You are on page 1of 6

MULTI-OBJECTIVE EVOLUTIONARY OPTIMIZATION PID TUNING FOR

LONGITUDINAL MOVEMENT OF AN AIRCRAFT.

Garbelini, Matheus, Reynoso-Meza, Gilberto



Polytechnic School, Pontifical University of Parana (PUCPR), Curitiba, CO 80215-901 BRA

PPGEPS, Pontifical University of Parana (PUCPR), Curitiba, CO 80215-901 BRA

Emails: matheus.gabelini@pucpr.edu.br, g.reynosomeza@pucpr.br

Abstract Longitudinal control tuning is one of the principal tasks in the control system of an aircraft as its
pitch movement is affected by different flight conditions and most of the times, aggressive environment. In order
to get a better flight stability, multi-objective evolutionary optimization techniques can be applied to the tuning
procedure in order to depict the trade-off between conflicting objectives. In this paper, we use such techniques
to appreciate the trade-off of the aircraft longitudinal angle response. We compare two different algorithms
in a four-objective problem that lead longitudinal movement control trough a standard proportional-integral-
derivative (PID) controller.

Keywords Multi-objective optimization, PID controller, Longitudinal movement.

1 Introduction When the multi-objective optimization is


united with the already defined closed loop model,
Aerial systems nowadays are demanding a growing design criteria must be well defined in order to
capability in its stabilization control. Both small correctly get the best performance. For this rea-
and big planes require a controller that is capa- son, adequate design objectives that can provide
ble of delivering the desired performance under robust and meaningful indicators can be challeng-
different operation conditions (Duarte-Mermoud ing to define depending on the complexity of the
et al., 2005). PID controllers are widely used problem.
for this approach due to its simplicity. Further-
more, if multi-objective optimization is used along The more clear the objectives are, the easier
to bring the optimum trade-off for the decision the decision-making processes gets. Even though,
maker, the results can get very satisfactory when the algorithms that are used in the evolutionary
related to the design preferences. processes also needs to be fit correctly the prob-
The use of multi-objective optimization evolu- lem. In most cases the behavior of a certain al-
tionary algorithms (MOEAs) corresponds a very gorithm is not known for a specific design prob-
known category for such multi-objective optimiza- lem due to the objectives number or complex-
tion techniques (Das et al., 2016). Its use extends ity. Hence, a comparison between different evo-
in many different areas where there is the necessity lutionary algorithms needs to be done before pre-
to find the trade-off between the designed problem ceding to the multi-criteria decision make stage
and its system requirements. (MCDM). The algorithm trade-off convergence or
Several control literatures report the com- divergence rate between different objectives can
bined use of a feedback controller with MOEAs be acquired by analyzing the hypervolume graph.
for parameters tuning (Reynoso-Meza et al., 2014,
and the references therein). In this context, This paper presents a simple multi-objective
fuzzy controller and MOEAs was used to aug- statement and uses reference case parameters to
ment the flight control for an F16 aircraft in improve pertinence of the solutions and also to
(Stewart et al., 2010), the transition process in facilitate the direct comparison between the eval-
flight mode of a tiltwing aircraft in (Holsten uated evolutionary algorithms.
and Moormann, 2015), the control of a commer-
cial Cessna airplane in (Yamina Boughari and The remainder of this paper is organized as
Theel, 2014) and further. follows: The second section of this paper present
(Duarte-Mermoud et al., 2005) study the im- the background of the multi-objective problem,
pact and result of a PID and CMRAC controllers depicting its model, control loop and the mul-
in the pitch of Cessna 182 airplane according to tiobjective problem (MOP) statement. In sec-
its lifter angle (e ). This paper approach is to di- tion 3 to 5, will be described the fundamentals
rectly compare the results from the reference PID steps to implement an optimization problem with
parameters used in the previous paper to the re- multi-objective approach: The MOP definition,
sults obtained after the same PID goes through the multi-objective optimization (MOO) process
MOEAs optimization. Further, it can show its and the MCDM stage. The results are validated
improvement and viability in real aerial applica- in section 6 and a conclusive view of this work is
tions generally. commented in section 7.
2 Overview transfer function C(s) that is described in equa-
tion 2.
To understand the tuning approach with MOEAs,
kp
its necessary firstly to understand the require- C(s) = kp + + kp d s (2)
i s
ments of the aerodynamic system. Many variables
are involved in the design problem, but the prin- where kp is the proportional gain, i is the
cipal challenge is to wisely choose which relevant integral time and d the derivative time.
information use in MOEAs decision variables or
objectives.

2.1 Longitudinal movement control of a plane


The airplane aerodynamic can be described in dif- Figure 2: Control loop
ferent ways, but is convenient to use a less com-
plex model that keep a certain simplicity and also Due to the system behavior dependency of
provides the necessary accuracy about the aero- the PID controller, the MOP resides in obtaining
dynamic movement in a MOP statement. The the best parameters from C(s) such as propor-
Figure 1 depicts a basic plane longitudinal move- tional, derivative and proportional gain in order
ment that needs to be carefully analyzed with the to achieve the desired output O(s) performance.
designed optimization problem. Evolutionary multiobjective optimization (EMO)
techniques could be useful for the purpose in tun-
ing these parameters.

2.3 Multiobjective problem statement


According to (Miettinen, 2012) a MOP with n ob-
jectives can be stated such as:

min J() = [J1 (), ..., Jn ()] (3)



Figure 1: Airplane axis movement
where = [1 , 2 , . . . , n ] is the decision vec-
tor and J() the objective vector subjected to the
When the airplane is flying, its speed and sev-
inequality and equality vectors K() and L() re-
eral external ambient conditions can interfere in
spectively (see equation 4 and 5).
its stabilization. Due to the necessity in keeping
the simplicity, these conditions can be abstracted K() 0 (4)
and assumed constant1 without compromising the
model accuracy. L() = 0 (5)
In this paper, the airplane is assumed already
flying and with its roll and yaw axis in steady = [i i i ]; i = [1, 2, . . . , n] (6)
state. By doing this, the MOP can be focused to The decision vector i is limited to the upper
the desired task such as pitch axis tunning. and lower bound vector i and i respectively.
The Airplane can perform the pitch move- As in MOPs there are different conflicting ob-
ment when the lever is moved along a desired in- jectives, theres not just a single solution, but in-
clination. The flight control system is responsible stead, a set of solutions called Pareto set P . The
for the lever positioning system. objective vector J associated with each solution
As in (Duarte-Mermoud et al., 2005) we use in Pareto set approximate an optimal Pareto front
a transfer function P (s) from equation 1 to sim- JP . All individuals in Pareto front consists of non-
ulate the response of pitch and also the influence
dominated solutions (see Figure 3).
of mechanical constraints of the lever itself.
An objective vector (Miettinen, 2012) J1(1 )
P (s) = 347.3544(s+0.05902)(s+2.001)
(1) dominates a second objective vector J(2 ) (de-
(s+10)(s2 +0.04417s+0.02933)(s2 +8.902s+27.79)
noted as J(2)  J(1)) if:
2.2 PID controller tuning Ji (1 ) Ji (2 ) , i [1, 2, . . . , n]

The control loop depicted in Figure 2 is responsi- Jl (1 ) < Jl (2 ) : l [1, 2, . . . , n] (7)
ble for controlling the lever angle O(s) through a
desired reference I(s). It comprises a simple PID In order to find the best PID parameters for
1 More information about Cessna plane flight conditions
the longitudinal control, the fundamentals steps
such as air humidity, average temperature and pressure,
to implement an optimization problem with multi-
plane average speed and further can be found in (Duarte- objective approach are described in the following
Mermoud et al., 2005). sections.
Figure 3: Pareto front concept Figure 4: Rise time, settling time and overshoot
concepts
3 Multiobjective problem definition
3.1 Constraints
As in (Duarte-Mermoud et al., 2005), the evalua-
tion of the controller performance in the required Due to the known behavior of theses variables in
longitudinal movement of the closed loop (Figure control literature, it is possible to easily determi-
2) is assumed as a unitary step response. In order nate bounds to them. As Kp and d can affect
to ensure that the PID controller provide better the behavior of the system, they must generally
characteristics to the MOP, four objectives J are be lower.
considered in the signal O(s), which are depicted The decision vector is limited to decision
in Figure 4 and described as follows: bounds and that are given by:

Settling Time JST (): Given a differential = [5, 100, 2]


function F , the settling time at a tolerance = [0, 0, 0] (13)
% is presented in the equation 8.
JST () = lim F (a, t) (8) Thus the decision vector become:
(a,t)(|Y (t)Yref |% ,)

= [0 kp 5, 0 i 100, 0 d 2] (14)
Rise time JRT (): Given an initial time t0
and the time tref which the signal Y (t) cross As the objectives are meaningful values from
the reference, the Rise time can be obtained the problem perspective, it is a common practice
by the equation 9. to adopt objectives constraints in order to ensure
performance and to match design requirements of
JRT () = tref t0 (9) the MOP. These constraints are proposed as fol-
lows:
Overshoot JOV (): Indicates the relative Settling time must be 0.75 seconds, given
error of the maximum deviation YOV from a tolerance condition of 2%.
the reference signal Yref where Its given by
the equation 10. Rise time must be 0.5 seconds.

YOV () Yref Overshoot must be 4%.


JOV () = 100 (10)
Yref Ms must be held between an absolute value
of 1.2 and 1.8 to allow solutions with a slightly
Maximum value of sensitivity function weak, robust or nearly strong PID controller.
JM S (): Indicates how sensible a closed loop
is in terms of a possible peak variation in a
4 MOO algorithm
dynamic system. Its given by equation 11.

1 A multi-objective optimization algorithm (MOO)


JM S (ref ) = max | | (11) must offer to the designer a good quality set of
0s=jw< 1 + C(s) P (s)
solutions and be very descriptive, allowing more
With these definitions, the objective vector flexibility in the decision process. Such require-
J() can be defined such as: ments is presented as challenges that the algo-
rithm must overcome (Reynoso-Meza, 2009):
J() = [JST (), JRT (), JOV (), JM S ()] (12)
To avoid earlier convergence in an optimal lo-
In which the decision variables comprehends cal space, losing generalization of the Pareto
the PID parameters = [kp , i , d]. front JP .
To guarantee the diversity of solutions in
order to improve the representation of the
Pareto front JP . 1 Initial population P (0) is initialized with
random N individuals selected from
In this paper, Spherical pruning multi-
searching space ;
objective optimization differential evolution (sp-
2 Evaluate P (0);
MODE) and Indicator based multi-objective opti-
3 Search for non-dominated solutions in
mization differential evolution (ib-MODE) are dis-
P (0) to get D(0);
cussed.
4 Apply spherical pruning in D(0) to get
A(0) and store the solutions;
4.1 sp-MODE 5 for i=1:max. gen. or convergence reached.
sp-MODE is a Differential Evolution (DE) based do
from (Reynoso-Meza, 2009), which is a stochastic 6 Random select subpopulation of NS (i)
algorithm. Its main characteristics are discussed individuals with solutions in P (i) and
as follows: A(i);
7 Apply mutation and crossover
To improve the convergence rate, the best so- operations on NS (i) to get the
lutions participate trough the entire evolution offspring O(i) (Fixing boundary
process. violations if needed);
8 Evaluate O(i);
Spherical pruning is used in the objective 9 if child < parent then
space to maintain diversity of solutions, re- 10 parent = child;
duction of cardinality of Pareto set P and 11 end
to avoid non-dominated solutions to be lost 12 Apply dominance on A(i) O(i) to get
in the evolution process. D(i);
Due to MOP approach, sp-MODE uses just 13 Apply spherical pruning on D(i) to get
two operators from DE: Mutation and Crossover1 . A(i + 1);
The main steps of sp-Mode are described in 14 Store A(i + 1);
algorithm 1. 15 end
16 Solution in A with the lower |J| is
4.2 ib-MODE proposed as the single run of sp-MODE.;

Ib-MODE is DE based in the indicator-based evo-


lutionary algorithm (IBEA). The algorithm take
in account preferences of DM to drive the environ-
mental and mating selection (Thiele et al., 2009).
Its main characteristics are: 1 Initial population P (0) initialized with size
and generation count m = 0;
Uses a reference point as a pertinence mecha- 2 Calculate fitness value of all point in
nism to focus in the DMs solution space pref- Pareto set P using indicator I;
erence and to generate a local approximation 3 for i=1:size of population do
of Pareto optimal set P . 4 Choose a point in P ;
5 Remove from the population;
Uses an indicator I (Zitzler E., 2004) to gen-
6 Update fitness value of the remaining
erate a fitness value in order to select the de-
individuals using indicator I;
sired individuals from a population .
7 end
The main steps of the ib-MODE are described 8 while m < max. gen. or convergence
in algorithm 2. criterion do
9 Perform binary tournament selection
5 Multi-criteria decision-making stage with replacement on P in order to
0
fill the temporary pool P ;
Given that in this case, there are four objec- 10 Apply recombination and mutation
0
tives, further tool are required in the multi-criteria operators to the mating pool P and
decision making stage (MCDM) of a MOOD. add the resulting offspring to P .
As its not possible to directly visualize a four- Increment the generator count m;
dimensional data, there is are alternatives such 11 end
as parallel coordinates2 (Inselberg and Dimsdale, 12 set A to points in P that represent
1 More information about mutation and crossover oper-
non-dominated solutions;
ators can be seen on (Storn and Price, 1997).
2 The tool used can be found on http://www.xdat.org/
1991), which will be used to better depict the S = [3.987, 10.328, 0.144]
trade-off between the solutions in both optimiza- J(S ) = [0.729, 0.213, 1.404, 1.760] (15)
tion algorithms.
Even tough in this paper we select one prefer-
able solution, it can be argued that all the final
6 Proposal validation
solutions: preferable, reference and aggressive ac-
The simulations and Pareto front approximations complish the design requirements. The DM can
were carried by Matlab and performed in a stan- chose which is adequate and if is necessary, even
dard personal computer with Intel Core i5-4210U, the aggressive solution can be considered in a spe-
1.7GHz, 8GB RAM. cific situation in which the aircraft wouldnt be
As the MOEAS adds stochasticity, 51 runs susceptible to an aggressive environment.
were performed on both MOP algorithms. In or-
der to evaluate the Pareto front approximation
performance of the algorithms, the median value
of hypervolume indicator (The Nadir point is used
along with the objectives bounds as its reference)
is generated and depicted in Figure 5. From
this distribution is possible to affirm that that ib-
MODE is preferable for this MOP as its conver-
gence is less variable than sp-MODE.
The parallel coordinate of the Pareto front ap-
proximation from both algorithms is also shown in
Figure 7, in which is possible to notice that the so-
lutions on ib-MODE have more quality due to its Figure 5: Hypervolumes distribution.
convergence. For this reason, ib-MODE Pareto
front approximation will be chosen to select the
desired controller structure and the following cri-
7 Conclusion and future works
teria are adopted:
Those controllers with JST () 0.73 are fil- The fundamentals steps to implement a MOP
tered and excluded from the analyses as they have shown their usefulness in the controller tun-
are very near to the objective constraint and ing applications such as the control of longitudinal
could represent risk solutions. movement of an aircraft. They allow the possibil-
ity to appreciate the trade-offs between usual ob-
Those controllers with JOV () 1.5 are dis- jectives such as settling time, rise time and over-
carded, since their performance is outside shoot. The MCDM stage fitted well when more
DMs preferences. than 3 objectives are involved and allows flexibil-
ity to the DM.
Those controllers with JM S () 1.76 are also
It is important to mention that even tough we
discarded, since the MOP needs less aggres-
used 4 objectives, rise time and overshoot were
sive controllers.
observed to be not-conflicting objectives, which
After this constraints, a total of 4 possible so- could reduce the MOP to a 3 objectives prob-
lutions are chosen to be evaluated. To help in fil- lem, thus reducing the complexity and computa-
tering the remain solution, a parallel coordinates tion time of the problem. Furthermore, in pro-
plot is generated with an additional axis contain- posal validation, although sp-mode showed to be
ing the integral derivative time Td , which can in- more viable than ib-mode for PID tunning in this
fluence the sensitivity of the closed loop system problem, there are MOPs with different cost func-
due to noise issues, thus, needed to be low. By tions and plant models that ib-mode can perform
doing this, its possible to acquire the preferable better than the former. Thus, a comparative ap-
solution. proach must be made for each case.
Such solution response is shown in Figure Future works involve the use of new heuristics
6. For further analysis, the response of a more and changes in the multi-objective optimization
aggressive controller with an objective vector algorithm that could offer better Pareto approx-
J(agr ) = [0.75, 0.194, 0.455, 1.799] and agr = imation. Furthermore, in order to better reflect
[4.0, 17.819, 0.130] is also included. the designers preference, a restructuration of the
Given the reference controller ref = design objectives could be made.
[2.400, 5.5996, 0.1750] with J(ref ) =
[0.750, 0.510, 3.900, 1.514] as described in (Duarte- Acknowledgements
Mermoud et al., 2005), the preferable solution
shows the following decision variables S and This work was partially supported by the Con-
objective vector J(S ): selho Nacional do Desenvolvimento Cientfico
Figure 6: Step response results.

Figure 7: Parallel coordinate of the Pareto front approximation.

e tecnologico of Brazil (CNPq) by fellowships tion and controller tuning, Masters thesis, Uni-
PIBIC-104678/2015-1 and BJT-304804/2014-2. versidad Politecnica de Valencia. .
Reynoso-Meza, G., Blasco, X., Sanchis, J. and MartA-
References nez, M. (2014). Controller tuning using evo-
lutionary multi-objective optimisation: Current
Das, S., Mullick, S. S. and Suganthan, P. N. (2016). trends and applications, Control Engineering
Recent advances in differential evolutionan up- Practice 28: 5873.
dated survey, Swarm and Evolutionary Compu- Stewart, P., Gladwin, D., Parr, M. and Stewart, J.
tation 27: 130. (2010). Multi-objective evolutionary-fuzzy aug-
Duarte-Mermoud, M. A., Rioseco, J. S. and Gonzalez, mented flight control for an f16 aircraft, Pro-
R. I. (2005). Control of longitudinal movement ceedings of the Institution of Mechanical Engi-
of a plane using combined model reference adap- neers, Part G: Journal of Aerospace Engineering
tive control, Aircraft Engineering and Aerospace 224(3): 293309.
Technology 77(3): 199213. Storn, R. and Price, K. (1997). Differential evolution
Holsten, J. and Moormann, D. (2015). Flight control a simple and efficient heuristic for global op-
law design criteria for the transition phase for a timization over continuous spaces, Journal of
tiltwing aircraft using multi-objective parameter global optimization 11(4): 341359.
synthesis, CEAS Aeronautical Journal 6(1): 17 Thiele, L., Miettinen, K., Korhonen, P. J. and Molina,
30. J. (2009). A preference-based evolutionary al-
Inselberg, A. and Dimsdale, B. (1991). Parallel co- gorithm for multi-objective optimization, Evolu-
ordinates, Human-Machine Interactive Systems, tionary Computation 17(3): 411436.
Springer, pp. 199233. Yamina Boughari, Ruxandra Mihaela Botez, G. G.
Miettinen, K. (2012). Nonlinear multiobjective opti- and Theel, F. (2014). Robust flight control of
mization, Vol. 12, Springer Science & Business the cessna citation x using evolutionary algo-
Media. rithms, Aerospace Systems and Technology Con-
ference 6: 1730.
Reynoso-Meza, G. (2009). Design, coding and imple-
Zitzler E., K. S. (2004). Indicator-based selection in
mentation of a multiobjective optimization algo-
multiobjective search, in Proceedings of Parallel
rithm based on differential evolution with spher-
Problem Solving from Nature (PPSN VIII) .
ical pruning: applications for system identifica-

You might also like