Professional Documents
Culture Documents
ARTICLE
DOI: 10.5772/63941
© 2016 Author(s). Licensee InTech. This is an open access article distributed under the terms of the Creative Commons Attribution License
(http://creativecommons.org/licenses/by/3.0), which permits unrestricted use, distribution, and reproduction in any medium, provided the
original work is properly cited.
Abstract 1. Introduction
In this work, we propose a new method for the optimal Fuzzy logic controllers have been commonly adopted in
design and tuning of a Proportional-Integral-Derivative many areas of engineering over the last few decades [1-4].
type (PID-type) interval type-2 fuzzy logic controller (IT2 This can be attributed to their linguistic-based structure,
FLC) for Delta parallel robot trajectory tracking control. which does not need a precise mathematical model of the
The presented methodology starts with an optimal design object and can handle the uncertainty of systems’ informa‐
problem of IT2 FLC. A group of IT2 FLCs are obtained by tion [5-8]. Some researchers [9, 10] have proved that FLCs
blurring the membership functions using a variable called are more robust and their performance is less sensitive to
blurring degree. By comparing the performance of the parametric variations than conventional controllers. H
controllers, the optimal structure of IT2 FLC is obtained. Ying et al. proved that FLC is a nonlinear, variable param‐
Then, a multi-objective optimization problem is formulated eter controller, which can effectively control nonlinear,
to tune the scaling factors of the PID-type IT2 FLC. The strong coupling with time-varying structure systems [11].
Non-dominated Sorting Genetic Algorithm (NSGA-II) is Due to the closed kinematic structure, Delta robots present
adopted to solve the constrained nonlinear multi-objective better performance in accuracy, rigidity and payload
optimization problem. Simulation results of the optimized capacity over their serial counterparts. All these advantag‐
controller are presented and discussed regarding applica‐ es make this robot a good platform in many areas of
tion in the Delta parallel robot. The proposed method engineering [12-15]. However, the Delta robot system is a
provides an effective way to design and tune the PID-type kind of nonlinear, strong coupling system with time-
IT2 FLC with a desired control performance. varying parameters, and traditional controllers cannot
provide a satisfactory control performance. So FLC is a
Keywords PID-type IT2 FLC, Blurring Degree, Scaling good candidate to control such a robot effectively, some‐
Factors, Multi-objective Optimization, NSGA-II thing that is much needed in engineering applications.
3
t = Aq&& + Tcc q& + Tg + åt M ,i (3)
Figure 2. Scheme of the Delta robot i =1
Based on the virtual work principle, the rigid-body where A is the inertia matrix:
dynamic model of the Delta robot can be expressed as:
A = I L + J T mn J (4)
3 3
t mn + åt L ,i + åt M ,i = 0 (1)
i =1 i =1 T cc is the Coriolis/centripetal matrix:
where τmn is the generalized force of the travelling platform, Tcc = J T mn J& (5)
τL ,i is the generalized force of the actuating arm i, and τM ,i
is the generalized force of the passive arm i. T g is the gravity vector:
3
&& - G - J TG + t In practice, the Coulomb friction and viscous friction of the
t = I Lq&& + J T mn X n L n å M ,i
i =1
(2)
robot joints cannot be neglected. Most of the frictions of the
% =
A ò ò
xÎX uÎ J x
1 / ( x , u) (10)
, Ku
e Ku
U
Ku
e
, Ku
e E
Kp Ki
U
Ku
e E
Kd Ka
optimization procedure
Figure 7. Design and
where * denotes a general t-norm, the product t-norm is where the switching points L and R can be calculated using
adopted in this paper. the Karnik-Mendel (KM) algorithms [38].
3. Perform type-reduction procedure to get the type- 4. The final crisp output can be computed as:
reduced set. In this paper, the centre-of-sets type-
reducer is considered. The type-reduced set can be yl + yr
computed as: y= (15)
2
N
3.3 The structure design of IT2 FLC for the delta parallel robot
åf n
yn
Ycos ( x ) = U n =1
N
= éë yl , yr ùû (12) For most fuzzy controllers, the error and its time derivative
f n ÎF n ( x )
y n ÎBn åf n
are usually chosen as the inputs of the controllers. Howev‐
n =1 er, it is difficult for the fuzzy PD type controller to remove
the steady-state error. For the purpose of improving the
performance of the IT2 FLC to handle steady-state error
The centroid of the resulting IT2 output fuzzy set is an
and transient response at the same time, the PID-type IT2
interval T1 fuzzy set, which can be described by its left and
FLC is proposed in this work. A schematic view of the PID-
right end points yl and yr :
type fuzzy logic control system connected to the robot
platform is presented in Figure 6.
å f yln + å n = L + 1 f n yln
L N
n
In this paper, the inputs of the controller were designed
n =1
yl = (13) using two trapezoidal interval type-2 fuzzy sets for
å f n + å n= L +1 f n
L N
n =1 describing the input signal error e(t) and its time derivative
ė(t). The error was calculated as the difference between the
actual and desired angle of the joint. The output of the
å f yrn + å n = L + 1 f n yrn
R n N
n =1
controller u(t) was modelled using four triangular interval
yr = (14)
type-2 fuzzy sets {Y 1, Y 2, Y 3, Y 4}. The fuzzy rule base of
å f n + å n= L +1 f n
L N
n =1
the controller is:
1 1
E E 2
1 1
0
0 0.2 0.4 0.6 0.8 1
1 1
Figure 9. The RMSE values of the three joints with variable blurring degrees
E E
E 1 E1
1 1
1 1
E 1
E1
E 1 E1 E 1 E1
0.5 0.5
1 1 0
-1.5 -1 -0.5 0 0.5 1 1.5
0
-1.5 -1 -0.5 0 0.5 1 1.5
Y1 Y2 Y3 Y4
antecedent membership functions for (a) input
Figure 8. Illustrations of the
0.5
to zero and K a, K p , K d are set to one during the first tuning Figure 10. The membership functions: (a) error input sets (b) error derivative
input sets (c) output sets
procedure. By changing the value of α we can get a group
of IT2 FLCs with different antecedent membership func‐ Figure 11 depicts the output control surfaces of different
tions. When the membership functions are blurred, the blurring degrees. Due to space limitations, only the most
controller could not provide enough output control signal representative three control surfaces are selected. It can be
when adopting the same inputs [40]. So the value K u is also observed that when α = 0.5, IT2 FLC offers substantially
tuned in this step to overcome this issue. The value of K u is smoother control performance than other values.
Output
Output
Output
out
out
out
0 0 0
-1 -1 -1
2 2 2
2 2 2
0 1 0 1 0 1
0 0 0
e e e
-1 -1 -1
Input2 -2 -2
e
Input1 Input2 -2 -2
e
Input1 Input2 -2 -2
e
Input1
control surface when: (a) α=0, (b) α=0.5 and (c) α=1
Figure 11. The output
4. Multi-objective Optimization of the Scaling Factors of sorting genetic algorithm with an elitist strategy, especially
PID-Type IT2 FLC for multi-objective optimization. Figure 12 shows the
process of NSGA-II, and the algorithm used in this work
4.1 Multi-objective optimization and NSGA-II can be stated as:
The process of optimizing a mathematical problem ex‐ 1. Generate a uniformly distributed parent population P0
pressed systematically and simultaneously involving a with the population size N .
collection of objective functions is called multi-objective
2. Using crossover, mutation and selection operations to
optimization. It is a kind of multiple-criteria decision
create an offspring population Q0 with the population
making and usually has a set of optimal solutions. Mathe‐
matically speaking, a multi-objective optimization problem size N .
consists of optimizing a vector of functions: 3. Combine the offspring and parent population to form
extended population Rn with the population size of
Opt( F( x )) = ( f1 ( x ), f2 ( x ),K , f k ( x )) 2N .
subject to: gi ( x ) £ 0, i = 1,2,K , q , (17) 4. Sort the extended population based on non-domina‐
h j ( x ) = 0, j = 1,2,K , r. tion.
Definition 3. Pareto set: A set of non-dominated feasible 4.2 Multi-objective optimization of the scaling factors
solutions is said to be a Pareto set.
The goal of multi-objective optimization is to determine the
Definition 4. Pareto front: The image of a Pareto set in the
scaling factors of the PID-type IT2 FLC to achieve a
objective space is called a Pareto front.
desirable control performance. As depicted in Figure 7,
In this paper, we chose a Non-dominated Sorting Genetic there are four scaling factors to be tuned at the second step,
Algorithm II (NSGA-II) to find the Pareto solutions for so the decision variables of the multi-objective problem can
multi-objective optimization. It is a fast non-dominated be defined as:
where e1, e2 and e3 are the errors of the three joints. The multi-objective optimization was carried out in a
desired trajectory given in Eq.(16). NSGA-II was imple‐
We need f 1 to be minimized for it represents the robot mented using MATLAB with the population=200, and
having the best tracking accuracy. If the mean values of the generation=200. Figure 13 shows the Pareto front of the
torques provided by the motors are minimized during the multi-objective scaling factors’ optimization results
trajectory tracking operation, the energy consumption of obtained by NSGA-II. Table 1 shows the Pareto optimal
the robot will be reduced, which is very necessary in solutions of the optimization. It is noticed that there is a
engineering applications. So the second objective function trade-off between tracking accuracy and energy consump‐
is given by: tion. In other words, if we want to have a smaller tracking
error, the mean values of the torques will be larger, which
means the robot will use more energy to do the same work
f2 =
ò
W
T1 dW
+
ò W
T2 dW
+
ò W
T3 dW
(20)
and vice versa.
ò
W
dW ò W
dW ò W
dW
12
ì10 £ K p £ 70 11.8
ï
ï0 £ K d £ 5
í (21) 11.75
ï0 £ K i £ 5
ï0 £ K £ 5
î a 11.7
11.65
To improve the computational efficiency, the objective 0 0.5 1 1.5 2 2.5 3 3.5 4 4.5 5
f1
functions should follow the following constrains:
Figure 13. The Pareto front obtained by NSGA-II
Torque(N)
0
0
-5
-0.005 -10
-15
-0.01
-20
-0.015 -25
0 1 2 3 4 0 0.5 1 1.5 2 2.5 3 3.5 4
Time(s) Time(s)
Figure 14. Response of the robot with PID type IT2 FLC with minimum f1
applied to the controller, the mean value of the torque is 2 10.0014 0.0807 0.0160 0.5000 4.5401 11.6663
smaller, which means less energy is needed when the robot
3 10.2366 0.2422 0.3244 0.7703 0.2688 11.9556
finishes the same operation. However, the tracking accu‐
racy will decrease. It also can be observed from Figure 13-15 4 10.0148 0.2008 0.1914 0.7876 0.4305 11.8495
that, in the Pareto solutions, torque changes less than the 5 10.0125 0.2038 0.1791 0.7817 0.4541 11.8421
errors ( f 1 is in the range of 0.2673 to 4.5401 and f 2 is in the
…
range of 11.6663 to 11.9733), which means if we want to
reduce energy consumption by reducing the mean torque 196 10.0032 0.3267 0.1488 0.5952 0.6949 11.7859
of the motors in a given trajectory, the tracking accuracy 197 10.0057 0.3120 0.1403 0.5919 0.7515 11.7760
will decrease more quickly. Once the trajectory of the Delta
198 10.5219 0.3095 0.1613 0.5951 0.5998 11.8073
robot is given, it is difficult to optimize energy by tuning
the controller’s scaling factors. 199 10.0301 0.2357 0.2762 0.7719 0.2988 11.9180
An optimized PID controller used to control the Delta robot 200 10.0048 0.3140 0.1334 0.5659 0.8851 11.7569
[42] is adopted to compare the trajectory tracking control Table 1. Results of the multi-objective optimization
performance with PID-type IT2 FLC proposed in this
paper. Here, the Particle Swarm Optimization (PSO)
algorithm is used to tune the parameters of the PID Results Kp Kd Ki Ka f1 f2
controller. The position of each particle is represented by Min f 1 10.2423 0.2419 0.3440 0.7715 0.2673 11.9733
the proportional, integral and derivative gains of the PID
Min f 2 10.0014 0.0807 0.0160 0.5000 4.5401 11.6663
controller. The fitness function of the PSO algorithm is
chosen the same as in Eq.(19). The response of the Delta Table 2. Results of minimum f1 and f2
robot using an optimized PID controller is presented in
Figure 16. From Figure 14-16, we can see that the PID-type We can also objectively choose other values in the Pareto
IT2 FLC has much higher control precision than the solutions according to the design requirements. For
optimized PID controller, but the control torque is slightly example, if we want to get a small trajectory tracking error
larger than its PID counterpart (whether use the scaling but not too much average torque during the operation, the
factors with minimum f 1 or minimum f 2). Compared with solution with a relatively small f 1 is selected: K p = 10.0184,
the regularly used PID controller, the PID-type IT2 FLC K d = 0.3242, K i = 0.1838, K a = 0.6394. The simulation results are
designed in this work can provide a better control perform‐ shown in Figure 17, we can see that the errors and torques
ance in Delta robot trajectory tracking control. This obser‐ create some trade-offs with each other, so we get a control‐
vation can be explained by the characteristics of the IT2 ler with different control behaviour: an acceptable accuracy
FLC. Analysis has shown that the IT2 FLC can be seen as a and not too much torque. However, we cannot say that one
PI controller with variable gains [43], which can effectively solution is better than the others, because each solution is
control the nonlinear and strong coupling systems, such as a trade-off between the two objective functions. The final
the Delta robot, compared with the regular optimized PID choice of solutions can be made according to specific
controller. requirements of engineering applications.
Torque(N)
0
-0.01
-2
-0.015
-0.02 -4
Joint1
-0.025 Joint2 -6
Joint3
-0.03 -8
0 1 2 3 4 0 0.5 1 1.5 2 2.5 3 3.5 4
Time(s) Time(s)
Figure 15. Response of the robot with PID-type IT2 FLC with minimum f2
0.005 4
Joint1 1
0 Joint2 2 2
Joint3 3
-0.005 0
Error(rad)
Torque(N)
-0.01 -2
-0.015 -4
-0.02 -6
-0.025 -8
0 1 2 3 4 0 0.5 1 1.5 2 2.5 3 3.5 4
Time(s) Time(s)
Torque(N)
-5
-0.005
-10
-0.01
-15
-0.015 -20
0 1 2 3 4 0 0.5 1 1.5 2 2.5 3 3.5 4
Time(s) Time(s)