Proceedings of the 3rd IFAC Conference on

Proceedings of the 3rd IFAC Conference on

Advances in Proportional-Integral-Derivative
Proceedings of the 3rd IFAC Conference on Control
Advances in Proportional-Integral-Derivative Control
Available online at www.sciencedirect.com
Ghent, Belgium,
Proceedings
Advances May
of the 9-11,
3rd IFAC2018
Conference on Control
in Proportional-Integral-Derivative
Ghent, Belgium, May 9-11, 2018
Advances in Proportional-Integral-Derivative
Ghent, Belgium, May 9-11, 2018 Control
Ghent, Belgium, May 9-11, 2018
ScienceDirect
IFAC PapersOnLine 51-4 (2018) 346–351
All
All Stabilizing
Stabilizing State
State Feedback
Feedback Controller
Controller
Allfor
Stabilizing
Inverted State Feedback
Pendulum Controller
Mechanism
Allfor
Stabilizing
Inverted State Feedback
Pendulum Controller
Mechanism
for Inverted Pendulum Mechanism
for Inverted
Rahman Pendulum
Bitirgen ∗∗ Muhsin Hancer Mechanism
∗∗ ∗
∗∗ Ismail Bayezit ∗
Rahman Bitirgen ∗∗ Muhsin Hancer ∗∗ Ismail Bayezit ∗
Rahman Bitirgen Muhsin Hancer ∗∗ Ismail Bayezit ∗
∗Rahman Bitirgen ∗ Muhsin Hancer ∗∗ Ismail Bayezit ∗
∗ Faculty of
∗
of Aeronautics
Aeronautics and and Astronautics,
Astronautics, IstanbulIstanbul Technical
∗ Faculty
University, Istanbul, Turkey (e-mail: bitirgen@itu.edu.tr,
Technical
Faculty of Aeronautics and Astronautics, Istanbul Technical
∗ University, Istanbul, Turkey (e-mail: bitirgen@itu.edu.tr,
Faculty
University,of Aeronautics
Istanbul, and Astronautics,
bayezit@itu.edu.tr)
Turkey Istanbul Technical
(e-mail: bitirgen@itu.edu.tr,
∗∗ University, Istanbul, bayezit@itu.edu.tr)
Turkey (e-mail: bitirgen@itu.edu.tr,
∗∗ Faculty of Aeronautical
∗∗ and Space
bayezit@itu.edu.tr) Sciences, Necmettin Erbakan
∗∗ Faculty of Aeronautical and Space Sciences, Necmettin
University,
Faculty of Konya,
Aeronautical bayezit@itu.edu.tr)
Turkey and (e-mail:
Space Sciences, Necmettin Erbakan
mhancer@konya.edu.tr) Erbakan
∗∗ University, Konya, Turkey (e-mail: mhancer@konya.edu.tr)
Faculty
University, of Aeronautical
Konya, Turkey and (e-mail:
Space Sciences, Necmettin Erbakan
mhancer@konya.edu.tr)
University, Konya, Turkey (e-mail: mhancer@konya.edu.tr)
Abstract: In the literature, finding all stabilizing controllers are widely studied. In this paper,
Abstract:
a Inofthe literature, finding all stabilizing controllers are
are widely studied. In
In this
this paper,
a case
case study
Abstract:
Abstract: studyIn In
inverted
ofthe
inverted
the
pendulum
literature,
pendulum
literature,
findingis
finding
considered
is all stabilizing
considered
all stabilizing
to test
test the
the methods
to controllers
controllers methods
are
for
for finding
widely
widely
studied.
finding
studied.
stabilizing
stabilizing
In this
state
paper,
state
paper,
afeedback
case study controller.
of inverted First, good setisofconsidered
pendulum state feedback to test gain
thematrix
methods is calculated, two of
for finding stabilizing feedback
state
feedback
aparameters
case study
feedback controller.
of inverted
controller.
are First,
fixedFirst, good
to pendulum
good the
reduce set of
setisof state
considered feedback
state feedback
computational to test gain
cost. the
gain The matrix
methods
matrix is
otheris two calculated,
for finding
calculated,
of feedbacktwo of
stabilizing
two of feedback
state
feedback
parameters
parameters
feedback
that make
parameters are
controller.
the
are fixed
system
fixed to
First,
to reduce
goodare
stable
reduce the
the computational
setthen
of state feedback
calculated
computational cost.
by gain
cost. The
gridding.
The other
matrix
other is two
The two of
boundary
of feedback
calculated, two
feedbackof parameters
of
the feedback
stability
parameters
that make the
parameters system
are system
fixedwith stable
tostable
reduce are then
thethen calculated by
computational cost. gridding.
The otherThetwo boundary
of feedback of the stability
parameters
region
that
region is
make
is calculated
the
calculated with the
the frequency
are
frequency method.
calculated
method. In
In byorder to
gridding.
order to have
have Thefast
fast response,
boundary
response, ofthe
the the eigenvalue
stability
eigenvalue
that
regionmake is theclosed
of calculated
the system loop
with stable
system
the are isthen
frequency definedcalculated
as the
method. Inbyleft gridding.
order handto havesideTheof the
fast boundary
−1 + jω
response, ofthe the
line stability
and the
eigenvalue
region
region
stability of
is the
the closed
calculated
ofregion closed loop
with
loop the
for feedback system
system is
frequency
gains defined
isare
defined as
as the
method.
calculated forleft
In
the thishand
order
left handto haveside
side of
scenario. fastthe
of the −1
−1 + + jω
response, jωthe line
line and
and the
eigenvaluethe
stability
region ofregion
stability the closed
region for
for feedback
loop system
feedback gains
gainsisare calculated
defined
are calculated as thefor this
forleft
thishandscenario.
side of the −1 + jω line and the
scenario.
© 2018, IFAC
stability region (International
for feedback Federation
gains are of calculated
Automatic Control) thisHosting by Elsevier Ltd. All rights reserved.
Keywords: all stabilizing controller, full stateforfeedback, scenario. PID controller, inverted
Keywords: all
Keywords: allstability
pendulum, stabilizing
stabilizing region,controller,
controller,
actuator full
full state
state feedback, PID controller,
saturation feedback, PID controller, invertedinverted
pendulum,
Keywords: allstability
pendulum, stability
stabilizing region,
region, actuator
controller,
actuator saturation
full state feedback, PID controller, inverted
saturation
pendulum, stability region, actuator saturation
1. INTRODUCTION time delay term, or for situations where time delay could
1. could
1. INTRODUCTION
INTRODUCTION time delay
delay term,
timeadequately
be or
or for
term,representedfor situations
situations
by a Padé where time
time delay
whereapproximation,delay could a
1. INTRODUCTION be
time adequately
simpler
be delayapproach
adequately represented
term,represented
ortoforthis by
situations
problem
by a
a Padé
where
could
Padé approximation,
time
be delay could
implemented
approximation, a
a
In the
In the literature,
literature, ratherrather than than finding
finding aa good good set set ofof con-
con- simpler approach to this problem could be implemented
be
by adequately
Munro
simpler (1999)
approach represented
and
to Munro
this byet aal.Padé
problem (1999).
could approximation,
be implemented a
troller
In the for
troller for aa given
givenrather
literature, system,
system, thanthefinding
the topic of
topic ofa finding
finding
good setall allofstabi-
stabi-
con- by Munro (1999) and Munro et al. (1999).
In the literature, rather than finding a good set of con- simpler
by Munro approach
(1999) to this
and Munroproblem
et al. could
(1999). be implemented
lizing PID
troller for acontroller
given system, is studied
the topic by manyof findingresearchers.
all stabi- In
lizing
troller
order to
lizing PID
for
PID controller
acontroller
test given possible
these is studied
system,
is studied
the topic by many
controllers,
by many
ofinvertedresearchers.
finding all stabi-
pendulum
researchers. In Munro
In
by Munro
Munro (1999)
(1999)
(1999) and
andand Munro
Munro
Munro et
etetal.
al.al.(1999)
(1999).suggested
(1999) suggested a a nu-
nu-
order to test these possible controllers, inverted pendulum merical
Munro
merical frequency
(1999)
frequency and domain
Munro
domain approach
et al.
approach in
(1999)
in order
order to
suggested
to find
find the
a nu-
the
lizing
has widely
order PID
widely
to testcontroller
used
theseby is studied
bypossible
many researchers by many
controllers, as a researchers.
inverted useful testbed
pendulum In Munroset of
merical (1999)
D-stabilizing
frequency anddomainMunro
low-order et compensators.
al. (1999)
approach in suggested
order Datta
to findaet nu-
al.
the
has
order toettest used many
thesebypossible researchers
controllers, as a
inverted useful testbed set of D-stabilizing low-order compensators. Datta et al.
Yadav
has widely al.used
(2016). Inverted
many pendulumas
researchers also haspendulum
a useful paramet-
testbed set merical
(2013),
of frequency domain
Ackermann
D-stabilizing and
low-order approach
Kaesbauer in order
(2001),
compensators. andto Bajcinca
Datta findet the
al.
Yadav
has et
widely al.
ric uncertainties
Yadav (2016).
used
et al. (2016). by Inverted
many
in both
Inverted pendulum
researchers
model parameters
pendulum as also
a
alsoand has
useful paramet-
testbed
has designing
paramet- set (2013), Ackermann and Kaesbauer (2001), and Bajcinca
(2006)of D-stabilizing
(2013), have
Ackermannshown low-order
that
and the compensators.
stabilizing
Kaesbauer region
(2001), Datta
and is et al.
defined
Bajcinca
ric uncertainties in both
both model parameters and designing
paramet- (2006) have shown that the stabilizing region
Yadav
controllers
ric et al.or(2016).
uncertainties compensator
in Inverted Yue
model pendulum
et
parameters alsoand
al. (2015). has designing (2013),
by a sethave
(2006) Ackermann
of convex and Kaesbauer
shown polygonal
that (2001),
slices normal
the stabilizing andKis
to
region
defined
ispBajcinca
axis in
defined
controllers
ric or
uncertainties
controllers compensator
in both model
or compensator Yue
Yue et et al. (2015).
parameters
al. (2015).and designing (2006) by a set of convex polygonal slices normal to K p axis in
the
by a(K set
phave
, Kofi , shown
and
convex K d that
)
polygonalthe stabilizing
parameter space
slices for
normal region
continuous
to K ispp defined
axistime
in
A certain set
controllers or ofcompensator
performanceYue criteria
et is(2015).
al. optimized in Malan the (Kpp , Kii , and Kdd ) parameter space for continuous time
AA certain
et certain
al. (1994), set of
set of performance
Saeki and Kimura
performance criteria is optimized
(1997),
criteria is optimized
Ohta et al. in Malan by
in(1997),
Malan PID
the a(K set
p ofi ,convex
controllers.
, K and K dpolygonal
) parameter slices
space normal
for to Kp axis
continuous in
time
PID(K
the controllers.
, K , and K ) parameter space for continuous time
Aet al.
Daley
et (1994),
certain andset
al. (1994), Liuof Saeki and
performance
(1999).
Saeki andBut Kimura
But
Kimura (1997),
criteria
this Ohta
is optimized
optimization
(1997), et al. (1997),
in(1997),
decreases
Ohta decreases
et al. Malan PID controllers.
p i d
the In this paper, the Quanser inverted pendulum platform
Daley and Liu (1999). this optimization the PIDthis controllers.
et
Daleyal. (1994),
uncalculatedand LiuSaeki
or(1999). andBut
unexpected Kimura (1997), Ohta
performance
this optimization loss.etSo,
al. (1997),
decreases the In
finding is taken
paper, the Quanser inverted pendulum platform
as
uncalculated
Daley
uncalculatedand Liu or
or unexpected
(1999).
unexpected But performance
this optimization
performance
all stabilizing PID controllers to design controller is help- In loss.
loss. So,
decreases
So, finding
the
finding
In
is this
taken
this as aa case
paper,
paper,
the study.
case
the
Quanser
study.
Quanser
The rest
rest of
Theinverted
inverted
the
the paper
of pendulumpaper is
pendulum is organ-
platform
organ-
platform
all stabilizing PID controllers to design (1994)
controller isfinding ised
is as
taken follows.
as a caseIn Section
study. 2,
The
help- ised as follows. In Section 2, the mathematical model the
rest mathematical
of the paper model
is organ- is
is
uncalculated
ful.stabilizing
all In this context,orPIDunexpected
Shafiei and
controllers performance
toShenton
design loss. and
controller So, is Shafiei
help- is taken
given
ised asand as a case
nonlinearities
follows. In study.
Section The
such
2, rest
as
the of the
trigonometric
mathematicalpaper is
terms organ-
model are
is
ful.
all
ful. In this
andstabilizing
Shenton
In context,
this context, Shafiei
PID controllers
(1997) Shafiei and
andto
have proposed
proposed Shenton
design
Shenton (1994)
controller
a graphical
graphical
(1994) and
and Shafiei
isShafiei
approachhelp- givengiven and nonlinearities such as trigonometric terms are
and Shenton (1997) have a approach ised asand
eliminated follows.
with In
nonlinearities
the Section
help of 2,
such the mathematical
as trigonometric
linearization. The modelare
terms
parameters is
ful.
using
and In this context,
(1997)Shafiei
D-partitioning
Shenton method
have andtoShenton
proposed determine (1994)
a graphical the and
bordersShafiei
approach of eliminated
given
of the
eliminatedand
system
with the help such
nonlinearities
with is assumed
the help
of linearization.
to
of be ascertain
trigonometric
linearization. and
The parameters
The termswith
obtained
parameters are
using
and
using
absolute D-partitioning
Shenton and (1997)
D-partitioning method
relativehave
method to
proposed
stability determine
to regionsa graphical
determine in the the borders
approach
the parameter
borders of
of of the system is the
assumed to linearization.
be certain and obtained with
eliminated
measurements with isand help of
with The parameters
absolute
using
space.
absolute and
D-partitioningrelative
Characterization
and relative stability
methodof all
stability regions
tothedetermine in
stabilizing
regions in the parameter
thegains
the borders
parameter using
of
of
the system
of measurements
the system and
is with identification
assumed
assumed
to be certaintest
identification
to be certain
and
test
and
cycles.
obtained
cycles.
obtained
The
The withcon-
con-
space. Characterization of all the stabilizing gains using trollability andand
measurements observability of the system
with identification The with
is tested.
test cycles. In
con-
absolute
a generalized
generalized
space. and Hermite-Biehler
relative stability
Characterization regions
theorem
of all theorem
the in provided
is
stabilizing thegains
parameterin
using trollability
Ho measurements
trollability
section and
andand
3, firstly, observability
with
observability
full of
identification
state the system is
test cycles.
of thecontroller
feedback system tested.
The con-
isistested.
proposed In
In
a
space. Hermite-Biehler
Characterization of all the is
stabilizing provided
gains in Ho
using section 3, firstly, full state feedback controller is proposed
aet generalized
al. (1996). Then, Ho et al. (1997b)
Hermite-Biehler theorem andisHo et al. (1997a)
provided in Ho trollability
in PD
section based
3, and with
firstly, observability
fullinstant
state of thecontroller
measurement
feedback systemof isistested.
displacement,
proposed In
aet
et
haveal. (1996).
(1996). Then,
generalized
al. extended this Ho
Hermite-Biehler
Then, Ho et
et al.
results al.to(1997b)
theorem
(1997b) and
andisHo
characterize Ho et
et al.
providedal. (1997a)
stabilizing (1997a) Ho in
inPID PD based
section
speed,
in PD 3, firstly,
angle,
based
withfull
and
with
instant
angular
measurement
staterate.
instant feedback
The
measurement
of displacement,
controller
system of has isfour
proposed
displacement, states
have extended this results to(1997b)
characterize stabilizing PID
way speed, angle,
et
have al. extended
(1996). Then,
compensators.
compensators. Ho et
this
Ho
Ho
et et al.
al.
results
al. (1998)
to
(1998) found
found
andaHo
characterize a
et al. (1997a)
systematic
stabilizing
systematic PID
way in
and PD
speed, based and
therefore
angle, with
the
and
angular
instantrate.
controller
angular
The system
measurement
has
rate. Thefour system of has four states
displacement,
parameters.
has four Finding
states
have
of findingextended
finding
compensators. this
the maximum
maximum
Ho et results and
al. and to characterize
(1998) minimum stabilizing
values
found avalues the PID
of the
systematic Kp ,, all
way and
speed, therefore
angle, and the controller
angular has
rate. Thefour parameters.
system has four Finding
states
of the minimum of K and therefore
stabilizing the
state controller
feedback has four
controllers parameters.
requires Finding
searching
compensators.
K
of , and
finding Kthe termsHo et
maximumto al. (1998)
guarantee
and found avalues
stability
minimum systematic
in the ofresulting
the way
K p
p, all
and stabilizing
therefore state
the feedback has
controller controllers
four requires searching
parameters. Finding
i
Kii ,finding d
and Kdd terms to guarantee stabilityvaluesin theofresulting p for
all all four
stabilizing parameters
state at
feedback the same
controllers time, so a
requiresconstraint
searching is
of
Ki , and Kthe
closed-loop maximum
terms
dsystem andgeneralization
to guarantee
based on minimum
stability in the Kp , for
of the Hermite-
resulting all all four parameters
stabilizing
introduced,
for all four twostate
of
parameters
at thecontrollers
feedback
its parameters
at the
same time,requires
same are fixed
time,
so a constraint
so and
a searching
the
constraint stabi-
is
is
closed-loop
Ki , andTheorem.
Biehler
closed-loop system
Kdsystem
termsMunrobased on
to guarantee
based generalization
(1999)
on stability
proposed
generalization inof
a
of the
the
new Hermite-
resulting
method
Hermite- introduced, two of its parameters are fixed and the stabi-
Biehler Theorem. Munro (1999) proposed a new method for all region
lizing
introduced,four parameters
for
two the
of its at the
other two
parameterssame
are time,
are fixedso and
calculated a constraint
by thegridding
stabi-is
closed-loop
to this
Biehler problemsystem
Theorem. based
based
Munro on on
the generalization
(1999) use of the
proposed of
Nyquist
a the
new Hermite-
plot.
method lizing regiontwo
introduced, for oftheitsother two areare
parameters calculated
fixed and bythe gridding
stabi-
to this problem based on the use proposed
of the Nyquist lizing
and region
frequency for the
methods. other A two
gain are
is calculated
then chosen by to gridding
test the
Biehler
to this problemTheorem. Munro
based (1999)
on the use of the Nyquist a newplot. method and
plot. lizingfrequency
response
and regionof
frequency formethods.
the the
system,
methods. other A two
Adue
gainare
gainto
is then chosenby
calculated
isactuator
then
to gridding
saturation,
chosen
test the
to test the
Munro and Soylemez (2000) claimed that Nyquist
plot. plot
plot response of the system, due to actuator saturation,
to
Munro this problem
and Soylemez based on the use
(2000) of thethat
claimed Nyquist Nyquist and
system
responsefrequency
is close
of themethods.
to A
unstability.
system, duegainTo
to isactuator
then chosen
overcome this test the
tosituation,
saturation, the
based
Munro
based approach
and
approach is
Soylemez computationally
(2000)
is computationally claimed much
much that faster
faster than
Nyquist
than that that
plot system is close to unstability. To overcome this situation,
Munro and Soylemez (2000) claimed that Nyquist plot response
the
systemgain is of
Kclosethe
is tosystem,
searched
unstability.due to
which makes
To actuator
all
overcome thesaturation,
roots
this stay
situation, the
on
of
based
of Ho
Ho et
et al.
approach
al. (1998)
(1998) and
and Shafiei
is computationally
Shafiei and
and Shenton
much
Shenton faster (1997),
than that
(1997), and
and the gain K is searched which makes all the roots stay on
system
the left is close
hand to
side unstability.
of −1 + jω Toaxis.overcome
Then athis
test situation,
point is
based
theHo
of approach
computing
et al. (1998) is computationally
time and Shafiei andmuch
is polynomic with
Shenton faster
system than
order,
(1997), that
andit the left hand side of −1 + jω axis. Then a test point on
gain K is searched which makes all the roots stay is
the
of Ho computing
et al. (1998) time is polynomic with system order, it chosen
the gain
left and K
hand is searched
results
side are
of −1 which
presented
+ jω makes in
axis. all
section
Then the 4.
a roots
test stay
point on
is
was
the
was recently
computing
recently timeand
realized
realized
Shafiei
isthat for
polynomic
that for
andwith
systems
systems
Shenton with
system
with
(1997),
no
no order, and
explicit
explicit it chosen and results are presented in section 4.
the computing time is polynomic with system order, it the
chosen leftand hand side are
results of −1 + jω axis.
presented in Then a4.test point is
section
was recently realized that for systems with no explicit
was recently realized that for systems with no explicit chosen and results are presented in section 4.
2405-8963 ©
Copyright © 2018
2018, IFAC
IFAC (International Federation of Automatic Control)
346 Hosting by Elsevier Ltd. All rights reserved.
Copyright
Peer review© under
2018 IFAC 346 Control.
responsibility of International Federation of Automatic
Copyright © 2018 IFAC 346
10.1016/j.ifacol.2018.06.089
Copyright © 2018 IFAC 346
IFAC PID 2018

Ghent, Belgium, May 9-11, 2018 Rahman Bitirgen et al. / IFAC PapersOnLine 51-4 (2018) 346–351 347

2. SINGLE INVERTED PENDULUM PLATFORM

Inverted pendulum with moving cart mechanism is a

widely-used platform to test new approaches in controller
design. Pendulum with moving cart on a track has two
equilibrium points: first, pendulum aligned with the grav-
ity vector pointing downwards, the other is pendulum
being again aligned with the gravity vector and in this
case pointing upwards. The first equilibrium point is in-
herently stable because the center of the gravity (CG)
of the pendulum is below the pivot point. In the second
equilibrium point, the CG of the pendulum is above the
pivot point and and marginally unstable. Therefore the
pendulum diverges and return the asymptotic stability
condition when it is disturbed with an insignificant force
or moment. The second condition is marginally unstable
due to this behaviour, controllers often implemented to
make the system stable in the inverted configuration.
In Fig. 1 and 2 the testbed configuration can be seen.
The cart can move left and right on the track, and the
pendulum can rotate around the pivot axis. This testbed is
a Single-Input Multiple-Output (SIMO) system, one input
is the motor voltage and two outputs are the position of the
cart and the pendulum angle. It has two high resolution
encoders to measure the displacement of the cart and the Fig. 2. Inverted Pendulum Platform
pendulum angle. There is a DC motor to move the cart
along the track which is the actuator to control the system. ηg Kg2 Jm
Jeq = Mc + 2
rmp
JT = Jeq Jp + Mp Jp + Jeq Mp lp2 (2)
Jκ = Jp + Mp lp2
Jξ = Jeq + Mp

2.2 State Space Form

The system can be represented in state space format. The

state vector:

T
x = [xc α x˙c α̇] (3)
The state space representation:

Fig. 1. Pendulum on a moving cart test platform. ẋ = Ax + Bu

(4)
y = Cx + Du
2.1 Mathematical Model where xc is the position of the cart, x˙c is the speed of the
cart, α is the pendulum angle, α̇ is the pendulum angular
The mathematical model of the system is derived from rate and u is the control signal which is voltage on the
the Newton’s second law of motion. The equations are motor. The system matrix:
nonlinear and needs to be linearized in order to design
linear controller. Assuming that the angle of the pendulum  
0 0 1 0
is small, the equations are linearized. 1 0 0 0 1 
A=  2 2  (5)
JT 0 Mp lp g −Jκ Beq −Mp lp Bp
1   0 Jξ Mp lp g −Mp lp Beq −Jξ Bp
x¨c = −Jκ Beq x˙c − Mp lp Bp α̇ + Mp2 lp2 gα + Jκ Fc
JT
1 1 T
α̈ = (− (Mp lp Beq ) x˙c − Jξ Bp α̇ + Jξ Mp lp gα + Mp lp Fc ) B= [0 0 Jκ Mp l p ] (6)
JT JT
(1)  
1 0 0 0
C= (7)
where 0 1 0 0

347
IFAC PID 2018
348 Belgium, May 9-11, 2018
Ghent, Rahman Bitirgen et al. / IFAC PapersOnLine 51-4 (2018) 346–351
00
D=
00
Here in the equations, Mp = 0.127(kg) is the mass of
the pendulum, Jp = 1.19 ∗ 10−3 (kg/m2 ), is the mass
moment of inertia of the pendulum about its center of
gravity, lp = 0.178(m) is the distance from the center
of gravity of the pendulum to the pivot point, g is the
gravitational acceleration, Beq = 5.4(N s/m) is the friction
coefficient between the cart and the track, Bp = 2.4 ∗
10−3 (N ms/rad), is the viscous damping coefficient at the
pendulum axis, Mc = 0.57(kg) is the mass of the cart,
ηg = 1 is the planetary gearbox efficiency, Kg = 3.71 is
the planetary gearbox gear ratio, Jm = 3.90 ∗ 10−7 (kgm2 )
rotor inertia, rmp = 6.35 ∗ 10−3 (m) is the motor pinion
radius. The terms Jeq , JT , Jκ and Jξ are related to inertias,
can be calculated with the given parameters.
The poles of the open loop system:
P = [0 6.38 − 6.38 − 12.47]
There is a pole on the right hand side of the s plane for
the open loop system, therefore it is proved that the open
loop system is unstable.
Controllability and observability are checked before de-
signing the controller, in order to inspect if all modes
are controllable and if all the states are observable. The
controllability matrix is calculated and it is seen that it
does not lose rank therefore the system is controllable.
The observability matrix does not lose rank, therefore it is
possible to observe all the states.
3. CONTROLLER DESIGN
The system has one input and two outputs, and therefore it
can be represented with two transfer functions. However,
when PID controller is designed for each of the transfer
function (TF) assuming each TF has a SISO behaviour,
the control input for each controller only takes into account
the position or the angle feedback individually. This make
it difficult to find a set of controller that both satisfies the
stability of position and angle. It is considered that, with a
full state-feedback controller, all poles of the system can be
placed anywhere in the left half plane to make the system
stable.
T
u = −Kx K = [k1 k2 k3 k4 ] x = [xc α x˙c α̇] (10)
It should be noted that, k1 and k3 behaves like a
Proportional-Derivative (PD) controller for the position
of the cart, and k2 , k4 pair behaves like a PD controller
for the pendulum angle. Pole placement method can be
used to find the stabilizing K when desired location of
the closed loop system poles are known. In this case, it is
known by both simulation and experiment on the actual
system, poles in (11) stabilizes the system.
T
P = [−2 − 3 − 18 + 10j − 18 − 10j]
The closed loop system matrix:
Ac = A − BK
The desired poles of the closed loop system is given, and by
(8) solving Ackermann’s formula, one can calculate the gain
K. In this case, K is calculated as:
K = [−42 115 − 46 15] (13)
In this paper, a good set of K value is calculated, then it is
to be found when k1 and k3 (the part of the K that controls
the position and velocity of the cart.) are fixed, what are
the possible sets of k2 and k4 (the part of the K that
controls the angle and angular velocity of the pendulum.).
Two methods are used to find all possible values of k2 and
k4 .
3.1 Gridding Method
In this brute force method, k2 and k4 are gridded between
the values of 0 and 250. k1 and k3 are fixed, for each and
every single point on the grid, it is possible to calculate
the roots of the characteristic equation. If the point on the
grid does not have any root on the right half plane, then
(9) that K is said to be possible stabilizing controller for the
system. When the gridding is dense enough, this method
can give all possible K that stabilizes the system when k1
and k3 are fixed. However, it should be noted that this is a
computationally expensive method, when the gridding is
dense, it takes time to give a solution.
k2 vs. k 4
250
200
150
k4
100
50
0
0 50 100 150 200 250
k2
Fig. 3. All Possible Stabilizing Values of k2 and k4 when
k1 = −42 and k3 = −46
3.2 Frequency Method
In Fig. 3, it can be seen that the 2D shape have boundaries.
On the stable side of the boundary, if a k2 and k4 pair is
chosen and roots are calculated, the roots are observed
to be close to the imaginary axis on the left half plane,
when on the border, two roots of the system are on the
imaginary axis.
In this sense, it can be computationally less expensive to
search the boundaries of the stabilizing k2 and k4 . Instead
(11) of s in the characteristic equation, jω is placed. Now in
this case, k2 and k4 are not gridded and are to be found.
In order for jω to be a root of the characteristic equation,
(12) real and imaginary part of the Pc (jω) should be equal to
348
IFAC PID 2018

Ghent, Belgium, May 9-11, 2018 Rahman Bitirgen et al. / IFAC PapersOnLine 51-4 (2018) 346–351 349

Cart Position a4 =1
80
Desired Position a3 = 1.56k3 + 6.78k4 + 12.42
Simulated Position (15)
60 a2 = 1.56k1 + 6.78k2 + 0.72k3 − 40.77
a1 = −66.45k1
40

In order for jω to be a root of the characteristic equation,

Cart Position (mm)

20
both real and imaginary parts of the Pc (jω) should be
equal to zero.
0

-20
Re {Pc (s)} = a4 ω 4 − a2 ω 2 + a0 = 0
  (16)
-40 Im {Pc (s)} = −ω a3 ω 2 − a1 ω = 0

-60 Solving (16) for k2 and k4 :

-80
0 1 2 3 4 5 6 7 8 9 10 ω 4 − b1 (k1 , k3 )ω 2 + b2 (k1 )
Time (s) k2 =
b3 ω 2
2
(17)
Fig. 4. Cart Position Time Response for K = −c1 (k3 )ω + c2 (k1 , k3 )
k4 =
[−42 180 − 46 50] c3 ω 2

Pendulum Angle where

3

b1 = 1.56k1 + 0.72k3 − 40.77

2
b2 = −66.45k1
b3 = 6.78
Pendulum Angle (Degree)

1 (18)
c1 = 1.56k3 + 12.42
c2 = 0.72k1 − 66.54k3 − 506.64
0 c3 = 6.78

-1
First limit values of k2 and k4 are calculated, when ω = 0
and ω → ∞. In order to calculate k2 and k4 only ω is
gridded between frequencies 0 and 1000.
-2
Table 1. Limit Values for k2 and k4
-3 ω k2 k4
0 1 2 3 4 5 6 7 8 9 10 0 ∞ ∞
Time (s) −c1
∞ ∞ c3

Fig. 5. Pendulum Angle Time Response for K =

In Fig. 3 the blue dots are calculated with the gridding
[−42 180 − 46 50]
method, and the red curve is the boundary of stability
zero. There comes two equations, one from the real part, calculated by the frequency method. In this figure, any
and one from the imaginary part. There are two unknowns value of k2 and k4 in the stability region makes the system
and two equations, thus k2 and k4 can be calculated given stable theoretically. A set of K is chosen and implemented
ω. on the simulation, the results are presented in the Fig.
4 for cart position, Fig. 5 for pendulum angle. These
In this method, only ω is gridded, and no root calculations figures inspire the control system engineer to revise the
are done, only two linear equations are solved to find k2 controller gains, in order to mitigate high oscillation rates
and k4 . It can be said that, this method, comparing it and prevent actuator saturation.
to the gridding method is computationally less expensive.
However, in this method, only the boundaries are calcu- The stability boundary jω could be shifted to further left
lated and it is not known which side of the boundary is −1 + jω to avoid such situations, in this case, any K set
stable. In this case a test set can be chosen to test whether that have at least one root between jω and −1 + jω lines
the region is stable or not. are disregarded.
Placing jω instead of s in the characteristic equation: In the gridding method, K values that make the closed
loop system to have roots on the left hand side of −1 +
jω are taken into the new stability region, and for the
Pc (s) = a4 s4 + a3 s3 + a2 s2 + a1 s + a0 frequency method, −1 + jω is placed instead of s in
(14)
Pc (jω) = a4 ω 4 − a3 ω 3 j − a2 ω 2 + a1 ωj + a0 the characteristic equation. The same approach is then
used to calculate the stability boundary via gridding the
where frequency.

349
IFAC PID 2018
350 Belgium, May 9-11, 2018
Ghent, Rahman Bitirgen et al. / IFAC PapersOnLine 51-4 (2018) 346–351

In Fig. 6, the blue dots are the k2 and k4 pairs that are Pole Spread of the System
calculated with the gridding method. The red curve is the 40
0.55
jω crossing values of k2 and k4 and the orange curve is the
−1 + jω crossing values for k2 and k4 . 30

It can be observed that the new stability region is signifi- 20

cantly smaller comparing it to Hurwitz stability and it is
also a bounded region. 10

Imaginary Axis
An arbitrary pair is chosen in this region. 0

K = [−42 145 − 46 25]

-10
The experiment is conducted on the actual system and
the results are presented in the Figs. 8, 9 and 10. The -20
time responses of cart position, pendulum angle and motor
voltage are better comparing it to the first case scenario. It -30
can be noticed that, for the cart position, there is a steady 0.55
state error. This is due to the fact that the mathematical -40
-30 -25 -20 -15 -10 -5 0
model of the system does not represent the system exactly. Real Axis
In the mathematical model, the cart can move very small
distances, however in the real system there is contact Fig. 7. Pole Spread of the system when controller is within
between the cart and the track and the minimum distance the boundaries of the stability region shown in Fig. 6
that the cart can travel depend on the dimension of the
gears. Therefore, when the cart and pendulum settles on Cart Position
a point on the track, the amount of control signal that 60
Desired Position
comes from the feedback of position and speed, is not Simulated Position
50
enough to change the position. Because, at the same time, Measured Position
the feedback from the angle and angular rate also act to 40
stabilize the pendulum in the upward configuration. This
30
can be eliminated with introducing an integral term for
Cart Position (mm)

the position part of the controller. 20

10

k vs. k 0
2 4
250
-10

-20

200 -30

-40
30 31 32 33 34 35 36 37 38 39 40
150 Time (s)
k4

Fig. 8. Cart Position Time Response for K =

100 [−42 145 − 46 25]

namely gridding and frequency methods. The latter is

50 observed to be less computationally expensive comparing
the first method.
Considering only Hurwitz stability, the stability region for
0 k2 and k4 might have some values of K that cannot drive
0 50 100 150 200 250
k2 the system to stability beacuse of unmodelled dynamics
and actuator saturation. In order to overcome this sit-
uation, the stability boundary is taken as the left half
side of −1 + jω line. Taking an arbitrary point from this
Fig. 6. Values of k2 and k4 when k1 = −42 and k3 = −46 region, the system is driven to stability without actuator
that places the poles on the left side of −1 + jω saturation.
For future research, the authors plan to take the friction
4. CONCLUSIONS coefficient as an uncertain parameter and then calculate all
stabilizing gains when the exact value of the parameter is
In this paper, a well-known test-bed is used to test unknown but is bounded. The length of the pendulum can
methods to find all possible controllers when two of its also be taken as an uncertain parameter, however in this
parameters are fixed. The stability region for the other case this uncertain parameter enters in the characteristic
two parameters are calculated using two different methods, equation, polynomic fractional fashion. It can also be

350
 IFAC PID 2018
Ghent, Belgium, May 9-11, 2018 Rahman Bitirgen et al. / IFAC PapersOnLine 51-4 (2018) 346–351 351

Pendulum Angle Bajcinca, N. (2006). Design of robust pid controllers using

1.5 decoupling at singular frequencies. Automatica, 42(11),
Simulated Angle 1943–1949.
Measured Angle
Daley, S. and Liu, G. (1999). Optimal pid tuning using
1 direst search algorithms. Computing & Control Engi-
neering Journal, 10(2), 51–56.
Pendulum Angle (Degree)

Datta, A., Ho, M.T., and Bhattacharyya, S.P. (2013).

0.5 Structure and synthesis of PID controllers. Springer
Science & Business Media.
Ho, M.T., Datta, A., and Bhattacharyya, S. (1996). A
0 new approach to feedback stabilization. In Decision and
Control, 1996., Proceedings of the 35th IEEE Confer-
ence on, volume 4, 4643–4648. IEEE.
-0.5 Ho, M.T., Datta, A., and Bhattachary, S. (1997a). Control
system design using low order controllers: constant gain,
pi and pid. In American Control Conference, 1997.
-1 Proceedings of the 1997, volume 1, 571–578. IEEE.
30 31 32 33 34 35 36 37 38 39 40
Time (s) Ho, M.T., Datta, A., and Bhattacharyya, S. (1997b). A
linear programming characterization of all stabilizing
Fig. 9. Pendulum Angle Time Response for K = pid controllers. In American Control Conference, 1997.
[−42 145 − 46 25] Proceedings of the 1997, volume 6, 3922–3928. IEEE.
Ho, M.T., Datta, A., and Bhattackaryya, S. (1998). Design
Motor Voltage
of p, pi and pid controllers for interval plants. In
8 American Control Conference, 1998. Proceedings of the
1998, volume 4, 2496–2501. IEEE.
6 Malan, S.A., Milanese, M., and Taragna, M. (1994). Ro-
bust tuning for pid controllers with multiple perfor-
4 mance specifications. In Decision and Control, 1994.,
Proceedings of the 33rd IEEE Conference on, volume 3,
Motor Voltage (Volt)

2
2684–2689. IEEE.
Munro, N. and Soylemez, M. (2000). Fast calculation
0
of stabilizing pid controllers for uncertain parameter
-2
systems. IFAC Proceedings Volumes, 33(14), 549–554.
Munro, N., Söylemez, M., and Baki, H. (1999). Compu-
-4 tation of d-stabilizing low-order compensators. Control
Systems Centre Report, 882.
-6 Munro, N. (1999). The systematic design of pid controllers.
Ohta, Y., Li, J., Tagawa, K., and Haneda, H. (1997).
-8 Robust pid controller design. In Proceedings of NOLTA,
30 31 32 33 34 35 36 37 38 39 40
1053–1056.
Time (s)
Saeki, M. and Kimura, J. (1997). Design method of
Fig. 10. Motor Voltage Time Response for K = robust pid controller and cad system. IFAC Proceedings
[−42 145 − 46 25] Volumes, 30(11), 1511–1516.
Shafiei, Z. and Shenton, A. (1994). Tuning of pid-type
studied on the steady state error of the cart position by controllers for stable and unstable systems with time
adding an integral term in K. delay. Automatica, 30(10), 1609–1615.
Shafiei, Z. and Shenton, A. (1997). Frequency-domain
design of pid controllers for stable and unstable systems
ACKNOWLEDGEMENTS with time delay. Automatica, 33(12), 2223–2232.
Yadav, M., Gupta, A.K., Pratap, B., and Saini, S. (2016).
The authors would like to thank to the BAP (Scientific Robust control design for inverted pendulum system
Research Projects) Office of ITU for funding the labora- with uncertain disturbances. In Power Electronics,
tory. Intelligent Control and Energy Systems (ICPEICES),
IEEE International Conference on, 1–6. IEEE.
The authors also would like to thank Prof. M. Turan Yue, M., Wei, X., and Li, Z. (2015). Zero-dynamics-based
Söylemez for providing vision throughout the entire course adaptive sliding mode control for a wheeled inverted
of Control of System with Parametric Uncertainity during pendulum with parametric friction and uncertain dy-
Fall Term of 2017-2018. namics compensation. Transactions of the Institute of
Measurement and Control, 37(1), 91–99.
REFERENCES
Ackermann, J. and Kaesbauer, D. (2001). Design of robust
pid controllers. In Control Conference (ECC), 2001
European, 522–527. IEEE.

351