Sunteți pe pagina 1din 5

ICROS-SICE International Joint Conference 2009

August 18-21, 2009, Fukuoka International Congress Center, Japan

Feedback Controller with Nonlinear Compensator Optimized by


Genetic Algorithm for Rotary Crane System
Hiroshi Kinjo1 , Kunihiko Nakazono2 , Kohei Ohnishi3 and Tetsuhiko Yamamoto4
1
Faculty of Engineering, University of the Ryukyus, Okinawa, Japan
(Tel: +81-98-895-8632; E-mail: kinjo@tec.u-ryukyu.ac.jp)
2
Faculty of Engineering, University of the Ryukyus, Okinawa, Japan
(Tel: +81-98-895-8633; E-mail: nakazono@tec.u-ryukyu.ac.jp)
3
Faculty of Science and Technology, Keio University, Yokohama, Japan
(E-mail: ohnishi@sd.keio.ac.jp)
4
Tokushima College of Technology, Tokushima, Japan
(E-mail: t.yamamoto@tokuco.ac.jp)

Abstract: The rotary crane is known to be a class of nonholonomic systems which is hard to control by classical
methods. The crane system has inherently nonlinear characteristics. In this paper, we propose a feedback controller with
an embeded nonlinear compensator for the rotary crane system. For the nonlinear compensation, we use a hyperbolic
fuction and some coefficients. Determinations of the feedback gains and the compensator coefficients are difficult. In this
paper, we apply GA optimization to determine the sets of parameters of the controller.

Keywords: Nonlinear control, Optimization, Intelligent systems, Automation technology, Industrial automation systems

1. INTRODUCTION parameters.
In this paper, we propose a feedback controller with
A rotary crane system is used to transport a load mass an embeded nonlinear compensator for the rotary crane
to a desired position through rotating, raising, and low- system. For the nonlinear compensation, we use a hyper-
ering the boom. This motion is accompanied by oscilla- bolic function and some coefficients. Determinations of
tion of the load, or load swing. In order to suppress such the feedback gains and the compensator coefficients are
load swing, rotary crane systems are commonly operated difficult. Thus, we apply GA optimization to determine
using both the rotation angle and the lean angle. Opera- the sets of parameters of the controller. For compari-
tion of the rotary crane system solely by rotation around son between our method and neurocontrol system, our
the vertical axis results in a nonholonomic system that method has less searching variables for the GA.
makes the control problem complex and it is necessary to
design time-varying feedback or discontinuous feedback 2. ROTARY CRANE SYSTEM
controllers [1].
Many vibration control methods of load swing in a ro-
tary crane system have been proposed [2-6]. An exam- z
ple is a control method in which the control law follows
the most suitable trajectory along which the load swing Jib
will be smallest [2]. The theory of the optimal regulator l
for linear systems, fuzzy reasoning, and the feedback law
h
for nonlinear systems are examples of such control meth-
ods. These control methods can fundamentally stabilize m
the load swing in the circumferential direction, but can-
not suppress swing in the radial direction. Some vibration y
control methods for the crane system with only control-
lable rotation operation have been reported. Anti-sway O θ
control method of the crane system based on a skillful op- r
erator’s knowledge has been introduced [3]. Load swing
x
suppression based on a linear feedback law by switching
the two modes in the radial and circumferential directions Fig. 1 Crane system.
has been presented [4,5]. However, all of these control
methods require knowledge of difficult control theories. Figure 1 shows a rotary crane system as an example
In contrast, a neuro-controller trained by an evolution- of nonholonomic control object. Here, x, y, z are the
ary computation technique such as a genetic algorithm coordinates of the load mass position, θ is the rotation
[7-9] is one of the simpler way to realize than conven- angle (control input), r is the radius of rotation, h is the
tional control systems [6]. However, the neuro controller height from the tip of the boom, m is the mass of the load,
based on evolutionary computation has many searching and � is the length of wire which is not elastic or slack.

- 1817 - PR0002/09/0000-1817 ¥400 © 2009 SICE


GA 4. CONTROLLER DESIGN
Xr + θ X 4.1 Design procedure
K Crane system

GA k1 ... kn c1 ... cn

Nonlinear Fig. 3 Genetic code for GA.


compensator
Figure 3 shows the genetic code that constructs param-
Fig. 2 Nonlinear control system with GA for crane eters of the feedback gain and nonlinear compensator.
system. The set of optimum values of the control parameters is
determined by the real coded GA [10]. The algorithm of
evolution process for the controller is as follows:
The control purpose is to suppress load swing from the STEP 1: Create initial parameters randomly. The initial
arbitrary position to the reference position by control of values are set in the range of [−0.3, 0.3] for parameter ki
crane rotation alone. and [0.0, 1.0] for ci .
Using a Lagrange equation for the constraint system, STEP 2: Calculate an error function E while demon-
the dynamic model of the rotary crane system can approx- strating control simulation for all controllers.
imately be described in terms of the following equations: STEP 3: Select the parent controllers by roulette wheel
⎡ ⎤ ⎡ ⎤⎡ ⎤ parent selection.
x 0 1 0 0 x
d ⎢ ⎥ ⎢ 0 ⎥ ⎢ ⎥ STEP 4: Perform a crossover operation of BLXα method
⎢ ẋ ⎥ = ⎢ −λ 0 0
2
⎥ ⎢ ẋ ⎥(1) to produce new controllers.
dt y⎣ ⎦ ⎣ 0 0 0 1 ⎦ ⎣ y ⎦
ẏ 0 0 −λ 0 2
ẏ STEP 5: Alternate the controllers including the new con-
⎡ ⎤ trollers to the next generation. Iterate from STEP 2 until
0 the evolution process reaches generation 100.
⎢ λ2 r cos θ ⎥
+⎢ ⎥ Further information regarding the parameters of the GA
⎣ 0 ⎦
is provided in Table 2.
λ r sin θ
2
� 4.2 Error function
where λ = g/l is natural frequency; θ is the control
input. Because if the vibration of the load swing is suffi- An error function E is used to evaluate the perfor-
ciently small, only the x–y plane needs to be considered. mance of each controller during the GA-based training
It is difficult for a designer to control the rotation an- process. The error function E is described by
gle of the rotary crane system by conventional methods E = (xr − xend )2 + (y r − y end )2 (5)
because the dynamic model is a type of nonholonomic
+(ẋr − ẋend )2 + (ẏ r − ẏ end )2
system. As in a nonholonomic system, the vibration con-
trol method using a static continuous state feedback can- +(θr − θend )2
not stabilize the load swing. It is necessary to design a where (·)r are the desired values of the control variables
time-varying feedback controller or a discontinuous feed- and control input. (·)end denotes the final values after
back controller. The controller is instead designed using performing the control simulation in the GA process. The
a nonlinear feedback with GA-based optimization in this error function is defined as a sum of squared errors of the
paper. desired values and final values of the variables and con-
trol input. The fitness of the controller is expressed in the
3. CONTROL SYSTEM term of the inverse of the error function E. In the GA
For the control object, we built a nonlinear feedback evolution, the parameters of the beedback controller are
control system as shown in Fig. 2. Control input is de- modified in order to maximize the fitness function deter-
scribed as follows: mined by the error function in Eq. (5).
θ(t) = K(X r − f (X(t))) (2)
5. NUMERICAL SIMULATIONS
where K = [k1 , k2 , · · · , kn ] is feedback gain vector;
X(t) = [x1 (t), x2 (t), · · · , xn (t)]T and X r are the state 5.1 Experimental parameters
vector and the desired value vector; f (·) is a nonlinear The validity of the proposed controller was veri-
vector function that uses a hyperbolic function as: fied through computer simulations using Runge-Kutta
f (X(t)) = [tanh(c1 x1 (t)), tanh(c2 x2 (t)), · · · , (3) method. Sampling time is 10 ms. The parameters of the
rotary crane system and the initial conditions are listed in
tanh(cn xn (t))]T ,
Table 1. In the GA simulation, the real coded GA param-
C = [c1 , c2 . · · · , cn ] (4) eters were set as shown in Table 2.
where ci , {i = 1, · · · , n} are coefficients. The goal of the evolution progressed by the GA is
In this research the gain vector K and coefficient vec- to obtain a feedback controller that suppresses the load
tor C are determined by GA. swing of the rotary crane system upon movement from

- 1818 -
Table 1 Parameters of rotary crane system and initial Figure 9 shows the evolutionary result for the case that
conditions. the initial angle is θ0 = π rad. Table 4 shows the values of
Load mass m 1.0 [kg] the optimized parameters by GA. Figures 10 to 13 show
the control resuts using the resulting controller for initial
Rotation radius r 1.0 [m]
angle θ0 = π rad. It can be seen that the proposed method
Length of wire � 2.0 [m] has good control performance.
Acceleration of gravity g 9.8 [m/s2 ]
Table 4 Optimized parameters (Initial θ0 = π rad).
Initial rotation angle θ0 π/2 or π [rad]
Reference position (xr , y r ) (1, 0) [m] i i=1 i=2 i=3 i=4
ki 0.5880 −1.1699 −1.1919 0.4129
Table 2 GA parameters. ci 10.5619 0.2830 0.3113 1.8956
No. initial population 50 individuals
No. children 30 individuals 6. CONCLUSION
Selection Roulette wheel
A simple GA-based approach was applied to realize a
Crossover BLXα nonlinear feedback controller for load swing suppression
α 1.0 in rotary crane system involving only rotation about the
vertical axis. Simulations have confirmed that the real
Final generation 100
coded GA scheme is effective for searching the controller
Final time of the control simulation T 10.0 [s] parameters for the rotary crane system and the proposed
controller can control the system effectively.

the initial position (x0 , y0 ) to the reference position REFERENCES


(xr , y r ). [1] M. Sampei, “Feedback control of nonholonomic sys-
5.2 Results tem”, Journal of the SICE, Vol. 36, No. 6, pp. 396–
403, 1997 (in Japanese).
Figure 4 shows the evolutionary result for the case that [2] S. Yamazaki, T. Itoh, and T. Hisamura, “Theoretical
the initial angle is θ0 = π/2 rad. It can be seen that the Consideration on a Control of Jib-Crane”, Transac-
GA is able to search the well suitable values of the con- tion of the SICE, Vol. 15, No. 6, pp. 826–832, 1982.
trol parameters. Table 3 shows values of the optimized [3] M. Ohto and S. Yasunobu, “A Rotary Crane Control
parameters by GA. Figures 5 to 8 show the control resuts Method Based on a Skillful Operator’s Knowledge”,
using the resulting controller. Figure 5 shows the control Transaction of the SICE, Vol. 33, No. 9, pp. 923–929,
trajectory, where the solid line is for the load and the dot- 1997. (in Japanese)
ted line is for the jib of the crane. Figures 6 and 7 are the [4] R. Kondo and S. Shimahara, “Anti-Sway Control of
values of the load positions x, y and the velocities. Figure a Rotary Crane via Two-Mode Swiching Control”,
8 shows the control input. It can be seen that the control Transaction of the SICE, Vol. 41, No. 4, pp. 307–313,
command curve θ(t) is smooth. We can observe that the 2005. (in Japanese)
proposed control system has good control performance. [5] R. Kondo, “Vibration Control of Rotary Crane via
Two-Mode Swiching Control”, Journal of the SICE,
Vol. 45, No. 7, pp. 632–637, 2006. (in Japanese)
1
[6] K. Nakazono, K. Ohnishi, H. Kinjo, T. Yamamoto,
10−1 “Vibration Control of Load for Rotary Crane System
Using Neural Network with GA-based Training”, Ar-
10−2 tificial Life and Robotics, Vol. 13, pp. 98–101, 2008.
E

10−3 [7] L. Davis, Handbook of genetic algorithms, Van Nos-


trand Reinhold, New York, 1991.
10−4 [8] D. B. Forgel, Evolutionary Computation, IEEE
10−5
Press, 1995.
0 20 40 60 80 100 [9] D. E. Goldberg, Genetic Algorithms in Search, Op-
Generation timization & Machine Learning, Addison Wesley,
Fig. 4 Evolutionary result (Initial θ0 = π/2 rad). 1989.
[10] L. J. Eshelman, D. J. Schaffer, Real-Coded Genetic
Table 3 Optimized parameters (Initial θ0 = π/2 rad). Algorithms and Interval-Schemata, Foundations of
i i=1 i=2 i=3 i=4 Genetic Algorithms 2, Morgan Kaufmann Publishers,
ki −0.0981 −2.225 −0.3261 0.1746 pp. 187-202, 1993.
ci 3.3305 0.8419 0.1323 11.2624

- 1819 -
2

1 Start
y [m]

0
End

-1

-2
-2 -1 0 1 2
x [m]
Fig. 5 Control result (Trajectory) (Initial θ0 = π/2 rad).

2
x
1.5 y
1 1
0.5
x,y [m]

0 10−1
-0.5
-1 10−2
E

-1.5
10−3
-2
0 2 4 6 8 10
10−4
t [s]
Fig. 6 Control result (Positions) (Initial θ0 = π/2 rad). 10−5
0 20 40 60 80 100
. Generation
2 x.
1.5 y Fig. 9 Evolutionary result (Initial θ0 = π rad).
1
x,y [m/s]

0.5
0
-0.5
. .

-1
-1.5
-2

0 2 4 6 8 10
t [s]
Fig. 7 Control result (Velocities) (Initial θ0 = π/2 rad).

2
θ [rad]

-2

-4
0 2 4 6 8 10
t [s]
Fig. 8 Control result (Control input) (Initial θ0 = π/2
rad).

- 1820 -
2

End
y [m]

0
Start

-1

-2
-2 -1 0 1 2
x [m]
Fig. 10 Control result (Trajectory) (Initial θ0 = π/2
rad).

2
x
1.5 y
1
0.5
x,y [m]

0
-0.5
-1
-1.5
-2
0 2 4 6 8 10
t [s]
Fig. 11 Control result (Positions) (Initial θ0 = π rad).

.
2 x.
1.5 y
1
x,y [m/s]

0.5
0
-0.5
. .

-1
-1.5
-2

0 2 4 6 8 10
t [s]
Fig. 12 Control result (Velocities) (Initial θ0 = π rad).

2
θ [rad]

-2

-4
0 2 4 6 8 10
t [s]
Fig. 13 Control result (Control input) (Initial θ0 = π
rad).

- 1821 -

S-ar putea să vă placă și