Beruflich Dokumente
Kultur Dokumente
Abstract—This paper deals with the problem of optimizing a based on genetic algorithms to adjust off-line the input and
fuzzy self-tuning PID controller for robot manipulators. Fuzzy output Membership Functions (MFs) of a fuzzy self-tuning
PID controllers have been developed and applied in many fields PID controller for robot manipulators. The performance of
in the last fifteen years. However, there is no systematic method
to design Membership Functions (MFs) for these controllers. the proposed fuzzy scheme has been verified by means of
We propose a simple method based on Genetic Algorithms simulation tests on a two degree of freedom robot arm. The
(GA) to find optimal input and output MFs of a fuzzy self- results of the proposed optimization method are compared with
tuning PID controller. The stability via Lyapunov theory for the those reported in [5].
closed loop control system is also analyzed and shown that is The remainder of this work is organized as follows: Section
asymptotically stable for a class of gain matrices depending on
the manipulator states. To show the usefulness of the proposed 2 is devoted to the robot dynamics. Section 3 presents the
approach, simulation results using a two degree of freedom robot structure of the linear PID controller with nonlinear gains and
arm are presented. recalls the stability analysis. Fuzzy approach for self-tuning
the PID controller gains, presented in Section 4. Section 5
I. I NTRODUCTION describes the application of the genetic algorithm to adjust the
Most of the present industrial robots are controlled through Self-tuning PID controller. A simulation evaluation to verify
local PID controllers. The PID controller has been shown the theoretical results is presented in Section 6. Finally, in
in practice to be effective for set point control of robot Section 7 we give some conclusions to our work.
manipulators. However, the performance of PID controllers
depends heavily on the operating parameters of the system. II. ROBOT DYNAMICS
Once these parameters change, a significant amount of effort In the absence of friction and other disturbances, the dynamics
is required to manually tune the PID controllers. In order to get of a serial n-link rigid robot can be written as [7]:
high performance it may be necessary to have variable gains
M (q)q̈ + C(q, q̇)q̇ + g(q) = τ (1)
for these controllers [1]. Several practical techniques have been
suggested to choose adequate values for the controller gains where q is the n × 1 vector of joint displacements, q̇ is the
which depend on the robot configuration such as gain schedul- n × 1 vector of joint velocities, τ is the n × 1 vector of
ing, fuzzy control and neural networks approach [2], [3], [4]. applied torques, M (q) is the n×n symmetric positive definite
We use the potential of fuzzy self-tuning schemes in order to manipulator inertia matrix, C(q, q̇) is the n × n matrix of
design a methodology for on-line selection of the proportional, centripetal and Coriolis torques, and g(q) is the n × 1 vector
derivative and integral gains for the PID controller. In addition, of gravitational torques obtained as the gradient of the robot
to guarantee Lyapunov stability for the closed-loop system, ∂U (q )
potential energy U(q), i.e. g(q) = ∂ q
this approach ensures also practical performances beyond the
standard non-fuzzy scheme. The development of the self- III. S TABILITY A NALYSIS OF PID C ONTROLLERS WITH
tuning PID scheme is based in [5]. In recent years fuzzy N ONLINEAR G AINS
logic has been widely used as successful practical approach PID controller is a well known set point control strategy
for designing and implementing control systems. Fuzzy logic for manipulators which ensures asymptotic stability for fixed
controllers have three significant advantages over conventional symmetric positive definite gain matrices. In order to improve
techniques, such as PID control, they are cheaper to develop, the performance, it may be necessary to have variable gains.
they cover a wider range of operating conditions, and they are
more readily customizable in natural language terms. Fuzzy A. PID Control with Nonlinear Gains
PID controllers have been developed and applied in many In this section we introduce a new PID controller whose
fields in the last 30 years. However, there is no systematic main feature is that stability hold even though the parameters
method to design Membership Functions and rules for these depend on the robot state. A generalization of the classical
controllers [6]. In this paper we propose a simple method linear PID controller can be obtained by allowing nonlinear
22
αM (q)]q̇. Suppose that Kvρ (q̃) is chosen such that Kvρ (q̃) −
αM (q) is positive definite, i.e.:
kvρli
Pn > α, i = 1, · · · , n, (11)
j=1 maxq |Mij (q)|
2
V̇ (q̃, q̇, ω) ≤ − [λm {Kvρl } − αkc kq̃k] kq̇k (12)
Ki (ω ) ∂ g (ϑ)
h i
T
−αq̃ Kp (q̃) − α + | q̃
∂ ϑ ϑ=ξ
23
Fig. 3. Block diagram fuzzy self-tuning PID control
24
where q 0d (t) is the response of a first order system to the given
set point and q(t) is the actual output. This error is used to q1
6 [deg]
obtain an IAE parameter which is aimed to minimize .. .... .... .... ....
Z ∞ 90 ....................................................................................
.. ... ... .......
.
... .... ... ... ......
... ... ... ... .......
IAE = |e(t)| dt . (18) 75 ... ... ...
... .... ...
.. ....
. . ..
... ... ..
... .... .. ... ... ... Fuzzy self-tuning PID
o .. ..... .. ... ..
... ....... ... .... ...
... ...... ... .... ...
Another parameter for the evaluation of individuals is 60 ... ...... ... ... ...
... ... ... .... ..
the maximum torque required by the controller. The fitness .. .... ... .... ... Optimal Fuzzy self-tuning PID
...... .. ... ..
. .
............................................................................ ..........................................................................................................................................................................................................................
function penalizes those individuals who exceed the maximum 45 ..
.... ...
.... ...
... ...
torque as follows 30
.. ..
... .
... .
... ..
... ...
.
.. ...
....
...
IAE, if |TM | ≤ TL 15 .....
......
F itness = (19) ...
.. t [sec]
(IAE + |TM |)|TM |3 , if |TM | > TL ..
...
... -
0
where |TM | is the maximum absolute torque used by the 0 2 4 6 8 10
controller in the simulation time. TL is the maximum torque
each actuator can provide. Torque penalization assures that Fig. 4. Desired and actual positions link 1, for the fuzzy self-tuning PID
even if an individual has an error very close to zero, if it and optimal fuzzy self-tuning PID.
exceeds the maximum torque it will anyway have a poor
evaluation.
The minimization problem is turned to a maximization q2
6 [deg]
problem if the fitness is multiplied by −1. Using fitness 90 ........................................................................... ..........................................................................................................................................................................................................................................................................................
...
...... ... ........ .
...... ... ....
.... .. . ... ..... .... .
normalization before the artificial reproduction process allows ...
... ..
.....
......
... .... ....
Optimal Fuzzy self-tuning PID
..... .. ......... .
. . ..
.
.
. . ....... ... . .. ..
... ... ...
to have negative fitness. 75 .
.. ..
... ..
..... ...
.. .. ..
.. .... .
. ... .
.
... .... ...
.
... ..
A population with a fixed number of 50 individuals was ... ..
..
... ... ....
.. .. .. ... .......
... .......
Fuzzy self-tuning PID
.. .. . .. .
... .... ... .....
evolved along 30 generations, the mutation probability was 60 ... ..
.
.. . ... .... ...
.
....
..........
... .. .. ...
.
.
.
.. .
. .
.
. . ... ..... ..
... .. .... .....
... ....................... .... .... .... .......
pm = 0.02. Two-point crossover [14] is used to create new 45
..
.. ..
.. .
...............................................................
.... ...
individuals. Elitism [15] is implemented, after reproduction ..
... ..
.. .
.. ...
.. .
the worst individual is replaced by the best found so far in 30 ... ...
..
......
......
each generation. ........
......
15 ....
..
.....
.....
VI. S IMULATION R ESULTS ...
......
t [sec]
0 .... -
To illustrate the performance of the self-tuning PID Con- 0 2 4 6 8 10
troller optimized by the vgGA with respect to the empirically
adjusted, some experiments were carried out with the robot Fig. 5. Desired and actual positions link 2, for the fuzzy self-tuning PID
arm model used in [5]. The maximum torque provided by the and optimal fuzzy self-tuning PID.
actuators is given by |τ1 | ≤ 200 [Nm] and |τ2 | ≤ 15 [Nm].
To compare its performance in the same conditions the
system is given
the following references:
45[deg] if 0 ≤ t < 2 sec τ1 [Nm]
6
200
qd1 (t) = 90[deg] if 2 ≤ t < 4 sec
....
45[deg] if 4 ≤ t < 10 sec .
..... Fuzzy self-tuning PID
150 ... ...
..... ..
90[deg] if 0 ≤ t < 2 sec ...
... ...
...... .......
.
.
Optimal Fuzzy self-tuning PID
... ... ..... .........
... ..
... .... ..... ..........
qd2 (t) = 45[deg] if 2 ≤ t < 4 sec 100 .... .........
..
.... ..........
........ .........
..........
.
.... ........ ..........
................
.... .........
90[deg] if 4 ≤ t < 10 sec . .. .......
... ...... .......................
............. ...................... ..... .....
.......... ..... . ......
50 ........ ................ ...................................................................................... .... ....
........................... ............................................................... ... ..................... ... ... .. ..
These position references really demand large torques to . .. .... . .. ........................ .. ..
............. ........ ..........
.......... .. ....
........ ...
... ..
.. ...
. . .. ...
...................... ......... ....... .....................................................................................................................................................................
.......... .. ...............................................................................................................................................................................
... . .. .....
.... .................. ...... .
.. . ....... .
reach the amplitude of the requested step. The FLTs for each 0 ... ............ ... ... ... ... ... ... ... ...... ... ... ... ... ... ... ... ... ... ... ... ....................... ... ... ... ... ... ... ... ... ...... ... ... ... ... ... ... ... ... ... ... ... ... ... ... ... ... ... ... ... ... ... ...
....
... ..................
.... .....
....... .......
link were adjusted separately by the vgGA. .
...... ........
..
... ........
... ........
Tables I and II show the fuzzy partitions of the universes of −50 . ..
.... .......
..
... ....
..
discourse for the input and output of each FLT, parameters are
−100
adjusted by trial and error and by the vgGA. These parameters
change significantly in most cases. t [sec]
−150 -
Fig. 4 and 5 illustrate the position of link-1 and link-2. From 0 2 4 6 8 10
those figures we can observe that the transient responses for
both links using the fuzzy self-tuning PID controller adjusted Fig. 6. Applied torque joint 1, for the fuzzy self-tuning PID and optimal
by the vgGA are much better than those of the controller fuzzy self-tuning PID.
adjusted by trial and error.
25
loop system for set-point control. There is no need to know
τ [Nm] the plant parameters, only the desired positions and joint
62
20 displacements for each link.
The stability via Lyapunov theory for the closed loop control
15 ......
..
Fuzzy self-tuning PID system is also analyzed and shown that is asymptotically stable
.
......
.........
...
.. for a class of gain matrices depending on the manipulator
...... ..... Optimal Fuzzy self-tuning PID
10 ........ .... states.
.......... .... ..
.......... ... .........
............ ..... ........
This procedure has been illustrated for a two degrees-of-
........ .. ..
. ......... ...
.................... ........ ............ .
5 .... .............................. ... . ... . ....
.... ..
..................... .................................................. . the freedom robot and tested via computer simulations. From
..... ....................................................................................................................................... ............ .......................... ..... ..........................................................................................................................................................................................
............... ....... ..... .. .............................................................................................................................................................................................
... ................. .... .......................... .. .......... . . . . . .............
..... ............
................. ..........
................ ... ............... ....
the results, it is clear to show that a satisfactory control
............................ .. ........ .
..... ................. ... ... ... ... ... ... ....................................... ... ... ......................................................................................... ... ... ... ... ... ... ... ... ... ...... ... ... ... ... ... ... ... ... ... ... ... ... ... ... ... ... ... ...... ... ... ...
0 ..... ............ ................................................................ ....
....
.....
........... ............... . .
.
... .
............
...
.....
performance can be achieved by using the proposed algorithm.
.... ...... ....
.... ....
−5 ...
..
...
...
...
... ACKNOWLEDGMENTS
..... ....
.... ....
... .
−10
..
.....
.... This research was funded by CONACyT México postdoc-
...
. toral fellowship grant No. 290536, CONACyT SNI Project
t [sec]
−15 - No.89605, ITESM eRobots Research chair and Dirección Gen-
0 2 4 6 8 10 eral de Educación Superior Tecnológica (DGEST) México.
R EFERENCES
Fig. 7. Applied torque joint 2, for the fuzzy self-tuning PID and optimal
fuzzy self-tuning PID. [1] Kelly R., Haber R., Haber R. E. and Reyes F., “Lyapunov stable
control of robot manipulators: A fuzzy self-tuning procedure”, Intelligent
Automation and Soft Computing, vol. 5, No. 4, pp.313-326, 1999.
[2] J. C. Shen,“Fuzzy Neural Networks for Tuning PID Controller for Plants
The torque delivered by the controllers is shown in Fig. with Underdamped Responses”, IEEE Transactions on Fuzzy Systems,
vol. 9, No. 2, April 2001.
6 and 7. The maximum torques required by the controller [3] C. H. Tsai, C. H. Wang, and W. S. Lin,“Robust Fuzzy Model-Following
adjusted by trial and error are higher than those of the adjusted Control of Robot Manipulators”, IEEE Transactions on Fuzzy Systems,
by the vgGA. In all cases the torque remains within the vol. 8, No. 4, August 2000.
[4] H. B. Kazemian, “The SOF-PID Controller for the Control of a MIMO
actuator limits. Robot Arm”, IEEE Transactions on Fuzzy Systems, vol. 10, No. 4, August
2002.
VII. C ONCLUSIONS [5] J. L. Meza, V. Santibañez, R. Soto and M. A. Llama, “Stable Fuzzy
Self-Tuning PID Control of Robot Manipulators”, IEEE International
In this paper we have addressed the optimal fuzzy self- Conference on Systems, Man and Cybernetics, Oct. 2009 (To Appear).
tuning PID control for robot manipulators. We have proposed [6] K. K. Ahn, D. Q. Truong, “Online tuning fuzzy PID controller using
a genetic-algorithm based method for the optimal adjustment robust extended Kalman filter”, Journal of Process Control 19, pp. 1011-
1023, 2009.
of membership functions of a fuzzy self-tuning PID controller [7] M. Spong, S. Hutchinson and M. Vidyasagar, Robot Modeling and
which guarantees asymptotic stability of the overall closed- Control, John Wiley and Sons, 2006.
[8] R. Ortega, A.Lorı́a, P. Nicklasson, and H. Sira–Ramirez, Passivity–based
control of Euler–Lagrange systems Springer–Verlag, 1998.
TABLE I [9] V.M. Hernandez-Guzman, V. Santibañez and R. Silva-Ortigoza, “A new
F UZZY PARTITIONS FOR THE INPUT OF EACH FLT ADJUSTED BY TRIAL Tuning Procedure for PID Control of Rigid Robots”, In Advanced
AND ERROR COMPARED WITH THOSE ADJUSTED BY THE VG GA. T HE Robotics (22), pp. 1007-1023, 2008.
UNITS ARE [ DEG ]. [10] Li-Xi Wang, “A course in fuzzy systems and control”, Prentice Hall,
1997.
[11] K. Valarmathi, D. Devaraj, T.K. Radhakrishnan, “Real-Coded genetic
Empirically adjusted Adjusted by vgGA
algorithm for system identification and controller tuning”, Applied Math-
Pkp (|q̃1 |) = {0, 2, 4, 10, 30, 180} {0, 31, 38, 53, 63, 180} ematical Modelling, pp. 3392-3401, 2008.
Pkp (|q̃2 |) = {0, 2, 4, 10, 30, 180} {0, 18, 23, 32, 46, 180} [12] Devinder Kaur, Sam A. McGilvery, Adel A. Ghandalky, “Genetic Fuzzy
Pkv (|q̃1 |) = {0, 2, 4, 10, 30, 180} {0, 41, 47, 60, 72, 180} Gain Schedule for a Non-Linear Valve”, Fuzzy Information Processing
Pkv (|q̃1 |) = {0, 2, 4, 10, 30, 180} {0, 13, 27, 37, 45, 180} Society, NAFIPS, pp. 96-101, 2006.
Pki (|ω1 |) = {0, 1, 6, 14, 29, 86} {0, 13, 32, 61, 78, 180} [13] Manuel Valenzuela, “The Virtual Gene Genetic Algorithm”, LNCS,
Pki (|ω2 |) = {0, 6, 29, 57, 115, 286} {0, 2, 11, 39, 62, 180} Springer-Berlag Berlin Heidelberg, pp. 1457-1468, 2003.
[14] William M. Spears, Kenneth A. De Jong, “An Analysis of Multi-Point
TABLE II Crossover”, In Rawlins, “Foundations of Genetic Algorithms”, pp. 301-
F UZZY PARTITIONS FOR THE OUTPUT OF EACH FLT ADJUSTED BY TRIAL 315, 1991.
AND ERROR COMPARED WITH THOSE ADJUSTED BY THE VG GA. T HE
[15] D. E. Goldberg, “Genetic Algorithms in Search, Optimization and
UNITS ARE [N M / DEG ].
Machine Learning”, Reading, MA: Addison Wesley, 1989.
26