Vous êtes sur la page 1sur 9

ISA Transactions 58 (2015) 6775

Contents lists available at ScienceDirect

ISA Transactions
journal homepage: www.elsevier.com/locate/isatrans

L1 adaptive control of uncertain gear transmission servo systems


with deadzone nonlinearity$
Zongyu Zuo a,b,n, Xiao Li c, Zhiguang Shi d
a
The Seventh Research Division, Beihang University (aka Beijing University of Aeronautics and Astronautics, BUAA), Beijing 100191, China
b
Science and Technology on Aircraft Control Laboratory, Beihang University (BUAA), Beijing 100191, China
c
Mechanical Science and Engineering, University of Illinois at Urbana-Champaign, Urbana, IL 61801 USA
d
The Institute of Information Technology, China Aerospace Science and Industrial Corporation, Beijing 100070, China

art ic l e i nf o a b s t r a c t

Article history: This paper deals with the adaptive control problem of Gear Transmission Servo (GTS) systems in the
Received 5 January 2015 presence of unknown deadzone nonlinearity and viscous friction. A global differential homeomorphism
Received in revised form based on a novel differentiable deadzone model is proposed rst. Since there exist both matched and
29 June 2015
unmatched state-dependent unknown nonlinearities, a full-state feedback L1 adaptive controller is
Accepted 13 July 2015
Available online 3 August 2015
constructed to achieve uniformly bounded transient response in addition to steady-state performance.
This paper was recommended for publica- Finally, simulation results are included to show the elimination of limit cycles, in addition to
tion by Y. Chen. demonstrating the main results in this paper.
& 2015 ISA. Published by Elsevier Ltd. All rights reserved.
Keywords:
Adaptive control
Deadzone
Gear transmission servo systems
Nonlinear control

1. Introduction backlash characteristic description with separated parameters via


using appropriate switching and internal functions. However, these
Backlash nonlinearity (e.g., deadzone, hysteresis) due to the papers deal with an exact backlash model. For the systems in the
mechanical transmission via gears is ubiquitous in mechanical presence of parametric uncertainties, neural network adaptive con-
transmission systems, which may deteriorate, if not considered in trol [5,6], fuzzy logic adaptive control [7,8], and sliding mode control
the controller design, the dynamic performance and steady preci- [9] have been employed in control system design. The system with
sion of the controlled systems. backlash nonlinearity between states, such as the Electric Mechanical
In the past decades, many methods have been addressed in a Actuator (EMA) system, can be divided into two subsystems with
signicant literature on the backlash. The hysteresis model [1] is backlash. In consideration that it is more realistic to take the
usually employed to describe the backlash nonlinearity at the input moments of inertia of both subsystems into account, the deadzone
or output, with the assumption that the driven part of the controlled model [1] for backlash description is widely used due to its high
system keeps unchanged when the driving part is in the backlash coincidence with the physical reality. This deadzone model well
period. The system with input/output hysteresis can be transformed describes the backlash nonlinearity due to the transmission torque
into a pseudo-linear system through inverse model compensation. between the driving and driven parts of controlled systems, which,
Using suboptimal convex optimization and Lyapunov tools, Tarbour- however, signicantly complicates the control design problem.
iech et al. [2,3] provided a constructive solution and sufcient Although the inverse model compensation has been applied to such
conditions for the state feedback stabilization of systems with input systems as in [10,11], it is impossible for the input of the deadzone
backlash. The results in [4] propose an analytic expression of model (i.e., the displacement difference between the two sides) to
jump instantly due to the existence of inertia. Therefore, the inverse
model compensation for deadzone model is not preferable in

This work was supported by the National Natural Science Foundation of China practice. The results in [12] put forward a new smooth inverse for
(Nos. 61203022, 61473010 and 61203231). the backlash and proposed a robust adaptive output controller, but
n
Corresponding author at: The Seventh Research Division, Beihang University
this inverse model requires the differential signal of the control input.
(aka Beijing University of Aeronautics and Astronautics, BUAA), Beijing 100191,
China. An alternative is to establish a multi-model switching system [9,14
E-mail address: zzybobby@buaa.edu.cn (Z. Zuo). 16], which consists of several subsystems with controllers designed

http://dx.doi.org/10.1016/j.isatra.2015.07.009
0019-0578/& 2015 ISA. Published by Elsevier Ltd. All rights reserved.
68 Z. Zuo et al. / ISA Transactions 58 (2015) 6775

separately. Su et al. [18] proposed a continuous dynamic model to where Jm, m and cm are the moments of inertia, position and
describe a class of backlash-like hysteresis and its solution has a viscous friction coefcient of the driving side, respectively, Jl, l
linear form with respect to the input signal, which facilitatesthe and cl are the moments of inertia, position and viscous friction
controller design, such as the adaptive backstepping designs in coefcient of the load side, respectively, N0 is the reduction ratio, u
[18,19]. However, the parameter in this hysteresis model cannot be is the torque input, m  N 0 l is the relative displacement,
chosen freely since it is subject to certain approximation require- Dead; ; k is the transmission torque, dened by
ment. In [13], the hysteresis operator is transformed into an 8
equivalent linear time-varying system with uncertainties, where < k 
> if Z ;
the L1 adaptive control can be applied. Similarly, taking into account Dead; ; k 0 if j j o ; 2
>
: k
for the unknown parameters in the deadzone model, adaptive if r  ;
control [14,15], optimal predictive control [16] and sliding mode
control [9] have been adopted in the literature to deal with the with k being the rigidity coefcient and the width between
parametric uncertainties, which considerably complicates the control gears, respectively.
design of the systems with backlash nonlinearity between states.
Motivated by the deciency of the existing methods and the real-
life engineering practice, this paper studies the adaptive control Remark 1. The deadzone description of the transmission torque
problem of GTS systems in the presence of both matched and in (2) is typically non-differentiable. In the real-life engineering
unmatched state-dependent unknown nonlinearities due to the practice, the parameters of transmission gears (i.e., , k and N0)
unknown deadzone nonlinearity and viscous friction. To eliminate can be looked up in the operator's manual. However, the exact
the limit cycles phenomenon [1], a newly developed differential values of these parameters, especially the rigidity coefcient and
deadzone model [17] is employed such that the system dynamics width between gears, are hard to derive, and these parameters,
can be transformed into a canonical form with the nonlinear uncer- including the viscous friction coefcient, may vary due to concus-
tainties bounded by linear growth conditions. Since the unmatched sion or abrasion.
nonlinear uncertainties cannot be addressed by recursive design
methods (e.g., Backstepping) developed for strict feedback systems
or pure feedback systems and conventional model reference adaptive 2.2. Differentiable deadzone model
control (MRAC) for systems with matched uncertainties, the L1
control structure is thus employed as an alternative method to achieve To overcome the non-smooth property of traditional deadzone
uniform transient response and steady state tracking for GTS systems. in (2) and avoid the collision at the backlash end, a new
Note that the L1 adaptive control has been applied for the uncertain differentiable deadzone model [17] is adopted here for controller
hysteretic systems in [13], where the Preisach-type hysteresis in input design, given by
and only matched uncertainty are considered. Different from the work
"   #
in [13], the key features studied in this paper include several aspects: k cosh 
(i) the uncertain deadzone nonlinearity appears between states which T s ; ; k k ln    ; 3
2 cosh
introduces both unmatched and matched uncertainties; (ii) the control
design, based on the new deadzone model with arbitrary approxima-
where and k are dened in (2), and is a positive adjustable
tion accuracy, becomes easier and requires less control effort com-
parameter (refer to as soft degree).
pared to the one using relatively coarse approximation; (iii) limit
Using (2) and (3), the approximation error can be derived as
cycles can be eliminated effectively due to the elasticity (or smooth
property) of this new model with an extra design degree of freedom. 8
>
> k
The remainder of this paper is organized as follows: Section 2 >
>  k  if Z ;
>
> 2
presents the GTS system dynamics and a newly developed differ- >
>
< k
entiable deadzone model. Section 3 presents the global coordinate 9 Dead; ; k  T s ; ; k  k  2 if j j o ; 4
>
>
transformation and the controller structure, while Section 4 analyzes >
>
>
> k
>
: k  2 if r  ;
the closed-loop transient and steady state performances. In Section 5, >
an illustrative simulation example is discussed. Finally, the paper is
ended by concluding remarks in Section 6.
where 9 lncosh  =cosh . It has been proved
in [17] that
2. Problem statement
kln2
j j o and lim j j 0; 5
This section investigates the GTS system dynamics with deadzone
- 1

nonlinearity, and introduces a newly developed differential deadzone


which means the non-smooth property of backlash nonlinearity
model to overcome the non-differential property of the conventional
can be smoothed to any arbitrary precision by the parameter , as
deadzone model which greatly facilitates the control law design.
illustrated in Fig. 1.

2.1. Dynamic model of GTS systems


2.3. State space description of GTS systems
A general dynamical model for GTS systems can be derived as
follows [1]: Let x x1 ; x2 ; x3 ; x4 T l ; l ; m ; m T A R4 be the state vector of
8 a GTS system, where l _ l and m _ m denote the speed of the load
d m d m
2
>
>
> Jm
< cm u  Dead; ; k; side and the driving side, respectively. It follows from (3) and (4) that
dt 2 dt
1
> d l
>
2
d k
>
: J l 2 cl l N0 Dead; ; k; Dead; ; k T s ; ; k kx x ;
dt dt 2
Z. Zuo et al. / ISA Transactions 58 (2015) 6775 69

Fig. 1. Deadzone approximation with different soft degrees: k 2, 1. (a) 1 and (b) 5.

where x lncoshx  =coshx . Thus, the state 3. Main results


space equations of the GTS system in (1) can be formulated as
2_ 3 2 x2
3 2 3
0
We start with a global coordinate transformation for control
x1 law design. After that a full-state feedback L1 adaptive control
6
6 x_ 2 7 6 J0 x3  1 x J 7
N k N 0 x 607
7 6 7
6 7 6 l 7 6 7u; structure is proposed.
6 _ 76
l

4 x3 5 4 x4 7 607
5 4 5
x
x_ 4  2 x  J J
1
3.1. Global coordinate transformation
m m

y 1 0 0 0x x1 ; 6 There exists a global differential homomorphism


2 3
where 1 x 9N 20 k=J l x1 cl =J l x2 N 0 k=2J l x , 2 x 9 kx 2 3
z1
x1
6 7
cm x4 kx =2=J m and x 9 x3 N 0 x1 . 6 z2 7 6 x2 7
In view of Remark 1, let k, , c l and c m be the known nominal 6 7 6 7
z6 76 N 0 T s  c l x2 7 10
4 z3 5 66 Jl 7
7
values of the cor responding parameters in (1). Let the nominal 4 N0 kx4  N0 x2 c l N 0 T s  c l x2 5
differentiable deadzone model T s be denoted by z4 J lx  J 2
l l
"   #
k cosh  such that the system in (8) can be transformed into the following
T s ; ; k k ln    7
2 cosh form:
2 3
2 3 2 32 3 0 2 3
We rewrite (6) as z_ 1 0 1 0 0 z1 6 7 0
6_ 7 6 6 q 7
2_ 3 2 3 2 3 2 3 6 z2 7 6 0 0 1 0 7 7
6 z2 7 6
6 7
2
7 6 07
x2 0 0 6 76 6 76 c 76 7
6
x1 6 z_ 3 7 4 0 0 0 1 7 5 4 z 5 6  J l q2 7 4 05
7v
6 x_ 2 7 6 7 6 N0 x N0 T s  T s  cl  c l x2 7 6 0 7 4 5 6 2 7
3 l
6
6 7 6 Jl 3
N0 k
x  x 7 6 J J J 7 6 7 4 2 5
6 _ 76
1 76 l l l
7 6 7u z_ 4 0 0 0 0 z4 c l  N 0 kJ l lx
q2 N0 klx
q4 1
4 x3 5 4 x4 7 6 0 7 607 J2 J
5 4 5 4 5 l l

x_ 4  2 x  x cm  c m
J m  J m T s T s  J m x4
1 1
Jm
9 Az f z bm v; 11

9px qx bu; where q2 and q4 are the second and the fourth element of q in (8),
respectively, lx 9 1 tanhx  =2 tanhx =2 veri-
y 1 0 0 0x 9 hx; 8 fying

where 1 x 9 N 20 k=J l x1 c l =J l x2  N 0 k=2J l x , 2 x 9 1  tanh o lx o 1; 12


kx c m x4 k x =2=J m , and x 9 lncosh x  = cosh and the virtual control input v is dened as
x  veries
v L4p hx uLb L3p hx: 13
1  2 o x o 2  1: 9

In (8), p(x) stands for the known nominal dynamics, while q(x) for Remark 2. It can be straightforwardly veried from (10) that
the uncertain dynamics. Further, a reasonable assumption for j z=xj N0 k=J l lx 2 4 0 holds. Thus, the coordinate transforma-
robust analysis is needed in the sequel. tion (10) is global. Note that the real control input can be
computed from (13) as u v  L4p hx=Lb L3p hx.
Assumption 1. For the GTS systems in (1), the nominal values of
the model parameters satisfy j k  k j r k , j  j r , j cl c l j r 3.2. L1 adaptive controller
cl and j cm  c m j r cm , where k, , cl and cm are known upper-
bounds of these approximation errors. The full-state feedback control structure for the GTS systems is
dened as
The control objective is to design a full-state feedback adaptive
v vb va ; 14
controller such that (i) all the closed-loop signals of the system in
(1) are bounded, and (ii) the output y(t) tracks a given bounded where vb 9  z denotes the basic nonadaptive component and va
reference signal r(t) in both transient and steady states. is the adaptive component to be designed later, T A R4 is a
70 Z. Zuo et al. / ISA Transactions 58 (2015) 6775

feedback gain vector such that Am 9 A  bm is a Hurwitz matrix has DC gain 1 and C0 1. The adaptive component va in (14) is
specifying the desired closed-loop dynamics. Substituting (14) into dened as the output of the following low pass lter:
(11) one obtains va s KDs^ s; 30
z_ t Am zt bm va t f 1 z Bum f 2 z; 15
where ^ s is the Laplace transform of the signal
T
yt c zt; z0 z0 ; 16 ^ t va t ^ 1 t ^ 2m t K g rt
1
where c 1; 0; 0; 0T , Bum A R43 is a constant matrix such that with K g  1=cT Am bm , ^ 2m s H m
1
sH um s^ 2 s and with
T
bm Bum 0 and rankbm ; Bum  4, f 1 : R4 -R and f 2 : R4 -R3 are ^
^ t and ^ t being dened as ^ t t J z J ^ t, i 1,2.
1 2 i i L1 i
unknown model dynamics, i.e., Toward this end, the L1 adaptive controller for the GTS systems
c 2l N 20 kJ l lx N0 klx consists of (14), (23)(28) and (30).
f1 q2 q4 ; 17
J 2l Jl Remark 3. Notice that the transfer function between ^ 2 s and
1
2 3 va(s), CsH m sH um s, is proper and stable, which guarantees the
0
6 q 7 existence of J CsH m 1
sH um s J L1 and v in (56). Thus, there is no
f2 6
4 c 2 5:
7 18 singularity and differentiation involved in the generation of the L1
 J l q2 component va(s). If setting Cs 1, i.e., MRAC law va s
l

^ 1 s ^ 2m s  K g rs, the transfer function between ^ 2 s and


It can be veried that 1 1
" # va(s), H m sH um s, is not proper and thus J H m sH um s J L1 is
f1  1 undened, which implies that the conventional MRAC is not
bm Bum f;
f2 applicable here.

where
3.3. Model uncertainties analysis
2 3
1 0 0
60 1 07 Dene the following positive constants:
6 7
Bum 6 7:
40 0 15 k kln2
a0 2k k k;
0 0 0 2
In this formulation, f1 represents the matched unknown model c
a1 cl k l ;
dynamics and Bum f 2 represents the unmatched uncertainties. Jl Jlk
Let k
a2 ;
H xm s sI  Am 1
bm ; 19 k
1 k c l
a3 cm N 0 ;
H xum s sI  Am  1 Bum ; 20 Jm N0 k
!
1 k J l cm c l
T 1 a4 ;
H m s c sI  Am bm ; 21 J m N0 k N0 k1  tanh
cm J l
H um s cT sI  Am  1 Bum : 22 a5 :
J m N 0 k1  tanh
Consider the following state predictor:
 
z^_ t Am z^ t bm va t ^ 1 t J zt J L1 ^ 1 t
Lemma 1. If Assumption 1 holds, then we have
  j f 1 j rK 1 J zt J 1 B1 ; 31
Bum ^ 2 t J zt J L1 ^ 2 t ; 23
J f 2 J 1 r K 2 J zt J 1 B2 ; 32
^ cT z^ t;
yt z^ 0 z0 ; 24 where
(



)
where ^ 1 t A R, ^ 1 t A R, ^ 2 t A R3 and ^ 2 t A R3 are adaptive
c 2  N 2 kJ 1  tanh

c 2  N2 kJ




l
0 l 0 l
N0 k
K 1 max
l
;

a1 a2 a3 a4 a5 ;
estimates.
J 2

J 2
Jl
l l
The adaptive laws are dened as
c
_ K 2 max 1; l a1 a2 ;
^ 1 Proj^ 1 ;  bTm P z~ t J z J L1 ; 25 J
(
l


) !

c 2  N 2 kJ 1  tanh

c 2 N 2 kJ


l 0 l

l 0 l
k N 0 a0
^_ 1 Proj^ 1 ;  bTm P z~ t; 26 B1 max

;

;

J2

J2
Jm Jl
l l

_ N 0 a0
^ 2 Proj^ 2 ;  BTum P z~ t J z J L1 ; 27 B2
Jl
:

^_ 2 Proj^ 2 ;  BTum P z~ t; 28 Proof. First, it follows from Assumption 1, (5) and (9) that

where z~ t z^ t  zt, A R is the adaptive gain, P P T 4 0 is


k k

j T s  T s j
k  kx x  x

the solution of the Lyapunov equation ATm P PAm  Q for arbi-


2 2

trary Q Q T 4 0 and Proj;  denotes the projection operator [22].  


o k j x j k k : 33
Let K be a positive gain and D(s) represents a strictly proper
transfer function, chosen to ensure that Using the coordinate transformation (10), we obtain
KDs Jl cl
Cs 29 j x j o j z3 j j z2 j ; 34
1 KDs N0 k N0 k
Z. Zuo et al. / ISA Transactions 58 (2015) 6775 71



For the transient and steady-state response analysis we con-

Jl cl

j x4 j
z4 N 0 z2 z3
sider the following reference system:

N 0 klx N 0 klx

8
> z_ ref t Azref t bm vref t f 1 zref Bum f 2 zref ;
Jl cl >
>
r j z4 j N 0 j z 2 j j z3 j : >
< vref t  zref t varef t;
N 0 k1  tanh N0 k1  tanh
39
35 >
> v s Cs1ref s H m 1
sH um s2ref s  K g rs;
> aref
>
: y t cT z t; z 0 z ;
ref ref ref 0
Thus, we have


where iref s is the Laplace transformations of the signal

N 0 N0 c  c l

j q2 j

x T s  T s  l x2
1ref t f i zref t, i1,2.
Jl Jl Jl
Dene
N 0 k kln2 N 0 k N 0 k N 0 k cl
o j x j j z2 j vr J J 1 r J Cs J L1 K 1 r B1 J CsH m 1 sHum s J L1 K 2 r
2J l Jl Jl Jl Jl
B2 J CsK g J L1 J r J L1 : 40
N 0 a0
r a1 j z2 j a2 j z3 j ; 36
Jl
Lemma 2. For the closed-loop reference system in (39), subject to the
and L1 norm condition (38), if J z0 J r 0 , then


1 1 cm c m

J zref J L1 o r ; J vref J L1 o vr : 41
j q4 j

 x  T s T s  x4

Jm J m J m

k kln2 k k cm Proof. It follows from (39) that


o j x j k j x4 j
2J m Jm Jm Jm Jm zref s sI  Am  1 bm varef s sI  Am  1 bm 1ref s
a0 sI  Am  1 Bum 2ref s zin s
r a3 j z2 j a4 j z3 j a5 j z4 j : 37
Jm
H xm svaref s H xm s1ref s H xum s2ref s zin s
It follows from (17) and (18) that Gm s1ref s Gum s2ref s H xm sK g rs zin s; 42
(



)

c 2  N 2 kJ 1  tanh

c 2  N 2 kJ
N 0 klx where zin(t) is the signal with Laplace transform zin s

0 l

l 0 l

j f 1 j r max
l ;



j q2 j j q4 j

J2

J2
Jl sI  Am  1 z0 . Since Am is Hurwitz and z0 is bounded, then
l l
J zin J L1 r in . Then, for all t A 0;  we have
rK 1 J z J 1 B1 J zref s J L1 r J Gm s J L1 J 1ref s J L1 J Gum s J L1 J 2ref s J L1

c J H xm sK g J L1 J rs J L1 in : 43
J f 2 J 1 r max 1; l j q2 j r K 2 J z J 1 B2 ;
Jl
We prove the bounds in (41) by contradiction. Assume that the
where (12), (36) and (37) are used. rst bound in (41) does not hold. Since J zref 0 J 1 J z0 J
1 r 0 o r and zref t is continuous, then there exists a time

Remark 4. Actually, Lemma 1 ensures that the nonlinear model 1 A 0;  such that
uncertainties f1 and f2 satisfy the Lipschitz condition. In addition, if J zref 1 J 1 r while J zref t J 1 o r ; 8 t A 0; 1 ;
there exist no parametric uncertainties, i.e., k cm cl 0,
which implies that
we have ai 0, i 1; 2; 3; 4, and lim-1 a0 0. Then it follows
that K 1 K 2 0 and lim-1 B1 lim-1 B2 0. J zref 1 J L1 r : 44

It has been veried in Lemma 1 that


4. Closed-loop performance analysis J iref J L1 r K i J zref J L1 Bi r K i r Bi : 45

A reference system is constructed for the proofs of performance Substituting (45) into (43) yields
bounds, which assumes perfect identication of all uncertainties, J zref 1 J L1 r J Gm s J L1 K 1 r B1 J Gum s J L1 K 2 r B2
but cancels them only within the bandwidth of the low pass lter J H xm sK g J L1 J rs J L1 in r J Gm s J L1
C(s) dened previously. Then we show that the states and the J Gum s J L1 0 K 1 r B0 J Hxm sK g J L1 J rs J L1 in : 46
control signal of the actual system are uniformly bounded and the
performance bounds of the closed-loop system with L1 controller The condition in (38) can be solved for r to obtain the bound
are computable. J zref 1 J 1 o r , which leads to a contradiction with (44) and thus
proves the rst bound in (41). This further implies that the upper
bounds in (45) hold with strict inequality, i.e.,
4.1. Closed-loop reference system J iref J L1 o K i r Bi : 47

The initial condition z0 is assumed to be inside an arbitrary The bound on vref(t) follows from (39) and (47),
large known set, i.e., J z0 J 1 r 0 o1 for some 0 40. For the J vref J L1 o J J 1 r J Cs J L1 K 1 r B1 J CsH m
1
H um s
proofs of stability and performance bounds, the choice of K and D
J L1 K 2 r B2 J CsK g J L1 J r J L1 vr :
(s) also needs to ensure that, for a given 0, there exists r 4 in
such that the following L1 norm condition holds

r  J Hxm sCsK g J L1 J rs J L1  in
J Gm s J L1 J Gum s J L1 0 o ; 4.2. Prediction error signal
K 1 r B 0
38
According to Lemma 1 and Lemma A.8.1 [23] for the system in
where 0 K 2 =K 1 , B0 maxfB1 ; B2 =0 g, in J ssI Am 1
J L1 0 , (15), if v(t) is continuous and moreover the following bounds hold
1
Gm s H xm sI  Cs and Gum s I  H xm sCsH m scT H xum s. J z J L1 r ; J v J L1 r v ;
72 Z. Zuo et al. / ISA Transactions 58 (2015) 6775

then, for all t A 0; , there exists continuous 1 t A R, 1 t A R, Theorem 1. Let the adaptive gain be lower bounded as in (51) and
2 t A R3 and 2 t A R3 with (piecewise-)continuous derivative the projection be conned to the bounds in (52). Given the system in
that (11) and the L1 adaptive controller via (23)(30) subject to the
L1 norm condition in (38), and the closed-loop reference system in
J i t J 1 o bi ; J _ i t J 1 o di ;
(39), if J z0 J r 0 , then we have
J i t J 1 o bi ; J _ i t J 1 o d i ;
J zt J L1 o ; 57
f i zt i t J z J L1 i t;
J vt J L1 o v ; 58
for i1,2, where bi K i , bi Bi , di and d i are computable
bounds, and 4 0 is an arbitrary constant. Thus, the system in (15) J z~ t J L1 o 0 ; 59
can be rewritten over t A 0;  as
J zref t  zt J L1 o 1 ; 60
z_ t Am zt bm va t 1 t J z J L1 1 t Bum 2 t J z J L1 2 t

yt cT zt; z0 z0 : 48 J vref t  vt J L1 o 2 ; 61

Let J yref t  yt J L1 o J cT J 1 1 : 62
~ i ^ i  i ; ~ i ^ i  i ; i 1; 2:
Proof. Assume the bounds in (60) and (61) do not hold. Then,
Using the above notations, the following error dynamics can be
since J zref 0  z0 J 1 0 o 1 , J vref 0  v0 J 1 0 o 2 , and z(t),
derived from (23) and (48):
zref t, v(t) and vref(t) are continuous, there exists such that
z~_ t Am z~ t bm ~ 1 t Bum ~ 2 t; 49 J zref  z J 1 1 or J vref  v J 1 2 ;
where z~ t z^ t  zt, ~ i t ^ i t  i t with i t i t J z J L1 while
i t, i 1,2.
J zref t  zt J 1 o 1 ; J vref t  vt J 1 o 2 ; 8 t A 0; :
Dene
  max P  This implies that at least one of the following equalities holds
m 4 2b1 2b1 32b2 3 2b2 4 d b1 d 1 3b2 d2 3 b2 d 2 :
min P b1 1 J zref  z J L1 1 or J vref  v J L1 2 : 63
50
Lemma 2 implies that
The following lemma shows that if the adaptive gain is chosen
J zref J L1 r r ; J vref J L1 r vr : 64
to verify the lower bound:
m Using the denitions of and v in (55) and (56), it follows from
4 ; 51 the bounds in (64) and (63) that
min P 2
0
J z J L 1 r r 1 ; J v J L1 r vr 2 v :
and the projection is conned to the bounds:
Hence, with the selection of the adaptive gain in (51) and the
J ^ i t J 1 r bi ; J ^ i t J 1 r bi ; i 1; 2; 52 projection bounds in (52), Lemma 3 ensures that
then the prediction error z~ t can be systematically reduced J z~ J L1 o 0 : 65
in both transient and steady-state by increasing the adaptation
gain. Let ~ t ~ 1 t ~ 2m t, where ~ 2m s is the signal with the Laplace
transform ~ 2m s H m 1
sH um s~ 2 s. We rewrite (30) as
Lemma 3. Let the adaptive gain be lower bounded as in (51) and the
projection be conned to the bounds in (52). Given the system in (15) va s Cs1 s H m
1
sH um s2 s  K g rs ~ s: 66
and the L1 adaptive controller via (23)(30) subject to the L1 norm The actual closed-loop system in (15) takes the form
condition in (38), if
zs Gm s1 s Gum s2 s  H xm sCs~ s H xm sCsK g rs zin :
J z J L1 r ; J v J L1 r v ; 53
67
then we have
From (67) and (42) we have
J z~ J L1 o 0 ; 54
zref s  zs Gm s1ref s  1 s Gum s2ref s  2 s Hxm sCs~ s:
where 0 was dened in (51).
Moreover, from the prediction error dynamics (49) we have
The proof follows the same lines of the Lemma 3 in [24]. 1
Hm scT z~ s ~ 1 s ~ 2m s ~ s;
4.3. Transient and steady-state performance which leads to
zref s zs Gm s1ref s  1 s Gum s2ref s  2 s
Let denote an arbitrarily small positive quantity and let 1
H xm sCsH m scT z~ s:
r 1 ; 55 Therefore, we have
v vr 2 ; 56 J zref  z J L1 r J Gm s J L1 J 1ref s  1 s J L1 J Gum s
J L1 J 2ref s  2 s J L1 J H xm sCsH m
1
scT J L1 J z~ J L1 : 68
where
Since the following equalities always hold
1 T
J H sCsH m c J L1
1  xm  ; J iref  i J 1 J f i zref f i z J 1 r K i J zref z J 1 ; i 1; 2; 69
1  K 1 J Gm s J L1 J Gum s J L1 0 0
2 J J 1 J Cs J L1 K 1 1
J CsH m
H um s J L1 K 2 1 from (68) we have
J CsH m c J L1 0 :
1 T
J zref  z J L1 r J Gm s J L1 K 1 J zref  z J L1 J Gum s J L1 K 2 J zref  z J L1
Z. Zuo et al. / ISA Transactions 58 (2015) 6775 73

1
J H xm sCsH m scT J L1 J z~ J L1 :

The upper bound in (65) and the L1 norm condition in (38) lead
to the upper bound
1
J H xm sCsH m scT J L1
J zref  z J L1 r   J z~ J L1
1  K 1 J Gm s J L1 J Gum J L1 s0
1
J H xm sCsH m scT J L1
r   1  o 1 : 70
1  K 1 J Gm s J L1 J Gum J L1 s0 0

On the other hand, it follows from (14), (30) and (39) that
vref v  zref  z varef  va :

One can write


Fig. 3. Phase portrait: step reference.
J vref  v J L1 r J J 1 J zref  z J L1 J Cs J L1 J 1ref  1
1
J L1 J CsH m sH um s J L1 J 2ref  2 J L1 J CsH m
1
scT
J L1 J z~ J L1 r J J 1 J zref  z J L1 J Cs J L1 K 1 J zref  z
1
J L1 J CsH m sH um s J L1 K 2 J zref  z
1
J L1 J CsH m scT J L1 J z~ J L1 :

The bounds in (59) and (70) lead to


 
J vref  v J L1 r J J 1 J Cs J L1 K 1 J CsH m
1
sH um s J L1 K 2 1

1
J CsH m scT J L1 0 2  o 2 : 71

Finally, we note that the upper bounds in (70) and (71) contradict
the equalities in (63), which proves the bounds in (60) and (61). The
results in (57), (58), and (62) follow directly from the bounds in (64)
and (59) and from the fact that yref t  yt cT zref t  zt. Fig. 4. Control input: step reference.

5. Simulation studies
To implement the L1 adaptive controller, we set
To evaluate the effectiveness of the proposed adaptive controller,
the parameters of a GTS system employed in the simulation are 625; 500; 150; 20; K 748:8; Q 10  3 I4 ; 105 ;
shown in Table 1. The initial states in the simulation are x1 0rad,
x2 0rad=s, x3 0rad and x4 0rad=s. The soft degree 10 in Ds
1
; 1
K g  cT Am bm  1 625:
(3) is selected with sufcient accuracy (refer to Fig. 1). s5 17:7s4 179:8s3 1516s2 2213s

By Lemma 1, the projection bounds in (52) are calculated as


b1 K 1 0:98; b2 K 2 0:03; b1 B1 3; b2 B2 1:
Table 1
GTS system parameters.
Consider the step reference signals given by r 1 1rad,
System parameters Actual value Nominal value Unit
r 2 2rad and r 3 3rad, respectively. Fig. 2 shows the scaled
responses of the output (i.e., the position of load side) for scaled
Load inertia Jl 3 kg m2 input signals. Figs. 35 display the simulation result under reference
Motor inertia Jm 1.8 kg m2 input r1. The phase portrait in Fig. 3 demonstrates that no limit cycle
Gear ratio N0 5 occurs due to the introduction of the novel smooth differentiable
Gear coefcient k 2.02 k 2 N m=rad
deadzone model in (3) for the controller design. The time history of
Backlash gap 0:001 0:002 rad
Viscous friction coefcient cl 0.45, cm 0.4 c l 0:4, c m 0:35 N m=rad
the torque input is plotted in Fig. 4. Fig. 5 shows the online
estimates of both matched and unmatched model uncertainties.
A performance comparison with the traditional PD controller,
_ kp et, is carried out to highlight the
dened by ut kd et
control design in this paper, where et rt  x1 t, kp and kd
represent the proportional and the derivative gain, respectively.
Figs. 6 and 7 show the system output response and the load phase
portrait, from which we easily nd that the limit cycle appears and
the transient response is more aggressive with a big overshoot.
To further demonstrate the effectiveness of the differential dead-
zone model, a higher gear mesh stiffness (i.e. k 25) is considered. In
addition, a periodic square wave signal with the holding time being
10 s is selected as the reference input to test our L1 controller. It is
shown from Figs. 8 and 9 that the responses of the closed-loop
system are still smooth with such a higher torsional stiffness
coefcient k.
Clearly, the simulation results validate that the L1 adaptive
Fig. 2. System output: step reference. controller, based on the new differentiable deadzone model,
74 Z. Zuo et al. / ISA Transactions 58 (2015) 6775

Fig. 5. Estimated model uncertainties: step reference. (a) Matched model uncertainty and (b) unmatched model uncertainty.

Fig. 8. System output: square wave reference.

Fig. 6. System Output: PD Controller.

Fig. 9. Control input: square wave reference.

deadzone and viscous coefcients. A full state feedback controller


augmented with L1 adaptive control law is provided to ensure
guaranteed transient and steady state performance. A simulation
example is presented to demonstrate that the closed-loop system
suppresses both the matched and unmatched nonlinear model
uncertainties efciently, and also achieves the steady state refer-
ence input tracking.

Fig. 7. Phase portrait: PD controller. Acknowledgments

ensures uniform transient and steady state performance in the The authors would like to thank Professor Naira Hovakimyan
presence of uncertain backlash nonlinearity and viscous friction. from University of Illinois Urbana-Champaign for her insights and
valuable comments.

6. Conclusions References

This paper presents a novel differentiable deadzone model for [1] Nordin M, Gutman PO. Controlling mechanical systems with backlasha
controller design of GTS systems in the presence of uncertain survey. Automatica 2002;38(10):163349.
Z. Zuo et al. / ISA Transactions 58 (2015) 6775 75

[2] Tarbouriech S, Pieur C, Queinnec I. Stability analysis for linear systems with [13] Zou X, Luo J, Cao C. Adaptive control for uncertain hysteresis systems. ASME J
input backlash through sufcient LMI conditions. Automatica 2010;46 Dyn Syst Meas Control 2014;136(1) 011011-1011011-7.
(11):19115. [14] Coelho L, Cunha M. Adaptive cascade control of a hydraulic actuator with an
[3] Tarbouriech S, Queinnec I, Pieur C. Stability analysis and stabilization of adaptive dead-zone compensation and optimization based on evolutionary
systems with input backlash. IEEE Trans Autom Control 2014;59(2):48894. algorithms. Expert Syst Appl 2011;38(10):122629.
[4] Vrs J. Modeling and identication of systems with backlash. Automatica [15] Khan MB, Malik FM, Munawar K. Switched hybrid speed control of elastic
2010;46(2):36974. systems with deadzone. In: Proceedings of IEEE international conference on
[5] Guo J, Yao B, Chen QW, Wu XB. High performance adaptive robust control for industrial engineering and engineering management (IEEM); 2010. p. 16414.
nonlinear system with unknown input deadzone. In: Proceedings of IEEE [16] Rostalski P, Besselmann T, Bari M, Van Belzen F, Morari M. A hybrid approach
conference on decision and control; 2009. p. 76759. to modelling, control and state estimation of mechanical systems with back-
[6] Selmic RR, Lewis FL. Deadzone compensation in systems using neural motion
lash. Int J Control 2007;80(11):172940.
control networks. IEEE Trans Autom Control 2000;45(4):60213.
[17] Shi Z, Zuo Z. Backstepping control for gear transmission servo systems with
[7] Shahnazi R, Pariz N, Kamyad AV. Adaptive fuzzy output feedback control for a
backlash nonlinearity. IEEE Trans Autom Sci Eng 2015;12(2):7527.
class of uncertain nonlinear systems with unknown deadzone-like hysteresis.
[18] Su CY, Stepanenko Y, Svoboda J, Leung TP. Robust adaptive control of a class of
Commun Nonlinear Sci Numer Simul 2010;15(8):220621.
nonlinear systems with unknown backlash-like hysteresis. IEEE Trans Autom
[8] Jang JO. Lee PG, Park SB, Ahn IS. Deadzone compensation of systems using
fuzzy logic. In: Proceedings of American control conference; 2001, p. 47889. Control 2000;45(12):242732.
[9] Corradini ML, Manni A, Parlangeli G. Variable structure control of nonlinear [19] Zhou J, Wen C, Zhang Y. Adaptive backstepping control of a class of uncertain
uncertain sandwich systems with nonsmooth nonlinearities. In: Proceedings nonlinear systems with unknown backlash-like hysteresis. IEEE Trans Autom
of IEEE conference on decision and control; 2007. p. 20238. Control 2004;49(10):17517.
[10] Ahmad NJ, Ebraheem HK, Alnaser MJ, Alostath JM. Adaptive control of a DC [22] Pomet JB, Praly L. Adaptive nonlinear regulation: estimation from the
motor with uncertain deadzone nonlinearity at the input. In: Chinese control Lyapunov equation. IEEE Trans Autom Control 1992;37(6):72940.
and decision conference; 2011. p. 42959. [23] Hovakimyan N, Cao C. L1 adaptive control theory: guaranteed robustness with
[11] Yu P, Zhao ZB, Bao GJ. Inverse compensation method for the executor with fast adaptation. Philadelphia, PA: SIAM; 2010.
deadzone based on adaptive inverse control. In: Proceedings of international [24] Cao C, Hovakimyan N. L1 adaptive controller for a class of systems with
conference on machine learning and cybernetics; 2007. p. 5837. unknown nonlinearities: Part I. In: Proceedings of american control confer-
[12] Zhou J, Zhang C, Wen C. Robust adaptive output control of uncertain nonlinear ence; 2008. p. 40938.
plants with unknown backlash nonlinearity. IEEE Trans Autom Control
2007;52(3):5039.

Vous aimerez peut-être aussi