Vous êtes sur la page 1sur 12

Received October 15, 2017, accepted November 23, 2017, date of publication December 6, 2017, date of current version

March 9, 2018.
Digital Object Identifier 10.1109/ACCESS.2017.2780118

A Fuzzy Model Predictive Control Based Upon


Adaptive Neural Network Disturbance Observer
for a Constrained Hypersonic Vehicle
YU MA AND YUANLI CAI
Institute of Control Engineering, Xi’an Jiaotong University, Xi’an 710049, China
Corresponding author: Yu Ma (mayu_xjtu@163.com)
This work was supported by the National Natural Science Foundation of China under Grant 61463029 and Grant 61308120.

ABSTRACT A fuzzy model predictive control scheme based upon adaptive neural network disturbance
observer is proposed for the longitudinal dynamics of a constrained hypersonic vehicle (HV) in the presence
of diverse disturbances. First, an equivalent disturbed fuzzy dynamic model with the varying parameters is
constructed to approximate the nonlinear dynamics, where the inevitable lumped disturbances, including
the fuzzy modeling error, extraneous disturbances, and model uncertainties caused by aerodynamic uncer-
tainties, need to be suppressed. Subsequently, according to the parameter-dependent Lyapunov function,
the proposed scheme taking the varying parameters into account is developed to explicitly handle the
constraints of fuel equivalence ratio, elevator deflection, and angle of attack. Furthermore, based on the
strong nonlinear approximation ability of neural network (NN), an adaptive neural network disturbance
observer with the adaptive laws of NN’s weight matrixes is established to estimate lumped disturbances, and
then an additional compensator formulated by integrating the estimations of lumped disturbances and the
corresponding compensation gain matrix is appended to the proposed method for suppressing the lumped
disturbances directly. Finally, the comparative simulation results for tracking the reference commands of
velocity and altitude demonstrate that the proposed method provides a satisfactory tracking performance
even when HV is in the presence of lumped disturbances and constraints.

INDEX TERMS Hypersonic vehicle, adaptive neural network disturbance observer, fuzzy model, fuzzy
model predictive control, parameter dependent Lyapunov function.

I. INTRODUCTION Over the past decades, numerous control methods have


Hypersonic vehicle (HV) primarily powered by the advanced been applied to obtain the control system with satisfactory
scramjet engine provides a promising access to sustained stability, performance and robustness. Due to the complex
hypersonic flight in near space compared with the conven- nonlinear dynamics of HV, Lie derivative notation could be
tional turbojet-powered aircraft whose extreme velocity is employed to linearize the nonlinear dynamics of velocity
less than Mach 5. Hence, HV will make the intercontinental and altitude, and then an input-output linearization model
voyage and global military attack promptly, and occupy a is obtained to facilitate the flight control system design.
pivotal role in space routine of civilian and military [1], [2]. On the basis of the input-output linearization model which
With the development of scramjet engine, the flight control could be regarded as a nonlinear dynamic inversion structure,
system considered as another crucial component of HV has nonlinear robust control [6] and sliding mode control [7], [8]
attracted much attention in recent years [3]. However, there were implemented. Furthermore, considering that the distur-
exist highly nonlinear dynamics and significant couplings bance observer widely used in engineering fields [9]–[12]
between propulsive and aerodynamic forces, so the flight con- could be employed to handle system uncertainties, composite
trol system design is still a challenging task due to the afore- predictive flight control [13] and anti-windup control [14]
mentioned inherent peculiarities and the devastating impacts provided superior uncertainties rejection and tracking per-
caused by the aerodynamic uncertainties and extraneous formance under uncertain conditions. Although the above
disturbances [4], [5]. nonlinear control methods based on the input-output

2169-3536
2017 IEEE. Translations and content mining are permitted for academic research only.
VOLUME 6, 2018 Personal use is also permitted, but republication/redistribution requires IEEE permission. 5927
See http://www.ieee.org/publications_standards/publications/rights/index.html for more information.
Y. MA, Y. CAI: Fuzzy Model Predictive Control Based Upon ANNDO for a Constrained HV

linearization method have provided superior control perfor- systems [16], [18], [25] exploited to deal with the constraints
mance, there exist strict requirements on the accuracy and of fuel equivalence ratio and elevator deflection, RMPC [29]
relative order of the model. Therefore, due to the cascade was utilized to explicitly handle the constraints of fuel equiv-
structure of HV dynamics, the strict-feedback formulations alence ratio, elevator deflection and AOA at the stage of
of the nonlinear dynamics were obtained for the purpose control system design. In addition, the lumped disturbances
that back-stepping methods [15], [16] were applied to design which mainly consist of the fuzzy modeling error, extraneous
control system. To solve the problem of ‘‘explosion of terms’’ disturbances, and model uncertainties caused by aerodynamic
existed in the traditional back-stepping method, tracking uncertainties were not taken into account in the control
differentiator [17] and dynamic surface control methodol- schemes [32], [33], but may degrade the control performance
ogy [18], [19] were integrated with back-stepping method. and destroy the stability of the system. Hence, handling
Besides, nonlinear disturbance observer [20] and extended constraints and suppressing lumped disturbances which are
state observer [21], [22] were employed to enhance the regarded as attractive issues are addressed in this paper, and
robustness against diverse uncertainties. Recently, instead consequently, a fuzzy model predictive control scheme based
of HV dynamics in the strict-feedback form, neural net- upon adaptive neural network disturbance observer is pro-
works (NNs) [23]–[25] were employed to approximate the posed on the basis of the equivalent disturbed fuzzy dynamic
nonlinear dynamics due to NNs’ strong nonlinear approx- model with the varying parameters which is constructed to
imation ability [26], and the control strategies based on approximate HV nonlinear dynamics. Considering that the
NNs achieved excellent robustness performance. Especially, general fuzzy model predictive control methods [34], [35]
a minimal-learning-parameter (MLP) [25] algorithm utilized merely obtained conservative control performance due to the
to simplify the control structure only regulated the maximum ignorance of the varying parameters, so inspired by the linear
norm rather than the elements of NN’s weight vector. Com- parameter varying model predictive control (LPVMPC) [36]
pared with the two methods mentioned to simplify the nonlin- which was derived by using parameter dependent Lyapunov
ear dynamics for designing control system, small perturbation function (PDLF), a fuzzy model predictive controller which
method employed to directly linearize the nonlinear system at takes the varying parameters into account and explicitly
a specific trim condition is more easily and commonly used handles the constraints of fuel equivalence ratio, elevator
in engineering areas [3]. Then on the basis of the linearized deflection and AOA is proposed on the basis of the equivalent
model obtained by the small perturbation method, diverse lin- fuzzy dynamic model. Furthermore, to counteract the effect
ear control methods, such as robust control [27], guaranteed of lumped disturbances directly, an adaptive neural network
cost control with poles assignment [28] and robust model pre- disturbance observer (ANNDO) based on the strong non-
dictive control (RMPC) [29], have been investigated to design linear approximation ability of NN is proposed to estimate
control system. However, because the linearized model could lumped disturbances, and then an additional compensator
only approximate the nonlinear dynamics in the neighbor- is formulated by integrating the estimations of lumped dis-
hood of the specific trim condition, one drawback to these turbances and the corresponding compensation gain matrix.
linear control methods is that expected control performance Therefore, the fuzzy model predictive control scheme based
cannot be achieved under the flight condition far from the upon ANNDO can make the velocity and altitude track their
specific trim condition. To overcome this drawback, Takagi- reference commands effectively in the presence of inevitable
Sugeno (T-S) fuzzy model utilized to approximate smooth lumped disturbances.
nonlinear systems in [30] and [31] has attracted tremendous The contributions of this paper are mainly summarized
attention recently. Then the T-S fuzzy model constructed as follows: 1. Because the equivalent fuzzy dynamic model
by a set of linearized models with fuzzy If-Then rules was with the varying parameters measured online can approach
investigated to approximate the nonlinear dynamics, and then the nonlinear dynamics of HV more accurately, the fuzzy
fuzzy control schemes, such as fuzzy guaranteed cost tracking model predictive control scheme based upon ANNDO takes
control [32] and fault-tolerant control [33] were proposed the measured parameters and the modeling error into account,
on the T-S fuzzy model. Nevertheless, the constraint task and provides more satisfactory tracking performance than
of the fuel equivalence ratio, elevator deflection and angle RMPC [29] developed on a single linear model; 2. Com-
of attack (AOA), which was not taken into account in the pared with the control schemes [32], [33], the fuzzy model
control schemes [32], [33], has been considered as a crucial predictive control scheme based upon ANNDO addresses
issue to be addressed directly in the control system design. the attractive issues of handling constraints and suppressing
It is because the fuel equivalence ratio for the propulsion lumped disturbances, and provides effective tracking per-
system is constrained to ensure normal operation of scram- formance of the velocity and altitude even when HV is in
jet engine without the influence of thermal choking, the the presence of lumped disturbances and constraints. The
elevator deflection is constrained due to the limitation of remainder of this paper is organized as follows. Section II
the control surface displacement, and AOA which mainly presents the problem formulation including model descrip-
determines the air flow properties into the scramjet engine of tion and fuzzy model construction. Then, the detailed proce-
HV is constrained to ensure that the performance of scramjet dure of the fuzzy model predictive control scheme based upon
engine is not degraded. Generally, in addition to auxiliary ANNDO is established in Section III. Finally, Section IV

5928 VOLUME 6, 2018


Y. MA, Y. CAI: Fuzzy Model Predictive Control Based Upon ANNDO for a Constrained HV

demonstrates the simulation studies, and Section V provides In actual flight, the aerodynamic coefficients are uncertain,
conclusions. so their uncertain forms considered in this paper are expressed
as follows
II. PROBLEM FORMULATION
A. MODEL DESCRIPTION
CL (α) = (1 + 1CL ) CL (α)
In order to facilitate controller design and stability analy- CD (α) = (1 + 1CD ) CD (α)
sis of HV, the model constructed by five order nonlinear CT (β) = (1 + 1CT ) CT (β)
differential equations was developed to describe the motion CM = (1 + 1CM ) CM (4)
of HV’s longitudinal dynamics at NASA Langley Research
Center [6]. Considering HV is affected by the extraneous dis- where CM = CM (α) + CM (q) + CM (δe , α); 1CL , 1CD , 1CT
turbances di (i = 1, 2, . . . , 5), then disturbed model including and 1CM denote the corresponding uncertain terms.
velocity V , flight-path angle γ , altitude h, AOA α and pitch
rate q is expressed as follows B. FUZZY MODEL CONSTRUCTION
(T cos α − D) µ sin γ In order to facilitate controller design, we define the state
1
V̇ =
m

r2
+ d1 vector as x = [x1 , x2 , x3 , x4 , x5 ]T = [V , γ , h, α, q]T , and
1
L + T sin α (µ − V 2 r) cos γ input vector as u = [β, δe ]T , then the disturbed model (1)
γ̇ = − + d2 with the aerodynamic forces (2) and (3) can be rewritten as
mV V r2
ḣ = V sin γ + d3 the following form of MIMO nonlinear system
α̇ = q − γ̇ + d4 ẋ = f (x) + g(x)u + d + 11
q̇ = Myy /Iyy + d5 (1) y = Cx (5)
where m, µ, Iyy denote vehicle mass, gravitational constant, where
moment of inertia, respectively; r = h + R is radial distance 
−D m − µ sin γ r 2
  
from Earth’s center, and R is radius of Earth; For the sake  L (mV ) − (µ − V 2 r) cos γ (V r 2 )


of systemic analysis, the lift L, drag D, thrust T , pitch- 
f (x) = 
 V sin γ


ing moment Myy are approximated by the following curve-  q − L (mV ) + (µ − V 2 r) cos γ (V r 2 ) 
  
fitted functions that depend on state variables ( V , γ , h, ρV 2 S c̄[CM (α) + CM (q) + CM ,1 (α)] (2I )

α, q) and control inputs (fuel equivalence ratio β, elevator
deflection δe ) = [f1 (x), f2 (x), f3 (x), f4 (x), f5 (x)]T ,
0.01288ρV 2 S cos α m
  
0
L = 0.5ρV 2 SCL (α)  0.01288ρVS sin α m 0 
D = 0.5ρV 2 SCD (α)
 
g(x) =  0 0 ;
 −0.01288ρVS sin α m
  
T = 0.5ρV 2 SCT (β) 0 
Myy = 0.5ρV 2 S c̄[CM (α) + CM (δe , α) + CM (q)] (2) 0 ρV 2 S c̄ce (2Iyy )
d = [d1 , d2 , d3 , d4 , d5 ]T
with the following aerodynamic coefficients simplified at the
nominal cruising flight denotes the extraneous disturbances imposed on HV;
2 11 denotes model uncertainties caused by the uncertainties
CD (α) = 0.645α + 0.0043378α + 0.003772
of the aerodynamic coefficients in (4); Because we con-
β<1

0.02576β sider the tracking task of the altitude and velocity in this
CT (β) =
0.0224 + 0.00336β β > 1 paper,the output of system is y = [V , h]T by defining
CM (α) = −0.035α 2 + 0.036617α + 5.3261 × 10−6 10000
C = .
CM (q) = (c̄/2V )q(−6.796α 2 + 0.3015α − 0.2289) 00100
CM (δe , α) = CM ,1 (α) + CM ,2 (δe ) In addition, the constraints of fuel equivalence ratio, eleva-
tor deflection and AOA considered in this paper are the same
CM ,1 (α) = −ce α, CM ,2 (δe ) = ce δe , CL (α) = 0.6203α as that adopted in [29], namely
(3)
xmin ≤ x4 ≤ xmax
where CL (α), CD (α) are the functions of α, and denote the
umin ≤ u ≤ umax (6)
coefficients of lift and drag; the thrust coefficient CT (β) is a
function of β; The moment coefficient consists of the coef- where xmin = −8◦ and xmax = 8◦ ; umin = [0, −20◦ ]T and
ficients CM (α), CM (q), CM (δe , α) which are the functions umax = [3, 20◦ ]T .
depending on α, q, and δe ; ρ, S, c̄ represent the air density, Note that the control goal for the cruise phase of HV is
reference area, and mean aerodynamic chord, respectively; to steer the altitude and velocity from the initial values to
ce denotes the scaling factor. The values of above parameters the desired trim conditions [3], so the reference command
are the same as that in [7]. vector yr = [Ve , he ]T considered in this paper is assumed

VOLUME 6, 2018 5929


Y. MA, Y. CAI: Fuzzy Model Predictive Control Based Upon ANNDO for a Constrained HV

as a given set of constants. According to yr , we could obtain


the equilibrium point (xe , ue ) at the desired trim conditions,
meanwhile, f (xe ) + g(xe )ue = 0 and yr = Cxe , then y is
driven to track yr by controlling system to the desired trim
conditions.
1 1
Define x̃ = [x̃1 , x̃2 , . . . , x̃5 ]T = x − xe and ũ = u − ue
as the shifted state vector and shifted input vector, then the
nonlinear model (5) is transformed into

x̃˙ = f̃ (x̃) + g̃(x̃)ũ + d + 11 (7)

Where f̃ (x̃) = f (x̃ + xe ) + g(x̃ + xe )ue and g̃(x̃) = g(x̃ + xe );


x̃min ≤ x̃4 ≤ x̃max , ũmin ≤ ũ ≤ ũmax , x̃min = xmin − xe,4 ,
x̃max = xmax − xe,4 , ũmin = umin − ue and ũmax = umax − ue .
Considering that HV’s longitudinal dynamics can be
approximated more accurately by a T-S fuzzy model [37]
than by a single linear model [29], so we construct a fuzzy
dynamic model to approximately describe the local nonlinear
dynamics of nonlinear model (7). In this paper, we select FIGURE 1. Block diagram of the proposed control system. (a) Overall
control scheme. (b) Architecture of ANNDO’s NN with adaptive laws.
x̃1 , x̃4 , x̃5 as three premise variables zj , j = 1, 2, 3, namely
z1 = x̃1 , z2 = x̃4 , z3 = x̃5 . Furthermore, assuming z1 = x̃1 ∈
[δx 1 , δ x̄1 ], z2 = x̃4 ∈ [δx 4 , δ x̄4 ] and z3 = x̃5 ∈ [δx 5 , δ x̄5 ],
Finally, the fuzzy dynamic model equivalent to the nonlin-
two fuzzy sets are chosen for each premise variable: fuzzy
ear model (7) is constructed as
sets ‘‘about δx 1 ’’ and ‘‘about δ x̄1 ’’ are denoted as F11 and
F12 , respectively; fuzzy sets ‘‘about δx 4 ’’ and ‘‘about δ x̄4 ’’ x̃˙ = A(ω)x̃ + B(ω)ũ + 1 (8)
are denoted as F21 and F22 , respectively; fuzzy sets ‘‘about 8 8
δx 5 ’’ and ‘‘about δ x̄5 ’’ are denoted as F31 and F32 , respectively. ωi (z)Ai , B(ω) = ωi (z)Bi , z =
P P
where A(ω) =
Then the membership functions are 1
 defined as F1 (z1 ) =
i=1 i=1
[z1 , z2 , z3 ]T and ω = [ω1 , ω2 , . . . , ω8 ]T ; Set i = 4(i1 −
1 − 1 1 + e−0.125[z1 +(δ x̄1 −δx 1 )/2] , F12 (z1 ) = 1 − F11 (z1 ),

1) + 2(i2 − 1) + i3 , i1 , i2 , i3 = 1, 2, and define ,ϕi (z) =
F21 (z2 ) = 1 − 1 1 + e−67.45z2 , F22 (z2 ) = 1 − F21 (z2 ),
 
8
i
F1i1 (z1 )F2i2 (z2 )F33 (z3 ), then we have ωi (z) = ϕi (z) ϕi (z)
P
F31 (z3 ) = 1 − 1 1 + e−46.35z3 , F32 (z3 ) = 1 − F31 (z3 ).
 
i=1
Moreover, let i = 1, 2, . . . , 8, and define µi1 , µi2 , µi3 as 8
which satisfy 0 ≤ ωi (z) ≤ 1 and ωi (z) = 1 ; In addition,
P
follows:
i=1
F1 , i = 1, 2, 3, 4 F2 , i = 1, 2, 5, 6 considering that the modeling error of fuzzy model is 12 =
 1  1
µi1 = µi2 = f̃ (x̃) + g̃(x̃)ũ − A (ω) x̃ − B (ω) ũ, so 1 = 11 + 12 + d
F12 , i = 5, 6, 7, 8, F22 , i = 3, 4, 7, 8,
denote the lumped disturbances primarily comprising three
F3 , i = 1, 3, 5, 7
 1
µi3 = categories: extraneous disturbances d, model uncertainties
F32 , i = 2, 4, 6, 8. 11 and modeling error of fuzzy model 12 .
Hence, the fuzzy dynamic model is constructed by eight Remark 1: Note that A(ω) and B(ω) are time varying due to
(23 ) fuzzy rules, and the i th rule is given as following form. the varying parameters ωi (z) which can be measured online,
Rule i: so compared with the single linear model adopted to represent
If z1 is µi1 , z2 is µi2 , z3 is µi3 the nonlinear model (7) in [29], the equivalent fuzzy dynamic
Then x̃˙ = Ai x̃ + Bi ũ + d + 11 model (8) can approach the dynamic motion of the nonlinear
where eight corresponding operating points are chosen as model (7) more accurately, and additionally, take modeling
δ x̃1 = [δx 1 , 0, 0, δx 4 , δx 5 ]T , δ x̃2 = [δx 1 , 0, 0, δx 4 , δ x̄5 ]T , error into account.
δ x̃3 = [δx 1 , 0, 0, δ x̄4 , δx 5 ]T , δ x̃4 = [δx 1 , 0, 0, δ x̄4 , δ x̄5 ]T ,
δ x̃5 = [δ x̄1 , 0, 0, δx 4 , δx 5 ]T , δ x̃6 = [δ x̄1 , 0, 0, δx 4 , δ x̄5 ]T , III. CONTROLLER DESIGN
δ x̃7 = [δ x̄1 , 0, 0, δ x̄4 , δx 5 ]T , δ x̃8 = [δ x̄1 , 0, 0, δ x̄4 , δ x̄5 ]T . In this paper, a fuzzy model predictive control scheme based
The linearization approach proposed in [38] is uti- upon ANNDO is proposed to design a composite controller
lized to obtain the corresponding linear models [Ai , Bi ]: u for the purpose that the altitude and velocity are con-
Ai = [ai1 , ai2 , ai3 , ai4 ,. ai5 ]T and air = ∇ f˜r (δ x̃i ) + trolled to track their reference commands effectively even
2 when HV suffers from inevitable lumped disturbances and
(f˜r (δ x̃i ) − δ x̃Ti ∇ f˜r (δ x̃i ))δ x̃i kδ x̃i k , r = 1, 2, . . . , 5, in the constraints of control inputs and AOA. As depicted in
which f˜r is the r th element of f̃ in (7), and ∇ f˜r (δ x̃i ) is the Fig. 1 (a), the composite controller u consists of nominal
gradient of f˜r at δ x̃i ; Bi = g̃(δ x̃i ). fuzzy model predictive controller unc and compensator ucc .

5930 VOLUME 6, 2018


Y. MA, Y. CAI: Fuzzy Model Predictive Control Based Upon ANNDO for a Constrained HV

For the equivalent fuzzy dynamic model in absence of lumped Assumption 2: Assuming 3 in (13) is bounded, and
disturbances, it is obvious that the varying parameters ωi (z) then there exists a small positive upper bound ξ such that
are measurable, so we firstly propose that the LPVMPC k3k2 ≤ ξ .
strategy [36] taking ωi (z) into account is utilized to design Define the observer error of the ANNDO (9) as e = x̃ − χ ,
the nominal fuzzy model predictive controller unc for dealing subtract (9) from (8), then we obtain the following observer
with the constraints explicitly. In fact, lumped disturbances 1 dynamic system
may deteriorate nominal performance of unc , or even lead to
ė + λe = 1̃ (14)
instability. Hence, in order to counteract the effects of lumped
disturbances, the ANNDO is constructed to estimate lumped Theorem 1: Consider the observer dynamic system (14)
disturbances for formulating the compensator ucc with the under Assumption 1 and Assumption 2, the adaptive laws of
compensation gain matrix κ(ω). w̄i and v̄j are proposed as follows
 T T
A. ANNDO CONSTRUCTION w̄˙ i = −w̃˙ i = η1 [φ(V̄ s)ei − φ 0 V̄ sei − θ1 w̄i ] (15)
˙v̄j = −ṽ˙ j = η2 [sj (eT W̄ T φ 0 ) − θ2 v̄j ]
Based on the strong nonlinear approximation ability of NN,
the following ANNDO depicted in Fig. 1 (b) is designed to where ei is the i th element of e, and sj is the j th element of
estimate the inevitable lumped disturbances 1 in the equiva- s; ηl and θl , l = 1, 2, are the positive adjusted coefficients to
lent fuzzy dynamic model(8). be designed properly.
The signals involved in the Lyapunov function (16)
χ̇ = λ(x̃ − χ ) + A(ω)x̃ + B(ω)ũ + 1̂ (9)
are semi-globally uniformly ultimately bounded since the
where χ is inner state vector of ANNDO; λ is the gain Lyapunov function is bounded for all initial conditions.
parameter, and satisfies λ > 21 ; 1̂ defined as the estimation Proof: Choose a Lyapunov function as
of 1 is the output of NN, namely 1 T 1 T 1 T
T T
V (e, W̃ , Ṽ ) = e e+ tr(W̃ W̃ ) + tr(Ṽ Ṽ ) (16)
1̂ = W̄ φ(V̄ s) (10) 2 2η1 2η2

where s = [1, x̃T ]T is the input of NN; V̄ = The time derivative along V (e, W̃ , Ṽ ) is expressed as
[v̄T1 , · · · , v̄Tj , · · · , v̄Tn1 ]T ∈ Rn1 ×n2 and W̄ = [w̄1 ,· · · ,w̄i , · · · , 1 T ˙ 1 T
V̇ (e, W̃ , Ṽ ) = eT ė + tr(W̃ W̃ ) + tr(Ṽ˙ Ṽ ) (17)
w̄n3 ] ∈ R(n2 +1)×n3 are the weight matrixes to be adjusted η1 η2
adaptively; w̄i is the i th column of W̄ , v̄j is the j th row T ˙ n3 T n1
w̃Ti w̃˙ i and tr(Ṽ˙ Ṽ ) = ṽ˙ j ṽTj ,
T
P P
of V̄ , and vp is defined as the p th column of V̄ ; φ(V̄ s) = Note that tr(W̃ W̃ ) =
i=1 j=1
.φ1 , · · · , φp , · · · , φn2 ] , in which φp is chosen as φp =
[1, T
substituting (13)-(15) into (17)yields
1 (1 + exp(−vTp s)).
V̇ (e, W̃ , Ṽ )
Assumption 1: According to the universal approximation n3 n1
∗ ∗ 1 X ˙i + 1
X
theorem [39], the ideal weight matrixes V̄ and W̄ are = eT (−λe + 1̃) + w̃Ti w̃ ṽ˙ j ṽTj
assumed to be bounded, η1 η2
and there exist positive upper bounds i=1 j=1
∗ 2 2 n3 n1
(0v̄ , 0w̄ ) such that v̄ ≤ 0 2 and w̄∗ ≤ 0 2 . The ideal X X
j i j v̄j i w̄i = −λkek2 + eT 3 + θ1 w̃Ti w̄i + θ2 v̄j ṽTj
∗T ∗T
estimation is 1∗ = W̄ φ(V̄ s), and then we have i=1 j=1
T T T 0 T T T
∗T ∗T + e (W̃ φ(V̄ s) − W̃ φ V̄ s) + e (W̄ φ 0 Ṽ s)
T T
1 = 1∗ + ε = W̄ φ(V̄ s) + ε (11) n 3 n1
X T T X T
− w̃Ti [φ(V̄ s)ei − φ 0 V̄ sei ] − [sj (eT W̄ φ 0 )]ṽTj
where ε is the approximation error, and there exists an suffi-
i=1 j=1
ciently small constant ε̄ such that kεk ≤ ε̄. (18)
∗ ∗
Define Ṽ = V̄ − V̄ and W̃ = W̄ − W̄ , then subtract (10) n1
T
from (11), we obtain the estimated error 1̃ T
Since eT (W̄ φ 0 Ṽ s) −
T
[sj (eT W̄ φ 0 )]ṽTj
P
= 0,
∗T ∗T T T j=1
1̃ = W̄ φ(V̄ s) − W̄ φ(V̄ s) + ε (12) n3
T T T T T T
w̃Ti [φ(V̄ s)ei −φ 0 V̄ sei ]−eT (W̃ φ(V̄ s)− W̃ φ 0 V̄ s) =
P
∗T T T
where φ(V̄ s) = φ(V̄ H s) + φ 0 Ṽ s + O
denotes the first ; φ0 i=1 2
∗T
derivative of φ(V̄ s), and OH is the higher order term in the 0, 2w̃Ti w̄i ≤ w̄∗i − kw̃i k2 ≤ 0w̄2 i − kw̃i k2 , 2v̄j ṽTj ≤
2
∗T
0v̄2j − ṽj and 2eT 3 ≤ kek2 + k3k2 ≤ kek2 + ξ , we

Taylor expansion of φ(V̄ s). Then 1̃ can be expressed as
∗T T ∗T T T T
have
1̃ = W̄ φ(V̄ s) + W̄ φ 0 Ṽ s − W̄ φ(V̄ s)
∗T V̇ (e, W̃ , Ṽ )
+W̄ OH + ε 1 1 X 3
1 X 1 n
2 n
T T T
= W̃ φ(V̄ s) + W̄ φ Ṽ s − W̃ φ V̄ s + 3 (13)0 T T 0 T ≤ −(λ − )kek2 − θ1 kw̃i k2 − θ2 ṽj + σV
2 2 2
i=1 j=1
T ∗T ∗T
where 3 = W̃ φ 0 V̄ s + W̄ OH + ε. ≤ −ζ V (e, W̃ , Ṽ ) + σV (19)

VOLUME 6, 2018 5931


Y. MA, Y. CAI: Fuzzy Model Predictive Control Based Upon ANNDO for a Constrained HV

where ζ = min {2λ − 1, θ1 η1 , θ2 η2 } > 0 and σV = the compensation gain matrix κ(ω) is properly designed as
n3 n1
θ1 P 2
2 0 w̄i + θ2 P 2
2 0v̄j + ξ2 . follows
n o−1
i=1 j=1
According to the standard Lyapunov extension theo- κ(ω) = − C[A(ω) + B(ω)K(ω)]−1 B(ω)
rem [40], solving the inequality (19) yields ·C[A(ω) + B(ω)K(ω)]−1 (25)
σV
 
0 ≤ V (e, W̃ , Ṽ ) ≤ V (e, W̃ , Ṽ )|t=0 − exp(−ζ t) Then we obtain the compensator ucc = κ(ω)1̂ to directly
ζ counteract the effects of lumped disturbances.
σV To sum up the above analysis, the fuzzy model predic-
+ , ∀t ≥ 0 (20)
ζ tive control scheme based upon ANNDO is summarized as
follows.
Therefore, e, W̃ and Ṽ are proved to be semi-globally uni-
Design Procedure:
formly ultimately bounded since V (e, W̃ , Ṽ )|t=0 is bounded.
Step 1. Select x̃1 , x̃4 , x̃5 as three premise variables zj ,
Moreover, e, W̃ and Ṽ can be sufficient small by selecting
j = 1, 2, 3, the equivalent fuzzy dynamic model (8) is
proper values of λ, θ1 , θ2 , η1 and η2 , so 1̃ is sufficient small,
obtained by defining the membership functions Fjl (zj ), l =
and there exists a sufficient small upper bound 01̃ such that
1, 2 and Rule i, i = 1, 2, . . . , 8, which are discussed exhaus-
||1̃|| ≤ 01̃ . Thus, the proof is complete.
tively in section II.B.
Step 2. At each control period, considering the nomi-
B. FUZZY MODEL PREDICTIVE CONTROLLER BASED
UPON ANNDO
nal fuzzy dynamic model with measurable  parameter ωi (z),
define a PDLF V (k + i|k) = x̃(k + i k)T P(ω)x̃(k + i k),
Considering the nominal fuzzy dynamic model in absence 8
ωi (z)P i , and then a state feedback control law
P
of lumped disturbances, A(ω) and B(ω) are linear combi- P(ω) =
8 i=1
ωi (z)Ai and
P
nations of Ai and Bi such that A(ω) = 8
ωi (z)K i is design to ensure that the constraints
P
i=1 K(ω) =
8 i=1
ωi (z)Bi in which the varying parameters ωi (z) (23),(24) and the following inequality are satisfied.
P
B(ω) =
i=1  2
8 V (k + i + 1|k) − V (k + i|k) = − x̃(k + i k) 2
satisfy 0 ≤ ωi (z) ≤ 1 and ωi (z) = 1. Accord-
P
 2
i=1 − ũ(k + i k) R (26)
ingly, the nominal fuzzy dynamic model could be regarded
as a polytopic uncertain system with the polytope  = Assuming that there exists an upper bound ϕ such that
Co {[A1 , B1 ], [A2 , B2 ], . . . , [A8 , B8 ]}, so according to the J∞ (k) ≤ V (k|k) ≤ ϕ, then we define symmetric matrices
LPVMPC strategy [36], we could design the nominal fuzzy Qi = ϕP −1 −1
i and K i = Y i Gi , Y i and Gi are fully parame-
model predictive controller terized matrices, and can be obtained by solving the follow-
ing convex optimization problem which minimizes ϕ with
unc = ũ + ue LMI constraints.
8
ũ = K(ω)x̃, K(ω) =
X
ωi (z)K i (21) min ϕ (27)
ϕ,Y i ,Gi ,Qi
i=1
1 x̃T
 
s.t ≥ 0, ∀i = 1, 2, . . . , 8 (28)
which minimizes the upper bound on the worst case perfor- x̃ Qi
mance criterion 
Gi + GTi − Qi ∗ ∗ ∗

 Ai Gi + Bi Y i Ql ∗ ∗ 
min max J∞ (k)  ≥ 0,
2 Gi ϕI
 1/2
ũ(k+i/k), i≥0  0 ∗ 
∞ h
 i R1/2 Y i 0 0 ϕI
x̃(k + i k) 2 + ũ(k + i k) 2
X 
J∞ (k) = 2 R
(22) ∀i = 1, 2, . . . , 8, ∀l = 1, 2, . . . , 8 (29)
i=0  
X ∗
≥ 0, ∀i = 1, 2, . . . , 8,
subject to Y Ti Gi + GTi − Qi
ũr (k + i k) ≤ ũmax,r , X rr ≤ ũ2max,r , r = 1, 2 (30)

r = 1, 2 (23)
 
Z

x̃4 (k + i k) ≤ x̃max (24) ∗
≥ 0,
(Ai Gi + Bi Y i )T C Tx Gi + GTi − Qi
where 2 and R are the symmetric weighting matrices.
∀i = 1, 2, . . . , 8, 2
Znn ≤ x̃max , n = 4. (31)
Furthermore, note that it is difficult to guarantee the stable
tracking of altitude and velocity only by the robustness of where C x = [0, 0, 0, 1, 0]; Symbol ∗ denotes transpose of the
the nominal fuzzy model predictive controller when HV is in lower block part in (29)-(31).
the presence of diverse lumped disturbances, so the proposed A toolbox for optimization in MATLAB [41] is used
ANNDO is constructed to estimate lumped disturbances, and to solve the convex optimization problem (27)-(31), and

5932 VOLUME 6, 2018


Y. MA, Y. CAI: Fuzzy Model Predictive Control Based Upon ANNDO for a Constrained HV

then we obtain the feasible decision variables, ϕ, Qi , Y i Invoking the compensation gain matrixκ(ω), we obtain
and Gi . Accordingly, the state feedback controller is ũ = B (ω) κ(ω) + I = 0, and the output y is represented as
K(ω)x̃, and the nominal fuzzy model predictive controller is
y = C[A (ω) + B (ω) K(ω)]−1 ẋ + Cxe
unc = ũ + ue .
Remark 2: According to the results in [36], the LMI (29) + C[A (ω) + B (ω) K(ω)]−1 B (ω) κ(ω)1̃ (37)
ensures that inequality (26) holds by applying the Schur In [36], it was proved that the state feedback controller
complement. Hence, the PDLF V (k|k) is strictly decreasing, ũ = K(ω)x̃ could asymptotically stabilize the closed-loop
and the closed-loop system is proved to be asymptotically system, so the state x ultimately converges to a steady state
stable. Meanwhile, the LMIs of (30) and(31) ensure that the such that lim ẋ → 0. In Theorem 1, we have proved
constraints of (23) and (24) hold. t→∞
Remark 3: Because the measurable parameters ωi (z) are that 1̃ is sufficient small, and there exists a sufficient
taken into account, the LPVMPC strategy derived by using small upper bound 01̃ such that ||1̃|| ≤ 01̃ . Let σ =
PDLF achieves less conservative results as compared with −C[A (ω) + B (ω) K(ω)]−1 1̃, and assume that there exists
RMPC adopted in [29] and derived by using a single quadratic a small positive constant σ̄ such that ||σ || ≤ σ̄ . Therefore
Lyapunov function. when t → ∞, y(∞) will be
Step 3. Define the observer error as e = x̃ − χ and set the y(∞) = yr + σ (38)
input of NN as s = [1, x̃T ]T , ANNDO (9) with the adaptive
laws (15) is proposed to estimate the lumped disturbances, Define ye = y − yr as the tracking error vector of velocity
then we obtain the compensator ucc = κ(ω)1̂ by inte- and altitude, so when t → ∞, ye ultimately converges
grating the estimation 1̂ and the compensation gain matrix into a small neighborhood ψ = ye | ye ≤ σ̄ around

κ(ω) in (25). Finally, the following composite controller u is zero. Hence, it is proved that the proposed composite con-
implemented. troller (32) could make the velocity and altitude track their
reference commands effectively even when HV is in the
u = K(ω)x̃ + ue + κ(ω)1̂ (32) presence of lumped disturbances and constraints.
Theorem 2: Consider HV longitudinal model (5) in the
IV. SIMULATION RESULTS
presence of lumped disturbances and constraints, the com-
To verify the effectiveness of the previously proposed
posite controller (32) designed by the fuzzy model predictive
method, we carry out two numerical simulations for HV
control scheme based upon ANNDO is proposed to guarantee
model (1)-(3) in the MATLAB environment where a fourth
that the velocity and altitude track their reference commands
order Runge-Kutta algorithm is employed with the fixed-
effectively.
step 0.01s. The detailed parameters of simulation model
Proof: Note that x̃ = x − xe , ũ = u − ue , the fuzzy
could refer to [7]. In both numerical simulations for the
dynamic model (8) equivalent to model (5) can be represented
cruise phase, the velocity tracks a step command of 100 ft/s
as
from initial value 14960 ft/s to final value 15060 ft/s,
ẋ = A(ω)x + B(ω)u − A(ω)xe − B(ω)ue + 1 (33) meanwhile, the altitude tracks a step command of 180 ft
from initial value 109820 ft to final value 110000 ft.
Furthermore, considering 1̃ = 1 − 1̂, the composite And then the initial and final trim conditions listed in
controller (32) is rewritten as Table 1 are obtained. For constructing the equivalent fuzzy
dynamic model(8), the he upper bounds and lower bounds
u = K(ω)x̃ + ue + κ(ω)1̂ of the premise variables are selected as δx 1 = −100,
= K(ω)(x − xe ) + ue + κ(ω)(1 − 1̃) (34) δ x̄1 = 0, δx 4 = −0.1, δ x̄4 = 0.1, δx 5 = −0.15
and δ x̄5 = 0.15. In the performance criterion(22), we
Substituting (34) into the model (33) yields select 2 = diag(0.005, 0.005, 0.5, 0.005,0.005) and R =
diag(0.01, 0.01). To set the proper parameters of ANNDO,
h i
ẋ = B(ω) K(ω)(x − xe ) + ue + κ(ω)(1 − 1̃)
we firstly select the nodes in each layer of NN as n1 = 6,
+ A(ω)x − A(ω)xe − B(ω)ue + 1 n2 = 11 and n3 = 5, then initialize the weights W̄ and V̄
as zero matrixes with appropriate dimensions. Accordingly,
= [A(ω) + B(ω)K(ω)] x − B(ω)K(ω)xe the adjustment coefficients of the adaptive laws in (15) are
− A(ω)xe + (B(ω)κ(ω) + I)1 − B(ω)κ(ω)1̃ (35) selected as η1 = η2 = 250, θ1 = 0.25 and θ2 = 0.01, and
additionally, the gain parameter in (9) is chosen as λ = 0.51.
Then we have Case 1. In order to verify the superiorities of the fuzzy
model predictive control scheme based upon ANNDO
x = [A(ω) + B(ω)K(ω)]−1 ẋ − [A(ω) + B(ω)K(ω)]−1
h i (FMPC-ANNDO), the RMPC [29] is introduced for compar-
· (B(ω)κ(ω) + I)1 − B(ω)κ(ω)1̃ + xe (36) ison. In the same nominal conditions without the considera-
tion of extraneous disturbances and model uncertainties, the
where I is an identity matrix. comparative simulation results under FMPC-ANNDO and

VOLUME 6, 2018 5933


Y. MA, Y. CAI: Fuzzy Model Predictive Control Based Upon ANNDO for a Constrained HV

FIGURE 2. Tracking results in Case 1. (a) Velocity. (b) Altitude.

FIGURE 3. Control inputs in Case 1. (a) Fuel equivalence ratio. (b) Elevator deflection.

FIGURE 4. Other states in Case 1. (a) Flight-path angle. (b) Pitch rate. (c) AOA.

TABLE 1. Initial condition for state variables and final trim condition for AOA vary in the given bounds, but FMPC-ANNDO makes
state variables and control inputs.
the velocity and altitude track their desired command more
quickly, in particular, provides much less fluctuation when the
altitude approaches its desired command. Moreover, because
FMPC-ANNDO employs the equivalent fuzzy dynamic
model to approach the nonlinear dynamics and takes the
modeling error into account directly, the tracking errors under
RMPC are shown in Figs. 2-4. It is observed from Fig. 2, FMPC-ANNDO are smaller than that under RMPC. In addi-
Fig. 3 and Fig. 4(c) that both methods achieve the satisfactory tion, Fig. 4 reveals that compared with RMPC, the response
tracking performance of velocity and altitude in the condi- curves of the flight-path angle, pith rate and AOA provided
tion that the fuel equivalence ratio, elevator deflection and by FMPC-ANNDO fluctuate with smaller ranges.

5934 VOLUME 6, 2018


Y. MA, Y. CAI: Fuzzy Model Predictive Control Based Upon ANNDO for a Constrained HV

FIGURE 5. Tracking results in Case 2. (a) Velocity. (b) Altitude.

FIGURE 6. Control inputs in Case 2. (a) Fuel equivalence ratio. (b) Elevator deflection.

FIGURE 7. Other states in Case 2. (a) Flight-path angle. (b) Pitch rate. (c) AOA.

Case 2. To facilitate this simulation case for practice flight, in [32] for comparison with FMPC-ANNDO. Figs. 5-9 show
the uncertainties of aerodynamic coefficients and the extra- the comparative simulation results under FMPC-ANNDO
neous disturbances are taken into account. The uncertainties and FGCTC. It is apparent from Fig. 5 that due to the
of aerodynamic coefficients in (4) are set as 1CL = 0.04, excellent estimations of lumped disturbances by ANNDO,
1CD = 0.04, 1CT = −0.04 and 1CM = 0.04. Addition- the proposed FMPC-ANNDO performs an excellent robust-
ally, the extraneous disturbances imposed on the nonlinear ness against lumped disturbances, and guarantees that the
system of HV are set as d1 = −8, d4 = 0.005sin(2t) when velocity and altitude asymptotically converge to their desired
t ≥ 10, and d5 = 0.05sin(0.2π t + π ) when t ≥ 13. On the commands such that the tracking errors are bounded with the
basis of the same fuzzy dynamic model, we introduce the sufficiently small regions. However, FGCTC cannot make the
fuzzy guaranteed cost tracking control (FGCTC) presented velocity and altitude track their desired commands any more

VOLUME 6, 2018 5935


Y. MA, Y. CAI: Fuzzy Model Predictive Control Based Upon ANNDO for a Constrained HV

FIGURE 8. Results of estimations and curve of kσ k in Case 2. (a)-(e) 1i and estimations 1̂i , i = 1, 2, . . . , 5. (f) kσ k and its upper bound σ̄ .

9. Adaptive adjustment results of NN’s weight matrixes in case 2. (a) w̄i , i = 1, 2, . . . , 5.



FIGURE
(b) v̄j , j = 1, 2, . . . , 6.

in the presence of lumped disturbances. Figs. 6 and Fig. 7(c) still tend to be stable under FMPC-ANNDO, whereas they
illustrate that FMPC-ANNDO dealing with constraints dur- are fluctuated with larger amplitudes under FGCTC.
ing control design procedure guarantees that fuel equivalence Figs. 8(a)-(e) show the estimations of lumped disturbances,
ratio, elevator deflection and AOA vary in the given bounds, which are obtained by ANNDO, and particularly, the estima-
which avoids the repeated parameter adjustment of FGCTC. tions of model uncertainties and modeling errors in altitude
Consider that FMPC-ANNDO’s compensator for suppress- and flight-path angle loops are depicted in Figs. 8 (b)-(c).
ing lumped disturbances is constructed by the estimations It is apparent that there exists an adaptive adjustment process
of lumped disturbances, so the control inputs, the elevator due to the parameters of ANNDO which need to be regulated
deflection in particular, are fluctuated subject to the vary- online at the beginning. Fig. 8(f) verifies that when t → ∞,
ing lumped disturbances. Accordingly, Fig. 7 shows that the σ in (38) is bounded, and there exists a small positive constant
flight-path angle, AOA and pith rate are also affected, but σ̄ = 0.11 such that kσ k ≤ σ̄ . Hence, the tracking errors

5936 VOLUME 6, 2018


Y. MA, Y. CAI: Fuzzy Model Predictive Control Based Upon ANNDO for a Constrained HV

of velocity and altitude are bounded such that they are suf- [11] W.-H. Chen, ‘‘Nonlinear disturbance observer-enhanced dynamic inver-
ficiently small. In addition, as shown in Fig. 9, we can see sion control of missiles,’’ J. Guid., Control, Dyn., vol. 26, no. 1,
pp. 161–166, Jan. 2003.
that the adaptive laws in (15) guarantee NN’weight matrixes [12] X. Y. Diao, Y. Jin, L. Ma, S. H. Ding, and H. B. Jiang, ‘‘Composite active
bounded. front steering controller design for vehicle system,’’ IEEE Access, vol. 5,
It can be concluded that compared with FGCTC, the pro- pp. 6697–6706, 2017.
[13] J. Yang, Z. Zhao, S. Li, and W. X. Zheng, ‘‘Composite predictive flight
posed FMPC-ANNDO provides the satisfactory tracking per- control for airbreathing hypersonic vehicles,’’ Int. J. Control, vol. 87, no. 9,
formance of velocity and altitude even when HV is in the pp. 1970–1984, Mar. 2014.
presence of lumped disturbances and constraints. [14] H. An, J. Liu, C. Wang, and L. Wu, ‘‘Disturbance observer-based anti-
windup control for air-breathing hypersonic vehicles,’’ IEEE Trans. Ind.
Electron., vol. 63, no. 5, pp. 3038–3049, May 2016.
V. CONCLUSION [15] B. Xu, F. Sun, H. Liu, and J. Ren, ‘‘Adaptive Kriging controller design for
In this paper, a fuzzy model predictive control scheme based hypersonic flight vehicle via back-stepping,’’ IET Control Theory Appl.,
vol. 6, no. 4, pp. 487–497, Mar. 2012.
upon ANNDO has been proposed for HV, where the attrac- [16] Q. Zong, F. Wang, R. Su, and S. K. Shao, ‘‘Robust adaptive backstepping
tive issues of handling constraints and suppressing lumped tracking control for a flexible air-breathing hypersonic vehicle subject to
disturbances have been addressed. On one hand, based on the input constraint,’’ Proc. Inst. Mech. Eng. G, J. Aerosp. Eng., vol. 229, no. 1,
pp. 10–25, Jan. 2015.
equivalent fuzzy model constructed to approach the nonlinear
[17] X. Bu, X. Wu, R. Zhang, Z. Ma, and J. Huang, ‘‘Tracking dif-
dynamics of HV, the proposed scheme taking the varying ferentiator design for the robust backstepping control of a flexible
parameters of fuzzy model into consideration explicitly deals air-breathing hypersonic vehicle,’’ J. Franklin Inst., vol. 352, no. 4,
pp. 1739–1765, Apr. 2015.
with the constraints of the fuel equivalence ratio, elevator
[18] Q. Zong, F. Wang, B. Tian, and R. Su, ‘‘Robust adaptive dynamic surface
deflection and AOA. On the other hand, due to the proposed control design for a flexible air-breathing hypersonic vehicle with input
ANNDO constructed to estimate the diverse lumped distur- constraints and uncertainty,’’ Nonlinear Dyn., vol. 78, no. 1, pp. 289–315,
bances excellently, an additional compensator is appended to Oct. 2014.
[19] D. D. Lei, T. T. Wang, D. Cao, and J. T. Fei, ‘‘Adaptive dynamic surface
enhance the robustness of the proposed method in order to control of MEMS gyroscope sensor using fuzzy compensator,’’ IEEE
counteract the effect of lumped disturbances directly. There- Access, vol. 4, pp. 4148–4154, 2016.
fore, the proposed method makes the velocity and altitude [20] G. H. Wu and X. Y. Meng, ‘‘Nonlinear disturbance observer based robust
backstepping control for a flexible air-breathing hypersonic vehicle,’’
asymptotically converge to their desired commands such that Aerosp. Sci. Technol., vol. 54, pp. 174–182, Jul. 2016.
the tracking errors are bounded with the sufficiently small [21] Z. Q. Pu, R. Y. Yuan, X. M. Tan, and J. Q. Yi, ‘‘Active robust
regions. The effectiveness of the proposed method is vali- control of uncertainty and flexibility suppression for air-breathing
hypersonic vehicles,’’ Aerosp. Sci. Technol., vol. 42, pp. 429–441,
dated by the simulation results. In this paper, all states of May 2015.
HV are assumed to be available, but flight-path angle and [22] Z. Pu, X. Tan, G. Fan, and J. Yi, ‘‘Uncertainty analysis and robust trajectory
AOA are difficult to be measured in practice. In our future linearization control of a flexible air-breathing hypersonic vehicle,’’ Acta
Astronaut., vol. 101, pp. 16–32, Sep. 2014.
research, we will improve our proposed method purely on the
[23] B. Xu, C. G. Yang, and Y. Q. Pan, ‘‘Global neural dynamic surface
basis of partially measurable states. tracking control of strict-feedback systems with application to hypersonic
flight vehicle,’’ IEEE Trans. Neural Netw. Learn. Syst., vol. 26, no. 10,
pp. 2563–2575, Oct. 2015.
REFERENCES
[24] B. Xu, ‘‘Robust adaptive neural control of flexible hypersonic flight vehi-
[1] F. R. Chavez and D. K. Schmidt, ‘‘Uncertainty modeling for multivariable- cle with dead-zone input nonlinearity,’’ Nonlinear Dyn., vol. 80, no. 3,
control robustness analysis of elastic high-speed vehicles,’’ J. Guid. Con- pp. 1509–1520, 2015.
trol Dyn., vol. 22, no. 1, pp. 87–95, 1999. [25] X. Bu, X. Wu, D. Wei, and J. Huang, ‘‘Neural-approximation-based robust
[2] M. A. Bolender, ‘‘An overview on dynamics and controls modelling of adaptive control of flexible air-breathing hypersonic vehicles with para-
hypersonic vehicles,’’ in Proc. IEEE Amer. Control Conf., St. Louis, MO, metric uncertainties and control input constraints,’’ Inf. Sci., vols. 346–347,
USA, Jun. 2009, pp. 2507–2512. pp. 29–43, Jun. 2016.
[3] B. Xu and Z. K. Shi, ‘‘An overview on flight dynamics and control [26] J. P. Cai, L. T. Xing, M. Zhang, and L. J. Shen, ‘‘Adaptive neural network
approaches for hypersonic vehicles,’’ Sci. China Inf. Sci., vol. 58, no. 7, control for missile systems with unknown hysteresis input,’’ IEEE Access,
p. 070201, Jul. 2015. vol. 5, pp. 15839–15847, 2017.
[4] A. A. Rodriguez et al., ‘‘Modeling and control of scramjet-powered hyper- [27] C. I. Marrison and R. F. Stengel, ‘‘Design of robust control systems for
sonic vehicles: Challenges, trends, and tradeoffs,’’ in Proc. AIAA Guid., a hypersonic aircraft,’’ J. Guid. Control Dyn., vol. 21, no. 1, pp. 58–63,
Navigat. Control Conf. Exhib. Honolulu, HI, USA, Aug. 2008, p. 6793. Feb. 1998.
[5] B. Fidan, M. Mirmirani, and P. A. Ioannou, ‘‘Flight dynamics and control [28] H. Y. Li, Y. L. Si, L. G. Wu, X. X. Hu, and H. J. Gao, ‘‘Guaranteed
of air-breathing hypersonic vehicles: Review and new directions,’’ in Proc. cost control with poles assignment for a flexible air-breathing hypersonic
12th AIAA Int. Space Planes Hyperson. Syst. Technol., Norfolk, VA, USA, vehicle,’’ Int. J. Syst. Sci., vol. 42, no. 5, pp. 863–876, May 2011.
Dec. 2003, p. 7081. [29] H. Gao et al., ‘‘Nonlinear disturbance observer-based model predictive
[6] Q. Wang and R. F. Stengel, ‘‘Robust nonlinear control of a hypersonic control for a generic hypersonic vehicle,’’ Proc. Inst. Mech. Eng., I, J. Syst.
aircraft,’’ J. Guid., Control, Dyn., vol. 23, no. 4, pp. 577–585, 2000. Control Eng., vol. 230, no. 1, pp. 3–12, Jan. 2016.
[7] H. Xu, M. D. Mirmirani, and P. A. Ioannou, ‘‘Adaptive sliding mode control [30] Z. Zhong, Y. Zhu, and T. Yang, ‘‘Robust decentralized static output-
design for a hypersonic flight vehicle,’’ J. Guid., Control, Dyn., vol. 27, feedback control design for large-scale nonlinear systems using
no. 5, pp. 829–838, 2004. Takagi-Sugeno fuzzy models,’’ IEEE Access, vol. 4, pp. 8250–8263,
[8] Q. Zong, J. Wang, and Y. Tao, ‘‘Adaptive high-order dynamic sliding 2016.
mode control for a flexible air-breathing hypersonic vehicle,’’ Int. J. Robust [31] Z. X. Zhong, Y. Z. Zhu, and M. Karimi, ‘‘Dynamic output feedback
Nonlinear Control, vol. 23, no. 15, pp. 1718–1736, Jul. 2013. fuzzy control of large-scale nonlinear networked systems: A two-channel
[9] T. D. Do, ‘‘Disturbance observer-based fuzzy SMC of WECSS without triggering approach,’’ IEEE Access, vol. 5, pp. 12428–12438, 2017.
wind speed measurement,’’ IEEE Access, vol. 5, pp. 147–155, 2017. [32] X. Hu, L. Wu, C. Hu, and H. Gao, ‘‘Fuzzy guaranteed cost tracking control
[10] Z. Q. Liu, ‘‘Ship adaptive course keeping control with nonlinear distur- for a flexible air-breathing hypersonic vehicle,’’ IET Control Theory Appl.,
bance observer,’’ IEEE Access, vol. 5, pp. 17567–17575, 2017. vol. 6, no. 9, pp. 1238–1249, Jun. 2012.

VOLUME 6, 2018 5937


Y. MA, Y. CAI: Fuzzy Model Predictive Control Based Upon ANNDO for a Constrained HV

[33] Q. K. Shen, B. Jiang, and V. Cocquempot, ‘‘Fault-tolerant control for YU MA was born in Shaanxi, China, in 1988. He
T-S fuzzy systems with application to near-space hypersonic vehicle with received the B.S. degree in automation engineering
actuator faults,’’ IEEE Trans. Fuzzy Syst., vol. 20, no. 4, pp. 652–665, from Chang’an University, Xi’an, China, in 2011.
Aug. 2012. He is currently pursuing the Ph.D. degree in con-
[34] S. Mollov, R. Babuska, J. Abonyi, and H. B. Verbruggen, ‘‘Effective trol science and engineering with Xi’an Jiaotong
optimization for fuzzy model predictive control,’’ IEEE Trans. Fuzzy Syst., University, Xi’an. His current research interests
vol. 12, no. 5, pp. 661–675, Oct. 2004. include nonlinear control system, model predictive
[35] T. Zhang, G. Feng, and J. Lu, ‘‘Fuzzy constrained min-
control, intelligent control, and flight control.
max model predictive control based on piecewise Lyapunov
functions,’’ IEEE Trans. Fuzzy Syst., vol. 15, no. 4, pp. 686–698,
Aug. 2007.
[36] N. Wada, K. Saito, and M. Saeki, ‘‘Model predictive control for lin- YUANLI CAI was born in Guizhou, China,
ear parameter varying systems using parameter dependent Lyapunov in 1963. He received the B.S. and Ph.D. degrees
function,’’ IEEE Trans. Circuits Syst., vol. 53, no. 12, pp. 1446–1450, in aerospace engineering from Northwestern Poly-
Dec. 2006. technical University, Xi’an, China, in 1984 and
[37] H.-N. Wu, Z.-Y. Liu, and L. Guo, ‘‘Robust L∞ -gain fuzzy disturbance in 1990, respectively.
observer-based control design with adaptive bounding for a hypersonic From 1993 to 1998, he was an Associate Pro-
vehicle,’’ IEEE Trans. Fuzzy Syst., vol. 22, no. 6, pp. 1401–1412, fessor with the School of Electronics and Infor-
Dec. 2014. mation Engineering, Xi’an Jiaotong University,
[38] M. C. M. Teixeira and S. H. Zak, ‘‘Stabilizing controller design for
Xi’an. From 1998 to 1999, he was invited as
uncertain nonlinear systems using fuzzy models,’’ IEEE Trans. Fuzzy Syst.,
a Senior Scholar with the Chinese Academy of
vol. 7, no. 2, pp. 133–142, Apr. 1999.
[39] F. L. Lewis, A. Yesildirek, and K. Liu, ‘‘Stabilizing controller design for Sciences. From 2003 to 2004, he visited the University of California at
uncertain nonlinear systems using fuzzy models,’’ IEEE Trans. Neural Riverside, Riverside, CA, USA. Since 1999, he has been a Professor with
Netw., vol. 7, no. 2, pp. 388–399, Mar. 1996. the School of Electronics and Information Engineering, Xi’an Jiaotong
[40] M. Krstic, I. Kanellakopoulos, and P. V. Kokotovic, Nonlinear and Adap- University. His research interests include guidance, control and dynamics of
tive Control Design. New York, NY, USA: Wiley, 1995. aircraft, intelligent information processing system and technology, modeling
[41] J. Lofberg, ‘‘YALMIP: A toolbox for modeling and optimization in and simulation of complex systems, and nonlinear theory and application.
MATLAB,’’ in Proc. IEEE Int. Symp. Comput. Aided Control Syst. Design, He is currently a Senior Member of the AIAA.
Sep. 2004, pp. 284–289.

5938 VOLUME 6, 2018

Vous aimerez peut-être aussi