Professional Documents
Culture Documents
Chattering-Free Neuro-Sliding
Mode Control of 2-DOF Planar
Parallel Manipulators
Regular Paper
DOI: 10.5772/55102
© 2013 Le et al.; 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 This paper proposes a novel chattering free Keywords Parallel Manipulators, Neural Network,
neuro‐sliding mode controller for the trajectory Sliding Mode Control, Chattering‐Free, Uncertainty
tracking control of two degrees of freedom (DOF) Compensator
parallel manipulators which have a complicated
dynamic model, including modelling uncertainties,
frictional uncertainties and external disturbances. A 1. Introduction
feedforward neural network (NN) is combined with an
Parallel manipulators are closed‐loop kinematic chain
error estimator to completely compensate the large
mechanisms which have such advantages as high
nonlinear uncertainties and external disturbances of
accuracy, high stiffness, high payload capability and low
the parallel manipulators. The online weight tuning
moving inertia, etc. They are widely used in numerous
algorithms of the NN and the structure of the error applications, such as humanoid robots, simulators,
estimator are derived with the strict theoretical medical robots and micro‐robots, and they are becoming
stability proof of the Lyapunov theorem. The upper increasingly popular in the machine‐tool industry [1].
bound of uncertainties and the upper bound of the Compared to serial manipulators, the dynamic model of
approximation errors are not required to be known in parallel manipulators is significantly complicated by the
advance in order to guarantee the stability of the presence of multiple closed‐loop chains and singularities.
closed‐loop system. The example simulation results As a result, the control of parallel manipulators needs
show the effectiveness of the proposed control strategy more advanced control techniques than that of serial
for the tracking control of a 2‐DOF parallel manipulators.
manipulator. It results in its being chattering‐free, very
small tracking errors and its robustness against The motion control of parallel manipulators has attracted
uncertainties and external disturbances. many researchers in studying its potential performance.
As the parallel manipulator operates on a horizontal in which = (a,p)T 4 is the vector of the joint angles;
planar plane, the Lagrangian functions only contain the t = (a,p)T 4 is the joint vector; and Ft = (Fa,Fp)T 4 is
kinetic energy of the mechanism: the friction torque vector of the equivalent tree‐structure
open chain system. a = (a1,a2)T and p = (p1,p2)T = (0,0)T
1 1 are the input torque vectors of the active joints and
Li m ( x 2 y i21 ) I zi1ai2
2 i1 i1 2 passive joints respectively. Fa = (fa1,fa2)T and Fp = (fp1,fp2)T =
(0,0)T are the friction force vectors of the active joints and
1 1 passive joints respectively. Here, we assume that the
mi 2 ( x i22 y i22 ) I zi 2pi2 (2)
2 2 effect of the friction force on the passive joints is much
smaller than that on the active joints. Thus, in order to
where mi1 and mi2 are the masses of the links of the serial simplify the dynamic model, only the disturbances on the
chain i; Izi1, Izi2 are the inertia tensors of the links of the active joints are considered.
serial chain i (i=1,2).
The inertia matrix Mt 44 and the Coriolis and
Letting ri1 and ri2 be the distance from the joints to the centrifugal force matrix Ct 44 in (7) are expressed by:
centre of mass for each link, we have:
1 0 1c ap1 0
x i1 ri1 sin aiai ,
0 2 0 2 c ap 2
Mt , (8)
y i1 ri1 cos aiai , 1c ap1 0 1 0
0 2 c ap 2 0 2
x i 2 li1 sin aiai ri 2 sin pipi ,
0 0 1sap1p1 0
y i 2 li1 cos aiai ri 2 cos pipi . 0 0 0 2 sap 2p 2
Ct
(9)
1sap1 a1 0 0 0
By substituting the above equations into (2), the 0 2 sap 2a 2 0 0
Lagrangian function becomes:
τ a ΨT τt (10) C θ F τ
M aθ (17)
a a a a a
Property 1: M̂ a is symmetric and positive definite. 3. The feedforward neural network architecture
The input layer: The input vector of the NN is denoted by: s e Λe θ a (θda Λe ) θ a θ ar (28)
V v1 , v2 ,..., vNh
Ni N h
, 1 d 2
s i si , i=1,2 (29)
2 dt i
T
vi vi1 , vi 2 ,..., viN Ni , i 1, N h (21)
i where i is a strictly positive constant. Equation (29)
indicates that the energy of s should decay so long as s is
The inputs and outputs of the hidden layer are, not zero.
respectively, presented as:
The transfer function in the hidden layer is used as a where the first term eq 2 is the equivalent control which
sigmoid function: keeps the trajectory of the system state on the sliding
surface; and the second term sw 2 is the discontinuous
1 exp( z) control which drives the system states toward the sliding
g( z ) (24)
1 exp( z) surface when they are deviating from this surface.
The output layer: The weight matrix connecting the hidden The equivalent control is considered for the nominal
and output layers is expressed by: system in the absence of the uncertainties and external
disturbances:
T
W w1 , w2 ,..., wN N h 2 ,
h
ˆ θ
τ eq M ˆ (31)
a ar C aθar
wi wi1 , wi 2 , i 1, N h2
(25)
The discontinuous control is designed as:
The outputs of the NN are expressed by:
τ sw Ksign(s) (32)
Nh
yk w G, k 1,2 (26)
i 1 ik i where K = diag(k1,k2), k1 and k2 are positive constants.
τ NN
s
θda , θ da τa θa , θ a
s
s e Λe T
fest
s τ eq
The proposed controller is expressed by the following V*); N 2 is the bounded reconstruction error due to the
equation: inadequate number of neurons in the hidden layer of the
NN.
ˆ θ
τa M ˆ
a ar Caθar fNN fest Ts (35)
For convenience, equation (36) is rewritten as:
where fNN 2 is the output of the NN whose structure is
described in section 3 for the online approximation of the W
τ a fNN W*T G ˆ ε
TG (37)
N
unknown vector a 2. Since the output of the NN is
not able to approximate a accurately, the error and W
where the symbols G * , Ĝ , G are, respectively,
estimator fest 2 in (56) is used to attenuate the * *
defined as G G( x , V ) Nh, G ˆ G( x , Vˆ ) Nh,
approximation errors. The term Ts is used in enhancing ˆ Nh and W
G* G W* Wˆ Nh2.
G
the robustness of the control system. Moreover, T =
diag(T1, T2) is a diagonal positive definite matrix in which for a given x can be
The Taylor series expansion of G
T1 and T2 are positive constants.
written as follows:
The inputs of the NN are chosen as the errors and
T G G VT
derivative of errors: x e1 , e1 , e2 , e 2 4 . The NN is 1 1
VT
G2 G
used to approximate the unknown uncertainties a
online. The approximation error is denoted as:
G
2
ˆ Ο( V
V* V ˆ T x)
G
Nh G VT
ˆ T G( x,V
τ a fNN W*T G( x,V* ) W ˆ )ε (36)
ˆ
|V V Nh |V Vˆ
N
T 1
G GN V 1 s T M ˆ s sT M
ˆ s sT M ˆ s (46)
GV 1 , G2 , , N h Nh a a a
h
2
T T
V x V x VT x |V Vˆ
Properties 1 and 2 in section 2 give us:
ˆ
V V
V * 4 N h
ˆ s sT M
s T M ˆ s (47)
a a
ˆ T x) Nh is a vector of higher‐order terms and it is
Ο( V
assumed to be bounded. ˆ s 2sT C
sT M ˆ s (48)
a a
By substituting (38) into the approximation error By substituting (47) and (48) into (46) we obtain:
equation (37), we have:
V 1 sT C ˆ s
ˆ sM
τ a fNN *T
W GV V ˆ W*T Ο( V
TG
Tx W ˆ T x) ε a a
N
ˆ sM
sT C ˆ θ ˆ
ˆ TG V
W Tx W ˆ δ
TG (39) a a a M aθar
V
where: ˆ sτ C
sT C ˆ θ τ M
ˆ θ (49)
a a a a a a ar
TG V
δW ˆ T x) ε
T x W *T Ο( V (40)
V N Now, substituting the proposed controller (35) into (49)
we obtain:
Next, we design an error estimator fest to estimate and
compensate for the error vector 2. The online V 1 sT fNN τ a fest Ts (50)
learning algorithms of the NN and the structure of the
error estimator fest are derived in the next section From (39) and (50), we have:
following the Lyapunov method.
ˆ TG V
V 1 sT W ˆ δ f Ts
TG
Tx W (51)
4.3 Stability analysis V est
According to the Lyapunov stability analysis, the system From (41) and (51), we have the derivative of the
is stable if the Lyapunov function candidate is positive Lyapunov candidate function:
definite and its derivative is negative semidefinite.
V V 1 V 2 V 3 V 4
Consider the Lyapunov function candidate:
ˆ TG V
sT W Tx W ˆ δ
TG Ts
V
V V1 V2 V3 V4 (41)
where: T 1W
tr W
ˆ tr V ˆ δ
T 1V T ξ 1f
est (52)
1 Tˆ
V1 s M as (42) As V is desired to be at least negative semidefinite, we
2
choose the online update laws for the NN and design the
estimator as follows:
V2
1
2
T 1W
tr W
(43)
ˆ Gs
ˆ T
W (53)
V3
1
2
tr V
T 1V
(44)
ˆ
V̂ ‐ x G v Ws
T
(54)
1 T 1
V4 δ ξ δ (45) fest ξs (55)
2
in which and are the positive learning rates; is a fest ξ sdt Γ (56)
diagonal positive constant matrix of the error estimator
(56); and δ δ fest is the estimated error. where equation (56) is derived from equation (55).
Moreover, the constant matrix is eliminated from the
Obviously, V1, V2, V3 and V4 are positive definite integration in (56) since the recursive estimation algorithm
functions. Therefore, V is a positive definite function. can recover it. The constant vector is chosen as zero.
ˆ TG V
V sT W ˆ δ
TG
Tx W Ts 5
ˆ T tr V
T
T
T Gs
tr W T ˆ
x G v Ws δ s
-5
ˆ TG V
sT W Tx W ˆ δ
TG Ts
V -10
sT W ˆ sT W
TG ˆ TG V T s
Tx δ 0 1 2 3 4 5 6 7 8 9 10
V Time [s]
(a)
sT Ts 0 (57)
-3
x 10
In (57), since T is a diagonal positive definite matrix, 1
V 0 only when s = 0. Therefore, we can see that the
Or, equivalently: -2
0 1 2 3 4 5 6 7 8 9 10
Time [s]
lim e 0 lim θa θda
t t (b)
(59)
tlim e 0 lim θ a θ da Conventional SMC using BLM
t
Adaptive Fuzzy SMC
Proposed controller
Thus, it is proved that, with the proposed controller (35), the
actual active joint positions converge on their desired values. Figure 6. Tracking errors of the end‐effector in case 1: (a) X–
direction and (b) Y‐direction.
5. Simulation study
In practice, it is difficult to measure the distances from the
Simulation studies were conducted on Matlab‐Simulink and joint to the centre of mass and the inertia tensors of the
the mechanical model of the 2‐DOF planar parallel links. As such, we conducted the simulations with
manipulator was built in SimMechanics toolbox. The link different parameters, both in the mechanical model of the
parameters in the mechanical model are set as follows: l11 = robot and in the controllers, as follows:
l21 = 0.102(m), l12 = l22 = 0.18(m) and l0 = 0.132(m) are the link
rˆi1 0.9ri1 and rˆi 2 0.9ri 2 , i = 1,2 (60)
lengths; r11 = 0.05(m), r21 = 0.055(m), r12 = r22 = 0.09(m) are the
distances from the joint to the centre of mass of the links; m11
where rˆi1 and rˆi 2 were used for calculating M̂ a and Ĉ a
= 0.8(kg), m21 = 0.78(kg), m12 = 1.17(kg), m22 = 1.2(kg) are the
in the controllers.
masses of the links; Iz11 = 0.0027(kg.m2), Iz21 = 0.0031(kg.m2),
Iz12 = Iz22 = 0.0013(kg.m2) are the inertia tensors of the links.
The friction models of the system, including the viscous
friction and the Coulomb friction torques, are defined as
follows:
0.22
A0
0.2
where the coefficients are chosen as Fc1 = Fc2 = 0.3 and Fv1 =
Y[m]
0.16
Fv2 = 2.
0.14
10
(c)
(d)
5
0
-5 0
0 1 2 3 4 5 6 7 8 9 10
Time [s]
-5
(a) 0 1 2 3 4 5 6 7 8 9 10
Time [s]
(b)
Control input active joint 2 [Nm]
10
5
(d)
3.6
5 4.5
3.5
4
3.4
0
3.5 3.3
3 3.2
-5
0 1 2 3 4 5 6 7 8 9 10 3.1
2.5
Time [s] 3
2
(b) 4.5 4.6 4.7 5.9 6 6.1 6.2
Time [s] Time [s]
(c) (d)
Control input active joint 2 [Nm]
Control input active joint 1 [Nm]
5 3.6
Figure 8. Case 1 when applying the adaptive fuzzy SMC: (a)
4.5 3.5 Control input of the active joint 1; (b) Control input of the active
3.4 joint 2; (c) Enlargement of the localized region in (a); (d)
4
3.3
Enlargement of the localized region in (b).
3.5
3.2
The parameters in the proposed controller are set as: T =
3
3.1 diag{10, 10}; the number of neurons in the hidden layer of
2.5 3 the NN is 10; the learning rates are chosen as = 10‐3 and
4.5 4.6
Time [s]
4.7 5.9 6 6.1
Time [s]
6.2
= 210‐3; the initial weight matrices are W(0) = 0.001
I410 and V(0) = 0.001 I102; the constant matrix in the
(c) (d)
error estimator (50) is chosen as = diag{5000, 5000}.
Figure 7. Case 1 when applying the conventional SMC using
BLM: (a) Control input of the active joint 1; (b) Control input of
In simulations, the 2‐DOF parallel manipulator is
the active joint 2; (c) Enlargement of the localized region in (a);
(d) Enlargement of the localized region in (b). disturbed by step external disturbance forces da1(t) = [2,
0]T at t = 2.5s (on active joint 1) and periodic external
+ An adaptive fuzzy sliding mode controller which was disturbance forces da2(t) = [0, 2sin(1.7t)]T at t = 5s (on
proposed in [22]: active joint 2).
ˆ θ
τa M ˆ (63)
a ar C aθar As K fuzzy First, in case 1, we conduct the simulation when the initial
position of the end‐effector E(x,y) of the parallel
where A is the diagonal positive parameter matrix chosen manipulator lies on the top of the reference circular
as A = diag{10, 10}; and Kfuzzy = [kfuzzy1, kfuzzy2]T in which each trajectory A0(0.066, 0.21). The centre coordinates of the
kfuzzyi is estimated by an individual SISO fuzzy system (i = reference circular trajectory are (0.066, 0.16) and the
1, 2). radius is 0.05. The end‐effector is driven to track the
10 0.18
(c)
Y [m]
0.16
5
0.14
0
0.12
-5
0 1 2 3 4 5 6 7 8 9 10 0.1
0 0.02 0.04 0.06 0.08 0.1 0.12
Time [s]
X[ ]
(a) Desired trajectory
Conventional SMC using BLM method
Adaptive Fuzzy SMC
10 Proposed controller
Control input active joint 2 [Nm]
-3
0 x 10
1
Tracking error X‐direction [m]
0
-5
-1
0 1 2 3 4 5 6 7 8 9 10
Time [s] -2
(b) -3
Conventional SMC using BLM
-4 Adaptive Fuzzy SMC
5 -5 Proposed controller
3.6
Control input active joint 2 [Nm]
Control input active joint 1 [Nm]
-6
4.5 3.5 0 1 2 3 4 5 6 7 8 9 10
Time [s]
4 3.4
(a)
3.5 3.3
3 3.2 -3
x 10
2.5 3.1 1
Tracking error Y‐direction [m]
0
2 3
4.5 4.6 4.7 5.9 6 6.1 6.2 -1
Time [s] Time [s]
-2
(c) (d) -3
Conventional SMC using BLM
Figure 9. Case 1 when applying the proposed controller: (a) -4 Adaptive Fuzzy SMC
Control input of the active joint 1; (b) Control input of the active -5 Proposed controller
(b)
Figure 6 shows the tracking errors of the end‐effector in
Figure 11. Tracking errors of the end‐effector in case 2: (a) X–
the X‐direction and in the Y‐direction in case 1. As can
direction and (b) Y‐direction.
be seen from the figure, the tracking errors caused by
the adaptive fuzzy SMC are a little bit smaller than the The control inputs of the active joints 1 and 2 in case 1
errors associated with the conventional SMC using of the conventional SMC using BLM are shown in
BLM. In particular, the proposed controller brings about Figure 7. From the enlargements of the localized
the smallest tracking errors (almost converging on zero) regions, it can clearly be seen that the chattering
compared with the conventional SMC using BLM and phenomenon remains. If we increase the boundary
the adaptive fuzzy SMC. In addition, it can be seen that layer thickness or decrease the switching gains for
the large model uncertainties and external disturbances reducing greater reduction of the chattering, the
are greatly compensated for using the proposed tracking errors will be increased and the robustness of
controller. the system will not be guaranteed.
5
-0.5
-0.2
Sliding surface 2
-0.6
Conventional SMC using BLM (d)
-0.8 5
Adaptive Fuzzy SMC
-1 Proposed controller
-1.2 0
0 1 2 3 4 5 6 7 8 9 10
Time [s]
(b) -5
Figure 12. Sliding surfaces of (a) active joint 1 and (b) active joint 0 2 4 6 8 10
Time [s]
2 in case 2.
(b)
10
Control input active joint 1 [Nm]
(c)
2.8
5.4
5
5.2 2.6
0 5 2.4
4.8 2.2
-5
0 1 2 3 4 5 6 7 8 9 10 4.6
Time [s] 2
(a) 4.4
4.1 4.2 4.3 4.4 4.5 6.2 6.25 6.3
10 Time [s] Time [s]
Control input active joint 2 [Nm]
(c) (d)
(d)
5 Figure 14. Case 2 with the adaptive fuzzy SMC: (a) Control input
of the active joint 1; (b) Control input of the active joint 2; (c)
0
Enlargement of the localized region in (a); (d) Enlargement of the
localized region in (b).
2.8
5.6
2.6
can completely compensate for the uncertainties and any
external disturbances; hence, the tracking errors in the
5.4 2.4 case of using the proposed controller are almost reduced
5.2 2.2 to zero, as shown in Figure 6.
2
5 Next, a simulation is carried out to investigate the control
6.2 6.25 6.3
performance when the end‐effector E(x,y) of the parallel
4.1 4.2 4.3 4.4 4.5
Time [s] Time [s] manipulator does not lie on the reference circle. In this
(c) (d) case, we can demonstrate the convergence of the tracking
Figure 13. Case 2 with the conventional SMC using BLM: (a) errors and sliding functions. Figure 10 shows the
Control input of the active joint 1; (b) Control input of the active comparison of the trajectory tracking among the
joint 2; (c) Enlargement of the localized region in (a); (d) conventional SMC using BLM, the adaptive fuzzy SMC
Enlargement of the localized region in (b). and the proposed controller. The centre coordinates of the
10
(c) regions, it can clearly be seen that the chattering
phenomenon remains in the case of the conventional
5
SMC using BLM but that it is eliminated in the cases of
the adaptive fuzzy SMC and the proposed controller.
0
It can be concluded from the above‐mentioned simulation
results that the proposed controller is highly efficient in
-5
0 1 2 3 4 5 6 7 8 9 10 the control of the 2‐DOF planar parallel manipulators.
Time [s]
(a) 6. Conclusions
10
neuron sliding mode controller for tracking the control of
(d) 2‐DOF parallel manipulators. The proposed controller is
5
based on the combination of three ingredients. The first
ingredient is the equivalent control, which is derived
0
from the dynamic model of the 2‐DOF parallel
manipulators. The second one is a feedforward NN used
-5
to adaptively learn the large nonlinear uncertainties and
0 1 2 3 4 5 6 7 8 9 10
Time [s] external disturbances of the parallel manipulators. The
(b) final part is an error estimator for compensating the
approximation errors of the NN and the higher‐order
5.6
2.8
terms in Taylor series expansion. The online weight
tuning algorithms of the NN and the structure of the
Control input active joint 1 [Nm]
5.4
2.6 error estimator are derived with the strict theoretical
stability proof of the Lyapunov theorem. The connection
5.2
2.4 weights of the NN can be adapted online during the
5
trajectory tracking control of the parallel manipulators
2.2
without any offline training phase. The main advantages
4.8 2
of the proposed controller in comparison with the
existing SMC methods are as follows: (1) The proposed
4.1 4.2 4.3 4.4 4.5 6.2 6.25 6.3 controller can completely compensate the large nonlinear
Time [s] Time [s]
uncertainties and external disturbances of parallel
(c) (d)
manipulators. (2) The proposed control strategy does not
Figure 15. Case 2 with the proposed controller: (a) Control input need to know either the upper bounds of any
of the active joint 1; (b) Control input of the active joint 2; (c) uncertainties or the bound of any approximation errors.
Enlargement of the localized region in (a); (d) Enlargement of
Its structure is simple, easy to implement and yet leads to
the localized region in (b).
the acquisition of good results. These advantages mean
The results of the tracking errors of the end‐effector in that the proposed controller is suitable in application to
case 2 on the X‐direction and the Y‐direction ‐ which are those tracking control parallel manipulators which have a
shown in Figure 11 ‐ prove the excellence of the control complicated dynamic model and huge uncertainties.
performance of the proposed controller in comparison
with the conventional SMC using BLM and the adaptive Simulation results demonstrated the effectiveness of the
fuzzy SMC. It can be seen that the proposed controller proposed controller in the trajectory tracking control of a
brings about the smallest and the fastest convergence of 2‐DOF parallel manipulator. It has been shown that the
tracking errors. proposed controller brings about the smallest tracking
errors compared with the conventional SMC using BLM
Figure 12 shows the comparison of the convergence of the and the adaptive fuzzy SMC. The chattering in the control
sliding functions among all three controllers. As can be inputs is eliminated and the control system is highly
seen from the figure, the sliding functions in the case robust against uncertainties and external disturbances.