You are on page 1of 19

ARTICLE IN PRESS

Control Engineering Practice 16 (2008) 241–259


www.elsevier.com/locate/conengprac

Dynamic modelling of a TRMS using analytical and


empirical approaches
A. Rahideha,b, M.H. Shaheeda,, H.J.C. Huijbertsa
a
Department of Engineering, Queen Mary, University of London, London E1 4NS, UK
b
School of Electrical and Electronics Engineering, Shiraz University of Technology, Shiraz, Islamic Republic of Iran
Received 13 October 2006; accepted 23 April 2007
Available online 15 June 2007

Abstract

This investigation presents 1 degree of freedom (1DOF) dynamic modelling of an experimental aerodynamic test rig, a twin rotor
multi-input–multi-output system (TRMS) using analytical as well as empirical approaches. The TRMS is a highly nonlinear system
which can be considered as an experimental model of a complex air vehicle. Such vehicles are required to be modelled precisely to ensure
satisfactory control performance to meet the demand for automation and sophistication. The TRMS is modelled in terms of vertical and
horizontal 1DOF dynamics using Newtonian and Lagrangian methods based analytical approaches and neural networks based empirical
approaches. The analytical modelling is carried out in two phases. In the first phase the interface circuit, DC motors and propulsive
forces due to these motors are modelled. Thereafter, the dynamic equations for the remaining parts are formulated taking all the effective
forces into account. Two neural network based models are developed using Levenberg–Marquardt (LM) and gradient descent (GD)
algorithms. The responses of all the analytical and empirical based models are compared with that of the real TRMS to validate the
accuracy of the models. The performances of the models are also compared with respect to each other.
r 2007 Elsevier Ltd. All rights reserved.

Keywords: Dynamic modelling; Helicopter; Lagrange method; Newton method; TRMS; Neural networks

1. Introduction and Miller (1990). This investigation deals with the


development of a dynamic model and design of a feedback
The TRMS is a laboratory set-up developed by Feed- controller for reducing torque oscillations of a side-by-side
back Instrument Ltd. (Feedback Co., 1998) for control rotor tandem helicopter, which is an MIMO system. Wu,
experiments. The system is perceived as a challenging Ignatov, Muenst, Imaev, and Zhu (2002) have presented a
engineering problem due to its high nonlinearity, cross- nonlinear flight controller design for a helicopter model by
coupling between its two axes and inaccessibility of some of trajectory linearization method. In that paper, the authors
its states and outputs for measurements. Accurate dynamic have presented the development of a nonlinear dynamic
model of the system is, thus, required to be developed to model of a 3 degree of freedom (3DOF) flight control
achieve control objectives satisfactorily. experiment apparatus, a helicopter, with three propellers
Substantial investigations are reported to have addressed driven by DC motors. Exact linearization and non-
the modelling and control of TRMS and similar systems interacting control of a four rotors helicopter via dynamic
using various analytical and artificial intelligence (AI)- feedback has been reported in Mistler, Benalleque, and
based empirical approaches. For instance, general dynamic Msirdi (2001). The authors have presented a nonlinear
modelling and feedback control of a side-by-side rotor dynamic model for a four rotors helicopter in a form suited
tandem helicopter has been proposed in Rao, Biswas, Butz, for control design. Hover control via Lyapunov approach
for an autonomous model helicopter has been proposed by
Corresponding author. Mahony, Hamel, and Dzul (1999). In that investigation a
E-mail address: m.h.shaheed@qmul.ac.uk (M.H. Shaheed). dynamic model based on physical insight has been

0967-0661/$ - see front matter r 2007 Elsevier Ltd. All rights reserved.
doi:10.1016/j.conengprac.2007.04.008
ARTICLE IN PRESS
242 A. Rahideh et al. / Control Engineering Practice 16 (2008) 241–259

Nomenclature Jtr/mr moment of inertia in tail/main DC motor


(kg m2)
g gravitational acceleration (m/s2) Jv moment of inertia about horizontal axis (kg m2)
h length of pivoted beam (m) Lah/v armature inductance of tail/main motor (H)
iah/n armature current of tail/main motor (A) Mcable(ah) torque of the flat cable force (N m)
ka j torque constant (N m/A) Meh/v electro-magnetic torque of tail/main motor
kah/v, kfhp, kfhn, kfvp, kfvn, kth/v positive constants (N m)
lb length of counterweight beam (m) Mfric,h torque of the friction force in horizontal plane
lcb distance between the counterweight and the (N m)
joint (m) Mfric,v torque of the friction force in vertical plane
lm length of main part of the beam (m) (N m)
lt length of tail part of the beam (m) MLh/v load torque of tail/main motor (N m)
lT1 centre of gravity of the free–free beam (m) Mprop,h propulsive torque due to tail propeller rotation
lT2 centre of gravity of the counterbalance beam (N m)
(m) Mprop,v propulsive torque due to main propeller rota-
m mass (kg) tion (N m)
mb mass of the counterweight beam (kg) Pn nominal power (W)
mcb mass of the counterweight (kg) Rah/v armature resistance of tail/main motor (O)
mh mass of the pivoted beam (kg) T kinetic energy (J)
mm mass of main part of the beam (kg) T1 kinetic energy of the free–free beam (J)
mmr mass of the main DC motor (kg) T2 kinetic energy of the counterbalance beam (J)
mms mass of the main shield (kg) T3 kinetic energy of the pivoted beam (J)
mt mass of tail part of the beam (kg) Uh/v horizontal/vertical voltage control input (V)
mtr mass of the tail DC motor (kg) Vh/v horizontal/vertical motor terminal voltage (V)
mts mass of the tail shield (kg) V potential energy (J)
mT1 total mass of the free–free beam (kg) V1 potential energy of the free–free beam (J)
mT2 total mass of the counterbalance beam (kg) V2 potential energy of the counterbalance beam (J)
rms radius of the main shield (m) V3 potential energy of the pivoted beam (J)
rts radius of the tail shield (m) Vn nominal voltage (V)
Btr/mr viscous friction coefficient of the tail/main DC ah horizontal position of TRMS beam (rad)
motor (kg m2/s) av vertical position of TRMS beam (rad)
Eah/v electro-motive force of tail/main motor (V) jh/v magnetic flux of tail/main motor (Wb)
Fh/v nonlinear function of aerodynamic force from oh/v rotational speed of tail/main rotor (rad/s)
tail/main rotor (N) on nominal speed (rpm)
J1 moment of inertia of the free–free beam (kg m2) Oh angular velocity of TRMS beam in horizontal
J2 moment of inertia of the counterbalance beam plane (rad/s)
(kg m2) Ov angular velocity of TRMS beam in vertical
J3 moment of inertia of the pivoted beam (kg m2) plane (rad/s)

developed for a reduced scale autonomous helicopter using Barrientos (2005) have proposed identification of a small
Newton’s equations. Real-time stabilization and tracking unmanned helicopter model using genetic algorithm (GA).
of a four-rotor mini rotorcraft has been described in A hybrid (analytic and empiric) model of a small
Castillo, Dzul, and Lozano (2004). The authors have helicopter, the stability, response of the model and
developed a controller and implemented it on a mini parameter identification using GA have been presented in
rotorcraft. The dynamic model of the four-rotor rotorcraft that study. Mathematical modelling of an unmanned aerial
has been obtained via a Lagrange approach. Nonlinear vehicle (UAV) during take off with nose-wheel off the
dynamics and control of a wind-milling gyroplane rotor ground has been discussed (Song-Yun, Ji-Hong, & Chun-
has been proposed by Somov and Polyntsev (2003). In that Hua, 2004). The model is formulated according to New-
study the mathematical models of auto-rotation and ton–Euler method. Jun, Roumeliotis, and Sukhatme (1999)
flapping the wind-milling rotor have been extracted and have presented the state estimation of an autonomous
used for nonlinear dynamic modelling. The development of helicopter using Kalman filtering. The authors have
a nonlinear model and nonlinear control strategy for a proposed a technique to estimate the state of a robot
Vario scale model helicopter has been presented (Avila- helicopter using combination of gyroscopes, acceler-
Vilchis, Brogliato, Dzul, & Lozano, 2003). The model is ometers, inclinometers and global positioning system
based on Lagrangian equations. Cerro, Valero, and (GPS).
ARTICLE IN PRESS
A. Rahideh et al. / Control Engineering Practice 16 (2008) 241–259 243

Modelling of TRMS and similar system using neural and Mao (2002) have investigated the application of
networks (NNs) are also abundantly reported in the wavelet based NNs for identifying the model of a robot
literature. For example, nonlinear modelling of a TRMS manipulator. The authors have introduced an algorithm
using radial basis function (RBF) networks has been for constructing the wavelet network, and backpropaga-
addressed in Ahmad, Shaheed, Chipperfield, and Tokhi tion method has been used for network training. The
(2000), which presents nonlinear system identification application of NNs in the identification of models for
method for modelling air vehicles of complex configura- underwater vehicles can be found in Van de Ven, Johansen,
tion. Ahmad, Chipperfield, and Tokhi (2000) have carried Sorensen, Flanagan, and Toal (2007). In that method,
out dynamic modelling and optimal control of a TRMS. knowledge regarding the various parts of the model has
The extracted model is employed in the design of a been used to apply NNs for those parts of the model that
feedback linear quadratic Gaussian (LQG) compensator. are most uncertain.
Performance analysis of four types of conjugate gradient This investigation addresses the 1DOF horizontal and
algorithms in the nonlinear dynamic modelling of a TRMS 1DOF vertical dynamic modelling of a TRMS using
using feedforward NNs has been reported (Shaheed, 2004). analytical and empirical approaches. The analytical
Dynamic modelling of a TRMS has also been presented models are developed considering all the effective
in Mat-Darus, Aldebrez, and Tokhi (2004), which has forces into account. Such investigations on dynamic
investigated the utilization of NNs and parametric linear modelling of TRMS or similar systems are very limited
approaches for modelling the system in hovering position. in the literature. It is also evident from the literature that
Aldebrez, Mat-Darus, and Tokhi (2004) have proposed the system has also not been modelled using effective
a parametric modelling of a TRMS using GA. In neural learning algorithms like that of Levenberg–
their approach the global search technique of GA has Marquardt.
been used to identify the parameters of the TRMS based The paper is organized as follows: Section 2 describes
on one-step-ahead prediction. It is noted that NNs have the TRMS and presents the dynamic models developed
also been used to model various other systems. For using Newtonian approach. Section 3 presents the
instance, an efficient NN model has been proposed for model based on Lagrangian method. The NN-based
real-time path planning with obstacle avoidance of modelling is presented in Section 4. Section 5 presents
car-like robots in dynamic environment (Yang & Meng, the results of all the developed models and their
1999). In that investigation the stability and convergence of comparative performances. Finally, the paper is concluded
the NN system has been proved using Lyapunov stability in Section 6.
analysis. Mohd-Hashim, Tokhi, and Mat-Darus (2004)
have investigated a nonlinear dynamic modelling of a
flexible beam structure using NN. A comparative perfor- 2. Newtonian based modelling
mance of the model and conventional recursive least-
square scheme has been carried out in the time and The TRMS as introduced in Section 1 is shown in Fig. 1.
frequency domains. Dynamic modelling of flexible-link It is driven by two DC motors. Its two propellers are
manipulators using NNs with application to the space perpendicular to each other and joined by a beam pivoted
station remote manipulator system has been proposed in on its base that can rotate freely in the horizontal and
Talebi, Patel, and Asmer (1998). The method can be vertical plane. The joined beam can be moved by changing
considered both as an on-line identifier that can be used as the input voltage in order to control the rotational speed of
a basis for designing NN controllers as well as an off-line the propellers. The system is equipped with a pendulum
learning scheme to compute deflections due to link
flexibility for evaluating forward dynamics. Karras (2004) Tail rotor
has proposed NN models based on regularization techni-
ques for off-line robot manipulator path planning. Tail shield
The method involves a regularized Kohonen feature Main shield
map which aims at quantizing the input variable space Main rotor
into smaller regions representative of the input space DC motor+ Pivot
probability distribution. Modelling and control of a tacho
pneumatic manipulator based on dynamic NN has been
Free-free beam
presented in Wang and Peng (2003). The proposed method
consists of a three-layer feedforward NN as controller
and a diagonal recurrent NN as model predictor. Model- Counterbalance DC-motor+
Tower tacho
ling of robot manipulator dynamics based on a multi-
dimensional RBF-like NN has been presented (Krabbes &
Doschner, 1999). The NN is hierarchically organized to TRMS 33-220
reach optimal adjustment to the common structural
knowledge about the identification problem. Sun, Wang, Fig. 1. The twin rotor MIMO system.
ARTICLE IN PRESS
244 A. Rahideh et al. / Control Engineering Practice 16 (2008) 241–259

Table 1
The main differences between a helicopter and a TRMS

TRMS Helicopter

Location of pivot Midway between The main rotor head


point two rotors
Lift generation or Speed control of Collective pitch controla
vertical control main rotor
Yaw is controlled Tail rotor speed Pitch angle of the tail rotor
by blades
Cyclical control No Yes (for directional control)
a
Pitch angles of all the blades of the main rotor are changed but at
constant rotor speed.

counterweight hanging from the beam and is used for


balancing the angular momentum in steady state or
with load. In certain aspects, its behaviour resembles
that of a helicopter. For example, it possesses a strong
Fig. 2. Relationship of Uv and Vv.
cross-coupling between the collective (main rotor) and the
tail rotor like a helicopter. However, the TRMS is
substantially different from a helicopter. Table 1 lists the
main differences between a helicopter and a TRMS.
The crude dynamic model of the system supplied by the
manufacturer does not represent the system dynamic
precisely as all the effective forces are not taken into
consideration. However, accurate dynamic modelling of
the system is a prerequisite to achieve satisfactory control
performance. In this study the system is modelled using
Newtonian and Lagrangian equations taking all the
effective forces into account as well as using NN-based
modelling approaches. The analytical modelling is carried
out in two stages. In the first stage, the interface circuit, DC
motors and propulsive force due to these motors are
modelled. Thereafter the dynamic equations for the
remaining parts are developed.

2.1. Interface circuit


Fig. 3. Relationship of Uh and Vh.

It is noted that the magnitude of the input voltage of the


main motor (Uv) in the MATLAB/Simulink environment Lah/v Rah/v
and that of the motor terminal voltage (Vv) are not same,
iah/v
and the relationship between these two is nonlinear. The
relationship is obtained by changing the input voltage in Eah/v P
Vv/h N
the MATLAB/Simulink environment, and measuring the
relative terminal voltage as shown in Fig. 2. Like the main h/v
motor the magnitude of the input voltage of the tail motor
(Uh) in the MATLAB/Simulink environment and that of
the motor terminal voltage (Vh) are different. Fig. 3 shows Fig. 4. Circuit diagram of a DC motor.
the relationship between Uh and Vh.
in Eqs. (1)–(5). The motor parameters are listed in Table 2:
2.2. DC motors diah=v
V h=v ¼ E ah=v þ Rah=v iah=v þ Lah=v , (1)
dt
The TRMS possesses two permanent magnet DC
motors; one for the main and the other for the tail E ah=v ¼ kah=v jh=v oh=v , (2)
propelling. The motors are identical with different mechan-
ical loads. The mathematical equations governing the doh=v
M eh=v ¼ M Lh=v þ J tr=mr þ Btr=mr oh=v , (3)
main and tail motors, as shown in Fig. 4, are presented dt
ARTICLE IN PRESS
A. Rahideh et al. / Control Engineering Practice 16 (2008) 241–259 245

Table 2 dav
The motors parameters
¼ Ov . (8)
dt
Parameter Value

Pn (W) 6
2.4. 1DOF TRMS modelling in horizontal plane
Vn (V) 18
on (rpm) 8300 The mathematical model of the remaining parts of the
Jmr (g cm2) 1272 system in horizontal plane is described in Eqs. (9)–(11) (see
Jtr (g cm2) 248 Fig. 6). In Eq. (9) the first term is the torque of propulsive
Ra (O) 8
La (mH) 0.86
force due to the tail rotor, the second term implies the
Ka.j (N m/A) 0.0202 torque of the friction force, and the third term refers to the
kth 3.6  107 torque of the flat cable force that is completely nonlinear
ktv 8.7  107 and can be obtained by point-by-point measurement. Fig. 7
shows the torque of the friction force that covers viscous,
coulomb and static frictions:
− v Vertical plane
Tail rotor lt dOh l t F h ðoh Þcos av  M fric;h  M cable ðah Þ
lm ¼ , (9)
dt D cos2 av þ E sin2 av þ F
where
g(mtr + mts) Fv (v)
mtg av ¼ cte,
lcb Main rotor
m  m 
mmg m t
lb − lcb
gmb
D¼ þ mmr þ mms l 2m þ þ mtr þ mts l 2t ,
g(mmr + mms) 3 3
gmcb mb 2 2 2 mts 2
E¼ l þ mcb l cb ; F ¼ mms rms þ r ,
3 b 2 ts
TRMS 33-220 (
kfhp joh joh for oh X0;
F h ðoh Þ ¼ (10)
kfhn joh joh for oh o0;
Fig. 5. Gravity forces and propulsive force in the vertical plane.

dah
¼ Oh . (11)
M eh=v ¼ kah=v jh=v iah=v , (4) dt

M Lh=v ¼ kth=v joh=v joh=v . (5)


Horizontal plane
Tail rotor
TRMS 33-220

2.3. Vertical 1DOF model of the TRMS


h Main rotor
Fh (h)
The model of the remaining parts of the system in
vertical plane is described in Eqs. (6)–(8) (see also Fig. 5).
In Eq. (6) the first term denotes the torque of the
propulsive force due to the main rotor, the second term Horizontal axis
refers to the torque of the friction force, and the torque of
gravity force is shown in the third term. Fig. 6. Propulsive force in the horizontal plane.
dOv l m F v ðov Þ  M fric;v þ g½ðA  BÞcos av  C sin av 
¼ ,
dt Jv
(6) Mfric,v/h

where
m  m 
t m
A¼ þ mtr þ mts l t ; B ¼ þ mmr þ mms l m ,
2 2
m  Ωv/h
b
C¼ l b þ mcb l cb ,
2
(
kfvp jov jov for ov X0;
F v ðov Þ ¼ (7)
kfvn jov jov for ov o0;
Fig. 7. The friction torque profile.
ARTICLE IN PRESS
246 A. Rahideh et al. / Control Engineering Practice 16 (2008) 241–259

3. Lagrangian based modelling


Tail rotor

TRMS 33-220
In developing Lagrangian based model the TRMS h R O
3 y
configuration has been divided into three subsystems; the
P3 Main rotor
first one consists of the free–free beam (tail and main
beams), tail rotor, main rotor, tail shield and the main
P1
shield, the second one comprises the counterbalance beam
h
and weight, and finally the third one is a pivoted beam.
x
3.1. Coordinates
Fig. 10. Top view of TRMS.
Assume ½rx ðR1 Þ; ry ðR1 Þ; rz ðR1 Þ denotes the coordinate of
point P1 on the free–free beam parameterized in the
distance R1 from O1 (that means P1O1 ¼ R1). And also It should be noted that ah has no effect on the rz(R)’s
assume OO1 ¼ h in which O is the original of the and for simplicity it can be assumed to be zero as shown in
coordinates. It is noted that in order to simplify the figure Fig. 9.
the x- and y-axes have been drawn from O2. Let ½rx ðR2 Þ; ry ðR2 Þ; rz ðR2 Þ denotes the coordinate of
According to Figs. 8–10 the following equation can be point P2 on the counterbalance beam parameterized in
obtained: the distance R2 from O1 (that means P2O1 ¼ R2). Accord-
8
ing to Fig. 8 the following equation can be obtained:
< rx ðR1 Þ ¼ R1 sin ah cos av þ h cos ah ;
>
8
ry ðR1 Þ ¼ R1 cos ah cos av  h sin ah ;
>
: r ðR Þ ¼ R sin a :
(12) < rx ðR2 Þ ¼ R2 sin ah sin av þ h cos ah ;
>
z 1 1 v ry ðR2 Þ ¼ R2 cos ah sin av  h sin ah ; (13)
>
: r ðR Þ ¼ R cos a :
z 2 2 v
Tail rotor z
For more accuracy the point P3 can be considered with
Pivoted beam
the coordinate ½rx ðR3 Þ; ry ðR3 Þ; rz ðR3 Þ on the pivoted beam
P3 O where R3 is the distance between P3 and O:
O1 Free-free beam 8
h
< rx ðR3 Þ ¼ R3 cos ah ;
>
Counter P1 Main rotor ry ðR3 Þ ¼ R3 sin ah ; (14)
balance
P2
>
: r ðR Þ ¼ 0:
beam z 3

−v

3.2. 1DOF TRMS modelling in horizontal plane


ry (R1)
O2
y In this case the vertical angle is constant and so the
velocities related to each part can be obtained by
rx (R1) differentiating Eqs. (12)–(14) with respect to time and the
square magnitude of the velocity of Pi is given by
P′1
v2 ðRi Þ ¼ v2x ðRi Þ þ v2y ðRi Þ þ v2z ðRi Þ; i ¼ 1; 2; 3. (15)
x
Substitution and simplification yield the following
Fig. 8. Twin rotor MIMO system.
equations:
z v2 ðR1 Þ ¼ R21 a_ 2h cos2 av þ h2 a_ 2h , (16)
R1
v2 ðR2 Þ ¼ R22 a_ 2h sin2 av þ h2 a_ 2h , (17)
O1
y
−v
Tail rotor R2 v2 ðR3 Þ ¼ R23 a_ 2h . (18)
P1
P2
The kinetic and potential energies are shown in the
Main rotor following equations, respectively:
Z
T ¼ 2 v2 ðRÞ dmðRÞ;
1
(19)

TRMS 33-220 Z
V ¼g rz ðRÞ dmðRÞ. (20)
Fig. 9. Front view of TRMS with ah ¼ 0.
ARTICLE IN PRESS
A. Rahideh et al. / Control Engineering Practice 16 (2008) 241–259 247

Z
The kinetic and potential energies of the free–free beam mh 2
J3 ¼ R23 dmðR3 Þ ¼ h ,
are also obtained as follows: 3
T 1 ¼ 12J 1 a_ 2h cos2 av þ 12h2 a_ 2h mT1 , (21) V 3 ¼ 0. (26)
Z m  The Lagrangian equation defines L as
t
J1 ¼ R21 dmðR1 Þ ¼ þ mtr þ mts l 2t
3 X
3 X
3
m  mms 2 L¼ Ti  Vi (27)
m
þ þ mmr þ mms l 2m þ r þ mts r2ts , i¼1 i¼1
3 2 ms
Z and the equation of motion is given by
 
mT1 ¼ dmðR1 Þ ¼ mt þ mtr þ mts þ mm þ mmr þ mms , d qL qL X
 ¼ M ih . (28)
dt q_ah qah i
V 1 ¼ gmT 1 l T 1 sin av , (22)
Therefore by substituting Eqs. (21)–(26) into Eqs. (27)
R and (28) and simplification, the following equation is
R1 dmðR1 Þ
l T1 ¼ R obtained:
dmðR1 Þ X
ððmm =2Þ þ mmr þ mms Þl m  ððmt =2Þ þ mtr þ mts Þl t ½J 1 cos2 av þ J 2 sin2 av þ h2 ðmT 1 þ mT 2 Þ þ J 3 €ah ¼ M ih .
¼ . i
mT 1
(29)
The kinetic and potential energies of the counterbalance P
In Eq. (29), i M ih is the sum of applied torques in the
beam are demonstrated as
horizontal movement, and can be summarized as
T 2 ¼ 12J 2 a_ 2h sin2 av þ 12h2 a_ 2h mT2 , (23) X
M ih ¼ M prop;h  M fric;h  M cable ðah Þ, (30)
Z i
mb 2
J2 ¼ R22 dmðR2 Þ ¼ l þ mcb l 2cb ,
3 b M prop;h ¼ l t F h ðoh Þcos av .
Z
mT2 ¼ dmðR2 Þ ¼ mb þ mcb , 3.3. 1DOF TRMS modelling in vertical plane

V 2 ¼ gmT2 l T2 cos av , (24) This time the horizontal angle is constant and so the
velocities related to each part can be obtained by differentiating
R
R2 dmðR2 Þ ðmb =2Þl b þ mcb l cb Eqs.(12)–(14) with respect to time and the square magnitude of
l T2 ¼ R ¼ . the velocity of Pi in vertical plane is given by Eq. (15).
dmðR2 Þ mT2
Substitution and simplification yield the following equations:
The following equations are the kinetic and potential
energies of the pivoted beam, respectively: v2 ðR1 Þ ¼ R21 a_ 2v , (31)

T 3 ¼ 12J 3 a_ 2h , (25) v2 ðR2 Þ ¼ R22 a_ 2v , (32)

Fig. 11. A typical MLP network with R inputs and three layers i.e. M ¼ 3.
ARTICLE IN PRESS
248 A. Rahideh et al. / Control Engineering Practice 16 (2008) 241–259

v2 ðR3 Þ ¼ 0. (33) The equation of motion is given by


 
The kinetic and potential energies of each part can be d qL qL X
 ¼ M iv . (39)
written as follows: dt q_av qav i

T 1 ¼ 12J 1 a_ 2v , (34) So by substituting Eqs. (34)–(38) into Eqs. (27)–(39) and


simplifying the following equation is obtained:
V 1 ¼ gmT 1 l T 1 sin av , (35)   X
ðJ 1 þ J 2 Þ€av þ g mT 1 l T 1 cos av þ mT 2 l T 2 sin av ¼ M iv .
i

T2 ¼ 1 _2
2 J 2 av , (36) (40)
P
In Eq. (40), i M iv is the sum of applied torques in the
V 2 ¼ gmT 2 l T 2 cos av , (37) vertical movement, and can be expressed as
X
M iv ¼ M prop;v  M fric;v , (41)
T 3 ¼ V 3 ¼ 0. (38) i

M prop;v ¼ l m F v ðov Þ.
It can be seen both Newtonian and Lagrangian approaches
lead to the similar equations; however, the assumptions of
them are a little different, that means the pivoted beam has
been considered in the Lagrangian approach but it has been
neglected in the Newtonian method.

Fig. 13. The structure of the NN-based model in terms of 1DOF


horizontal.

Fig. 12. The flowchart of the Levenberg–Marquardt learning approach. Fig. 14. The structure of the NN-based model in terms of 1DOF vertical.
ARTICLE IN PRESS
A. Rahideh et al. / Control Engineering Practice 16 (2008) 241–259 249

0.6
100
Real TRMS
0.4 Model
10−1
0.2
10−2
Yaw angle (rad)

Magnitude
PSD of real TRMS
0
10−3 PSD of the model

−0.2
10−4

−0.4 10−5

−0.6 10−6

−0.8 10−7
0 10 20 30 40 50 60 70 0 0.5 1 1.5 2 2.5
time (sec) Frequency (Hz)

Fig. 15. The yaw angle of the beam in response to sine input with frequency of 0.2 Hz and amplitude 0.7 V, and PSD.

0.8
101
Real TRMS
0.6 Model
100
0.4
10−1
Yaw angle (rad)

PSD of real TRMS


0.2 PSD of the model
Magnitude

10−2
0
10−3
−0.2

−0.4 10−4

−0.6 10−5

−0.8 10−6
0 20 40 60 80 100 0 0.5 1 1.5 2 2.5
time (sec) Frequency (Hz)

Fig. 16. The yaw angle of the beam in response to square input with frequency of 0.1 Hz and amplitude 0.3 V, and PSD.

0.25
Real TRMS 10−1
0.2 Model
10−2
0.15
10−3
Pitch angle (rad)

PSD of real TRMS


PSD of the model
Magnitude

0.1
10−4
0.05
10−5

0 10−6

−0.05 10−7

−0.1 10−8
0 10 20 30 40 50 0 0.5 1 1.5 2 2.5
time (sec) Frequency (Hz)

Fig. 17. The pitch angle of the beam in response to sine input with frequency of 0.1 Hz and amplitude 0.5 V, and PSD.
ARTICLE IN PRESS
250 A. Rahideh et al. / Control Engineering Practice 16 (2008) 241–259

4. NN-based modelling ments arranged in patterns similar to biological NNs.


These computational models have now become exciting
Artificial neural networks (ANNs) are computational alternatives to conventional approaches in solving a variety
models comprising numerous nonlinear processing ele- of engineering and scientific problems. Fig. 11 shows a

0.4
100
Real TRMS
0.3 Model
10−1
0.2
10−2
Pitch angle (rad)

PSD of real TRMS

Magnitude
0.1
10−3 PSD of the model

0
10−4

−0.1 10−5

−0.2 10−6

−0.3 10−7
0 20 40 60 80 100 0 0.5 1 1.5 2 2.5
time (sec) Frequency (Hz)

Fig. 18. The pitch angle of the beam in response to square input with frequency of 0.05 Hz and amplitude 0.5 V, and PSD.

0.6
NN−based model 100
Real TRMS PSD of NN−based model
0.4
10−1 PSD of real TRMS

0.2
10−2
Yaw angle (rad)

Magnitude

0
10−3

−0.2
10−4

−0.4 10−5

−0.6 10−6

−0.8 10−7
0 10 20 30 40 50 60 70 0 0.5 1 1.5 2 2.5
time (sec) Frequency (Hz)

Fig. 19. The yaw angle of the beam in response to sine input with frequency of 0.2 Hz and amplitude 0.7 V, and PSD (LM).

0.8
101
NN−based model
PSD of NN−based model
0.6 Real TRMS
PSD of real TRMS
100
0.4
10−1
Yaw angle (rad)

0.2
Magnitude

10−2
0
10−3
−0.2

−0.4 10−4

−0.6 10−5

−0.8 10−6
0 20 40 60 80 100 0 0.5 1 1.5 2 2.5
time (sec) Frequency (Hz)

Fig. 20. The yaw angle of the beam in response to square input with frequency of 0.1 Hz and amplitude 0.3 V, and PSD (LM).
ARTICLE IN PRESS
A. Rahideh et al. / Control Engineering Practice 16 (2008) 241–259 251

typical multi-layer perceptron (MLP) network with R this part the LM learning algorithm is briefly explained.
inputs, M layers (here M ¼ 3), S1 neurons in the first layer, A full description can be found in Hagan and Menhaj
S2 neurons in the second layer and SM neurons in the Mth (1994). Assume all network weights and biases are
layer. The final layer is called the output layer due to the considered in a vector as
fact that it produces the network output. The other layers
are hidden layers, e.g. in Fig. 11 the network has two
w ¼ ½w11;1 . . . w1S1;R b11 . . . b1S1 w21;1 . . . w2S2;S1 b21 . . . b2S2 . . . . . .
hidden layers. As mentioned before LM and GD learning
M M T
algorithms have been used to train the empirical models. In wM M
1;1 . . . wSM;SM1 b1 . . . bSM 

Auto–correlation of residuals Cross–correlation of input and residuals


1 1
0.8 0.8
0.6 0.6
0.4 0.4
0.2 0.2
0 0
−0.2 −0.2
−0.4 −0.4
−0.6 −0.6
−0.8 −0.8
−1 −1
−500 0 500 −500 0 500
lag lag

Cross–correlation of input square and residuals Cross–correlation of input square and residuals square
1 1
0.8 0.8
0.6 0.6
0.4 0.4
0.2 0.2
0 0
−0.2 −0.2
−0.4 −0.4
−0.6 −0.6
−0.8 −0.8
−1 −1
−500 0 500 −500 0 500
lag lag

Cross–correlation of residuals and (input*residuals)


1
0.8
0.6
0.4
0.2
0
−0.2
−0.4
−0.6
−0.8
−1
−500 0 500
lag

Fig. 21. Correlation validation of 1DOF horizontal NN-based model with LM training algorithm.
ARTICLE IN PRESS
252 A. Rahideh et al. / Control Engineering Practice 16 (2008) 241–259

with n elements (the elements can be renamed as w(t+1) ¼ w(t)+Dw(t) and m is decreased using m ¼ m/b.
w ¼ ½ w1 w2    wn T ) and the total sum square error The new iteration is started if the stop criteria are not
P
is defined as V ðwÞ ¼ N 2
i¼1 ei ðwÞ, where N ¼ Q  SM, Q is
satisfied, otherwise m is increased using m ¼ mb and Dw(t) is
the number of patterns used for training and SM is the recalculated.
number of network outputs. The error vector is considered
as e ¼ ½ e1 e2    eN T . Note that each element of 4.1. 1DOF horizontal
error vector (e.g. ei) represents the difference between one
of the network outputs and the corresponding desired In order to model the TRMS in terms of 1DOF
value for only one patterns of training set. The flowchart of horizontal dynamics an MLP NN model of 6  3  1
LM learning approach is illustrated in Fig. 12. configuration has been used. In other words, the NN-based
LM learning algorithm can be considered as a modified model has six inputs, three neurons in hidden layer and one
version of Gauss–Newton approach. Like all other training neuron in output layer. To find a suitable configuration it is
algorithms the first step is initialization. The Jacobian common to start from a simple configuration, usually only
matrix is calculated using the differentiation of each error one hidden layer, and then increase the number of neurons
with respect to each weight or bias and consequently the and even the number of layers if necessary. The inputs are
result will be an N  n matrix. A set of equations are tail rotor voltage at present time, Vh(t), tail rotor voltage at
defined in order to find Dw(t). The new sum square error is previous time, Vh(t1), tail rotor voltage at two samples
then computed and compared to the previous one, if the before, Vh(t2), yaw angle of the beam at previous time,
performance becomes better, i.e. the new error is less than ah(t1), yaw angle of the beam at two samples before,
the previous one, then the new parameters are defined as ah(t2), and yaw angle of the beam at three samples

0.6 100
PSD of NN−based model
NN−based model
PSD of real TRMS
0.4 Real TRMS 10−1

0.2 10−2
Yaw angle (rad)

Magnitude

0 10−3

−0.2 10−4

−0.4 10−5

−0.6 10−6

−0.8 10−7
0 10 20 30 40 50 60 70 0 0.5 1 1.5 2 2.5
time (sec) Frequency (Hz)

Fig. 22. The yaw angle of the beam in response to sine input with frequency of 0.2 Hz and amplitude 0.7 V, and PSD (GD).

0.8
NN−based model
101
PSD of NN−based model
0.6 Real TRMS PSD of real TRMS
100
0.4
10−1
Yaw angle (rad)

0.2
Magnitude

10−2
0
10−3
−0.2

−0.4 10−4

−0.6 10−5

−0.8 10−6
0 20 40 60 80 100 0 0.5 1 1.5 2 2.5
time (sec) Frequency (Hz)

Fig. 23. The yaw angle of the beam in response to square input with frequency of 0.1 Hz and amplitude 0.3 V, and PSD (GD).
ARTICLE IN PRESS
A. Rahideh et al. / Control Engineering Practice 16 (2008) 241–259 253

before, ah(t3). Fig. 13 shows the structure of the NN- Levenberg–Marquardt and gradient descent (GD) training
based model in terms of 1DOF horizontal dynamics. It is methods. It is noted that all data for training and testing
noted that the activation functions used in the hidden layer have been extracted from the real TRMS.
and output layer are logarithmic and pure linear,
respectively. The number of patterns for training and 4.2. 1DOF vertical
testing are 8290 and 850 sets, respectively. The sample time
of data is set to be 0.2 s which implies that the frequency of In order to model the TRMS in terms of 1DOF vertical
sampling is 5 Hz. The NN has been trained with dynamics, an MLP NN model of 6  3  1 configuration

Auto–correlation of residuals Cross–correlation of input and residuals


1 1
0.8 0.8
0.6 0.6
0.4 0.4
0.2 0.2
0 0
−0.2 −0.2
−0.4 −0.4
−0.6 −0.6
−0.8 −0.8
−1 −1
−500 0 500 −500 0 500
lag lag

Cross–correlation of input square and residuals Cross–correlation of input square and residuals square
1 1
0.8 0.8
0.6 0.6
0.4 0.4
0.2 0.2
0 0
−0.2 −0.2
−0.4 −0.4
−0.6 −0.6
−0.8 −0.8
−1 −1
−500 0 500 −500 0 500
lag lag

Cross–correlation of residuals and (input*residuals)


1
0.8
0.6
0.4
0.2
0
−0.2
−0.4
−0.6
−0.8
−1
−500 0 500
lag

Fig. 24. Correlation validation of 1DOF horizontal NN-based model with GD training algorithm.
ARTICLE IN PRESS
254 A. Rahideh et al. / Control Engineering Practice 16 (2008) 241–259

has been used. In other words, the NN-based model has six comparing their responses with that of the real TRMS,
inputs, three neurons in the hidden layer and one neuron in comparing the power spectral densities (PSDs) of the
the output layer. The inputs are main rotor voltage at model and the real TRMS and performing a set of
present time, Vv(t), main rotor voltage at previous time, correlation tests. This section is divided into two subsec-
Vv(t1), main rotor voltage at two samples before, tions; one presenting the response of the analytical
Vv(t2), pitch angle of the beam at previous time, av(t1), modelling that is obtained using energy and mass equations
pitch angle of the beam at two samples before, av(t2), and and another related to the response of empirical NN-based
pitch angle of the beam at three samples before, av(t3). modelling with two learning algorithm approaches.
Fig. 14 shows the structure of the NN-based model in
terms of 1DOF vertical dynamics. It is noted that the
activation functions used in the hidden layer and the 5.1. Newtonian and Lagrangian based model
output layer are logarithmic sigmoid and pure linear,
respectively. The number of patterns for training and As mentioned before the Lagrangian and Newtonian
testing are set to be 10 700 and 750, respectively. The NN based model approaches lead to the similar equations, so
has been trained with GD as well as Levenberg–Marquardt the responses are also the same as shown here. For more
training methods. validation of the performance of the models, the PSD of
each response has been shown. Figs. 15 and 16 show the
5. Results yaw angle of the beam with respect to different input
signals and their corresponding PSDs. The pitch angle
Results of the developed models are discussed in this responses and corresponding PSDs are also shown in
section. Performances of the models are presented by Figs. 17 and 18.

0.25
NN−based model 10−1
Real TRMS PSD of NN−based model
0.2 PSD of real TRMS
10−2
0.15
Pitch angle (rad)

10−3
Magnitude

0.1
10−4
0.05

0 10−5

−0.05 10−6

−0.1 10−7
0 10 20 30 40 50 0 0.5 1 1.5 2 2.5
time (sec) Frequency (Hz)

Fig. 25. The pitch angle of the beam in response to sine input with frequency of 0.1 Hz and amplitude 0.5 V, and PSD (LM).

0.4
NN based model
100
Real TRMS PSD of NN
0.3
10−1 PSD of real TRMS

0.2
10−2
Pitch angle (rad)

Magnitude

0.1
10−3

0 10−4

−0.1 10−5

−0.2 10−6

−0.3 10−7
0 20 40 60 80 100 0 0.5 1 1.5 2 2.5
time (sec) Frequency (Hz)

Fig. 26. The pitch angle of the beam in response to square input with frequency of 0.05 Hz and amplitude 0.5 V, and PSD (LM).
ARTICLE IN PRESS
A. Rahideh et al. / Control Engineering Practice 16 (2008) 241–259 255

5.2. NN-based model those of previous section have been used in testing
procedure. Figs. 19 and 20 show the output of 1DOF
Two NNs with the same configurations have been horizontal NN-based model with LM training method with
developed for modelling the TRMS in terms of 1DOF respect to different input signals and their corresponding
vertical and horizontal dynamics. For validation, the PSDs. Correlation validation (Billings & Voon, 1986;
models’ dynamics have been tested with a data set that Billings & Zhu, 1994) of 1DOF horizontal NN-based
has not been used for the training purpose. In order to model with LM training algorithm is shown in Fig. 21. For
achieve a reasonable comparison, the inputs similar to comparison purpose, the results of the model developed on

Auto–correlation of residuals Cross–correlation of input and residuals


1 1
0.8 0.8
0.6 0.6
0.4 0.4
0.2 0.2
0 0
−0.2 −0.2
−0.4 −0.4
−0.6 −0.6
−0.8 −0.8
−1 −1
−500 0 500 −500 0 500
lag lag

Cross–correlation of input square and residuals Cross–correlation of input square and residuals square
1 1
0.8 0.8
0.6 0.6
0.4 0.4
0.2 0.2
0 0
−0.2 −0.2
−0.4 −0.4
−0.6 −0.6
−0.8 −0.8
−1 −1
−500 0 500 −500 0 500
lag lag

Cross–correlation of residuals and (input*residuals)


1
0.8
0.6
0.4
0.2
0
−0.2
−0.4
−0.6
−0.8
−1
−500 0 500
lag

Fig. 27. Correlation validation of 1DOF vertical NN-based model with LM training algorithm.
ARTICLE IN PRESS
256 A. Rahideh et al. / Control Engineering Practice 16 (2008) 241–259

the basis of standard GD algorithm have also been training set and that the model approximates the system
presented. Figs. 22 and 23 show the output of 1DOF well. It is common to check if the functions for lags in the
horizontal NN-based model with GD training method with interval t 2 ½20; 20 are zero within
pffiffiffiffi an asymptotical
pffiffiffiffi 95%
respect to different input signals and their corresponding confidence interval, i.e. if 1:96= moro1:96= m where r
PSDs. Correlation validation of 1DOF horizontal NN- is the correlation function and m is the number of data.
based model with GD training algorithm is shown in It is noted from Figs. 19 and 20 that the horizontal model
Fig. 24. with LM training algorithm performed so well that the
Figs. 25 and 26 show the output of 1DOF vertical NN- model and real system responses cannot be distinguished. It
based model with LM training method with respect to is also obvious from these figures that the PSDs of the model
different input signals and their corresponding PSDs. and system responses are closely overlapped especially at
Correlation validation of 1DOF vertical NN-based model low frequencies. It is noted that all the five correlation
with LM training algorithm is shown in Fig. 27. Figs. 28 functions; auto-correlation of residuals (Fig. 21a), cross-
and 29 show the output of 1DOF vertical NN-based model correlation of input and residuals (Fig. 21b), cross-correla-
with GD training method with respect to different input tion of input square and residuals (Fig. 21c), cross-
signals and their corresponding PSDs. Correlation valida- correlation of input square and residuals square (Fig. 21d),
tion of 1DOF vertical NN-based model with GD training and cross-correlation of residuals and input multiplied by
algorithm is shown in Fig. 30. residuals (Fig. 21e) are within the 95% confidence bands
If the residuals (model errors) contain no information indicating that the model is adequate, i.e. the model
about past residuals or about the dynamics of the system, it behaviour is closed to the real system performance.
is likely that all information has been extracted from the However, although Figs. 22 and 23 show that the horizontal

0.25
NN−based model 10−1
Real TRMS PSD of NN−based model
0.2 PSD of real TRMS
10−2
0.15
10−3
Pitch angle (rad)

Magnitude

0.1
10−4
0.05
10−5

0 10−6

−0.05 10−7

−0.1 10−8
0 10 20 30 40 50 0 0.5 1 1.5 2 2.5
time (sec) Frequency (Hz)

Fig. 28. The pitch angle of the beam in response to sine input with frequency of 0.1 Hz and amplitude 0.5 V, and PSD (GD).

0.3
100
PSD of NN−based model
0.2 PSD of real TRMS
10−1
0.1
10−2
Pitch angle (rad)

Magnitude

0
10−3

−0.1 10−4

−0.2 10−5

−0.3 NN−based model 10−6


Real TRMS
−0.4 10−7
0 20 40 60 80 100 0 0.5 1 1.5 2 2.5
time (sec) Frequency (Hz)

Fig. 29. The pitch angle of the beam in response to square input with frequency of 0.05 Hz and amplitude 0.5 V, and PSD (GD).
ARTICLE IN PRESS
A. Rahideh et al. / Control Engineering Practice 16 (2008) 241–259 257

Auto–correlation of residuals Cross–correlation of input and residuals


1 1
0.8 0.8
0.6 0.6
0.4 0.4
0.2 0.2
0 0
−0.2 −0.2
−0.4 −0.4
−0.6 −0.6
−0.8 −0.8
−1 −1
−500 0 500 −500 0 500
lag lag

Cross–correlation of input square and residuals Cross–correlation of input square and residuals square
1 1
0.8 0.8
0.6 0.6
0.4 0.4
0.2 0.2
0 0
−0.2 −0.2
−0.4 −0.4
−0.6 −0.6
−0.8 −0.8
−1 −1
−500 0 500 −500 0 500
lag lag

Cross–correlation of residuals and (input*residuals)


1
0.8
0.6
0.4
0.2
0
−0.2
−0.4
−0.6
−0.8
−1
−500 0 500
lag

Fig. 30. Correlation validation of 1DOF vertical NN-based model with GD training algorithm.

model with GD training algorithm has a reasonable real system responses are perfectly overlapped. It is also
performance with respect to overlapping and PSD tests, obvious from these figures that the PSDs of the model and
but according to Fig. 24 the correlation functions go beyond system responses are closely matched especially at low
the 95% confidence bands and therefore the model cannot frequencies. It is noted that all the five correlation
be considered adequate to be used for sophisticated control functions; auto-correlation of residuals (Fig. 27a), cross-
purpose. correlation of input and residuals (Fig. 27b), cross-
Figs. 25 and 26 show that the vertical model with LM correlation of input square and residuals (Fig. 27c),
training algorithm performed so well that the model and cross-correlation of input square and residuals square
ARTICLE IN PRESS
258 A. Rahideh et al. / Control Engineering Practice 16 (2008) 241–259

(Fig. 27d), and cross-correlation of residuals and input between the model output and the real system output, and
multiplied by residuals (Fig. 27e) are within the 95% reliability of the model in a range of necessary frequencies.
confidence bands indicating that the model is adequate.
However, Figs. 28 and 29 show that the vertical model with 6. Conclusion
GD training algorithm has a reasonable performance with
respect to overlapping and PSD tests, but according to In this investigation the dynamic modelling of an
Fig. 30 the correlation functions go beyond the 95% experimental system, TRMS, using both analytical and
confidence bands and therefore the model cannot be empirical based methods has been developed. In develop-
considered to be adequate. ing the analytical model of the system various parts of the
TRMS have been modelled separately and then integrated.
5.3. Comparative study The responses of the models are compared with that of the
real plant. It is shown that the models obtained using both
It can be observed from the results that the error of methods can generally be considered adequate in repre-
empirical model with efficient learning algorithms is less senting the system. However, the NN-based modelling
than the error of the analytical model. Although the approach with LM training algorithm appeared to be the
analytical model is less accurate it is valid for any input best in terms of accuracy, but analytical approaches have a
signal with any frequency. The empirical model is only better performance compared to the GD based NN
valid in the range of training data provided. On the other modelling approach.
hand, in the case of black-box modelling, the structure of
the process is lost and this is one of the drawbacks of References
empirical modelling. The significant advantage of
NN-based modelling is that the model is easy and fast to Ahmad, S. M., Chipperfield, A. J., & Tokhi, M. O. (2000). Dynamic
develop. Table 3 depicts the comparative performance of modelling and optimal control of a twin rotor MIMO system. In
Proceedings of the 2000 IEEE international conference on national
the analytical and empirical models. The responses of the
aerospace and electronics (pp. 391–398).
both Newtonian and Lagrangian models are satisfactory Ahmad, S. M., Shaheed, M. H., Chipperfield, A. J., & Tokhi, M. O.
and they follow the real TRMS response very closely. On (2000). Nonlinear modelling of a twin rotor MIMO system using radial
the other hand, although the NN-based models with LM basis function networks. In Proceedings of the 2000 IEEE international
training method follow the real TRMS response very conference on national aerospace and electronics (pp. 313–320).
Aldebrez, F. M., Mat-Darus, I. Z., & Tokhi, M. O. (2004). Dynamic
satisfactorily but the performance of the GD training
modelling of a twin rotor system in hovering position. In Proceedings
method is not as good as that of the LM training method. of the 2004 international symposium on control, communications and
Again, the response of the NN-based models with LM signal processing (pp. 823–826).
training method is better than that of the analytical models, Avila-Vilchis, J. C., Brogliato, B., Dzul, R., & Lozano, R. (2003).
but the responses of analytical models have a better Nonlinear modelling and control of helicopters. Automatica, 39(9),
1583–1596.
performance compared to the NN-based model with GD
Billings, S. A., & Voon, W. S. N. (1986). Correlation based model validity
training approach. As presented in Table 3, the mean tests for non-linear models. International Journal of Control, 44(1),
square errors (MSEs) of the yaw angle in the case of the 235–244.
Lagrangian and Newtonian models are 1.9  103 as Billings, S. A., & Zhu, Q. M. (1994). Nonlinear model validation using
against 1.35  105 and 1.74  102 in the case of NN- correlation tests. International Journal of Control, 60(6), 1107–1120.
Castillo, P., Dzul, A., & Lozano, R. (2004). Real-time stabilization and
based model with LM and GD training methods,
tracking of a four-rotor mini rotorcraft. IEEE Transaction on Control
respectively. Similar performance is achieved in the case Systems Technology, 12(4), 510–516.
of pitch angle. In this case the MSE 7.1  104 is obtained Cerro, J. D., Valero, J., & Barrientos, A. (2005). Identification of a small
with analytical models as opposed to 8.84  106 and unmanned helicopter model using genetic algorithms. In Proceedings
1.15  103 in the case of NN-based models with LM and of the 2005 IEEE international conference on intelligent robots and
systems (pp. 3360–3365).
GD training methods, respectively. Note that the criteria
Feedback Co., 1998. Twin rotor MIMO system 33-220 user manual.
for the model satisfaction are small amount of error Hagan, M. T., & Menhaj, M. B. (1994). Training feedforward networks
with the Marquardt algorithm. IEEE Transaction on Neural Networks,
Table 3 5(6), 989–993.
Comparative study of the analytical and empirical approaches Jun, M., Roumeliotis, S. I., & Sukhatme, G. S. (1999). State estimation of
an autonomous helicopter using Kalman filtering. In Proceedings of
Approach Model Yaw angle Pitch angle the 1999 IEEE international conference on intelligent robots and
MSE MSE systems, Vol. 3 (pp. 1346–1353).
Karras, D. A. (2004). Neural network models based on regularization
Analytical Newtonian 1.9  103 7.1  104 techniques for off-line robot manipulator path planning. In Proceed-
Lagrangian 1.9  103 7.1  104 ings of the 2004 IEEE international joint conference on neural networks,
Empirical NN(LM) Train 1.57  105 1.54  105 Vol. 1 (pp. 35–39).
Test 1.35  105 8.84  106 Krabbes, M., & Doschner, C. (1999). Modelling of robot dynamics based
NN(GD) Train 1.13  102 2.37  102 on a multi-dimensional RBF-like neural network. In Proceedings of the
Test 1.74  102 1.15  103 1999 IEEE international conference on information intelligence and
systems (pp. 180–187).
ARTICLE IN PRESS
A. Rahideh et al. / Control Engineering Practice 16 (2008) 241–259 259

Mahony, R., Hamel, T., Dzul, A. (1999). Hover control via Lyapunov Song-Yun, D., Ji-Hong, Z., Chun-Hua, H. (2004). Mathematical
control for an autonomous model helicopter. In Proceedings of the modelling of a UAV during take-off with nose-wheel off the
1999 IEEE international conference on decision and control, Vol. 4 ground. In Proceedings of the World automation congress, Vol. 15
(pp. 3490–3495). (pp. 359–364).
Mat-Darus, I. Z., Aldebrez, F. M., & Tokhi, M. O. (2004). Parametric Sun, W., Wang, Y., Mao, J. (2002). Using wavelet network for identi-
modelling of a twin rotor system using genetic algorithms. In fying the model of robot manipulator. In Proceedings of the
Proceedings of the 2004 international symposium on control, commu- fourth congress on intelligent control and automation, Vol. 2
nications and signal processing (pp. 115–118). (pp. 1634–1638).
Mistler, V., Benalleque, A., & Msirdi, N. K. (2001). Exact linearization Talebi, H. A., Patel, R. V., & Asmer, H. (1998). Dynamic modelling of
and non-interacting control of a 4 rotors helicopter via dynamic flexible-link manipulators using neural networks with application to
feedback. In Proceedings of the 2001 IEEE international workshop on the SSRMS. In Proceedings of the 1998 IEEE international conference
robot and human interactive communication (pp. 586–593). on intelligent robots and systems (pp. 673–678).
Mohd-Hashim, S. Z., Tokhi, M. O., & Mat-Darus, I. Z. (2004). Nonlinear Van de Ven, P., Johansen, T., Sorensen, A., Flanagan, C., & Toal, D.
dynamic modelling of flexible beam structures using neural networks. (2007). Neural network augmented identification of underwater vehicle
In Proceedings of the 2004 IEEE international conference on models. Control Engineering Practice, 15(6), 715–725.
mechatronics (pp. 171–175). Wang, X., Peng, G. (2003). Modelling and control for pneumatic
Rao, M., Biswas, S. K., Butz, B. P., & Miller, D. G. (1990). Dynamic manipulator based on dynamic neural network. In Proceedings of the
modelling and feedback control of a side-by-side rotor tandem 2003 IEEE international conference on systems, man and cybernetics
helicopter. In Proceedings of the 1990 IEEE international symposium Vol. 3 (pp. 2231–2236).
on intelligent control systems (pp. 780–785). Wu, X., Ignatov, R., Muenst, G., Imaev, A., & Zhu, J. J. (2002).
Shaheed, M. H. (2004). Performance analysis of 4 types of conjugate A nonlinear flight controller design for a UFO by trajectory
gradient algorithm in the nonlinear dynamic modelling of a TRMS using linearization method, Part I: Modelling. In Proceedings of the 2002
feedforward neural networks. In Proceedings of the IEEE international southeastern symposium on system theory (pp. 97–102).
conference on systems, man and cybernetics (pp. 5985–5990). Yang, X., & Meng, M. (1999). An efficient neural network model for path
Somov, Y. I., Polyntsev, O. Y. (2003). Nonlinear dynamics and control of planning of car-like robots in dynamic environment. In Proceedings of
a wind-milling gyroplane rotor. In Proceedings of the international the 1999 IEEE Canadian conference on electrical and computer
conference on physics and control, Vol. 1 (pp. 151–156). engineering (pp. 1374–1379).

You might also like