Professional Documents
Culture Documents
Mechatronics
journal homepage: www.elsevier.com/locate/mechatronics
a r t i c l e
i n f o
Article history:
Received 28 February 2013
Revised 3 May 2013
Accepted 6 May 2013
Available online 10 June 2013
Keywords:
Online auto-tuning
Flight control
PID rotor control
Optimization
a b s t r a c t
This paper presents a stochastic, multi-parameters, divergence optimization method for the auto-tuning
of proportionalintegralderivative (PID) controllers according to a fractional-order reference model. The
study aimed to approximate the step response of the real closed-loop ight control system to the
response of a theoretical reference model for a smoother and more precise ight control experience.
The proposed heuristic optimization method can auto-tune a PID controller without a precise plant
model. This is very advantageous when dealing with model and parameter uncertainties in real control
application and practice. Experimental study conrms the reference model driven auto-tuning of the DC
rotor prototype.
2013 Elsevier Ltd. All rights reserved.
1. Introduction
Today, there is increasing interest in electrically driven rotorcraft due to advances in high-speed DC motors and low-weight
battery technologies [1]. Electrically driven, multi-rotor, aerial
vehicle designs present several advantages compared to conventional aerial vehicles. They greatly reduce mechanical complexity
and the need for aerodynamic casings. They can maneuver by simply adjusting rotor speed, and thus, a well-designed multi-rotor
vehicle inherently provides more stable and reliable ight [2]. Mini
multi-rotor, unmanned aerial vehicles have great potential for new
applications, particularly in their use for indoor missions [3].
Flight control problems [410] involve serious complications,
due to the complex mechanisms of aircrafts, nonlinear and inaccessible aerodynamics, and frequent variability in ight conditions,
depending on altitude, payload, and weather conditions. Therefore,
online auto-tuning of controller parameters is needed to maintain
desired control responses under varying conditions of real ight.
Flight control of aerial vehicles demands smooth and precise
maneuvers because high acceleration and oscillations in motion
harm passengers and vehicle structure. High maximum overshoots
in the control of vertical and horizontal maneuvers may cause aerial vehicle crashes, especially when ight space is limited, such as
in landing or indoor ights of autonomous mini-aerial vehicles.
High overshoots in multi-rotor aerial vehicles also cause deteriora Corresponding author. Tel.: +90 55527134170.
E-mail address: baykant.alagoz@inonu.edu.tr (B.B. Alagoz).
0957-4158/$ - see front matter 2013 Elsevier Ltd. All rights reserved.
http://dx.doi.org/10.1016/j.mechatronics.2013.05.001
tion in ight stability due to the structural and aerodynamic coupling between rotors, which delays rotors settling into the
desired positioning. Moreover, high overshoots and oscillations before settling result in the consumption of more electrical energy
and reduce the ight range of vehicles. Consequently, smooth
and overshoot-free step responses are more appropriate for reliable and efcient ight control; vehicle acceleration should be kept
low and overshoots should be suppressed as much as possible [10
13]. Automated ight control systems should ensure smoother and
more precise maneuvers than human pilots alone could provide,
and increase the efciency in energy consumption by providing
proper power injection to the rotors, and accordingly increase
ight ranges. This study aimed to provide online auto-tuning of
PID controllers for the smooth and precise step response of rotors
for ight control practice.
Due to the widespread use of classical PID controllers in industry, several methods have been proposed to achieve practical tuning of classical PID controllers. These analytical controller-tuning
methods depend mainly on plant transfer function modeling and
presumptions [14,15]. However, unpredictable environmental factors affecting the system outputs and presumptions in system
modeling techniques can make it impossible to obtain an exact
theoretical modeling of real physical systems. Robustness margin
for attitude control of mini-rotorcraft known as X4-yer was presented for multivariable PD controller in order to stabilize the attitude of the aerial vehicle [10]. However, online auto-tuning of
controller coefcients is necessary to deal with the negative impacts of temporal changes and model uncertainties on controller
790
performance. Heuristic tuning methods are advantageous for online auto-tuning of real control systems because they can progressively improve their responses using previous control experiences.
Today, PID tuning is still a promising research area because it is
widely used in industrial applications. Several methods for selftuning controllers were developed in the literature. Methods for
PID tuning, based on genetic algorithms and differential evolution,
have been presented [1618]. A holistic, multi-objective optimization method, using a multi-objective evolutionary algorithm, has
also been proposed for controller tuning [19]. A self-tuning controller, based on control performance evolution, was developed
to improve both transient-state and steady-state control performance [20]. A mixture of conventional PID controller and the neural network was developed for capability of learning, adaptation
and tackling nonlinearity problems [21]. Radial Basis Function neural networks were used to design an adaptive PID controller [22].
Particle swarm optimization (PSO) algorithm was used for design
method for the self-tuning PID control in a slidercrank mechanism system [23]. A grey prediction model combined with a fuzzy
PID controller was used to change the grey step size in order to improve the force control quality for hydraulic load [24]. Nicol et al.
proposed a direct approximate-adaptive control, using CMAC nonlinear approximators, for quadrotor helicopter [9]. Barbosa et al.
presented a theoretical study for PID tuning according to Bodes
ideal control loop (BICL) reference model and discussed advantages
of the BICL model in control problems [25]. They reported that BICL
systems are robust for gain variations. BICL systems dene a fractional order system with iso-damping property [25,26]. Yeroglu
et al. investigated the dependence of BICL step responses on crossover frequency and fractional-order parameters [27].
This paper presents a reference model based optimization approach for the online auto-tuning of PIDs using the Stochastic Multi-Parameters Divergence Optimization (SMDO) method. SMDO
denes an optimization parameter space where the controller
parameters are represented by an optimization parameter vector.
The SMDO method stochastically redirects the optimization
parameter vector to a new parameter vector according to a predened objective function. The reference system is called the master
system, which yields the desired responses for the controller tuning. The slave system is the closed-loop PID control system. The
SMDO optimizer tunes the PID controller so that the response of
the slave system approximates the response of the master system.
The main advantage of masterslave system optimization is that
the master system can be a well-known and easily adjustable theoretical system for a desired system response, while the slave system is a real closed-loop PID control system with model
uncertainty and parameter uctuation. Thus, one can approximate
the responses of the real control system to a desired response of
the theoretical reference model. This study has two main contributions: It presents SMDO optimization method and employs this
method in a reference model based auto-tuning of PID controller
for the smooth control of TRMS rotors. Experimental results revealed that a low accelerated, precise and energy-efcient step response of rotors is possible by a conventional PID controller.
Bodes ideal control loop (BICL) [25,26] is used as a fractionalorder reference model in the proposed masterslave optimization
method so that desired step responses of the well-known BICL,
such as maximum overshoot, peak time, rise time, and settling
time, can be easily obtained [25,27].
The ight control performance of the adaptive PID controller
using the masterslave system optimization method was demonstrated in the experimental work conducted on a Twin-Rotor
MIMO (Multi-Input/Multi-Output) System (TRMS). The TRMS setup provides a nonlinear and complex helicopter ight control test
platform to simulate several ight modes such as hovering and
taking off in laboratory conditions [28]. Responses of the TRMS
Ls
x c
c
c 2 R;
Fig. 1. Basic block diagram of reference system guided PID controller tuning system
[25].
Ts
Ls
1
;
c
1 Ls
s
1
1 < c < 2:
xc
791
The SMDO method is based on consecutive, stochastic, bidirectional, componential set and trial (ST) operations. It searches for a
convenient divergence vector satisfying the condition given by Eq.
(3). Therefore, a component of X vector is tested initially for a random forward step. If it does not result in a convenient divergence
(DEn1;n P 0), then the component is tested for a random backward
step. If the backward step test also fails to yield a convenient divergence of Xn, the component keeps its original value.
In SMDO, the number of consecutively tested divergence vectors for a full search of all components is 2k for a k-parameter system. During the controller tuning process, controller parameters
are subsequently tested, and only the convenient divergences are
allowed to modify Xn vectors in the controller parameter space.
Let a divergence vector for a PID controller be written as
DX n /nv b:
EX opt min
fEX n g:
n
X 2X s
EX m EX n < 0;
792
n1
X /v b:
E
6
EX n /nv b EX n < 0;
X n1 X n /nv b:
EX n /nv b EX n < 0:
1
L
Om t Os t2 dt:
10
tL
Eq. (10) is a real valued error function dened to reduce dissimilarity between step responses of the reference model and the closedloop PID control system. This error function calculates squared difference of master and slave system outputs, Om and Os.
The SMDO algorithm tests all components, one by one, for the
both forward and backward directions. By calculating Eq. (10),
the method observes the impacts of random divergence on the error function for each component in both directions. Then, it decides
upon the best divergence, resulting in decrease of the error function, and continues searching, by using the best parameter vector.
The SMDO method presents the advantage of being independent from plant transfer functions modeling; however, it requires
a trial period (L h) in ST operations long enough to assess reliably
the effects of vector divergences on the error function. Use of
unnecessarily long trial periods has the side effect of slowing down
the tuning speed of the controller. A satisfactory tuning performance is possible when the trial period is slightly larger than the
periodicity of the reference signal (r(t)), in which case, the result
of the error function dened by Eq. (10) becomes consistent in each
trial.
2.4. A real-time instability prevention mechanism during online autotuning
Online auto-tuning of controllers, without pre-knowledge of
plant models, bears the risk of system instability during the optimization of the controller parameters. In practice, it may be difcult to specify parameter ranges, where the system stays stable.
Therefore, it is very important to develop a real-time instability
detection mechanism to avoid control system instability during
online auto-tuning by using the SMDO method. This study used a
simple error-threshold mechanism, based on the Lyapunov stability criterion. Hurwitz stability suggests that as poles of a system
793
(a) 0.8
(a) 150
iteration number
0.6
0.4
BICL response
100
50
0.2
10
15
20
auto-tuning seasons
0
10
20
30
-4
40
(b) 4
second
Emin
0.01
Tuning State
Tuned State
(b)
x 10
0.008
1
0.006
0
0.004
(c)
Kp
10
15
20
Fig. 5. Auto-tuning was repeated 20 times. (a) ST iteration numbers to reach tuned
state in each auto-tuning session. (b) Average errors in each auto-tuning session.
50
100
150
200
iteration
10
auto-tuning seasons
0.002
0
(a)
Tuned State
Ki
0
10
50
100
150
200
Tuned State
Kd
0
0
20
50
100
15
10
150
(b)
200
Tuned State
0
50
100
150
200
iteration
Fig. 4. (a) Step responses of initial PID controller (X0 = [5 10 8]), tuned PID controller
(X93 = [7.37 17.43 1.07]), and reference model BICL (c = 1.04, xc = 0.4). (b) Convergence of average error at 93th ST iteration. (c) Alteration of PID coefcients during
simulation.
approach the complex axis or right half plane, system output severely diverges from the reference signal r(t). This causes a sharp
rises in the error function. By using a predened error threshold
(ds), pole approximation to the complex axis was detected in
real-time by monitoring the condition of E(Xn) > ds according to
the Lyapunov stability criterion. When E(Xn) > ds occurred, the
SMDO method turned back to the best parameters from parameter
history, which provided a minimum error value during the optimization, as expressed below:
(
Xn
min fX u g EX n > ds
0<u<n
Xn
EX n 6 ds
)
:
11
(c)
Fig. 6. (a) Picture of the TRMS experimental setup. (b) Main parts of the TRMS.
(Label 1 indicates the bottom positioning and Label 2 indicates the top positioning
of the pendulum counterweight.) (c) Block diagram of the TRMS control setup using
masterslave SMDO.
794
tion model, and then the ne tuning of online PID was performed
on the TRMS experiment to obtain online ne tuned PID coefcients. As the masterslave system SMDO tunes PID coefcients,
the step responses of the real closed loop PID-controlled TRMS converged with the step responses of the BICL. Fig. 7a shows the step
responses of the main rotor after PID controller tuning with the reference system. In order to alter the TRMS systems operating conditions, the location of the pendulum counterweight was altered
from bottom to top positions, which considerably changed the vertical angular momentum of the main rotor during takeoff and hovering. Fig. 7b reveals the discrimination between the TRMS step
responses and the responses of the theoretical BICL reference model for changing the position of the pendulum counterweight. After
retuning of PID controller for new positioning of pendulum counterweight, step response of TRMS approximated again to the response of BICL as shown in Fig. 7c.
Unmanned multi-rotor aerial vehicles are largely driven by
step-wise control signals [24] because (i) processing delays occur
in communication and operators decisions for remote controlled
systems or (ii) computational delays occur in machine vision and
autonomous ight path decision algorithms for autonomous robotic systems [3033]. Fig. 8 shows the responses of the TRMS
main rotor to vertical angle control commands. Fig. 8a shows the
(a) 0.8
0.6
0.4
BICL
SMDO
r(t)
3. Experimental study
0.2
10
15
20
second
(b) 0.8
Undesired
Overshoot
0.6
0.4
BICL
SMDO
r(t)
0.2
10
15
20
second
(c) 0.8
0.6
0.4
BICL
SMDO
r(t)
0.2
10
15
20
second
This section presents experimental results demonstrating the
online auto-tuning of a PID controlled TRMS. In experiments, a reference signal in the form of a step function with amplitude of 0.5
was used to control the main rotor angle. Optimization initialized
with the PID coefcients, which were tuned off-line by the simula-
Fig. 7. (a) Main rotor step responses after online tuning with the pendulum
counterweight at low position (Label 1 in (b)). (b) Main rotor step responses after
pendulum counterweight changed to high position (Label 2 in (b)) for plant
perturbation. (c) Main rotor step responses after online tuning with the pendulum
counterweight at high position.
795
(a) 0.6
(b)
0.5
0.4
0.2
0.4
BICL
FD
r(t)
0.3
BICL
SMDO
r(t)
0.2
0.1
0
-0.2
0
20
60
40
80
100
20
40
60
80
100
second
second
(c)
0.5
(d)
0.4
0.4
0.3
0.3
0.5
0.2
BICL
SMDO
r(t)
0.1
0.2
BICL
SMDO
r(t)
0.1
0
-0.1
20
40
60
80
100
20
40
second
60
80
100
second
Fig. 8. (a) The responses of PID controller provided by Feedback Instruments Ltd. (FD) for step-wise control commands. (b) The responses of PID controller for step-wise
control commands after online ne auto-tuning of PID controller. (c) The responses of PID controller for step-wise control commands after perturbing the TRMS by changing
the position of the pendulum counterweight. (d) The responses of PID controller for step-wise control commands after online ne auto-tuning again.
0.8
0.6
0.4
SMDO
BICL
r(t)
0.2
10
15
20
25
30
second
Fig. 9. (a) Tail rotor step responses after online tuning with the pendulum
counterweight at the low position (Label 1 in Fig. 7b).
Table 1
Average errors and average accelerations.
PID tests
Position of
counter
weight
Average
squared
error
Average
acceleration
C FD s 5
8
10s;
s
12
C PL-SMDO s 7:49
1:07
17:35s;
s
13
C PH-SMDO s 8:14
0:96
18:50s:
s
14
Table 1 lists the average squared errors (|E|) and average accelerations (|A|) calculated by using the following discrete
approximations:
jEj
1
L
tL
Om Os 2 dt
nL
1X
Onm Ons 2 ;
L n
15
796
Z
1 t @ 2 Om
jAj
dt
L tL @t2
n
n
nL n
1X
Om n 1 Om n 1 2Om n:
2
L n
h
For any discrete time sampling of n and m for m > n > 0, a Dt time
interval can always be found such that Dt < mh nh. In this case,
one can easily write the following inequality,
16
lim
Dt!0
DE
< 0:
Dt
17
DE
DE
< lim
< 0:
m nh Dt!0 Dt
18
Then, for a state sampling at n and m for m > n > 0, DE = E(Xm) E(Xn) is considered, the condition of E(Xm) E(Xn) < 0 is valid. h
References
[1] Yang JH, Hsu WC. Adaptive backstepping control for electrically driven
unmanned helicopter. Control Eng Practice 2009;17:90313.
[2] Pounds P, Mahony R, Corke P. Modeling and control of a large quadrotor robot.
Control Eng Practice 2010;18:6919.
[3] Oung R, DAndrea R. Distrib Flight Array Mechatron 2011;21:90817.
[4] Song D, Han J, Liu G. Active model-based predictive control and experimental
investigation on unmanned helicopters in full ight envelope. IEEE Trans
Control Syst Technol 2012;1:17.
[5] Zhao X, Han J. Yaw control of helicopter: an adaptive guaranteed cost control
approach. Int J Innovative Comput, Inform Control 2009;5:226776.
[6] Yamamoto T, Mori S, Sakaguchi A. Data-driven Skill-Based PID Control of A
Pilot-Scale Helicopter Model. Int J Innovative Comput, Inform Control
2008;4:334858.
[7] Salazar-Cruz S, Lozano R, Escareno J. Stabilization and nonlinear control for a
novel trirotor mini-aircraft. Control Eng Practice 2009;17:88694.
[8] Ross M, Karpenko M. A review of pseudospectral optimal control: from theory
to ight. Ann Rev Control 2012;36:18297.
[9] Nicol C, Macnab CJB, Ramirez-Serrano A. Robust adaptive control of a
quadrotor helicopter. Mechatronics 2011;21:92738.
[10] Lara D, Romero G, Sanchez A, Lozano R, Guerrero A. Robustness margin for
attitude control of a four rotor mini-rotorcraft: case of study. Mechatronics
2010;20:14352.
[11] Vasconcelos JF, Silvestre C, Oliveira P, Guerreiro B. Embedded UAV model and
LASER aiding techniques for inertial navigation systems. Control Eng Practice
2010;18:26278.
[12] Shouzhaon S, AhmadMian A, Chao Bin ZJ. Autonomous take off and landing
control for a prototype unmanned helicopter. Control Eng Practice
2010;18:10539.
[13] Lee J, Yoo C, Park YS, Park B, Lee SJ, Gweon DG, et al. An experimental study on
time delay control of actuation system of tilt rotor unmanned aerial vehicle.
Mechatronics 2012;22:18494.
[14] Ang KH, Chong G, Li Y. PID control system analysis design and technology. IEEE
Trans Control Syst Technol 2005;13:55976.
[15] Meza GR, Nieto SG, Sanchis J, Blasco FX. Controller tuning by means of multiobjective optimization algorithms: a global tuning framework. IEEE Trans
Control Syst Technol 2012;21:44558.
[16] Saad MS, Jamaluddin M, Darus IZM. Implementation of PID controller tuning
using differential evolution and genetic algorithms. Int J Innovative Comput,
Inform Control 2012;9:776179.
[17] Vallejoa AJ, Morales-Menendezb R. Cost-effective supervisory control system
in peripheral milling using HSM. Annual Reviews in Control 2010;34:15562.
[18] Habera R, Schmitza U, Barsb R. Optimal choice of horizons for predictive
control by using genetic algorithm. Annual Reviews in Control 2004;28:538.
[19] Rani MR, Selamat H, Zamzuri H, Ibrahim Z. Multi-objective optimization for
PID controller tunning using the global ranking genetic algorithm. Int J
Innovative Comput, Inform Control 2012;8:26984.
[20] Tokuda M, Zamamoto T. Self-tuning PID controller based on control
performance evalutions. Int J Innovative Comput, Inform Control
2010;6:375162.
[21] Thanh TU DC, Ahn KK. Nonlinear PID control to improve the control
performance of 2 axes pneumatic articial muscle manipulator using neural
network. Mechatronics 2006;16:57787.
[22] Mizumoto I, Ikeda D, Hirahata T, Iwai Z. Design of discrete time adaptive PID
control systems with parallel feed forward compensator. Control Eng Practice
2010;18:16876.
[23] Kao CC, Chuang CW, Fung RF. The self-tuning PID control in a slidercrank
mechanism system by applying particle swarm optimization approach.
Mechatronics 2006;16:51322.
[24] Truong DQ, Ahn KK. Force control for hydraulic load simulator using selftuning grey predictor fuzzy PID. Mechatronics 2009;19:23346.
[25] Barbosa RS, Machado JAT, Ferreira IM. Tuning of PID controller based on Bodes
ideal transfer function. Nonlinear Dynam 2004;38:30521.
[26] Bode HW. Network analysis and feedback amplier design. New York: Van
Nostrand; 1945.
[27] Yeroglu C, Tan N. Classical controller design techniques for fractional order
case. ISA Trans 2011;50:46172.
[28] Rahideh A, Shaheed MH. Constrained output feedback model predictive
control for nonlinear systems. Control Eng Practice 2012;20:43143.
[29] Vachtsevanos G, Tang L, Drozeski G, Luis G. From mission planning to ight
control of unmanned aerial vehicles: strategies and implementation tools. Ann
Rev Control 2005;29:10115.
797