Proceedings of the 2007 IEEE International Symposium on FrBT1.

4
Computational Intelligence in Robotics and Automation
Jacksonville, FL, USA, June 20-23, 2007

A robust PID like state-feedback control via LMI approach: An

application on a Double Inverted Pendulum System
Akin Delibasi Ibrahim B. Kucukdemiral Galip Cansever

Abstract— This paper addresses the design method for robust The objective of this paper is to combine the effectiveness
PID like controllers which guarantee the quadratic stability, of the LMI based convex optimization methods with the
performance in terms of H2 and H∞ specifications, pole simple structure of the PID controllers in order to tackle the
locations and maximum output control. The approach is based
on the transformation of the PID controller design problem to aforementioned problems. The basic idea in this approach
that of state feedback controller design thereby the convex op- is to transform the PID controller design problem to that
timization approaches can be adapted. Real time experimental of state feedback controller design thereby the convex ap-
results on a double inverted pendulum system demonstrates the proaches can be adapted. Similar method has been consid-
validity and applicability of the proposed approach. ered in [9]. However the method in [9] is only valid for sec-
I. INTRODUCTION ond order systems. Compared with [9], our method does not
requires any restrictions on the order of the controlled plant.
Linear matrix inequalities (LMIs) have emerged recently as a Moreover, we completely validate the proposed technique on
useful tool for solving many convex control problems which physical experimental system: a double inverted pendulum. It
can not be sometimes trackable with algebraic tools. The is well known that the double inverted pendulum (DIP) which
basic idea of the LMI method is to translate the given control is indeed an extension of the inverted pendulum system is
problem into a semidefinite programming (SDP) problem so a very suitable experimental tool to investigate and verify
that it can be trackable with new generation computer tools different control methods for dynamic systems with higher-
which rely on interior point methods. Many control problems order nonlinearities. Its control is much more difficult than
such as joint minimization or maximization of multiple the single inverted pendulum because there are two linked
objectives [1], [2], control of some nonlinear plants[3], linear pendulums on a cart and we should consider to bring both
parameter varying [4], [5] and linear time varying systems, of the pendulums from the unstable hanging position to
some fuzzy systems [6] and gain-scheduling controllers [7] the stable upright position by only moving the cart on the
can all be handled easily by using LMI techniques. The horizontal plane. Besides these systems have fewer control
reader can refer to [3] for detailed information. inputs than degrees of freedom [10].
In spite of this much academical research on LMI Rest of the paper is organized as follows. In Section II,
based powerful optimal-suboptimal controllers, Proportional- the problem description and solution methods are presented.
Integral-Derivative (PID) controllers are still highly popular In Section III, dynamics of the DIP system is developed.
and widely used in many of the real control applications Application results are discussed in Section IV. Finally
in industry because of its simple structure and well known Section V concludes the paper.
tuning techniques [8] such as Ziegler-Nichols methods. How- Through out this paper, the identity and null matrices are
ever, there are numerous control problems where this simple denoted by I and 0, respectively. X > 0(≥, < 0) denotes
PID controller structure prove to be inadequate such as that X is a positive definite (positive semi-definite, negative
nonlinear systems, systems with relative degree higher than definite) matrix. X > Y means that X −Y is positive definite.
two, nonminimum phase systems etc. On the other hand, The trace of a matrix is denoted by trace(·), Cov(·) stands
practical control problems mostly deal with time varying for the convex polytope,  ·  refers to either the Euclidean
uncertain systems and require high performance such as vector norm or the induced matrix 2-norm and finally the
optimization of performance outputs, minimization of control notation  denotes the symmetric block in one symmetric
effort, fast and well-damped time responses, high disturbance matrix.
rejection ability even under the system uncertainty. However
simple PID controller scheme are not capable mostly to cope II. P ROBLEM D ESCRIPTION , (ROBUST PID- LIKE S TATE
with these problems all together, simultaneously. F EEDBACK C ONTROL )
Consider an uncertain nth order single input, multi output
This work was not supported by any organization
A. Delibasi is with Department of Electrical Engineering, Faculty of plant
Electrical–Electronics Engineering, Yildiz Technical University, 34349 Be-
siktas, Istanbul, Turkiye [email protected] ż = A(t)z + B(t)u
I.B. Kucukdemiral is with Department of Electrical Engineering, Faculty y = Cz (1)
of Electrical–Electronics Engineering, Yildiz Technical University, 34349
Besiktas, Istanbul, Turkiye [email protected] where the state vector z is assumed to be full
G. Cansever is with Department of Electrical Engineering, Faculty
of Electrical–Electronics Engineering, Yildiz Technical University, 34349 measurable. Here y is the output [A(t), B(t)] ∈ Ω =
Besiktas, Istanbul, Turkiye [email protected] Cov{[A1 , B1 ], [A2 , B2 ], . . . , [A p , B p ]} where p is the number

1-4244-0790-7/07/$20.00 ©2007 IEEE. 374

 FrBT1.4

of vertices, and Ω is the polytope. Assume that the state Q > 0 and R ≥ 0. It is well known that this problem is
vector is in the form of z = [z1 z2 z3 · · · zn/2 ż1 ż2 żn/2 ]T . This equivalent to the eigenvalue problem
type of system structure is widely faced in control systems.
If the control signal is chosen to be in the form of u = Kz minimize z̃T0 X −1 z̃0 ,
X,Y
then K simply contains proportional and derivative control ⎡ ⎤
gains for inner and outer feedback loops. However, in order Ãi X + X ÃTi + B̃iY +Y T B̃Ti XQ1/2 Y R1/2
subject to ⎣  −I 0 ⎦ ≤ 0,
to transform the system into a PID-like framework, extra
state variables must be included which are generally chosen   −I
to be the integral of necessary outputs which require to X > 0 and Y ∀ i = 1, . . . , p. (4)
have zero steady state error. For example, let z1 be the
selected output which requires to have a zero steady state Here z̃T0 X −1 z̃0
is the upper bound for J and the controller

error. If the augmented state is chosen to be τ := − (r − z1 ) that minimizes this performance index is K = Y X −1 . Mini-
where r is the desired trajectory for z1 output, then the PID mization of z̃T0 X −1 z̃0 can also be represented as an LMI such
like state-feedback control problem can be interpreted as in as
Fig. 1 where all the controller gains are analogous to state z̃T0 X −1 z̃0 ≤ γ , (5)
feedback gains. In this case, the plant and the controller can which is also equivalent to
be evaluated as 
γ I z̃T0
ż = A(t)z + B(t)u ≥ 0. (6)
 X
τ̇ = z1 − r Here γ is the specified upper bound to be minimized.
y = Cz (2)
 B. Bounding the control signal
z
u = [Kp1 Kp2 · · · Kp(n/2) Kd1 Kd2 · · · Kd(n/2) | KI ] , Assume z̃T0 X −1 z̃0 ≤ 1 is satisfied. If the closed-loop system
   τ

   is assumed to be stable, then this ellipsoid can be interpreted
K z̃
as an invariant set [3] which is labelled as ε in this paper.
where z̃ is the new augmented state vector. Consequently, one has
max u(t)2 = max Y X −1 z̃(t)2 ≤ max Y X −1 z̃2
t≥0 t≥0 z̃∈ε

= max(Y X −1/2 X −1/2 z̃)T (Y X −1/2 X −1/2 z̃)

z1
z2
−
 u(t) ... z̃∈ε
r τ − zn/2
−1 KI
+
+
−
Controlled Plant
ż1 = max(X −1/2 z̃)T (Y X −1/2 )T (Y X −1/2 )(X −1/2 z̃)
... ż2 z̃∈ε
Kp1
żn/2 ≤ λmax ((Y X −1/2 )T (Y X −1/2 ))(X −1/2 z̃)T (X −1/2 z̃)
= λmax ((Y X −1/2 )T (Y X −1/2 ))z̃T0 X −1 z̃0
+ Kp1
≤ λmax ((Y X −1/2 )T (Y X −1/2 )) ≤ u2max , (7)
+ Kp2
...
+ Kp(n/2) where umax is the maximum control signal available. The last
+ Kd1 inequality holds true if and only if the LMI

+ Kd2 ... X YT
≥ 0, (8)
Kd(n/2)
 u2max I
Fig. 1. General structure of the PID-like state feedback control. holds [9].
C. Pole placement in LMI regions
A. LMI based LQR design of the PID-like Structure Robust stability is not the only requirement in a control
In the previous section, the state-feedback formulation of system. A good controller should also provide sufficiently
a general Linear Time Varying (LTV) polytopic system was fast time responses under the influence of system parameter
obtained. In this section, we present the synthesis technique variations. A way to guarantee this, is to confine the closed
for the outlined PID like structure based on LMI framework. loop system poles to a predefined region in the complex plane
Lets consider the state feedback control of the LTV system where this region has the property of acceptable transient
response.
z̃˙ = Ã(t)z̃ + B̃(t)u In this part we give the general pole clustering techniques
y = Cz̃ , z̃(0) = z̃0 (3) for the proposed controller. The regions considered here are
LMI regions. These regions can be in various forms including
and let the control signal be u = K̃ z̃. The goal is to find half spaces, disks, conic sectors, vertical and horizontal
the control gain K̃ that minimizes the quadratic objective strips or the intersection of these regions. The key facts of
J(u) = sup 0∞ (z̃T Qz̃+uT Ru) over all state trajectories where using LMI regions are that: the intersection of LMI regions

375
 FrBT1.4

output matrix in compatible dimensions with the state vector

rd z̃. The system in (11) is said to be quadratically stabilizable
by the control vector u = K z̃ such that Hyw  < μ∞ if and
only if, there exist a symmetric matrix X and N such that
⎡ ⎤
ζ Ãi X + X ÃTi + B̃i N + N T B̃Ti B̃w XCT
X > 0, ⎣  −μ∞ I DTyw ⎦ < 0.
D   − μ∞ I
β
i = 1, . . . , p. (12)
φ Here Hk j stands for the transfer function from input j to
q  = NX T [3].
output k. If so a stabilizing control gain is K
E. H2 control of PID like state feedback system
Fig. 2. Region D. Let the system be modelled as in (11) where we assume
that Dyw = 0 and let the controller be in the structure of
u=K z̃. Then the system in (11) is said to be quadratically
also produce LMI regions and any convex region that is stabilizable by the state feedback control u = Kz̃ such that
symmetric with respect to real axis can be approximated Hyw 2 < μ2 if and only if there exist W and X such that
by an LMI region. An LMI region is any subset of D of
ÃXi + X ÃTi + B̃iW +W T B̃Ti + B̃w B̃Tw < 0,
a complex plane that can be defined as 
  V CX
D = z ∈ C : L + zM + z¯M T < 0 , (9) > 0, Trace[V ] < μ2 i = 1, . . . , p. (13)
 X
where L and M are real matrices such that L = LT . Here  =
If so, a quadratically stabilizing control gain is K
fD (z) := L + zM + z¯M T is called the characteristic function
W X −1 [3].
of D. In addition to this, the closed loop LTV system is
called quadratically D stable if it has all its eigenvalues in F. H2 /H∞ mixed control of PID like state feedback system
the D region which is equivalent to finding a symmetric Let the control system be
matrix X > 0 such that
z̃˙ = Ã(t)z̃ + B̃(t)u + B̃w w
L ⊗ X + M ⊗ (Ãi X + B̃iY ) + M T ⊗ (Ãi X + B̃iY )T < 0,
y1 = C1 z̃ + Dyw1 w (14)
i = 1, . . . , p. (10)
y2 = C2 z̃, z̃(0) = 0 (15)
Here the operator ⊗ is the so-called Kronecker product.
where y1 is the performance output for the H∞ optimization
Many convex regions can be readily depicted by fD (z). For
whereas y2 is the performance output that is related to
example the region D(rd , ζ , β , φ , q) which constrains the
the H2 optimization. Then the system in (14) is said to
minimum decay rate, minimum damping and the maximum
be quadratically stabilizable by the state feedback control
undamped natural frequency is shown in Fig. 2 [11]. This z̃ such that Hy1 w < μ∞ and Hy2 w < μ2 if and only
u= K
problem can be depicted by a set of 3 LMIs with a common
if there exist a symmetric matrix X and N and a symmetric
Lyapunov function such that:
matrix V such that
⎡ ⎤
L1 = 2φ , M1 = 1, Ãi X + X ÃTi + B̃i N + N T B̃Ti B̃w XC1T
 
−rd q 0 1 X > 0, ⎣  −I DTyw1 ⎦ < 0.
L2 = , M2 = ,
q −rd 0 0   − μ∞ I
 
−2β cos ζ 0 sin ζ cos ζ 
L3 = , M3 = , V C2 X
0 −2β cos ζ − cos ζ sin ζ > 0, Trace[V ] < μ2 , i = 1, . . . , p. (16)
 X

L j ⊗ X + M j ⊗ (Ãi X + B̃iY ) + M Tj ⊗ (Ãi X + B̃iY )T < 0,  = NX −1 [2].

If so, a stabilizing control gain is K
i = 1, . . . , p. j = 1, 2, 3. III. DYNAMICS OF D OUBLE I NVERTED P ENDULUM
D. H∞ control of the PID like state feedback system The DIP system which is used in our application consists of
two aluminium arms which are connected to each other on a
Let the system be modelled as
wheeled cart. The cart can move linearly along a horizontal
z̃˙ = Ã(t)z̃ + B̃(t)u + B̃w w track by a servo motor system. The goal of the controller is
y = Cz̃ + Dyw w , z̃(0) = 0 (11) to control the cart so that the two linked inverted pendulum
on the cart is balanced in perpendicular position. A principal
where w is the disturbance vector and B̃w is the correspond- structure of the pendulum system is shown in Fig. 3. Here
ing disturbance input matrix and Dyw is the disturbance respectively, α and θ are the angles of first and second

376
 FrBT1.4

θ m L2 m2
2 2

F4
mh
x mh
α m1 F3
α

mc F m1
F2 L1
Fig. 3. Double inverted pendulum. L1
F0 2

F1
pendulums with respect to the vertical axis, x stands for the x
position of the cart, mc is the mass of the cart, m1 stands mc
for the mass of the first link, m2 represents the mass of
second link, mh is the mass of the hinge and encoder, L1 Fig. 4. Transformation frames.
and L2 denote the length of the first and second pendulums,
respectively. I shows the physical values of these parameters.
frame F0 as:
Note that, the internal frictions are not considered in this ⎡ ⎤ ⎡ cg1 ⎤
work. pmx
1
T1 (1, 4) + x
The mathematical model of the DIP system can be derived ⎣ pm y ⎦=⎣
1
T1cg1 (2, 4) ⎦
using the Euler - Lagrange equation. The form of the Euler pzm1
T1cg1 (3, 4)
- Lagrangian equation used here is: ⎡ mh ⎤ ⎡ h ⎤
 px T1 (1, 4) + x
d ∂L ∂L ⎣ pm y ⎦=⎣
h
T1h (2, 4) ⎦
− = Qq (17) mh
dt ∂ q̇ ∂q pz T1h (3, 4)
⎡ m2 ⎤ ⎡ ⎤
where L = T −V is Lagrange operator, q = [x α θ ]T is gener- px (T1h xT2cg2 )(1, 4) + x
alized coordinates of system, T is system kinetic energy, V is ⎣ pm
y ⎦=⎣
2
(T1h xT2cg2 )(2, 4) ⎦
m2
system potential energy, Qq is the generalized force. In order pz (T1h xT2cg2 )(3, 4)
to derive the kinetic and potential energies of the moving Then the potential energies of the elements can be obtained
elements, we need to obtain their positions and velocities as:
relative to fixed reference frame. The transformations are PEm1 = m1 g pm
mh m2
z , PEmh = mh g pz , PEm2 = m2 g pz
1
defined as shown in Fig. 4. The transformation matrices On the other hand the kinetic energies of the moving ele-
between the frames are: ments are obtained
 m as:
⎡ ⎤ 2  m 1 2 
cos α 0 sin α L1 sin α d px 1
1
KEm1 = 2 m1 dt + d pdtz ,
⎢ 0 1 0 0 ⎥   m   m 2 
T1h = ⎢ ⎥
⎣ − sin α 0 cos α L1 cos α ⎦ (F1 → F3 ), d px h
2 h
KEmh = 12 mh dt + d pdtz ,
0 0 0 1  m   m  
d px 2
2 2 2
⎡ ⎤ KEm2 = 12 m2 + d pdtz , KEmc = 12 mc ẋ2 .
cos α 0 sin α L21 sin α dt
⎢ 0 1 0 0 ⎥ Then, these are all used in the Lagrange formulation to obtain
T1cg1 = ⎢ ⎥
⎣ − sin α 0 cos α L1 cos α ⎦ (F1 → F2 ), the nonlinear differential equations which are linearized to
2
0 0 0 1 obtain a linear model. The above system is modelled with
a force F as input. In order to design a feedback system,
⎡ ⎤
cos α 0 sin α L22 sin α we need to control the voltage to the motor. The force F is
⎢ 0 1 0 0 ⎥ related to motor supply voltage V through
T2cg2 = ⎢ ⎥
⎣ − sin α 0 cos α L2 cos α ⎦ (F3 → F4 ),
2 Km Kg Km2 Kg2
0 0 0 1 F= V− ẋ. (18)
Rr Rr2

From the transformation matrices one can derive the posi- Augmenting the system with the integrator state τ := − (r −
tions of relevant moving parts relative to the fixed reference x) to obtain a PID like framework, the new state vector be-

377
 FrBT1.4

TABLE I
DIP SYSTEM PARAMETERS .

Parameter Description Value

mc Mass of cart 0.7031kg
mh Mass of hinge and encoder 0.17kg
m1 Mass of pendulum1 0.04kg
m2 Mass of pendulum2 0.23kg
L1 Length of pendulum1 0.1524m
L2 Length of pendulum2 0.6413m
Km Motor Back-EMF Constant 0.0077V.s/rad
Kg Planetary Gearbox Gear Ratio 3.71
Ra Motor Armature Resistance 2.6Ω
r Motor Pinion Radius 0.0063m

comes z̃ = [x α θ ẋ α̇ θ̇ τ ]T . Substituting system parameters

in linearized plant, we obtain the nominal system matrices
for the augmented system as
⎡ ⎤
0 0 0 1 0 0
⎢ 0 0 0 0 1 0 ⎥
⎢ ⎥
⎢ 0 0 0 0 0 1 ⎥
⎢ ⎥
à = ⎢
⎢ 0 −5.9275 0.1759 −10.8681 0 0 ⎥ ⎥,
⎢ 0 109.001 −83.4691 75.2748 0 0 ⎥
⎢ ⎥
⎣ 0 −111.7275 153.1807 −77.1576 0 0 ⎦
1 0 0 0 0 0 Fig. 5. Experimental setup.

B̃ = [0 0 0 2.4228 − 16.78.12 17.201 0]T ,

150

B̃r = [0 0 0 0 0 0 − 1]T ,
100
where Ã, B̃ and B̃r satisfies
z̃˙ = Ãz̃ + B̃u + B̃r r. (19) 50

IV. A PPLICATION R ESULTS

Imaginary

In order to demonstrate the validity of the proposed

approach, the proposed control algorithm is applied to the −50

DIP experimental system where its linear nominal model

was presented in the previous section. The photograph of the −100

experimental setup is shown in Figure 5. We assumed that

the system has uncertainties only on the actuator dynamics. −150
−200 −150 −100 −50 0 50

These uncertain parameters are related with Ã(4, 4), Ã(5, 4),
Real

Ã(6, 4), B̃(4), B̃(5), B̃(6) in the nominal linear model of the
system. Here, we assume a 1% uncertainty on these com- Fig. 6. Distribution of closed-loop poles for the perturbed system.
ponents. The considered control problem requires  the min-
imization of the quadratic objective J(u) = sup 0∞ (z̃T Qz̃ +
uT Ru), bounding the control signal within −13V ≤ u ≤ and ζ = π /6. Since the number of uncertain components
+13V and placing the closed-loop system poles within the of the system matrices is 6, we have p = 64 vertex points.
LMI region which is shown in Fig. 2. We do not consider When solution of the LMIs (4), (5), (6), (8) and (11) is
H2 and H∞ optimization, however it can be easily taken into obtained for a common X > 0 and Y by using convex
consideration by including the LMI constraints (13), (12), optimization problem solver YALMIP/SEDUMI, we achieve
respectively or constraint (16) for mixed control. Here, the γ = 1.2490 ∗ 10−11 . Since K = Y X −1 one can readily obtain
weights for the quadratic objective function are chosen as K = [76.17 415.55 768.81 90.04 114.19 108.57 19.77].
Q = diag[0.1 10000 10000 0 0 0 2000], R = 4. Once this feedback gain is simulated on the vertex systems,
the perturbed poles that are shown in Fig. 6 are obtained.
The initial conditions for the state vector z̃ are simply chosen From this figure it is obvious that the controller which is
to be as z̃(0) = z̃0 = 0. Moreover, the LMI region that is obtained by the solution of the LMI problem can successfully
used for clustering the closed loop system poles has the confine the closed-loop system poles in the predefined region.
following specifications: φ = 0.28, q = 1, β = 4, rd = 190 In order to test the validity of the proposed method, the

378
 FrBT1.4

10
a
250
c
been proposed. The proposed method relies on reformulation
5
200
of the PID like feedback system as a state feedback system.
150
The method can take various multi objective optimization

x (milimeter)
α (degree)

0 100

−5 50
specifications into considerations such as linear quadratic
optimization, control signal gain reduction, pole placement
0
−10
−50

−15
0 5 10 15 20 25 30 35
−100
0 5 10 15 20 25 30 35 constraints, H2 and H∞ optimization in the existence of
time (second) time (second)

b d
plant uncertainties. Laboratory application on a DIP system
6

4
15
demonstrates the effectiveness and the validity of the ap-
10
2
5
proach.
θ (degree)

u (volt)
0
−2
−5
R EFERENCES
−4

−6 −10
[1] M. Chilali, P. Gahinet nad C. Scherer, ”Multi-objective Output Feed-
−8
0 5 10 15 20 25 30 35
−15
0 5 10 15 20 25 30 35 back Control via LMI Optimization”, in IFAC 13th Triennial World
time (second) time (second)
Congress, San Francisco, USA, 1996, pp. 249–254.
[2] P.P. Khargonekar and M.A. Rotea, Mixed H2 /H∞ Control:A Convex
Optimization Approach, IEEE Tran. Automat. Contr., vol. 39, 1991,
Fig. 7. Experiment 1 (a) Variation of α (b) Variation of θ (c) Variation pp 824–837.
of x(d) Variation of u . [3] S. Boyd, L.E. Ghaoui, E. Feron and V. Balakrishnan, Linear Matrix
Inequalities in System and Control Theory, Siam, Philadelphia; 1994.
a c
[4] F. Wu and K. Dong, Gain Scheduling Control of LFT Systems Using
3 100
Parameter-Dependent Lyapunov Functions, Automatica, vol. 42, 2006,
2
50
pp 39–50.
[5] C.W. Scherer, LPV Control and Full Block Multipliers, Automatica,
xa, xd(milimeter)

1
α (degree)

0 0 vol. 37, 2001, pp 361–375.

−1 [6] F. Khaber, K. Zebar and A. Hamzaoui, State Feedback Controller
−50
−2 Design via Takagi-Sugeno Fuzzy Model:LMI Approach, International
−3
0 10 20 30 40 50 60
−100
0 10 20 30 40 50 60
Journal of Computational Intelligence, vol. 2, 2005, pp148–153.
time (second) time (second)
[7] P. Apkarian and P. Gahinet, A Convex Characterization of Gain-
b d Scheduling H∞ Controllers, IEEE Tran. Automat. Contr., vol. 40,
1.5 15

1
1995, pp 853–864.
10
0.5
[8] K.J. Aström and T. Hägglund, Automatic Tuning of PID Controllers,
5
Instrument Society of America, Research Triangle Park NC, 1988.
θ (degree)

0
u (volt)

−0.5
0
[9] M. Ge, M.-S. Chiu and Q.-G. Wang, Robust PID Controller Design via
−1
−5
LMI Approach, Journal of Process Control, vol. 12, 2002, pp 3–13.
−1.5 −10
[10] W. Zhong and H. Röck, ”Energy and Passivity Based Control of
−2
0 10 20 30 40 50 60
−15
0 10 20 30 40 50 60 the Double Inverted Pendulum on a Cart”, in IEEE International
Conference on Control Applications, Mexico City, MExico, 2001, pp.
time (second) time (second)

896–901.
[11] M. Chilali and P. Gahinet, H∞ Design with Pole Placement Con-
Fig. 8. Experiment 2 (a) Variation of α (b) Variation of θ (c) Variation straints: An LMI Approach, IEEE Tran. Automat. Contr., vol. 41, 1996,
of x(red) and the desired trajectory (green) (d) Variation of u . pp 358–367.

solution K  of the control problem is applied to the real

time experimental system. During application, two different
experiments are performed. First experiment includes balanc-
ing the pendulums while holding the cart at the zero point
(mid point of the finite track) as close as possible under the
influence of disturbance. The variation cart, angular positions
of the pendulum 1 and 2 and the variation of the control
signal are shown in Fig. 7, where the disturbance is applied
to the system at t = 23s. Application results show that the
control system successfully suppresses the disturbance effect
and the control signal does not hit the signal bounds outside
the disturbance effect points.
As a second experiment, the trajectory following ability of
the system is investigated under a square wave form reference
trajectory. Fig. 8 clearly shows the time response of the
system. Again the results show that the controller exhibits
remarkably high trajectory following performance. Note that,
the control signal never exceeds the control signal bounds.
V. C ONCLUSION
In this paper, a novel approach for synthesis of the
controller gains for a quadratically robust PID controller has

379