Neurocomputing 199 (2016) 4049

Contents lists available at ScienceDirect

Neurocomputing
journal homepage: www.elsevier.com/locate/neucom

Adaptive ight control law based on neural networks and dynamic

inversion for micro-aerial vehicles
Romulus Lungu, Mihai Lungu n
University of Craiova, Faculty of Electrical Engineering, 107 Decebal Blvd., Craiova, Romania

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

Article history: The paper presents two new adaptive systems, for the attitude's control of the micro-aerial vehicles
Received 1 September 2015 (MAV's) insect type. The dynamic model describing the motion of MAV's with respect to the Earth tied
Received in revised form frame is nonlinear and the design of the new adaptive control system is based on the dynamic inversion
1 November 2015
technique. The inversion error is calculated with respect to the control law and two matrices (inertia and
Accepted 23 December 2015
Available online 16 March 2016
dynamic damping matrices) which express the deviation of the estimated matrices relative to the cal-
culated ones (the matrices from the nonlinear dynamics of MAV's) in conditions of absolute stability in
Keywords: closed loop system by using the Lyapunov theory. To completely compensate this error, an adaptive
MAV component (output of a neural network) is added in the control law. The system also includes a second
Dynamic inversion
order reference model which provides the desired attitude vector and its derivative. The two variants of
Neural network
the new adaptive control system are validated by complex numerical simulations.
& 2016 Elsevier B.V. All rights reserved.

1. Introduction
The results of the studies and research regarding the aero-
dynamics, the ight dynamics and the automatic control of the
MAV's (micro-aerial vehicles) are numerous. Very important
works for this research area have been obtained in [1] and [2].
Important results have been reported in the aerospace domain or
in other related areas: mechatronics, automation, or electronics
relative to the MAV's modeling, stabilization of the attitude, and
the ight control [212]. The micro-aerial vehicles can be regarded
sometimes as the physical models of the insects. Such mechanisms
generally consist of three subsystems: a command subsystem
(electrical engine or piezoelectric actuator), a wing actuation
equipment (the cinematic mechanism), and a controller. Consist-
ing of such advantages as limited volume, low mass, reasonable
cost and rapid transportation, MAV's have demonstrated success-
ful application in the military and civilian elds [13].
Progress has been also made towards: 1) the modeling and the
manufacturing of the miniaturized highly performing servo-
actuators (thorax of the MAV's insect type) to achieve the
wing's beat motion in the case of ying mini-robots and to com-
mand the MAV's motion through the modication of the wing's
beat and attack angles; 2) the manufacturing of miniaturized
n generated by the guidance system and to maintain steady condi-
Corresponding author.
E-mail addresses: romulus_lungu@yahoo.com (R. Lungu),
Lma1312@yahoo.com (M. Lungu).
sensors and transducers which merge the signals provided by the
accelerometers, gyros, optical or magnetic sensors; 3) the design
and software implementation of the linear or nonlinear observers
which serve the navigation and ight control systems [3,6,1416].
Because the atmospheric conditions and the dynamics of MAV's
are changing during the ight, it is difcult to use conventional
controllers. To design perfect conventional controllers, one has to
know the precise mathematical model of the system to be con-
trolled. Furthermore, the MAV's dynamics may vary with respect
to the altitude and the ight conditions; therefore, the adaptive
controllers are better choices. The adaptive control algorithms for
a large class of aerospace vehicles can be modied and adapted to
the stabilization and the ight control of the micro-aerial vehicles;
some of them are based on the usage of the dynamic inversion and
neural networks techniques [12,1719]. Generally, the dynamic
inversion relies on the philosophy of feedback linearization; the
plant nonlinearities are canceled and the closed loop plant
behaves like a stable linear system. The method is characterized by
simplicity in the control structure, ease of implementation, global
exponential stability of the tracking error and so on. On the other
hand, the strong point of the neural networks (NNs) is their
approximation ability, these being capable of approximating the
dynamics of unknown systems through learning.
The aim of the ight control system is to attain the commands
tions during ight [20]. A MAV insect type is highly susceptible
to atmospherics disturbances; in addition, the dynamics is highly

http://dx.doi.org/10.1016/j.neucom.2015.12.118
0925-2312/& 2016 Elsevier B.V. All rights reserved.
 R. Lungu, M. Lungu / Neurocomputing 199 (2016) 4049 41

nonlinear and time varying. Hence, neural networks based adap- The structure of the paper is the following one: the MAV's
tive ight controllers with the ability to achieve desired perfor- nonlinear dynamics is presented in the paper's Section 2; the
mance are essential for automating the MAV insect type. In order design of the new adaptive control system is achieved in Section 3;
to achieve this autonomous ight, adaptive ight controllers with in Section 4, complex simulations to validate the new designed
the ability to adapt to nonlinear dynamics of the MAV insect type adaptive control system have been performed and the obtained
are necessary. Thus the main objective of the paper is the design results are analyzed; nally, some conclusions are shared in Sec-
and software implementation of such adaptive controller using tion 5 of the paper.
either the quaternion vector or the vector of MAV's attitude; the
design of the control law is based on the dynamic inversion tech-
nique and Lyapunov theory, this process involving the calculation of 2. The dynamics of the micro-aerial vehicles
the controller's coefcients in conditions of stability.
Feedback linearization, in its various forms, is perhaps the most 2.1. The general dynamics of the micro-aerial vehicles
commonly employed nonlinear control method; to use it, all para-
metric plant uncertainties must appear in the same equation of the The dynamics of MAV's is generally nonlinear; in the case of the
state-space representation as the control (the main disadvantage of gliding ight, the dynamics is linear. The uncertainties cased by
the feedback linearization method). Feed-forward neural networks the MAV dynamic's incomplete knowing lead to the necessity of a
based on the back propagation learning algorithm have been used in robust controller. In these circumstances, the Lyapunov direct
[21]; the main disadvantage is that the neural networks require a method is used.
h iT
priori training on normal and faulty operating data. Other approa- Let us denote with  the attitude vector of
ches involves the usage of the time delay neural networks; a con- the MAV, where is the roll angle,  the pitch angle, and 
troller based on this type of neural networks has been designed in the ight direction angle. These angles express the angular posi-
[22], but its main drawback is related to the ight path track accuracy tion of the MAV (of the MAV's body tied frame oxyz ; ox  the
and to the fact that it is enable only under limited conditions. Several longitudinal axis (pointing the ight direction), oy  the axis
neural network control approaches have been proposed based on oriented towards the right wing, and oz  the axis perpendicular
Lyapunov stability theory [22,23]. The main advantage of these to the plane oxy and downward oriented with respect to the Earth
control schemes is that the adaptive laws were obtained from the tied frame (Darboux frame OXYZ) having the axis OX tangent to
Lyapunov synthesis and, therefore, guarantees the system's stability; the parallel and East oriented, OY  tangent to the meridian and
the disadvantage is that some conditions should be assumed; these North oriented, and OZ  the locus vertical line (Zenith oriented).
h iT
requirements are not easy to satisfy in practical control application Let us also consider _ _ _ _  the angular
p
[24]. Juang designed a new learning technique using a time delay
rate's vector of the MAV relative to the Earth tied frame. We
network or networks with back-propagation through time algo- h iT
rithms to control the landing [25]; the main drawbacks are: 1) the denote with b x y z  the vector of MAV's angular
number of hidden units was determined by trial; 2) the convergence rates with respect to the axes of the oxyz frame. The connection
between the above presented vectors is: p _ W  1 ;
time is high. Gain scheduling for PID controllers is the common b b
adaptive control technique used in [26]. An advantage of the gain _
W p W ; where the matrix W has the form:
scheduling technique is that it allows the parameters to be changed 2 3
1 0  sin
quickly with change in the plant dynamics. However, the application 6
W 40 cos sin cos 7
5: 1
of the technique is limited due to the absence of a learning process.
0  sin cos cos
The number of neural network based adaptive controllers
designed for MAV's ight control is extremely small; one of the The vector of the resultant moments acting upon the micro-
few such neural network based controller is presented in [20], but aerial vehicle is [14]:
the real-time implementation of this controller architecture is not
! !a !g !v !a
practical due to the following reasons: 1) the feedback from the M b M b M b M b M b; 2
identier model for the iterative change in the controller outputs !a !g !v
requires at least one sample time and 2) the computation time to where M b ; M b ; M b are the aerodynamic, the weight, and the
!g !v
obtain the new set of plant inputs is considerably high for multiple dynamic damping components, respectively; because M b and M b
! !a
iterations of training. In the time being, none of the neural net- have very small values, we can make the approximation M b M b :
work based adaptive controllers for MAV's ight control was !al !ar
We denote with F b and F b  the vectors of the aerodynamic
designed for the attitude's control of the MAV's insect type; this !
forces produced by the left and the right wings and with r l and
motivates our work. The recent studies and research have culmi- !
r r  the position vectors of the left and right wing's pressure
nated in the development of efcient ying robots, but, from our
centers with respect to the MAV's mass center (point o); the
information, none of them has adaptive controllers based on the
following equation connects these variables:
attitude or the quaternion vectors, Lyapunov theory, dynamic
compensators, reference models, and neural networks; this is !a !al !ar !a ! !al ! !ar
F b F b F b ; Mb r l  F b r r  F b : 3
achieved in this paper, being interesting to see if such adaptive
controllers can guarantee the control of MAV's attitude. Thus, our The vectors of the aerodynamic forces and moments produced
goal is to design new adaptive systems for the control of the by the two wings have the expressions [1]:
micro-aerial vehicles insect type by using the attitude vector or 2 l 3 2 3
F d cos L F rd cos R  F ld cos L F rd cos R
the quaternion one. Taking into account the advantages of these 6 l 7 a 6 7
elements (NNs, dynamic inversion, reference models etc.) and the F aa 6 7 6
4 F d cos L  F d cos R 5; M a r a 4  F d sin L  F d cos R 5;
r l l 7
l r l r
fact that, till now, no paper deals with the control of MAV's by Fl Fl Fd  Fd
means of neural networks, dynamic inversion concept, linear 4
dynamic compensators, reference models, and Lyapunov theory,
the present paper represents an absolute novelty in the search with r a  the position vector of the two wing's resultant aero-
!a !l !r !l !r
area of controller's design for the MAV's ight. dynamic force F a ; F l ; F l  the lift forces, F d ; F d  the wing's
42 R. Lungu, M. Lungu / Neurocomputing 199 (2016) 4049

drag forces, ?L and ?R  the angles of the wings in the where

beat plane (l, L left, r, R right). Using Eq. (4), one gets 2 3
 0 q3 q2
[1]: M ab  M? r ab F aa M? M aa ; where M? is the transformation   6 q  q1 7
1 x T 1 T x
F q I 3 q qq  1 q q I 3 ; q 4 3 0 5;
(rotation) matrix of the wing in the beat plane, while r ab is the posi- 2 2
!a  q2 q1 0
tion vector of the force F a relative to the mass center of the micro-
aerial vehicle. 13
The dynamics of the MAV's rotation is described by the equa- I 3 is the (3  3) identity matrix.
tion [14]: By time derivation of the Eq. (12), we get:
! _ F_ q ; q_ ;
K b ! ! !a q F q 14
b  K b M b; 5
t where F_ q ; q_ is obtained by time derivation of Eq. (13), i.e.:
with K b J b b  the rotation kinetic moment and J b  the inertia 1h x i
moment matrix with respect to the axes of the MAV's tied frame: F_ q; q_ q_ 2qq
_ T  q_ T qI 3 : 15
2
2 3
J xx 0  J xz By elimination of the vector between Eqs. (12) and (14), we
6 0 0 7
Jb 4 J yy 5: 6 obtain:
J xz 0 J zz
_ F_ q; q_ F  1 qq;
q F q _ 16
  1
The dynamics of the MAV's rotation (Eq. (5)) may be also By left multiplying the Eq. (16) with m F q ; the fol-
written as: lowing equation results:
     
K_ b  a
b K b Mb ; 7 m F  1 qq m _ m F  1 qF_ q ; q_ F  1 q q_ ; 17
with K b J b b J b W _ ; W  matrix whose form will be pre-    
or, by elimination of m m between Eqs. (9) and
_
sented later, while b has the form in [1], i.e.:

(17), having in mind that _ 12 F  1 q q_ ; we get:
2 3
0  z y M qq N q; q_ q_ M ab ; 18
b 6
4 z 0  x 7 _ :
5 W 8
where
 y x 0
  h   i
M q m F  1 q; N q; q_ n _ ;
_  M qF_ q; q_ F  1 q:
The second term from  the left
 side of Eq. (7) is negligible;
  _ or
Eq. (7) becomes: J b W _
Jb W 19
b
   
_ _ Concluding, the dynamics of the MAV is described either by the
m n ; b ; 9
Eq. (9) or the Eq. (18). Generalizing, the nonlinear dynamics of the
with MAV may be described by the equation:
        C yy Dy; y_ u; 20
_ J W
m Jb W ; n ; _ ;
_ ; 10
b
with u  the input vector and y  the output vector. Thus,
and    
2 3 u M ab ; y ; C y C m ;
     
0 0  _ cos _ n ; _ _
  6           7 Dy ; y_ D ; 21
_ _ 6  _ cos cos  _ sin sin 7
W ; 60 _ sin 7:
4       _    5
0  cos
_  sin cos  cos sin
_ or
u M ab ; y q; C y C q M q; Dy; y_ Dq; q_ N q; qq;
_ 22
   
2.2. The dynamics of the micro-aerial vehicles using the quaternion _
with m and n ; having the forms (10) and M q ; N
theory q ; q of forms (19).

Now, by using the quaternion theory (Euler rotational theory),

the MAV's attitude with respect to the Earth tied frame OXYZ 3. Design of the adaptive control system
may be modied through the rotation of MAV (considered to be
rigid solid) around a proper axis (Euler axis) with the angle ; the 3.1. The design of the general control law and of the adaptive control
components of the quaternion vector q~ are [4]: system

q1 1 sin ; q 2 sin ; q3 3 sin ; q4 cos ; 11
2 2 2 2 2 To compensate the inuence of the unknown or partially
known nonlinearities from the dynamics of MAV's, the conven-
where 1 ; 2 ; 3 are the directory cosines of the vector
  tional controllers should be replaced with adaptive architectures
1 2 3 T relative to the inertial frame; the following equa- based on dynamic inversion and neural networks. The design of
tions are veried: 21 22 23 1 ; q21 q22 q23 q24 1 : The rst three
  the general control law and of the neural network based control
components q1 ; q2 ; q3 of the vector q~ form the quaternion system (two variants) is presented in this section of the paper.
h iT
vector q which veries the equations: q q1 q2 q3 sin 2 and In order to obtain a linear control system of the MAV's attitude,
2 3 2 3 2 _ 3
0  3 2 1 we choose the control law [15]:
6  1 5; with p 4 2 7
7 6 6 7
q_ q; 4 3
x x 0 _ 5 4 _ 5:
u C yv Dy; y_ ; 23
 2 1 0 3 _
Another form for the differential equation of the quaternion's with a proportional-derivative (P.D.) pseudo-control law pro-
kinematics is [5]: vided by a linear dynamic compensator, i.e.:
q_ F q; 12 v kp e kd e_ v^ r ; v^ r y d ; 24
 R. Lungu, M. Lungu / Neurocomputing 199 (2016) 4049 43

with e yd  y ; while the variables having the index d are the To calculate the error ; buy using the Eqs. (23) and (30), we
imposed (desired) ones; the component v^ r y d leads to the con- get:
y d ; issue that will be justied later; kp and kd are
vergence y-  
C yy C yv C y v^ u^  Dy; y_ : 31
diagonal matrices having on their principal diagonal the positive
elements kp i ; kd i ; i 1; 3 : Replacing C y and Dy ; y_ forms (28) into Eq. (31), we
The system (20) may be described by the nonlinear input-state obtain:
and state-output equations [12]: h i
C  1 y C~ yv^ D~ y; y_ : 32
x_ f x; u; y hx; 25
With these, we obtain the system in Fig. 2 equivalent to the
where x is the system's state vector satisfying the conditions of the
one in Fig. 1. To calculate the matrices C~ and D ~ ; the method of
Hypothesis 1 in [17], i.e.: yr hr y ; y_ ; u; hr ddthr and h
r

u 0 ;
i

hr Lyapunov functions can be used. The system in Fig. 2 (the com-

i 0; r  1 ; u a 0 ; with r 2  the relative degree of the
ponent v^ a in this stage of design is neglected) has the input and
system.
the output y ; being described by the state equation:
According to (20) equivalent to the system described by (25)
 T  T
and to (23), it yields: E_ AE B; E E1 E2 e e_ : 33
y yr hr y; y;
_ u v: 26
The variable E_ 2 is successively expressed as: E_ 2 e y d  y
From (26) one obtains:     24
yd  v^ y d  vpd y d  kp e  kd e_  : Thus, the matri-
 
1
_ v:
u hr y; y; 27 0 I 0
ces A and B have the forms: A Dc A BDc ; A
  0 0 I
The automatic control of the vector y (based on the dynamic 0 I 0 h i
inversion technique) is achieved by means of the Eqs. (20), (24), ;B ; with Dc kp kd ; all the elements of the
0 0 I
and (27); the block diagram of the resulting control system is matrices A and B are 3  3 matrices; A is the matrix of the linear
presented in Fig. 1. system described by Eq. (26).
In order to effectively implement the control law (23), C y and We choose the Lyapunov function:
Dy ; y_ must be known with high precision; in reality, these
functions are not fully known. Let us consider that C^ y and D ^ 1 T 1

V a ET PE a~ a~ ; 34
y ; y_ are the estimates of the matrices C y and Dy ; y_ with the 2
approximation errors: with P (symmetric and positive dened matrix) solution of the
T
Lyapunov equation A P PA Q ; the matrix Q is chosen as a
C~ y C^ y  C y; D~ y; y_ D
^ y; y_  Dy; y_ : 28
symmetric and positive dened matrix (in particular it can be the
As a consequence, the control law (23) becomes: unity matrix). is a diagonal matrix whose elements are the gain
  coefcients, a~ a^  a ; with a^  the estimate of the vector a ;
^ y ; y_ h^  1 y ; y_ ; v^ ;
u^ C^ y v^ D 29
r h iT h iT h iT
a C ij Dij ; a^ C^ ij D ^ ij ; a~ C~ ij D~ ij ; i; j 1; 3; 35
with v^ of form (24). By dynamic inversion, the Eq. (23) gets the
form: where C ij and Dij are the elements of the matrices C y and Dy ; y_ ;
y v; v v^ ; 30 ^ ij are the elements of the matrices C^ y and D
C^ ij and D ^ y ; y_ ; while
C~ ij and D ~ ij are the elements of the matrices C~ y and D ~ y ; y_ ;
where is the vector of dynamic approximation error (error
respectively.
resulting from the approximation of the function hr y ; y_ ; u );
  By time derivation of the Eq. (34) and taking into account the
x ; u x y ; u y ; y_ ; u hr y ; y_ ; u  h^ r y ; y_ ; u^ : Eq. (33), we get:
In steady regime, imposing that y- y d -0 ; we obtain the
1  
component v^ r y d : V_ a  ET Q E ET PB a_~ a:
T 1
~ 36
To deduce the dependency between u and u^ ; we use the Taylor 2
series expansion of the function (27) around the set of variables By means of the notation E ET PB ; introducing the para-
  1 ^ 1  _ ^  d 1
y ; y_ ; v^ ; we get: _ _ meterization matrix Y  solution of the equation

u hr y; y;v h r y; y; v dvhr y; y; v
  1  
^
jv v^ v  v^ u^ dv^ hr y ; y_ ; v^ U :
d Y a~ E; 37

Fig. 1. The automatic system for the control of the vector y by using the dynamic inversion technique.
44 R. Lungu, M. Lungu / Neurocomputing 199 (2016) 4049

Fig. 2. The equivalent block diagram for the system in Fig. 1.

Fig. 3. The block diagram of the subsytem for the calculation of the error .

with of form (32), the Eq. (36) becomes: therefore, it is not sure that h^ r -hr : The block diagram of the
subsystem for the calculation of the error is presented in Fig. 3.
1
V_ a  ET Q E Y a~ a_~ a:
T 1
~ 38 To compensate this error, in the control law an adaptive compo-
2
nent va provided by a neural network (feed-forward type in this
We impose the fulllment of the condition:
paper) must be added; adding also a robustness component v ; we
Y  a_~
T 1
; 39 obtain: v^ a v  va : The expression of the component provided by
the neural network (NN ) has been borrowed from [16]:
which, by right multiplying with ; gets the form: c

a_~  Y T ; 40  
va W T V T ; 41
because : With (39), the Eq. (38) becomes V_ a  12ET Q E r
T

0 : Thus, the closed loop system is global asymptotically stable:  1

with  the sigmoid function having the form z 1 e  bz ;
E-0 (e and e_ tend to zero in the same time; y-yd and y- _ y_ d ). h iT
 
~
Also, a-0 ^
a-a : b b1 bn1 1 ; n1  the number of neurons in the neural
network's input layer, bj ; j 1 ; n1 1  the activation potentials
3.2. The design of the neural network based component of the gen-
of the neurons,  the input vector of NNc,
eral control law
  h iT h iT
From (37), when E-0 E-0 and a-0 ~ ; it results that it is not
sure that the approximation error converges to zero and,
1 v^ Tk yTk 1 I1 I2 I n1 ; 42
 R. Lungu, M. Lungu / Neurocomputing 199 (2016) 4049 45

h i h i imposed V_ d o 0 ; this system is described by the equation E_ AE

v^ k v^ t
T
v^ t  d v^ t  2d ; yTk yt yt  d yt  2d ;  
 B v^ a v^ r ; v^ a v  va : Taking into account this equation,
43 with the function f W ; V borrowed from [28], and imposing
with I j ; j 1 ; n1  the inputs of the neural network and d  the f_ W ; V 0 ; we obtained the Eq. (44) and V_ d ET PE o 0 ; P
T
simulation step [18]; it can also be regarded as a delay of the being the solution of the equation A P PA  Q : Thus, the
signals v^ t and yt ; in the simulations area, we considered d _
convergence E-0 is achieved; e and e_ simultaneously tend to zero
0:05 s : The two matrices in the pair W; V are the weights of the (y-yd and y-y_ d ).
_
) and the solutions of the
feed-forward neural network (NNc
equation's system [12]:
h   i 3.3. The design of the reference model
_  kW 2  0 V T E kW  W 0 ; W 0 W 0;
W

h i In Fig. 4 we present the block diagram of the adaptive system

V_  kV 2EW T 0 kV  V 0 ; V 0 V 0; 44 for the control of the vector y, based on the dynamic inversion
technique, with linear dynamic compensator (P.D. type), neural
with 0 ddz jz z0 ; kW ; kV and k  positive constants; E ET PB is the network, and reference model (command lter). The reference
training matrix of the NNc. The elements of the matrix W are the
model is a second order one; it provides the components (signals)
weights of the links between the n2 neurons of the hidden layer
yd ; y_ d ; y d ; y; y_ ; and y must converge to yd ; y_ d ; y d : The transfer
and the n3 neurons of the output layer; the elements of matrix V
matrix of the reference model is:
are the weights of the links between the n1 neurons of the input
layer and the n2 neurons of the hidden layer. The dimensions of
20
the two matrices are n2 1  n3 and n1 1  n2 ; respectively. H m s I ; 46
The robustness component of the control law is calculated by
0 0 s 20 3
s2 2
means of the formula borrowed from [19]:
with 0 0:7 and 0 2 f 0 ; f 0 150 s1 :
  E For the determination of the matrices kp and kd ; we choose the
vT kz Ff F E ke E; 45
E elements kp1 kp2 kp3 and kd1 kd2 kd3 as solutions of the

n o characteristic equations of the liner system in Fig. 4 having the
W 0
with kz and ke  positive gains, F ; F2f trace F T F
0 V
r F ; Ff is the Frobenius norm of the matrix F, while F is the
ideal matrix of the neural network. v should ensure the bound-
edness of the neural network's matrices W and V. The robustness
component (45) ensures the boundedness of the neural network's
matrices W and V. This equation has been deduced by Calise et al.
and used in the adaptive control of different aerospace vehicles,
using the dynamic inversion and neural networks [18,27].
The adaptive control law va  Eq. (41), with W and V solu-
tions of the system (44), together with the robustness component
v ; must compensate the approximation error of the function hr :
the adaptive control law (41) results from the condition that the
 
closed loop system (having the input va  v v^ r and the
 T
output E e e_ ) is global asymptotically stable; for this we
have chosen the Lyapunov function V d 12ET PE f W ; V and we
Fig. 4. The block diagram of the adaptive system for the control of the vector y ; based on the dynamic inversion technique, with linear dynamic compensator, neural
network, and reference model.
input yd and the output y ; with v^ a 0 and 0 ;
s2 kdi s kpi 0; i 1; 3: 47
Imposing
 that
 these equations have negative roots, the solu-
tions kpi ; kdi result. For example, we can choose kpi 20 and
kdi 20 0 ; i 1 ; 3 :
In Fig. 5 we present the subsystems for the calculation of the
matrices C y and Dy ; y_ using the Eqs. (29) and (30); we will
work with the mean values of the states variations (z ; z  the
mean value of z ; while z  the variation of the variable z); we
also must remark that in the previous calculations, the signs
and  have been omitted. The justication for the usage of the
mean values is related to the fact that the physical variables are
harmonic functions of period T 1=f 0 :
46 R. Lungu, M. Lungu / Neurocomputing 199 (2016) 4049

Fig. 5. The subsystems for the calculation of the matrices C y and Dy; y_ for: a) y ; b) y q :

4. Numerical simulation results
To study the functionality of the new designed adaptive control
systems, we consider a MAV insect type [2] having the mass m0
10  1 kg ; the following numerical values also characterizes the
MAV:
J xx 2 U10  7 kg m2 ; J xz 6 U10  7 kg m2 ;
We also considered the values yc 031 and a~ 0 0:0200  0:057
50:107100:15800:0854  0:12000:0384  0:086400:21060:1188 :
We software implemented in Matlab/Simulink the architecture
in Fig. 4 (two variants: a) y and b) y q) and we obtained
the time histories in Figs. 6 and 7.
From Fig. 6a we remark the cancel ofthe following variables:

the variations of the three Euler angles ; ; and of

J yy 18 U 10  7 kg m2 ; J zz 8 U10  7 kg m2 ; their corresponding angular rates _ ; _ ; _ ; this means

2 3 that the Euler angles (roll, pitch, yaw) and their corresponding
0:2 0 0:6 
6 angular rates tend to their desired values - d ; - d ; -
Jb 4 0 1:8 0 7 510
6
kg m2 :
d : On the other hand, it is conrmed that, the three components
0:6 0 0:8 of the vectors v ; v^ pd ; u^ ; v^ ; v ; va become null in steady regime
Using the formula kpi 20 and kdi 20 0 ; i 1 ; 3 ; we (in steady regime all the commands are canceled) Fig. 6b. Similar
determine the diagonal matrices kp and kd ; with these, the matrix conclusions can be drawn for the second variant of the system in
Dc is calculated and, then, the matrices A and B are determined; Fig. 4 the characteristics in Fig. 7; beside the variables plotted for
then, choosing the matrix Q 0:1 I 6 ; the solution of the Lyapunov the rst variant, we also provide here the three components of the
h iT
T
equation A P PA  Q is obtained. For the neural network (NNc) quaternion vector q q1 q2 q3 (Fig. 7a); the convergences
of the adaptive control systems we have chosen three layers of q1 -q1d ; q2 -q2d ; q3 -q3d are achieved with very small errors.
neurons: an input layer having n1 6 neurons, a hidden layer The adaptive system for the control of MAV's attitude without
having n2 7 neurons, and an output layer having n3 3 neurons. using the quaternion vector has smaller time regime period than
The delay d has been considered d 0:05 s : The matrices W 0 the system for the control of MAV's attitude which uses the qua-
W 0 and V 0 V 0 are null matrices; the other constants used for ternion vector. For the rst system, the variations of the three
the software implementation of the neural network are: kW kV Euler angles and of their corresponding angular rates have no
0:5 ; k 10 ; kz 0:01 ; ke 0:04 ; F 10 ; I 12 ; using these overshoots, while, for the second variant, the variations of the
values, we calculate the matrices of the neural network (W and V) angles and angular rates are characterized by overshoots. The
solution of the system (44), the adaptive signal va by means of characteristics in Figs. 6 and 7 prove the new architecture's sta-
Eq. (41), with of form (42), and having the form z bility and its small overshoots.
 1
1 e  bz ; the vector of activation potentials
 is chosen of form: In our paper, the modeling and the control of the MAV's atti-
b 1 0:9 0:8 0:7 0:6 0:5  0:4 : The initial  values of the tude is achieved either by means of the Euler angles or by the
vectors y and y_ d are: yd 0  2 3 2 deg and y_ d 0 quaternion vector. Comparing our results with the ones in the
 
d
2 5 3 deg=s ; the initial values of the vectors _
 y and y are: y paper [7], we can remark that in our work the nonlinearitie's
0  1:3  5 2 deg and y_ 0h 2 5 3 deg=s ; while the approximation error and the adaptive command are calculated
p i
initial quaternion vector is q T 0  0:55 0:7 1  0:552  0:72 . by using the estimates C^ y and D
^ y ; y_ of the matrices C y and
 R. Lungu, M. Lungu / Neurocomputing 199 (2016) 4049 47

Fig. 6. The time histories of the variables in Fig. 4 for the control of the MAV's attitude .
48 R. Lungu, M. Lungu / Neurocomputing 199 (2016) 4049
Fig. 7. The time histories of the variables in Fig. 4 for the control of the MAV's attitude by using the quaternion vector q.
Dy ; y_ or by using C~ y and D ~ y ; y_ (the components of the
matrix a) ~ from Eq. (20) of the MAV's dynamics; a~ is the solution of
the Eq. (40), with Y  the parameterization matrix ensuring that
the Lyapunov function (34) is negatively dened (asymptotically
stable system). The adaptive command in [7] is modeled by a
different neural network (with only one neuron's input layer and
one output layer); the output vector consists of the Euler angles.
The NN weight's matrices in [7] are not calculated as in this paper
 T
with respect to the vector E ; i.e. with respect to E e e_ ; but
with respect to the component vpd (the output of a linear dynamic
compensator); this leads to higher and faster jumps of the Euler's
angles. In our work, the dynamics of the states and command
variables are much slower, this being an advantage.
At rst sight, we considered in the simulation studies only the
initial state errors e0 and e_ 0 : Actually, the disturbances are
within e and e_ ; thus in E and E ; i.e. in the weights matrices W and
V the components of the command (41) and solutions of the Eq.
(44). Also, the disturbances (through e and e_ ) intervene in the
obtaining of the signal (see Fig. 3). Therefore, we can conclude
that the disturbances affecting the MAV's dynamics are taken into
account as well as the uncertainties which can be found in hr
 h^ r (with hr  known or partially known nonlinear function).
Because the adaptive control law is designed such that the closed
loop system is stable both in the absence and in the presence of
the disturbances, the robustness of the system is guaranteed.
5. Conclusions
In this paper we presented two neural network based systems
for the adaptive control of the MAV's attitude: by using the atti-
tude vector or the quaternion vector. The dynamic model
describing the motion of MAV's with respect to the Earth tied
frame is nonlinear and the design of the new adaptive control
system (having two variants) is based on the dynamic inversion
technique. The inversion error is calculated with respect to the
pseudo-control vector v^ and the matrices C~ y and D
~ y ; y_ which
^
express the deviation of the estimated matrices C y and D ^ y ; y_
 R. Lungu, M. Lungu / Neurocomputing 199 (2016) 4049
relative to the calculated ones C y and Dy ; y_ from the nonlinear [13] C.X. Miao, J.C. Fang, An adaptive three-dimensional nonlinear path following
description of the MAV, by imposing that the closed loop system is
global asymptotically stable (with the usage of the Lyapunov
function). The adaptive component v^ a of the pseudo-control v^
must compensate the error : The new adaptive system consists of
a linear dynamic compensator, a reference model and a neural
network providing the adaptive component v^ a ; the neural net-
work is trained by the attitude error vector. We obtained the
adaptive system for the attitude's control of the MAV's, its sub-
systems being the subsystem for the calculation of the approx-
imation error and the subsystem for the calculation of the
matrices C y and Dy ; y_ ; in two variants: y and y q : We (11) (2006) 18651869.
software implemented in Matlab/Simulink environment all the
components of the new adaptive system and we obtained the time
histories of the main variables; the obtained characteristics prove
the new architecture's correctness, stability, and its small
overshoots.
Acknowledgments
This work was partially supported by the project Computa-
tional Methods in Scientic Investigation of Space, Project no. 72/
29.11.2013, of the Romanian National Authority for Scientic
Research and Program for Research Space Technology and
Advanced Research STAR. It is also supported by the Grant no.
89/1.10.2015 of the Romanian National Authority for Scientic
Research and Innovation and CNCS UEFISCDI, Project code PN-II-
RU-TE-2014-4-0849.
References
[1] X. Deng, L. Schenato, W.C. Chung, S. Sastry, Flapping ight for biomimetic
robotic insects: Part. I system modeling, IEEE Trans. Robot. 22 (2006)
776786.
[2] X. Deng, L. Schenato, W.C. Chung, S. Sastry, Flapping ight for biomimetic
robotic insects: Part. II system modeling, IEEE Trans. Robot. 22 (2006)
789800.
[3] J. Fei, C. Batur, Robust adaptive control for MEMS vibratory gyroscope, Int. J.
Adv. Robot. Syst. 7 (4) (2010) 7782.
[4] J. Wie, B. Lu, Feedback control logic for spacecraft eigenaxis rotations under
slew rate and control constrains, J. Guid. Control Dyn. 18 (6) (1995).
[5] P. Tsiotras, Stabilization and optimality results for the attitude control pro-
blem, J. Guid. Control Dyn. 19 (14) (1996) 772779.
[6] X. Deng, L. Schenato, S. Sastry, Model identication and attitude control for a
micromechanical ying insect including thorax and sensors models, in: Pro-
ceedings of the IEEE International Conference on Robotics and Automation,
Taipei, Taiwan, 2003, pp. 11521157.
[7] B. Cheng, X. Deng, A neural adaptive controller in apping ight, J. Robot.
Mechatron. 24 (4) (2012) 602611.
[8] P. Chirarattananon, N.O. Perey-Arancibia, R.J. Wood, Wing trajectory control
for apping-wing micro robots using combined repetitive and minimum-
variance adaptive methods, in: Proceedings of the American Control Con-
ference Fairmout Queen Elizabeth, Montreal, Canada, 2012, pp. 38313838.
[9] P. Chirarattananon, K.Y. Ma, R.J. Wood, Adaptive control for takeoff, hovering
and landing of a robotic y, in: Proceedings of the International Conference on
Intelligent Robots and Systems, Tokyo, Japan, 2013, pp. 38083815.
[10] H. Mahjoubi, K. Byl, Steering and horizontal motion control in insect-inspired
apping-wing mavs: the tunable impedance approach. in: Proceedings of the
American Control Conference Fairmout Queen Elizabeth, Montreal, Canada,
2012, pp. 901908.
[11] H. Rifai, N. Marcand, G. Poulin, Attitude and position control of a apping
micro air vehicle. In: Intech (Ed.), Aerial Vehicle, 2009, pp. 555580.
[12] M. Lungu, Sisteme de conducere a zborului (Flight Control Systems), Sitech
Publisher, Craiova, 2008.
49
method for a x-wing micro aerial vehicle, Int. J. Adv. Robot. Syst. 9 (206)
(2012) 18.
[14] R. Lungu, L. Sepcu, M. Lungu, Four-bar mechanism's PD and neural adaptive
control for the thorax of the micromechanical ying insects, J. Dyn. Syst.,
Meas. Control (Trans. ASME) 137 (5) (2015).
[15] I.P. Murphy, Modeling and Control of Flapping Wing Robots (Master of Science
thesis), Virginia Polytechnic Institute University, Blacksburg, Virginia, 2013.
[16] M. Sitti, Piezoelectrically actuated four-bar mechanism with two exible links
for micromechanical ying insect thorax, IEEE/ASME Trans. Mechatron. 8 (1)
(2003) 2636.
[17] A. Isidori, Nonlinear Control Systems, Springer, Berlin, 1995.
[18] A.J. Calise, E.N. Johnson, M.D. Johnson, J.E. Corban, Applications of adaptive
neural-networks control to unmanned aerial vehicles, J. Harbin Inst. Technol. 3
[19] Q. Wang, B. Zhang, R. Liu, Nonlinear piezoelectric behavior of ceramic bending
mode actuators under strong electric elds, J. Appl. Phys. 86 (6) (1999)
33523360.
[20] D. Li, Z. Lu, X. Chen, Modeling and design adaptive double neural network
controller for eight-rotor micro aircraft vehicle, Int. J. Control Autom. 7 (11)
(2014) 169180.
[21] T. Wagner, J. Valasek, Digital autoland control laws using quantitative feedback
theory and direct digital design, J. Guid. Control Dyn. 30 (5) (2007) 13991413.
[22] J.O. Jang, G.J. Jeon, A parallel neuro-controller for DC motors containing non-
linear friction, Neurocomputing 30 (2000) 233248.
[23] S. Seshagiri, H.K. Khalil, Output feedback control of nonlinear systems using
RBF neural network, IEEE Trans. Neural Netw. 11 (2000) 6979.
[24] R.J. Wai, Development of new training algorithms for neuro-wavelet systems
on the robust control of induction servo motor drive, IEEE Trans. Ind. Electron.
49 (6) (2002) 13231341.
[25] J.G. Juang, K.C. Cheng, Application of neural networks to disturbances
encountered landing control, IEEE Trans. Intell. Transp. Syst. 7 (4) (2006)
582588.
[26] S. Bouabdallah, A. Noth, R. Siegwart, PID vs. LQ control techniques applied to
an indoor MicroQuadrotor, in: Proceedings of the IEEE International Con-
ference on Intelligent Robots and Systems, 2004, pp. 546551.
[27] E.N. Johnson, A.J. Calise, Pseudo-control hedging: a new method for adaptive
control. Navigation Guidance Control Technology Workshop. November, 2000,
pp. 12.
[28] R. Mori, S. Suzuki, Neural network modeling of lateral pilot landing control, J.
Aircr. 46 (2009) 17211726.
Romulus Lungu was born in Bistrita Nasaud, Romania,
in 1952. In 1976 he graduated the University of Craiova,
Faculty of Electrotechnics, specialization: Automatics
and Computers. In 1991 he received the Ph.D. degree
(Aerospatiale Engineering domain). Since 1997 he is
Professor at the University of Craiova, Avionics Division.
He is the founder of the specialization Board Equip-
ments and Devices at the University of Craiova and
since 2009 he is Ph.D. leader in the domain of Aero-
spatiale Engineering. His interest elds are: automatics
of the ying objects, rocket's control, auto-pilots,
gyroscopic equipment and systems etc.
Mihai Lungu was born in Craiova, Romania, in 1980. In
2004 he graduated the University of Craiova, Faculty of
Electrical Engineering, specialization: Board Equip-
ments and Devices. In 2008 he received the Ph.D.
degree in the domain of Aerospatiale Engineering (Ph.
D. thesis: Complex Adaptive and Optimal Systems for
the Stabilization, Navigation and Control of the Flying
Objects). Since 2008 he is Lecturer at the University of
Craiova, Electrical, Energetic, and Aerospatiale Engi-
neering Department. His elds of interest include: the
ight control systems, automatic pilots, gyroscopic
equipment and systems etc.