X. Zhang et al.: Robust Adaptive Control for Nonlinear Time-Varying Delay System
A3. The desired trajectory y
r
is smooth and available
with y
r
(0) at designer’s disposal; [y
r
,
y
r
,
y
r
]
T
belongs
to a known compact set for all t ≥ 0.
A4. The unknown time-varying state delays 𝜏
i
(t), i =
1, ···, n, satisfy the following inequalities
𝜏
i
(t) ≤ 𝜏
max
< 1. (5)
Remark 1. Assumption A1 is the same as that in [10,30],
which implies that the smooth functions g
i
(⋅), i =
1, ···, n, are strictly either positive or negative and it is
the controllable condition of system (1). Assumption A1
relaxes the assumption in [8,22], in which it was assumed
that 𝜑
i,j
(⋅) are known. Assumptions A2–A4 are common
in the dynamic surface control method.
2.2 Backlash-like hysteresis model
In this paper, the backlash-like hysteresis nonlinear-
ity is described by the following differential equation [5]:
dw
dt
= 𝛼
dv
dt
(𝜆v − w)+𝜓
dv
dt
, (6)
where 𝛼, 𝜆(> 0) and 𝜓 are unknown constant parameters
with 𝜆>𝜓. The solution of (6) is
w = 𝜆v + d(v) (7)
with
d(v)=(w
0
− 𝜆v
0
) exp[−𝛼(v − v
0
)sgn(
v)]
+exp(−𝛼vsgn(
v))
∫
v
v
0
(𝜓 −𝜆) exp[𝛼𝜉sgn(
v)]d𝜉,
(8)
where v
0
= v(t
0
) and w
0
= w(v
0
). It can be proved that
d(v) is bounded for any v ∈ R; furthermore,
lim
v→−∞
d(v)= lim
v→−∞
[w(v; v
0
, w
0
)−𝜆v]=(𝜆 − 𝜓)∕𝛼, (9)
lim
v→+∞
d(v)= lim
v→+∞
[w(v; v
0
, w
0
)−𝜆v]=−(𝜆 − 𝜓)∕𝛼.
(10)
That is, 𝛼 determines the rate at which w switches between
−(𝜆 − 𝜓)∕𝛼 and (𝜆 − 𝜓)∕𝛼: The larger the parameter 𝛼
is, the faster the transition frequency in w is going to be
[5]. Fig. 1 illustrates the class of backlash-like hysteresis
described by (6).
Now, taking (7) into consideration, (1) can be
rewritten as
x
i
=g
i
(
x
i
)x
i+1
+f
i
(
x
i
)+h
i
(
x
i𝜏
)+d
i
(t), i =1, ···, n−1,
x
n
=𝛽v + g
n
(
x
n
)d(v)+f
n
(
x
n
)+h
n
(
x
n𝜏
)+d
n
(t),
y= x
1
,
(11)
where
𝛽 = g
n
(⋅)𝜆, 𝛽>0, (12)
and d(v) is a bounded hysteresis term satisfying
d(v)
≤D, (13)
with D being a positive unknown constant.
A5. The signs of g
i
(
x
i
), i = 1, ···, n, are known. With-
out loss of generality, it is assumed that g
i
(
x
i
) > 0
and there exist two constants g
min
and g
max
satisfy-
ing 0 < g
min
≤
g
i
≤ g
max
.Also,thereexiststhe
constants 𝛽
min
, 𝛽
max
,andD
𝜆 max
, such that 𝛽
min
≤
𝛽(⋅) ≤𝛽
max
and D∕𝜆 ≤ D
𝜆 max
.
Remark 2. Due to 𝜆 being an unknown positive con-
stant, Assumption A5 is reasonable. Also, we emphasize
that g
min
, g
max
, 𝛽
min
, 𝛽
max
and D
𝜆 max
are not required in
implementation of the proposed control design. They are
used for analysis only.
2.3 Radial basis function neural network
(RBFNN) approximation
In this paper, the RBFNN is employed to approx-
imate a continuous function on a given compact set.
−8 −6 −4 −2 0 2 4 6 8
−10
−5
0
5
10
Hysteresis input v(t)
Hysteresis output w(t)
k=6.5
k=3.5
Fig. 1. Hysteresis curves given by (7), where the parameters
𝛼 = 1,𝜆 = 1.432,𝜓 = 0.105, and the input v(t)=k sin
(2.3t)
with k = 3.5 and k = 6.5, respectively.
© 2015 Chinese Automatic Control Society and Wiley Publishing Asia Pty Ltd