104 Y. Song et al. / Automatica 77 (2017) 103–111
In this work, we present a solution to the trajectory tracking
control problem of nth-order nonlinear system with uncertain
target trajectory and sensor faults. Four fundamental steps are
exploited to tackle this problem. First, introducing a filtered
variable to reduce the error dynamics from nth order to first order,
facilitating later control design and stability analysis. Second,
establishing an analytical model to reconstruct the polluted
or disguised desired trajectory, which generates the estimated
(predicted) trajectory with reflection of the actual target trajectory.
Third, an indirect NN-based method is used to approximate
the nonlinearities and uncertainties in the system to simplify
the online parameter updating. Fourth, the concept of virtual
parameter is used and blended into the skillfully chosen Barrier
Lyapunov Function (BLF) to guide the derivation of the tracking
control algorithms, here by keeping the BLF bounded, the compact
set condition for the NN input is always maintained during the
entire process of system operation as the barrier will never be
transgressed (He, Zhang, & Ge, 2014; Tee, Ge, & Tay, 2009; Xu
& Jin, 2013). With the help of these steps, we establish a UUB
stable tracking control solution for a class of MIMO nonlinear
systems that involve uncertain desired target trajectory, sensor
faults and unknown dynamics. This is the first reported work
explicitly addressing unknown trajectory tracking with low-cost
method.
The remainder of this paper is organized as follows: Section 2
formulates the problem and reviews some mathematical prelimi-
naries. Section 3 elucidates the indirect neuroadaptive control de-
sign process with formative stability analysis, where state tracking
is achieved. Section 4 presents the modified indirect neuroadaptive
controller based on BLF to ensure the compact set condition for the
NN unit to function reliably during system operation. An illustra-
tive example is included to demonstrate the effectiveness of the
proposed control method in Section 5. Finally, Section 6 concludes
the paper with suggestions for further study.
2. Problem statement and preliminaries
2.1. System description
Consider nonlinear MIMO systems of the following form
˙
x
i
= x
i+1
, i = 1, 2, . . . , n − 1
˙
x
n
= F (
¯
x) + G(
¯
x)u
(1)
where u = [u
1
, . . . , u
r
]
T
∈ R
r
is the control input vector (output
vector of the actuator),
¯
x = [x
T
1
, . . . , x
T
n
]
T
= [x
T
, · · ·, x
(n−1)
T
]
T
∈
R
mn
is the state of the system with x
j
= [x
j1
, . . . , x
jm
]
T
∈ R
m
, j =
1, 2, . . . , n and x
1
= x, F (·) ∈ R
m
is a smooth but unknown
function vector, and G(·) ∈ R
m×r
is the unknown and time-varying
control gain matrix of the system. Throughout this paper it is
assumed that the initial states of the system start from a compact
set large enough to cover the entire operational domain of interest
Ω
x
⊂ R
m
.
2.2. Unknown target trajectory
For a hidden (disguised) target, it is very difficult, if not
impossible, to get its trajectory (x
∗
) precisely in advance. Thus
any target acquisition system can only provide the estimated
(predicted) target trajectory x
d
capable of reflecting the desired
trajectory x
∗
to some extent. Based upon this observation, we
propose the following model to link x
d
with x
∗
via
x
d
(
t
)
= d
0
(
·
)
x
∗
(t) + ε
d0
(
·
)
0 < d
0
≤
∥
d
0
(
·
)
∥
≤
¯
d
0
< ∞
∥
ε
d0
(
·
)
∥
≤ ε
d
0
< ∞
(2)
where x
d
(t) = [x
d1
(t), . . . , x
dm
(t)]
T
∈ R
m
is the estimation
(prediction) of the target trajectory x
∗
(t) = [x
∗
1
(t), . . . , x
∗
m
(t)] ∈
R
m
, d
0
(·) ∈ R
m×m
is an unknown diagonal matrix and ε
d0
(·) ∈ R
m
is the estimation error. Here d
0
,
¯
d
0
, ε
d
0
are some unknown positive
constants, and
∥
•
∥
denotes the Euclidean norm of ‘‘•’’.
Similarly, the derivatives x
∗
(i)
, (i = 1, 2, . . . , n) and their
estimates (x
(i)
d
, i = 1, 2, . . . , n) are correlated by the following
relationship
x
(i)
d
(
t
)
= d
i
(
·
)
x
∗
(i)
(t) + ε
di
(
·
)
0 < d
i
≤
∥
d
i
(
·
)
∥
≤
¯
d
i
< ∞
∥
ε
di
(
·
)
∥
≤ ε
d
i
< ∞
(3)
where d
i
(·) ∈ R
m×m
is some unknown diagonal matrix, ε
di
(·) ∈ R
m
is the estimation error vector, and d
i
,
¯
d
i
, ε
d
i
are some unknown
constants.
In this work, we consider the scenario that, although x
∗
and
x
∗
(i)
, (i = 1, 2, . . . , n) are unavailable precisely, certain crude
information on x
∗
and x
∗
(i)
always exists and is reflected through
x
d
and x
(i)
d
to some extent. As such, it is reasonable to assume that
d
i
(·) is lower bounded by some non-zero unknown constant (the
case of
∥
d
i
(·)
∥
= 0 is not considered here because it implies that
the estimated trajectory x
d
does not carry (contain) any useful
information of the actual target trajectory at all), and that d
i
(·)
and ε
di
(·), (i = 0, 1, . . . , n) are upper bounded by some unknown
constants.
Remark 1. Most existing trajectory tracking control designs are
based on the assumption that x
∗
and x
∗
(i)
are known precisely
(Deng et al., 2008; Dong et al., 2012; Ge & Wang, 2004; Ge & Zhang,
2003; Li et al., 2016; Sanner & Slotine, 1992; Spooner & Passino,
1999; Zhang et al., 2000), which, however, might not be true
in practice. In fact, the target trajectory is generally unavailable,
especially for those hidden and sneaky target, where the target
trajectory x
∗
and x
∗
(i)
are buried within x
d
and x
(i)
d
. Apparently,
knowing x
d
and x
(i)
d
, i = 1, 2, . . . , n does not necessarily mean
knowing x
∗
and x
∗
(i)
, i = 1, 2, . . . , n, because d
0
, d
i
, ε
d0
and ε
di
are
both time-varying and unknown. In other words, from x
d
and x
(i)
d
one cannot recover or extract out the ideal trajectory x
∗
and x
∗
(i)
completely because of the uncertainties in d
0
, d
i
, ε
d0
and ε
di
, which
call for a more dedicated approach for tracking control design.
2.3. Sensor failures
Due to measurement noise and possible sensoring faults, one
can never get the precise measurement on x. Instead, it is only
possible to get
x
m
(
t
)
= x
(
t
)
+ ε
m0
(
·
)
∥
ε
m0
(
·
)
∥
≤ ε
m
0
< ∞
(4)
where x
m
(t) = [x
m1
(t), . . . , x
mm
(t)]
T
∈ R
m
is the measured value
of the system state x(t), ε
m0
(·) ∈ R
m
is the measurement error, and
ε
m
0
is a positive constant.
Similarly, for x
(i)
(t), (i = 1, 2, . . . , n) the following relationship
holds
x
(i)
m
(
t
)
= x
(i)
(
t
)
+ ε
mi
(
·
)
∥
ε
mi
(
·
)
∥
≤ ε
m
i
< ∞
(5)
where ε
mi
(
·
)
is the measurement error and ε
m
i
is some unknown
constant. The boundedness of ε
m
i
(i = 0, 1, . . . , n) is necessary
for the underlying problem to admit a feasible solution. Note that
neither x
(n)
nor x
(n)
m
is utilized for the later control design.