Sunteți pe pagina 1din 6

2015 European Control Conference (ECC)

July 15-17, 2015. Linz, Austria

Real-time Implementation of Mixing Adaptive Control on Quadrotor


UAVs
Kemal BÜYÜKKABASAKAL 1 , Barış FİDAN 2 , Aydoğan SAVRAN 1 and Nasrettin KÖKSAL2

Abstract— In this study, a novel multiple model adaptive parallel estimation schemes involving parameter identifiers
control scheme is designed and implemented for quadrotor or Kalman filters can be used for each candidate controller.
unmanned aerial vehicles (UAVs). The proposed scheme involves Mixing adaptive control is a novel robust MMAC approach
a mixing adaptive controller that blends a set of pre-designed
linear quadratic controllers. A particular goal of the design under the large parametric uncertainties [12]. It blends the
is guaranteeing robustness in lateral motion against modeling candidate controller signals considering the estimation of
uncertainties. The designed controller scheme is tested via real- unknown parameter. By using a mixing function, contribution
time experiments on Quanser Qball-X4 UAVs. The experimental of each candidate controller to the control input of the plant
results verify the efficiency of the proposed scheme. is determined. One of the main advantage of mixing adaptive
I. INTRODUCTION control is that, it gives opportunity to analyze the stability
of the system by using LTI tools. Mixing adaptive control
Dynamics of quadrotor vehicles are nonlinear and the is described in [12] with details for continuous time, and
corresponding states are coupled, which makes control de- extended to discrete time in [14].
sign for them more challenging. In the existing studies on In this study, we design a mixing adaptive controller for
quadrotor control, researchers have mostly focused on sta- attitude dynamics of the quadrotor. Our goal is to perform
bility control. Some classical methods such as proportional- lateral tracking of reference trajectories under inertial uncer-
integral-derivative (PID) and linear quadratic (LQ) control tainties.
are used for attitude stabilization [1]. Full state backstepping
technique for stabilizing the roll/pitch angles and increasing II. MODELLING OF QUADROTOR UAV S
the tracking performance is presented in [2]. In another A. Quadrotor Dynamics
study [3], an integral backstepping control is proposed
In this section, we derive equations of motion to be used
for controlling the attitude, altitude and position of the
in our control design later, considering the Qball-X4 (Fig. 1)
quadrotor. In [4], a feedback linearization based controller
as the benchmark system.
is designed together with high order sliding observer against
disturbances. In [5], a model predictive robust controller is
developed for path following tasks. In [6], a sliding mode zb
observer is applied in presence of model uncertainties and ψ

disturbances. An experimental model predictive controller for T3 T2


M3
reference tracking is designed in [7]. [8] presents a linear M2

matrix inequality based controller gain synthesis. A nonlinear T1 T4


adaptive controller for attitude regulation is presented in [9]. M1 M4

In this paper, we follow a multiple model adaptive con-


troller (MMAC) [10]–[12] design approach, considering the φ
xb yb
changes in the quadrotor system environment, aiming to θ
blend outputs of a set of controllers, each of which is mg
z
designed to provide desirable stability and performance x y

properties for a certain parametric setting of the system Global frame


environment. The MMAC approaches typically can be in two
different form regarding generation of the control signal. The Fig. 1. The Qball-X4 quadrotor UAV [15]
first one blends the signals of all candidate controllers. The
other MMAC form switches between candidate controllers. Quadrotors consist of four rotors on a rigid frame [16].
In the blending approach [11]–[13], candidate control signals They work as two pairs of propellers. These pairs rotate
are blended based on pseudo-probability calculations. Also in opposite directions. Quadrotor performs its motion with
1 Kemal the change of speed in these rotors. The whole quadrotor
Buyukkabasakal and Aydogan Savran are with Electrical
and Electronics Engineering Department, Ege University, Turkey motion dynamics is a 6 DOF system defined by the pitch,
<kemal.buyukkabasakal, aydogan.savran> roll, yaw angles (θ , φ , ψ respectively) and the global position
@ege.edu.tr [xc , yc , zc ]T of the frame.
2 Barış Fidan and Nasrettin Köksal are with Department of Mechanical
and Mechatronics Engineering, University of Waterloo, Canada <fidan, The corresponding nonlinear system equations are derived
nkoksal>@uwaterloo.ca in [17] using Newton-Euler methods as follows;

978-3-9524269-3-7 ©2015 EUCA 3597


where θ , φ , ψ are small. Linearized pitch, roll, yaw and
l(T1 −T2 ) (Jy −Jz )ψ̇ φ̇ altitude dynamics are given by the unified equation
θ̈ = Jx + Jx − dθ θ̇ ,
        
l(T3 −T4 ) (Jz −Jx )ψ̇ θ̇ η̇i 0 1 0 ηi 0 0
φ̈ = + − dφ φ̇ , η̈i  = 0 0
Jy Jy σi  η̇i  +  0  ug i + ξi  . (4)
v̇i 0 0 −w vi w 0
Kψ (T1 +T2 −T3 −T4 ) (Jx −Jy )ψ̇ φ̇
ψ̈ = Jz + Jz − dφ θ̇ , 2Kl 2Kl
(1) where η1 = θ , η2 = φ , η3 = zc , η4 = ψ, σ1 = Jx , σ2 = Jy ,
4K K
ẍc = (T1 +T2 +T3 +T4 )(sinψsinφ +cosφ sinθ cosψ)
− dxc ẋc , σ3 = 4KM , σ4 = Jz , ξ1 = ξ2 = ξ4 = 0, ξ3 = −g, ug 1 = uθ ,
ψ
m
ug 2 = uφ , ug 3 = uzc , ug 4 = uψ .
(T1 +T2 +T3 +T4 )(−sinφ cosψ+cosφ sinθ sinψ)
ÿc = m − dyc ẏc ,
uθ , uφ , uz and uψ in (4) are virtual control inputs for the
(T1 +T2 +T3 +T4 )(cosφ cosθ ) quadrotor dynamics. The actual PWM control signals are
z̈c = m − g − dzc żc , calculated by
where T1 , T2 , T3 , T4 are thrust forces; Jx , Jy , Jz are rotational
u1 = uθ + uψ + uz , u2 = −uθ + uψ + uz
inertia matrices; l is the distance between the center of (5)
u3 = uφ − uψ + uz , u4 = −uφ − uψ + uz
gravity and the propeller; dφ , dψ , dθ , dxc , dyc , dzc are drag
coefficients, and Kψ is thrust-to-moment gain. For disturbance rejection, we augment an integrator to (4),
At typical low speed operations, (1) can be simplified by and redefine the state variables, leading to
ignoring inertial, drag and coriolis effects [17]. Then, the ẋi = Ai xi + Bi ug i + ρi ,
simplified nonlinear equation system is obtained as follows:
xi = [ (ηi − ηi d ), ηi − ηi d , η̇i , vi ]T ,
R
l(T1 −T2 ) l(T3 −T4 ) Kψ (T1 +T2 −T3 −T4 )
θ̈ = , φ̈ = , ψ̈ = ,
Jx Jy Jz       (6)
0 1 0 0 0 0
(T1 +T2 +T3 +T4 )(sinψsinφ +cosφ sinθ cosψ) 0 0 1 0  0 0
ẍc = m Ai = 
0 0 0 σi  , Bi =  0  , ρi = ξi  .
    
(2)
(T1 +T2 +T3 +T4 )(−sinφ cosψ+cosφ sinθ sinψ) 0 0 0 −w w 0
ÿc = m ,
C. Uncertainty Effects and Problem Definition
(T1 +T2 +T3 +T4 )(cosφ cosθ )
z̈c = m −g The uncertainties in the inertia values for pitch and roll
dynamics, Jx and Jy , affect σi terms in the linearized dynam-
In [15], thrust model of each actuator is given by
ics (6). For the brevity, we drop the subscript i in (6) and
w rewrite this dynamics model in a form reflecting the inertial
Ti = K
ui , i = 1, .., 4, (3)
s+w uncertainty as
where ui is the pulse width modulation (PWM) voltage ẋ = A(α)x + Bug + ρ (7)
applied rotor i, K is the armature gain, and w is the bandwidth
of the actuator. The control scheme for the benchmark Qball- where α is an uncertain parameter representing the inertial
X4 system based on (3) is shown in Fig.2 uncertainty. The focus of our adaptive control design is per-
forming trajectory tracking tasks robustly to the uncertainty
in the system matrix A.
III. MIXING ADAPTIVE CONTROL DESIGN
A. Mixing Adaptive Control Approach
In this section, we summarize the mixing adaptive control
approach introduced in [12], [14]. Consider a linear SISO
plant which includes unknown parameters and disturbance,
y = G(s, α)u + d, G(s, α) = GN (s, α)(1 + 4m (s)),
(8)
N(s, α)
Fig. 2. Control scheme for Qball-X4 GN (s, α) =
D(s, α)
where α is unknown plant parameter, y is system output,
B. Linearization GN (s, α) is the transfer function of the nominal plant with
We linearize the nonlinear equations of motion (2) using unknown parameters, 4m (s) is the multiplicative model
the approximations sin θ ≈ θ , sin φ ≈ φ , sinψ ≈ ψ and uncertainty and d is the disturbance on the system. The
cosθ ≈ cosφ ≈ cosψ ≈ 1 for the typical operation points control objective is to regulate the output y to zero.

3598
As shown in [12], [14], to design an adaptive mixing
control scheme, there are some conditions to be satisfied:
The system needs to be such that (1) interval of the unknown
parameter α is known; (2) GN (s, α) is strictly proper; (3)
4m (s) is proper and analytic in Re(s) ≥ −δ2 0 ; (4) D(s, α) is
a monic polynomial with known degree.
A mixing adaptive control scheme has three main com-
ponents: (i) candidate controllers, (ii) adaptive law, and (iii)
mixing scheme. Each of the candidate controllers
Fig. 3. Mixing adaptive control scheme
u j = C j (s), j = 1, 2, ..., n (9)

is designed for meeting the desired control objective for a the observed data [18, p. 192], least squares algorithm with
pre-defined range Ω j of the uncertain parameter α. These parameter projection is used instead of gradient algorithm.
controllers are designed off-line. Let us consider the system in (7). Since the A(α) matrix
Adaptive law estimates the unknown parameter α. Mixing includes unknown parameter, mixing adaptive scheme is
scheme uses the estimation of α for blending candidate designed by considering estimation of unknown parameter
controllers. In [12], [14] gradient based method is presented α.
for the adaptive law. Assume that we have a priori knowledge on bounds of α
Mixing scheme, by using the α estimation, calculates the
necessary control signal for the system by blending candidate α ∈ Ω = [ᾱmin , ᾱmax ]. (14)
controllers [12]. For each range Ω j = [a j − b j , a j + b j ] of α, Let us define n subsets for the candidate controllers.
a mixing function is selected in the form Ω1 =[ᾱmin , α1 max ], Ω2 =[α2 min , α2 max ], Ω3 =[α3 min , α3 max ],

α −aj
 ..., Ωn =[αn min , ᾱmax ]
ψ j (α) = ψ (10) where α p min , α p max are lower and upper bounds for each
b
( −1 j candidate controller set with the condition;
ψ(ᾱ) = e 1−ᾱ 2 , if |ᾱ| < 1 (11) α(p+1) min < α(p) max for p ∈ ϒ = [1, 2, ..., n].
0 , otherwise Mixing functions corresponding to each subset of can-
didate controllers are selected as common given by
Candidate controller feedback gains are calculated using ψ1 , ψ2 , ψ3 , ..., ψn .
the formula Then set of candidate controllers are given as;
ψ j (α)
β j (α) = m , (12)
∑ j=1 ψ j (α) Λ , {C p } p∈ϒ (15)

where β j (α) = 0 if θ ∈
/ Ω j . The final output of the mixing where C p is the candidate controller corresponding to set
adaptive control scheme is given by Ωp.
Assume that candidate controllers are designed by using
m
state-feedback approach in the proposed scheme.
u(α) = ∑ u j (s)β j (α) (13)
j=1 Control law for candidate controllers are;

B. Proposed State-Space Design C(p) = −kT (p)x (16)

In [12], [14], mixing adaptive control is designed and where k(p) is state feedback gain vector for candidate
discussed on the plants with unknown parameter and multi- controller and p ∈ ϒ.
plicative perturbation. Also, [12] is studied based on regula- Let us consider the system defined in (7) with unknown
tion problem in multiple model adaptive control. A reference parameter α.
tracking problem is discussed for the discrete time systems in
     
0 1 0 0 0 0
[14]. The mixing adaptive controllers are designed by using 0 0 1 0  0 0
transfer function design in both [12] and [14]. ẋ = 
0 0 0 α  x +  0  ug + ξ 
     (17)
In this paper, we design a state-space approach based 0 0 0 −w w 0
mixing adaptive control scheme for reference tracking for
Assume that (17) can be rewritten in the linear parametric
the systems with unknown plant parameters.
form in [19]:
This paper contributes the mixing adaptive control theory
z = αµ (18)
in the sense of state-space approach and reference tracking.
State space approach based design is realized and imple- where α is the vector with all unknown parameters, z and µ
mented on a real system in this paper. Also, the gradient are signals which are measurable.
based algorithm with parameter projection is used as adaptive Since states x1 to x3 are measurable states of x and they
law both in [12] and [14]. In this study, since least squares are in diagonal form, we define the adaptive law using the
algorithm is less affected from the noise and inaccuracies in 3rd and 4th row of equation (17).

3599
We also selected common bump functions for these con-
ẋ3 = αx4 + ξ (19) trollers. Bump functions for pitch and roll controllers are
represented as ψ1 , ψ2 , ψ3 , ..., ψn respectively.
ẋ4 = −wx4 + wug (20)
To design the adaptive law for quadrotor system, para-
If the Laplace expression of (20) is written in (19) and metric models of each dynamics are derived. In this study
both sides of equation are filtered with a stable filter, the least squares algorithm with parameter projection is used as
linear parametric form of (17) is obtained as follows. adaptive law.
h sx i h wug i In the design of adaptive law for pitch and roll dynamics,
3
=α (21) let us recall the the state-space equation given in (6) and
s+λ (s + λ )(s + w)
| {z } | {z } (21). Then, linear parametric model for pitch/roll model can
z µ be expressed in form of (18) as follows:
where λ > 0 is a design constant. z= s
α = 1J , 2KLw
(s+λ ) η̇i , µ= (s+λ )(s+w) ugi (25)
For estimating the α in (21), least squares estimation
algorithm is used. Since we have priori knowledge about the where i = 1, 2.
bounds of α, we can constrain the estimation values within Altitude and Yaw Controller Design:
the bounds by combining the least squares adaptive law with Altitude and yaw controllers are designed based on the
projection operator. The proposed algorithm for the adaptive state feedback method considering the control law in (24).
law is as follows: Controller gains are calculated using linear-quadratic regula-
α̂˙ = PrΩ (pε µ) (22) tor. Since we do not have any uncertainty in these dynamics,
( controller gains hold constant during the experiments.
µ2
β p − p2 m2 , i f ᾱmin < α̂ < ᾱmax
ṗ = s (23) IV. IMPLEMENTATION AND EXPERIMENTAL
0, otherwise SETUP
where ε = z−ẑ
m2s
with ms = 1 + µ 2 , PrΩ (.) is a projection In this section, we will give the controller design param-
operator to keep α̂ in Ω. eters used during the experimental implementation.
In [15], theoretical value of Jx and Jy , which exists in
C. Design of Individual Controllers
A matrix of generic model, is given as 0.03. During the
Controllers for pitch, roll, yaw and altitude sub-dynamics experimental studies, we estimated the value of 1/J in
are designed considering the linearized dynamics in (6). the adaptive law, which will be explained in subsection
For increasing the lateral tracking performance of quadro- below. By considering these estimation, pitch and roll mixing
tor, since the pitch and roll dynamics directly determine the scheme updates the control law and supplies the necessary
lateral performance, we are interested in adaptive control adaptive control signals together with the fixed altitude and
of these dynamics. Accordingly, mixing adaptive controllers yaw controllers.
are designed for pitch/roll dynamics assuming the inertia in Quadrotor parameters are given as Jx = Jy = 0.03kg.m2 ,
these angular axis are unknown. Using the estimate of Jx and Jz = 0.04kg.m2 , K = 120N, Kψ = 4N.m, w = 15rad/sec, l =
Jy , candidate controllers supply the necessary pitch and roll 0.2m, m = 1.4kg in [15].
control signals.
For the yaw and altitude dynamics, we used non-adaptive A. Pitch and Roll Controller Design
state feedback controllers based on LQR. For designing scheme and candidate controllers for pitch
Note that for the pitch and roll controllers in this study, and roll, we selected number of candidate controllers as n =
reference angles are generated as given in [20]. 7.
Pitch and Roll Controller Design: Since the dynamic Since we estimate the value of 1/J in the adaptive law, we
model of pitch and roll includes inertial uncertainty, Jx and design our mixing scheme considering the 1/J. The nominal
Jy , an adaptive controller based on mixing adaptive control is value of 1/J is 33. For the candidate controller design, it is
designed for pitch and roll dynamics to increase the tracking assumed that 1/J is in the interval of Ω = [0, 50].
performance of motion along x − y axis.
Subsets considering the estimation of 1/J is are selected as
Since the theoretical values of inertia parameters and the
Ω1 = [0, 10], Ω2 = [7, 17], Ω3 = [14, 24], Ω4 = [21, 31], Ω5 =
state-space model of roll and pitch are same, we separate
[28, 38], Ω6 = [35, 45], Ω7 = [42, 50] for pitch/roll control
estimation value of Jx and Jy into common n subsets as
scheme.
Ω1 , Ω2 , Ω3 , ..., Ωn . Then the candidate controller law for
Candidate controller gains used in pitch/roll mixing adap-
these dynamics are selected as;
tive scheme are given in table below. Bump functions for
u = −kiT (p)xi (24) pitch and roll controllers are selected as ψ1 = ψ( α̂−55 ), ψ2 =
ψ( α̂−12
5 ), ψ 3 = ψ( α̂−19
5 ), ψ 4 = ψ( α̂−26
5 ), ψ5 = ψ( α̂−33
5 ),
where p ∈ ϒ = [1, 2, ..., n]. α̂−40 α̂−46
ψ6 = ψ( 5 ), ψ7 = ψ( 4 ), respectively.
Controller gains in (24) are calculated using LQR method. 1
While calculating the gains for each controller, the estimation Stable filter used in adaptive law is selected as (s+15) , cost
z−ẑ 2
value of inertia is considered. function is selected as J(α) = m2 with ms = 1 + µ .
s

3600
TABLE I
50
P ITCH /ROLL DYNAMICS C ANDIDATE C ONTROLLER G AINS

40
Candidate Controller k1 k2 k3 k4
Ω1 0.0644 0.0221 0.7547 0.0183
30
Ω2 0.0636 0.0194 0.8188 0.0183
Ω3 0.0626 0.0162 0.9235 0.0183
Ω4 0.0620 0.0141 1.0200 0.0183 20 Estimation of 1/Jx
Ω5 0.0616 0.0127 1.1000 0.0183 Estimation of 1/Jy
Ω6 0.0613 0.0117 1.1700 0.0183 10 Actual Value
Ω7 0.0611 0.0110 1.2400 0.0183
0

−10
B. Altitude Controller Design 0 20 40 60 80 100
Time (sec)
Fixed controller gains for altitude controller are calculated
using the LQR with the weighting matrices Q=diag(180 Fig. 5. Estimation results of 1/Jx and 1/Jy
0 25000 40) and R=12000. Note that weighting matri- 3d Trajectory
ces are determined by experimental studies. The feedback
gains for the altitude controller are calculated as k = 0.5
[0.1225, 0.5966, 1.4532, 7.2120]T .
0.4
C. Yaw Controller Design

Z (m)
0.3
Similar to the altitude controller, we have calculated
the feedback gains for yaw controller by using LQR. 0.2 Actual
Weight matrices are selected as Q= diag([400 0 15000 Desired
40]) and R=2000. The feedback gains for the yaw con- 0.1
1
troller used during the implementation are calculated as
k = [0.4472, 1.5655, 2.7402, 131.4320]T . 0.5
0 1
0.5
V. EXPERIMENTAL RESULTS 0
−0.5 −1 −0.5
Y (m)
Designed controllers are implemented to quadrotor for X (m)
measuring the performance of proposed control scheme. Fig. 6. x, y, z Tracking Performance
For the experimental implementation, we have applied a

seen that proposed scheme gives satisfactory results in sense


of tracking. Reference trajectories and measured coordinates

1
Actual
Desired
0.5
X (m)

−0.5
Fig. 4. Experimental implementation on Qball-X4

square trajectory along x − y axis to the quadrotor. Refer- −1


−0.6 −0.4 −0.2 0 0.2 0.4 0.6 0.8
ence trajectory for altitude is selected as constant as 0.45 Y (m)
meters. Finally, yaw reference is selected as zero degrees.
Experimental studies are realized for 93 seconds and the Fig. 7. x − y Tracking Performance
measurement results are obtained.
If we consider the estimation results of 1/J for pitch and along x, y and z axis are given in Figure 8. Reference angles
roll, we can say that least squares based estimation scheme generated corresponding to reference trajectories and their
converges the estimation values to their theoretical values. measurements can be seen in Figure 9.
Three dimensional representation of x − y − z experimental As seen from Figure 9, pitch, roll and yaw angles tracks
tracking performance is shown in Figure 6. x − y axis motion the general trend of reference signals with small deviations.
measurements for quadrotor is given in Figure 7. It can be These deviations are due to the coupling effects which

3601
1
control task. We have also seen that proper selection of
X (m)
Actual
0
Desired number of candidate controllers with parameter subsets is
−1 essential in performance of the mixing adaptive controller
0 20 40 60 80 100
Time (sec) scheme.
1
ACKNOWLEDGMENT
Y (m)

Actual
0
Desired This work is supported by the Canadian NSERC Discov-
−1
0 20 40 60 80 100 ery Grant 116806 and CFI Grant 31211. N. Koksal holds a
Time (sec) MEB (Turkish Ministry of National Education) scholarship.
Height (m)

0.5
Actual R EFERENCES
Desired
0 [1] S. Bouabdallah, A. Noth, and R. Siegwart. PID vs LQ control
0 20 40 60 80 100 techniques applied to an indoor micro quadrotor. In Proc. IEEE/RSJ
Time (sec) Intelligent Robots and Systems, Sep.-Oct. 2004, vol. 3, pp. 2451-2456.
[2] T. Madani and A. Benallegue. Control of a quadrotor mini-helicopter
Fig. 8. x, y, z Tracking Performance via full state backstepping technique. In Proc. 45th IEEE Conference
Angular Measurements θ − φ − ψ on Decision and Control, 2006, pp. 1515-1520.
0.2 [3] S. Bouabdallah and R. Siegwart. Full control of a quadrotor. In Proc.
Roll (rad)

Actual IEEE Intelligent Robots and Systems, 2007, pp. 153-158.


0 [4] A. Benallegue, A. Mokhtari, and L. Fridman. High-order sliding-mode
Desired
−0.2 observer for a quadrotor UAV. International Journal of Robust and
0 20 40 60 80 100 Nonlinear Control, vol. 18, no. 4-5, pp. 427-440, 2008.
Time (sec) [5] G.V. Raffo, M.G. Ortega, and F.R. Rubio. An integral predic-
0.2 tive/nonlinear H∞ control structure for a quadrotor helicopter. Auto-
Pitch (rad)

Actual matica, vol. 46, no. 1, pp. 29-39, Jan. 2010.


0 [6] L. Besnard, Y.B. Shtessel, and B. Landrum. Quadrotor vehicle control
Desired
−0.2 via sliding mode controller driven by sliding mode disturbance ob-
0 20 40 60 80 100 server. Journal of the Franklin Institute, vol. 349, no. 2, pp. 658-684,
Time (sec) March 2012.
10 [7] K. Alexis, G. Nikolakopoulos, and A. Tzes. Model predictive quadro-
Yaw (rad)

Actual tor control: attitude, altitude and position experimental studies. IET
0 Control Theory and Applications, vol. 6, no. 12, pp. 1812-1827, 2012.
Desired
−10 [8] T. Ryan and H. J. Kim. LMI-based gain synthesis for simple robust
0 20 40 60 80 100 quadrotor control. IEEE Tr. Automation Science and Engineering, vol.
Time (sec) 10, no. 4, pp. 1173-1178, 2013.
[9] W. Zeng, B. Xian, C. Diao, Q. Yin, H. Li, and Y. Yang. Nonlinear
Fig. 9. θ , φ , ψ Tracking Performance adaptive regulation control of a quadrotor unmanned aerial vehicle.
In Proc. IEEE Multi-Conference on Systems and Control, 2011, pp.
133-138.
[10] K.S. Narendra and J. Balakrishnan. Adaptive control using multiple
are ignored during the modeling of the system. Also, the models, IEEE Tr. Automatic Control, vol. 42, no. 2, pp. 171-187, Feb.
nonlinearity of the quadrotor system effects the tracking 1997.
performance. [11] S. Fekri, M. Athans, and A. Pascoal. Issues, progress and new results
in robust adaptive control. International Journal of Adaptive Control
If we consider the mean-square error of the reference and Signal Processing, vol. 20, pp. 519-579, 2006.
performance, MSE values for x and y axis are sufficiently [12] M. Kuipers and P. Ioannou. Multiple Model Adaptive Control with
small. MSE values for x, y and z axis are 0.1394, 0.1621 Mixing. IEEE Tr. Automatic Control, vol. 55, no. 8, pp. 570-578, 2010.
[13] R. L. Kosut and B. D. O. Anderson. Adaptive control via finite
and 0.0145, respectively. We can say that reference signal modelling and robust control. In Proc.1988 IFAC Workshop on Robust
tracking performance of the proposed method is satisfactory. Adaptive Control, Newcastle, Australia, 1988, pp. 91-95.
[14] S. Baldi, P. Ioannou, and E. Mosca. Multiple model adaptive mixing
VI. CONCLUSION control: The discrete-time case. IEEE Tr. Automatic Control, vol. 57,
no. 4, pp. 1040-1045, 2012.
In this study, we have proposed and analyzed a mixing [15] Quanser Inc. Quanser Qball-X4:User Manual. Doc. no. 888, 2013.
adaptive control based lateral motion controller with state [16] A. Tayebi and S. McGilvray. Attitude stabilization of a VTOL quadro-
tor aircraft. IEEE Transactions on Control Systems Technology, vol.
space dynamic formulation, focusing on implementation in 40, no. 3, pp. 562-571, May 2006.
quadrotor vehicle systems. Proposed method is implemented [17] T. Li, Y. Zhang, and B. Gordon. Fault tolerant control applied to a
on Quanser Qball-X4 quadrotor testbed. Experimental results quadrotor unmanned helicopter. In Proc. ASME/IEEE Int. Conference
on Mechatronic and Embedded System and Applications, 2011, pp.
show that the proposed method provides good performance 1-10.
and robustness against inertial uncertainties. Unknown iner- [18] P.A. Ioannou and J. Sun. Robust Adaptive Control. Dover Publications,
tia parameters are estimated correctly and estimate values 2012.
[19] P. Ioannou and B. Fidan. Adaptive Control Tutorial. Philadelphia,
converge close to their actual values. PA:SIAM, 2006.
It is observed that, due to neglecting coriolis and the cou- [20] I. Bayezit and B. Fidan. Distributed cohesive motion control of flight
pling effects between sub-dynamics in modeling, measured vehicle formations. IEEE Tr. Industrial Electronics, vol. 60, no. 12,
pp. 5763-5772, December 2013.
angular values have small difference from reference angles.
These neglected parameters effect the tracking performance
of lateral motion. Despite these coupling effects, in the lateral
motion, tracking errors are sufficiently small, satisfying the

3602

S-ar putea să vă placă și