in the appendix of [5].


3.2 Phase, Signature, Poles, Zeros, and Bode Plots
Let P denote a LTI plant and P (s) its rational transfer function with z + , p+ , (z − ,
p− ) denoting the numbers of RHP (LHP) zeros and poles, and n (m) the denominator
(numerator) degrees. Let the relative degree of P (s) be denoted rP and be defined
by
rP := n − m.

As defined earlier, the signature of P is

σ(P ) = (z − − z + ) − (p− − p+ ). (3.8)

Lemma 2 : [5]

1 dPdb (ω)
rP = − · (3.9)
20 d(log10 ω) ω→∞
2 ∞
σ(P ) = ∆ ∠P (ω) (3.10)
π 0

where
Pdb (ω) := 20 log10 |P (ω)| .

P roof : (3.9) states that the relative degree of P (s) is equivalent to the high
frequency slope of the Bode magnitude plot and (3.10) states that the signature can
be found from the net change in phase from the Bode phase plot.

13
Assuming P (s) has no poles or zeros on the ω axis, it can also be written

σ(P ) = −(n − m) − 2(z + − p+ ), (3.11)

or
σ(P ) = −rP − 2(z + − p+ ). (3.12)

Thus, if P (s) is stable, z + − p+ can be found using the Bode Diagram of P . The Bode
Diagram of P can be obtained experimentally by measuring the frequency response
of the system. For a stable plant, p+ = 0, thus z + can be determined from the Bode
frequency data.
If P is an unstable LTI plant with a rational transfer unknown to the control
design engineer, assuming P has no imaginary axis poles, also that a known feedback
controller C(s) stabilizes P , and the closed-loop frequency response can be measured
and is denoted by G(ω) for ω ∈ [0, ∞), then

G(ω)
P (ω) = (3.13)
C(ω)(1 − G(ω))

which is the frequency response of the unstable plant. Knowledge of C(s) and G(ω)
is sufficient to determine the numbers of z + and p+ , i.e. the number of RHP zeros
and poles for the unstable plant P (s). Now, let zc+ denote the number of RHP zeros
of the controller, C(s).
T heorem 1 : [5]
1
z+ = −rP − rC − 2zc+ − σ(G)

(3.14)
2
1
p+ = [σ(P ) − σ(G) − rC ] − 2zc+ . (3.15)
2

14
P roof : Given
P (s)C(s)
G(s) = ,
1 + P (s)C(s)

and since G(s) is stable,

σ(G) = (z − + zc− ) − (z + + zc+ ) − (n + nc )

= −rP − rC − 2zc+ − 2z + ,

which implies (3.14). Applying (3.8) to P (s) yields

σ(P ) rP
p+ = z + + + . (3.16)
2 2

Substituting (3.14) into (3.16) yields (3.15). It is important to note that the above
theorem assumes C(s) is a known stabilizing controller and the corresponding closed-
loop frequency response G(ω) for ω ∈ [0, ∞) can be measured. Thus, P (ω) can be
computed from (3.13). rP and σ(G) can both be obtained by applying the results of
Lemma 2 to P (ω) and G(ω), respectively. Since C(s) is known and rC and z + are
known, z + and p+ can be found.
3.3 Synthesis and Design of PID Controllers
In this section, the synthesis and design of PID controller for a continuous LTI
plant, with an underlying transfer function P (s) with n (m) poles (zeros) will be
described. It is assumed that the only information available for the control design
engineer is:

1. knowledge of the frequency response magnitude and phase, which is equivalent


to P (ω), ω ∈ [0, ∞) if the plant is stable;

2. knowledge of a known stabilizing controller and the corresponding closed-loop

15
frequency response, G(ω).

Such assumptions are safe for most systems. Also, it is reasonable to assume that
the plant has no ω poles or zeros so that the magnitude, its inverse and phase are
well-defined for all ω ≥ 0. From the preliminary discussion, the number of RHP
poles and zeros can be determined from the above data for any LTI plant and either
“divided out” or lumped together.
Let
P (ω) = |P (ω)|eφ(ω) = Pr (ω) + Pi (ω), (3.17)

where |P (ω)| denotes the magnitude and φ(ω) the phase of the plant at the frequency
ω.
Let the stabilizing PID controller be in the form of

Ki + K p s + K d s2
C(s) = , for T > 0, (3.18)
s(1 + sT )

where T is assumed to be fixed and small. This type of PID controller deters pure
differentiation when controller input signals are noisy. From this, the procedure for
determining the entire stability set can be developed.
Lemma 3 : [5] Let the characteristic equation of of the closed-loop system be

F (s) := s(1 + sT ) + (Ki + Kp s + Kd s2 )P (s)

and F̄ (s) = F (s)P (−s).

The closed-loop stability is equivalent to

σ(F̄ (s)) = n − m + 2z + + 2. (3.19)

16
P roof : Closed-loop stability is equivalent to the condition that all the zeros of F (s)
lie in the LHP, which is equivalent to the condition

σ(F (s)) = n + 2 − (p− − p+ ).

Considering the rational function

F̄ (s) = F (s)P (−s),

and noting that


σ(F̄ (s)) = σ(F (s)) + σ(P (s)),

the stability condition then becomes

σ(F̄ (s)) = n + 2 − (p− − p+ ) + (z + − z − ) − (p+ − p− )

= n + 2 + z + − z − = n − m + 2z + + 2.

From here, write

F̄ (ω) = ω(1 + ωT )P (−ω) + (Ki + ωKp − ω 2 Kd )P (ω)P (−ω)

= (Ki − Kd ω 2 )|P (ω)|2 − ω 2 T Pr (ω) + ωPi (ω) +ω (Kp |P (ω)|2 + Pr (ω) + ωT Pi (ω))
| {z } | {z }
F̄r (ω,Ki ,Kd ) F̄i (ω,Kp )

= F̄r (ω, Ki , Kd ) + ω F̄i (ω, Kp ).

T heorem 2 : [5] Let ω1 < ω2 < · · · < ωl−1 denote the distinct frequencies of odd
multiplicities which are solutions of

F̄i (ω, Kp ) = 0, or (3.20)

17
Pr (ω) + ωT Pi (ω)
Kp∗ = −
|P (ω)|2
cos φ(ω) + ωT sin φ(ω)
=−
|P (ω)|
=: g(ω) (3.21)

for fixed Kp = Kp∗ . Let ω0 = 0 and ωl = ∞, and j := sgn F̄i (∞− , Kp∗ ) . From here,
 

determine the string of integers

I = [i0 , i1 , i2 , . . . , il ],

with it ∈ {+1, −1} such that for n − m even

[i0 − 2i1 + 2i2 + · · · + (−1)l−1 2il−1 + (−1)l il ](−1)l−1 j = n − m + 2z + + 2 (3.22)

and for n − m odd

[i0 − 2i1 + 2i2 + · · · + (−1)l−1 2il−1 ](−1)l−1 j = n − m + 2z + + 2. (3.23)

Thus, for every fixed Kp = Kp∗ , the (Ki , Kd ) values corresponding to closed-loop
stability are given by
F̄r (ωt , Ki , Kd )it > 0, (3.24)

where the it ’s are taken from the string satisfying either (3.22) or (3.23), and the ωt ’s
are taken from the solutions of (3.21).
P roof : By Lemma 3, closed-loop stability has be reduced to the signature con-
dition given in (3.19). The theorem follows by applying Lemma 1 to determine the
signature of F̄ (s).

18
4 The Experimental Test-Bed
The Educational Control Products Industrial Emulator Apparatus may be trans-
formed into a variety of dynamic configurations which represent important classes of
real-world systems. The Industrial Emulator can represent many different physical
plants, including rigid bodies, having flexibility in drive shafts, gearing and belts, to
coupled discrete systems with disturbances in the form of vibrations, actuated at the
drive input having either collocated or noncollocated feedback. Several other impor-
tant non-ideal properties can be readily introduced and removed including backlash
and drive friction. The Industrial Emulator allows these properties to be charac-
terized in a controlled manner and facilitates study of control approaches to help
mitigate their effects. [4]

Figure 4.1: The ECP Systems Industrial Emulator Apparatus. [4]

19
4.1 The Electromechanical Plant
The Industrial Emulator Apparatus shown in Figure 4.1 is designed to emulate
a broad range of typical servo control applications. The Industrial Emulator elec-
tromechanical plant consists of a drive motor, which is servo-actuated, coupled to the
drive disk via a timing belt. Another timing belt couples the drive disk to the speed
reduction assembly (SRA), while a third belt completes the drive train to the load
disk. This is seen in Figure 4.2. The load and drive disks have variable inertia which
may be adjusted by moving or removing brass weights. Speed reduction adjustments
are accomplished by interchangeable belt pulleys in the SRA. Backlash may be in-
troduced via a mechanism located within the SRA, and flexibility may be introduced
by an elastic belt, either with a pulley or flexible belt which may be used to couple
the SRA with the load disk. The drive disk moves with a 1:1 gear ratio with the
drive motor. This allows for the drive inertia to be thought of as being collocated
with the drive motor. The load disk, however, will rotate at a different speed that
the drive motor due to the speed reduction. Drive flexibility and/or backlash may
be introduced between the load disk and the drive motor, thus the load inertia is
considered to be noncollocated with the drive motor. [4]
A disturbance motor, shown in Figure 4.2a, connects to the load disk via a 4:1
speed reduction, and is used to emulate viscous friction and disturbances at the
plant output. A brake below the load disk, also seen in Figure 4.2a, may be used to
introduce static or Coulomb friction at the plant output. Thus, friction, disturbances,
backlash, and flexibility may all be introduced, all in a controlled manner. These
properties or effects represent non-ideal conditions that are present to some degree in
virtually all physically realizable electromechanical systems. [4]
All rotating shafts of the mechanism are supported by ball bearings. Needle

20
(a) Emulator Apparatus: View 1 (b) Emulator Apparatus: View 2

Figure 4.2: Industrial Emulator Apparatus Layout. [4]

bearings in the SRA provide low friction conditions, with and without backlash.
High resolution, incremental encoders having 4,000 lines per revolution with 4×
hardware interpolation providing 16,000 counts per revolution for control coupled
directly to the drive (θ1 ) and load (θ2 ) disks providing position and derived velocity
feedback. The drive and disturbance motors are electrically driven by servo amplifiers
and power supplies in the Controller Box. The encoders are routed through the
Controller Box to interface directly with the digital signal processor (DSP) board via

21
a gate array that converts the encoder pulse signal into numerical values. [4]
Gear ratios are changed by adjusting gear combinations on the SRA. The fixed
gears at the drive and load disk have 12 and 72 teeth respectively. In combination
with the interchangeable gears at the SRA, gear ratios from 1.5:1 to 24:1 can be
achieved. [4] details how these various gear ratios can be achieved. Backlash can be
adjusted from 0 to 20 degrees at the SRA via a backlash screw. Static or Coulomb
Friction may be adjusted by tightening the screw in the friction brake. [4]
The system configuration utilized to collect data for this project was that of the
Industrial Emulator in its most basic form, i.e. the drive inertia disk decoupled from
the SRA with no brass weights attached.

22
5 Procedure for Data-Based Controller

Design
In this chapter, the procedure taken for data-based controller design will be given.
The techniques introduced in [5] will be applied to the Industrial Emulator Apparatus.
Section 5.1 describes the methods for determine the frequency response data for the
Industrial Emulator, and section 5.2 describes the design of a PID controller for the
system using the technique introduced in [5]. Testing of the stabilizing parameter set
will be presented in Chapter 6.
5.1 Collecting Frequency Response Data
First, time-domain data was collected by exciting the Industrial Emulator at indi-
vidual logarithmically incremented frequencies within its operable range. The time-
domain response to a sinusoidal input is shown in Figure 5.1, where the steady-state
output for the system is shown for a Sinusoidal input at 50.079 radians per second.

Figure 5.1: Industrial Emulator 50.079 rad/s Sinusoidal Response in Steady-State.

23
The frequency response for the system can be characterized by two properties at
any given frequency. The magnitude of a plant, |P (ω)|, at a given frequency, ω, is
given by
x1
|P (ω)| = , (5.1)
x0

and the phase, φ(ω), at a given frequency is given by

t1
φ(ω) = 360 .‘ (5.2)
t0

where phase is in degrees. Figure 5.2 gives the graphical representation showing how
to calculate the frequency response, i.e. the magnitude and phase, at a particular
frequency from the steady state output. This calculation is repeated as the we increase
the frequency of the sinusoidal input from low to high.

Figure 5.2: Graphical Representation of Magnitude and Phase Parameters.

From the time domain data, the magnitude and phase were collected when the

24
closed-loop response reached steady-state. In many cases, the zero-crossings and
peaks of the steady-state response can be compared and averaged to then be used
to calculate the magnitude and phase with (5.1) and (5.2). However, due to the
amount of noise present in the time domain data, it proved far more accurate to use
MATLAB’s Curve Fitting tools to determine the sinusoidal equation that described
the steady-state response of the system, for both the output and the input, noting
that in Figure 5.1a, the input is out of phase with a true sinusoid. A sinusoidal
equation of the form

y(t) = Astate sin(ωt + Bstate ) + Cstate , where state = {input, output}

is used, where C represents a slight shift that takes place at high frequencies for the
electromechanical plant. Thus, equation to determine the magnitude becomes

Aoutput
|P (ω)| = ,
Ainput

and the phase becomes


φ(ω) = Binput + Boutput .

Applying this procedure for all time domain sinusoidal responses collected yielded
the frequency response for the closed-loop system, G(ω). Thus, the open-loop plant
frequency data was then derived to produce the Bode Diagram shown in Figure 5.3.
From here the techniques described in [5] can be applied to determine the set of
stabilizing PID parameters.
5.2 Data-Based PID Controller Design
The open-loop frequency response data has be derived from the time-domain
output data. This is shown in Figure 5.3. First, note that the high frequency slope

25
Figure 5.3: Bode Diagram for Open-Loop Plant.

of the plant in Figure 5.3 is -20 db/decade, the relative degree must be n − m = 1.
The total change in phase, if more samples were taken would be -90 degrees, and so

π π
− = − (n − m) − 2(p+ − z + ) ,
2 2

26
and since the plant is stable, p+ = 0, giving z + = 0. The required signature for
stability can now be determined and is

σ(F̄ (s)) = (n − m) + 2z + + 2

= (1) + 2(0) + 2

=3

Since n − m is odd, we have

[i0 − 2i1 + 2i2 + · · · + (−1)l−1 2il−1 ](−1)l−1 j = 3, (5.3)

h i

 
where j := sgn F̄i (∞ , Kp ) = sgn lim g(ω) = 1. It is clear that only one term is
ω→∞

required to satisfy the equation above. From Figure 5.4, it is easy to see that (3.21)
has only one positive frequency as a solution and therefore i0 − 2i1 = 3. Thus, Kp
must be chosen so that F̄i (ω, Kp∗ ) = 0 has exactly one positive real zero. This gives
the feasible range of Kp values, i.e. Kp ∈ [−0.41195, 3.4926], to be the entire curve of
Figure 5.4, which depicts the function

cos φ(ω) + ωT sin φ(ω)


g(ω) = − . (5.4)
|P (ω)|

The feasible range of Kp is such that Kp intersects the graph of g(ω) only once. By
fitting the g(ω) curve, any value of Kp can be used to find the zeros of (3.20).
Now fix Kp∗ = 0.25, and compute for the ω value that satisfies

cos φ(ω) + ωT sin φ(ω)


− = 0.25.
|P (ω)|

The frequency satisfying the above equation is found by plotting g(ω) and determining

27
the horizontal axis value that corresponds with g(ω) = 0.25. This frequency is ω1 =
127.8 rad/s. From this, the only feasible string satisfying i0 − 2i1 = 3 is

F = {i0 , i1 } = {1, −1}

Figure 5.4: Graph of the function g(ω) in (5.4).

Thus, the following linear inequalities are obtained depicting the stability region:

Ki > 0,

40.28 − Ki + 16332.84Kd > 0. (5.5)

28
The complete set of stabilizing PID gains for Kp∗ = 0.25 is given in Figure 5.5. This set
is determined by finding the string of integers, {i0 , i1 } satisfying the stability condition
of (5.3). The corresponding linear inequalities, (5.5), determine the stabilizing set
shown in (Ki ,Kd ) space.

Figure 5.5: Complete set of stabilizing PID gains in (Ki ,Kd ) space when Kp = 0.25.

The stability set shown represents the stable parameters in (Ki , Kd ) space when
Kp = 0.25. This stability region extends throughout the entire first quadrant, for all
values of Kp inside the feasible range. The shaded areas depict imposed hardware
limitations. These hardware limitations are physical limitation given by the manu-
facturers of the Industrial Emulator. Exceeding these limits could result in damage

29
to the system. The intersection of the two shaded hardware limitation regions and
the stability region in (Ki ,Kd ) space when Kp = 0.25 represents the implementable
stability region. This small section is present on all levels of Kp . By attending to the
hardware limitation, given by

Kp ≤ 2,

Ki ≤ 0.4,

Kd ≤ 0.05,

the stability region found using the techniques introduced in [5] does indeed repre-
sent stabilizing PID controller parameters for the Industrial Emulator Apparatus,
determined completely without the use of analytical models.

30
6 Analysis of the Results
In this chapter, the result shown in Chapter 5 will be analyzed. First, a point
is taken in or around the region of stability, marked by a red asterisk, and then the
corresponding PID gains are implemented on the Industrial Emulator. To begin, a
point is taken from the edge of the stability region, safe within the both the hardware
limitations. Figure 6.1 depicts the point and the closed-loop step response with the
corresponding gains.

(a) Point taken within the Stability Region (b) Response with the Corresponding PID gain

Figure 6.1: Step Response with PID parameters, Kp = 0.25, Ki = 0.02936, and
Kd = 0.01252

In contrast, a point is taken from outside the stability region. The corresponding
PID gains implemented on the system cause an immediate plunge. This point rep-
resents an unstable PID controller, and which the response in Figure 6.2b exhibits.
It is important to note the relative position for both points with regard to the edge
of the stability region. The unstable point taken from outside the stability region is
relatively closer to the edge of stability region than the stable point taken within the

31
region. This indicates that the data-based design technique has captured the entire
set of stabilizing PID controller parameters. This can be attributed to the use of the
linear inequalities that define the bounds of the stability region.

(a) Point taken outside the Stability Region (b) Response with the Corresponding PID gain

Figure 6.2: Step Response with PID parameters, Kp = 0.25, Ki = −0.02202, and
Kd = −0.01565

Now, some tests are ran to determine the sensitivity of the given hardware limita-
tions in (Ki , Kd ) space when Kp = 0.25. The hardware limitations are characterized
by

Kp ≤ 2,

Ki ≤ 0.4,

Kd ≤ 0.05.

Figure 6.3 gives the systems response to various PID gains, with Kp = 0.25 and
Kd = 0.043818. A point within the stability region is tested first, seen in Figure
6.3a and 6.3b, and the corresponding PID gains generate a closed-loop response with

32
excellent tracking characteristic. Improved tracking is expected when differential
action is increased. The unstable response once again illustrates that points outside
the stability region will yield unstable responses. The unstable response is interesting
in that for about 20 seconds the system would seem stable. However, an exponential
plunge is soon after observed that shows the system is indeed unstable.

(a) Point taken within the Stability Region (b) Response with the Corresponding PID gain

(c) Point taken outside the Stability Region (d) Response with the Corresponding PID gain

Figure 6.3: Step Response with PID parameters chosen near the edge of Kd Hardware
Limitation, i.e. Kd = 0.043818.

33
Figure 6.4 depicts the response of the system near the hardware limits of Ki .
Integral action drives the input error signal to zero in steady-state. This is observed
in Figure 6.4b, where there is some initial tracking error, but this error is soon driven
to zero. Once again the point within the stability region generates a stable response
and the point outside the region generates an unstable response.

(a) Point taken within the Stability Region (b) Response with the Corresponding PID gain

(c) Point taken outside the Stability Region (d) Response with the Corresponding PID gain

Figure 6.4: Step Response with PID parameters chosen near the edge of Ki Hardware
Limitation, i.e. Ki = 0.39633.

34
The final point that validates the stability set generated by the data-based PID
controller technique introduced in [5] is at the edge of both hardware limitations.
Stability is still observed, while there is a slight attenuation of the controller at 3.5
seconds. This response shows that at the edges of the hardware limits, stable PID
controller parameters can still be selected, and the response generated from these
parameters will still exhibit desirable behavior. By testing parameters in and around
the stability region, verifying closed-loop stability for controller gains selected within
the stability set and instability for controller gains selected outside the stability set,
it can be said that the region generated by the data-based PID controller design
technique introduces in [5] is viable and valid for closed-loop stability.

(a) Point taken within the Stability Region (b) Response with the Corresponding PID gain

Figure 6.5: Step Response with PID parameters chosen near the edge of both Hard-
ware Limitations, Kp = 0.25, Ki = 0.4, and Kd = 0.046948

35
To further validate the stabilizing parameter set, various PID gains were tested
on the Industrial Emulator. The corresponding responses are shown in Figure 6.6b.
Close-loop stability is observed for all points taken and integral action soon reduces
steady-state error to zero.

(a) Various points taken within the Stability Region (b) The Response to the various points

Figure 6.6: Step Response with various PID parameters

36
7 Conclusion
In this project, the data-based PID design technique in [5] was experimentally
verified by using the ECP Industrial Emulator Apparatus. In the process, the entire
set of stabilizing PID gains for the Industrial Emulator was found. Time-domain
data, from which the frequency response data was derived, was collected from the
ECP Industrial Emulator Apparatus. The frequency response data was then used to
calculate the stabilizing PID parameter set. Finally, the stability set was validated
by testing points corresponding to PID gains in and around the stability set. Gains
selected within the stability set generated stable responses while gains selected outside
the stability set generated an unstable response. The determination of the stabilizing
PID parameter set was totally independent of analytical model. These results suggest
that data-based controller design is an attractive alternative to model-based controller
design for real-world systems.
Some open areas of research are a) determining stability sets for different config-
urations of the Industrial Emulator, b) determining performance measures such as
gain margin, phase margin, and H∞ norms, and c) robustness enhancements of the
designs with respect to the measured data. The progress of these and other topics is
necessary to further the frontiers of control theory currently defined by the limitation
of high-order model-based design.

37
Bibliography
[1] S. P. Bhattacharyya, A. Datta, and L. H. Keel. Linear Control Theory: Structure,
Robustness, and Optimization. Taylor and Francis Froup. LLC, 2009.

[2] H. W. Bode. Network analysis and feedback amplifier design. Van Nostrand,
Princeton, NJ, 1945.

[3] X. Chen, C. Wang, and H. Wu. Pid controller parameters: Self tuning algo-
rithm based on frequency characteristics. International Conference on Measuring
Technology and Mechatronics Automation, 2010.

[4] Education Control Products, 1 Buckskin Court, Bell Canyon, CA 91307. Indus-
trial Emulator / Servo Trainer, 1995.

[5] L. H. Keel and S. P. Bhattacharyya. Controller synthesis free of analytical models:


Three term controllers. IEEE Transations on Automatic Control, 53(6):1353 –
1369, July 2008.

[6] S. Mitra, L. H. Keel, and S. P. Bhattacharyya. Data-dased design of digital pid


controllers. Proceedings of the 2007 American Conrol Conference, 2007.

[7] H. Nyquist. Regulation theory. Bell Systems Technological Journal, 11:126 – 147,
1932.

[8] M. R. Katebi P. Balaguer, N. A. Wahab and R. Villanova. Multivariable PID


Control Tuning: A Controller Validation Approach. PhD thesis, Autonomous
University of Barcelona, Spain, 2007.

38
[9] T. Yamamoto, K. Takao, and T. Tamada. Design of a data-driven pid controler.
IEEE Transations on Control Systems Technology, 17(1):29 – 37, January 2009.

39

You might also like