See discussions, stats, and author profiles for this publication at: https://www.researchgate.

net/publication/379336034

Autotuning PID Controllers for Quadplane Hybrid UAV using Differential

Evolution Algorithm

Article in Journal of Aeronautics, Astronautics and Aviation, Series A · January 2024

DOI: 10.6125/JoAAA.202403_56(1S).23

CITATIONS READS

2 120

3 authors, including:

Mohammad Fahmi Pairan Syariful Shamsudin

Universiti Tun Hussein Onn Malaysia Universiti Tun Hussein Onn Malaysia
13 PUBLICATIONS 40 CITATIONS 38 PUBLICATIONS 419 CITATIONS

SEE PROFILE SEE PROFILE

All content following this page was uploaded by Syariful Shamsudin on 28 March 2024.

The user has requested enhancement of the downloaded file.

 Journal of Aeronautics, Astronautics and Aviation, Vol. 56, No. 1S, pp. 341 – 356 (2024) 341
DOI: https://doi.org/10.6125/JoAAA.202403_56(1S).23

Autotuning PID Controllers for Quadplane Hybrid

UAV using Differential Evolution Algorithm

Mohammad Fahmi Pairan1, Syariful Syafiq Shamsudin *, Mohd Fauzi Yaakub

Research Center for Unmanned Vehicles, Faculty of Mechanical and Manufacturing Engineering,
Department of Aeronautical Engineering, Universiti Tun Hussein Onn Malaysia, Johor, Malaysia

ABSTRACT

The quadplane has become the favorite UAV platform for drone delivery.
The dynamics and control of the quadplane UAV are extremely complex,
unstable, and nonlinear due to its required capability to operate in multiple
flight modes. The paper presents a method for autotuning attitude PID for a
quadplane UAV using differential evolution (DE), X-Plane simulation, and
neural network (NN)-based system identification. This study uses the DE
algorithm and a radial basis function neural network (RBF NN) to create an
autotuning PID and unified dynamics model for quadplanes, with the goal of
improving their performance in delivery missions. The unified RBF model is
then used for autotuning the PID controller using the DE algorithm. The
proposed RBF-NN system identification can successfully predict the attitude
dynamics of a quadplane in all flight modes with a coefficient of
determination of R2 greater than 0.9 and a computing mean time of less than
5 ms. By iteratively modifying the control settings to achieve optimal
performance, the DE algorithm replaces the requirement for manual PID
tuning, which can be time-consuming and suboptimal. Comparing DE tuning
to manual tuning, the results demonstrate a considerable improvement in
quadplane roll and pitch performance of about 58.5%, 65.0%, and 74.0% in
the overshoot, peak time, and rising time, respectively.

Keywords: Quadplane, System identification, Radial basis function, X-Plane,

Differential evolution

I. INTRODUCTION is considered a highly promising design [2,3]. This

platform is favored for autonomous parcel delivery
The utilization of autonomous delivery drones for the because of its simple transition mechanism, effortless
final stage of package and food delivery has been on the take-off and landing, and enhanced controllability [4,5].
rise as a result of the COVID-19 epidemic. This practice The quadplane UAV is commonly engineered with various
aids in mitigating the transmission of the virus through configurations and is capable of operating in numerous
conventional postal delivery [1]. Delivery drones are flight modes. Nevertheless, the shift from low to high
engineered to autonomously and securely transport speeds during flight mode changes also induces significant
packages to designated destinations. Ongoing research is aerodynamic changes. This aerodynamic effect makes the
centered around the configuration of drones for delivery dynamics of quadplane UAVs very complicated, unstable,
purposes, utilizing either vertical take-off and landing and nonlinear due to the combined fixed-wing and VTOL
(VTOL) designs such as helicopters and multirotors, or operations [6-8].
fixed-wing unmanned aerial vehicle (UAV) designs. A Accurate mathematical models are critical in-flight
quadplane, which integrates the structure and control applications such as stability control, system
characteristics of both a quadcopter and a fixed-wing UAV, verification, and simulation development [9]. System

Manuscript received, October 11, 2023, final revision, November 29, 2023
* To whom correspondence should be addressed, E-mail: [email protected]
342 Mohammad Fahmi Pairan Syariful Syafiq Shamsudin Mohd Fauzi Yaakub

identification is a technique for creating a quadcopter hybrid UAV systems with nonlinear and unknown
dynamics model from flight data, which improves dynamics.
prediction accuracy and provides a deeper understanding • The system identification approach based on automatic
of the system's characteristics [10,11]. To address the computer-generated frequency-sweep, which is based
requirements for future technological advancement and on simulation environments, reduces safety concerns by
commercial uses of quadplanes, a complete method is avoiding the risk to personnel and property while also
required to design a robust model-based controller with ensuring compliance with aviation regulations and
high mathematical model accuracy. Classical PID airspace restrictions, lowers the costs associated with
controller and linear quadratic regulator approaches have flight testing, and extends the flight duration due to an
been frequently used in hybrid UAVs [12]. PID controller unlimited power source.
has three important components: proportional, integral, • The autotuning PID approach utilizing the DE
and derivative control. These components control the algorithm helps to reduce tuning time and effectively
action based on previous control errors, current control addresses issues related to improper PID tuning gains in
errors, and expected future control errors [13]. Because of UAV control systems.
its simplicity, reliability, efficiency, and cost-effectiveness
without requiring considerable model knowledge, the PID II. MODEL DESCRIPTION
control method is an excellent starting point for hybrid
UAV controllers [14]. Nevertheless, PID performance 2.1 Air Vehicle
decreases when the UAV is outside of the linearization The UAV platform used in this research is a
point, and improper tuning of the PID gains can make Skywalker 1900 fixed-wing UAV with an H-configuration
performance improvement more difficult, time-consuming, quadcopter. Figure 1 illustrates the prototype's high-wing
and ineffective [15]. design, T-tail, single-boom, and pusher engine layout. It
Much study has been conducted in establishing should be noted that the wingspan is 1.9 m, and the UAV's
effective PID tuning methods, either through classical or maximum gross takeoff weight is 5 kg, allowing for a
intelligent optimization approaches, to address the issue of payload capacity of under 1 kg.
inappropriate tuning gains. Classic techniques such as
Ziegler-Nichols (ZN) and Cohen-Coon are frequently used
by engineers, especially when there is minimal knowledge
on the process model under control. However, determining
the ideal PID parameters using the ZN tuning method is
frequently difficult [11]. Different intelligent optimization
algorithms or modern heuristic optimization methods,
such as genetic algorithm (GA) [16], differential evolution
(DE) [13], and particle swarm optimization (PSO) [17], on
the other hand, have been widely studied and applied to
the automatic tuning of PID controllers [18].
DE is a highly promising evolutionary algorithm that
shares similarities with GA, particularly in its utilization
of crossover and mutation processes. The DE algorithm is
Figure 1 Quadplane UAV platform.
well-known for its precise optimization of the global
optimum, rapid convergence, simplicity, resilience, and
capability to handle cost functions that are non-
2.2 X-Plane Model
differentiable, non-linear, and multimodal [19,20]. Prior
The Plane Maker and Airfoil Maker create a complete
studies on DE have revealed a lack of attention given to its
quadplane UAV model for simulating flight dynamics and
utilization for hybrid UAV platforms in comparison to
serve for system identification purposes. The Plane Maker
quadcopter and fixed-wing UAV platforms. This motivates
application is embedded inside the X-Plane software, and
this study to find a comprehensive autotuning method for
it allows the modelling of user-designed aircraft.
quadplanes using the DE algorithm. In short, the research
Virtually any design can be built through this
work in this paper presents several key contributions,
software tool. After entering the aircraft's physical
including:
characteristics, such as weight, wingspan, controls, engine
• The neural network (NN) approach unifies the attitude
power, airfoil shapes, and others, the X-Plane simulator
dynamics model across the quadcopter, transition, and
replicates the aircraft's flight in a simulated environment
fixed-wing modes, significantly reducing the
by mirroring real-world conditions. Figure 2 compares the
development time, cost, and resource requirements for
simulation model of a hybrid quadplane UAV in X-plane
designing high-performance control systems for the
9 with the real quadplane UAV.
quadplane UAV. The NN model is also implemented for
autotuning PID using DE, making it applicable to
 Autotuning PID of Quadplane Hybrid UAV Using Differential Evolution Algorithm 343

(a) Actual quadplane (b) Constructed model in X-Plane

Figure 2 Comparison between the actual quadplane UAV and the constructed model in X-Plane 9.

2.3 SIL Simulation substituted with the models. Meanwhile, the "Quadplane
This section introduces a software-in-the-loop (SIL) Dynamics" block is replaced with a full-scale Quadplane
simulation method to evaluate and demonstrate quadplane UAV model in an X-plane flight simulator. The SIL
control. As depicted in Figure 3, the SIL test platform uses platform connects LabVIEW and X-Plane using the UDP
developed PID autopilot system models in the LabVIEW protocol, facilitating interaction between these two
environment. In this case, SIL's "Controller" block is environments.

Figure 3 SIL test platform.

The X-Plane flight simulator is crucial for developing LabVIEW computes the input controls for the quadplane,
this SIL and HIL test platform, which aids data exchange while X-Plane simulates UAV dynamics. Figure 4 shows
between X-Plane and LabVIEW. A LabVIEW-based PID the simulation environment in X-Plane with the LabVIEW
controller is applied to control the quadplane UAV during control interface.
simulation, managing its velocity, altitude, and attitude.

Figure 4 Quadplane flying in X-Plane environment with LabVIEW interface.

344 Mohammad Fahmi Pairan Syariful Syafiq Shamsudin Mohd Fauzi Yaakub

III. DYNAMIC EQUATION OF MOTION the right-hand rule, with the north, east, and down (NED)
directions. The quadplane inertia and body frame with
In this study, the inertial coordinate system's x, y, and z force and moment verctors are shown in Figure 5.
axes are defined using the Cartesian coordinate system of

Figure 5 Quadplane inertia and body frame with forces and moments vectors.

In analyzing the aircraft movements, a few coordinate used to obtain 12 nonlinear equations of aircraft motion as
systems are used, as follows: presented in Equations (3) to (5) and (8). In these equations,
• Inertial reference frame Fi (xn, ye, zd) Fb and Mb are the total force and moment that act on the
• Vehicle reference frame Fv (xv, yv, zv) plane in the body frame, m is aircraft mass, I is the inertia
• Body reference frame Fb (xb, yb, zb) matrix, and Rv/b is the rotation matrix from body to vehicle
• Wind reference frame Fw (xw, yw, zw) frame. The quadplane has one plane of symmetry, i.e., the
products of inertia 𝐼𝑥𝑦 are equal to zero. Both I and Rv/b
3.1 State Variable and Equations of Motion are given by Equation (6) and Equation (7), respectively.
The hybrid UAV is assumed to be a rigid body with
six degrees of freedom and 12 state variables. The state
vector of the system is described in Equation (1), where, 𝑢̇ 𝑝 𝑢
𝐹𝑏
𝑢, 𝑣, 𝑤, 𝑝, 𝑞, and 𝑟 are the projections of the translational [ 𝑣̇ ] = − ([𝑞 ] 𝑥 [ 𝑣 ]) (3)
𝑚
and rotational velocity in the body frame. The roll, pitch, 𝑤̇ 𝑟 𝑤
and yaw angles are respectively denoted by 𝜙, 𝜃, and 𝜓 to
determine the angular position of the plane relative to Fv,
𝑝̇ 𝑝 𝑝
while 𝑝𝑛 , 𝑝𝑒 , 𝑝𝑑 are position north, east, and down in
[𝑞̇ ] = 𝐼 −1 (𝑀𝑏 − ([𝑞 ] 𝑥 𝐼 [𝑞 ])) (4)
the Fi frame.
𝑟̇ 𝑟 𝑟
𝑋 = [𝑢 , 𝑣, 𝑤, 𝑝, 𝑞, 𝑟, ∅, 𝜃, 𝜓, 𝑝𝑛 , 𝑝𝑒 , 𝑝𝑑 ]𝑇 (1) 𝑝𝑛̇ 𝑢
[ 𝑞𝑒̇ ] = 𝑅𝑣/𝑏 [ 𝑣 ] (5)
In the meantime, the control vector is described in 𝑟̇𝑑 𝑤
Equation (2) to accommodate the flight mode operations.
In this equation, 𝑈1 , 𝑈2 , 𝑈3 , 𝑈4 are the control inputs for
thrust force, moment roll, moment pitch, and moment yaw, 𝐼𝑥𝑥 0 −𝐼𝑥𝑧
respectively. In addition, 𝛿𝑒𝑙𝑒 is elevator deflection input, 𝐼=( 0 𝐼𝑦𝑦 0 ) (6)
𝛿𝑎𝑖𝑙 is the flection of aileron, 𝛿𝑡ℎ𝑟 is the throttle input, −𝐼𝑥𝑧 0 𝐼𝑧𝑧
and 𝛿𝑟𝑢𝑑 is the deflection of rudder control surface.

𝑈 = [𝑈1 , 𝑈2 , 𝑈3 , 𝑈4 , 𝛿𝑒𝑙𝑒 , 𝛿𝑎𝑖𝑙 , 𝛿𝑡ℎ𝑟 , 𝛿𝑟𝑢𝑑 ]𝑇 (2) 𝑐𝜃 𝑐𝜓 𝑠∅ 𝑠𝜃 𝑐𝜓 − 𝑐 𝑠𝜓 𝑐∅ 𝑠𝜃 𝑐𝜓 + 𝑠∅ 𝑠𝜓

𝑅𝑣 = [𝑐𝜃 𝑠𝜓 𝑠∅ 𝑠𝜃 𝑠𝜓 + 𝑐 𝑐𝜓 𝑐∅ 𝑠𝜃 𝑠𝜓 − 𝑠∅ 𝑐𝜓 ] (7 )
𝑏
The equation of motion of a hybrid UAV is developed −𝑠𝜃 𝑠∅ 𝑐𝜃 𝑐∅ 𝑐𝜃
by applying Newton’s second law to both translational and
rotational equations of motion. Furthermore, it could be
 Autotuning PID of Quadplane Hybrid UAV Using Differential Evolution Algorithm 345

UAV. In the meantime, in Equation (12), the thrust motor

∅̇ 𝑝 + (𝑟 𝑐𝑜𝑠 ∅ + 𝑞 𝑠𝑖𝑛 ∅) 𝑡𝑎𝑛 𝜃 component is assumed to work along the z-axis of the body
[ 𝜃̇ ] = [ 𝑞 𝑐𝑜𝑠 ∅ − 𝑟 𝑠𝑖𝑛 ∅ ] (8) frame and is taken to be proportional to the sum of all
𝜓̇ (𝑞 𝑠𝑖𝑛 ∅ +𝑟 𝑐𝑜𝑠 ∅) 𝑠𝑒𝑐 𝜃 angular speeds of the propellers.
Meanwhile, the total moment acting on the aircraft
consists of the moments due to aerodynamic forces Mab
3.2 Forces and Moments Acting on Hybrid UAV
and motor thrust Mmotorb. Gravity acts at the center of mass
The force and moment components that occur at the
so that no moment is generated. The torque from the
center of gravity of an airplane are caused by gravity,
propeller rotation causes the moment in the engine. This
aerodynamic effects, engine thrust, and motor thrust.
effect is assumed to be relatively small and could be
Hence, the total force acting on the aircraft, as described
countered with aileron deflection. Thus, the moment due
in the body frame, consists of gravity force Fgb,
to engine components is considered negligible. Mab and
aerodynamic force Fab, engine thrust Fengb and motor thrust
Mmotorb are given by Equation (13) and Equation (14),
Fmotorb, which are given by Equations (9) to (12),
where ws is wingspan, c is mean chord, Cl, Cm, and Cn are
respectively. Note that in these equations, 𝜌 is the air
non-dimensional aerodynamic coefficients for roll, pitch,
density, 𝑉𝑎 is airspeed, 𝑆 is wing area, 𝐶𝑋, 𝐶𝑌, and 𝐶𝑍 are
and yaw moment, d is the drag coefficient, and l1,2 is the
the dimensionless force coefficients in the x, y, and z axes
distance of the motor center to the rotation axis.
denoted in the body frame, 𝑏 is thrust coefficient, and Ωi
is the angular speed of propeller i. 𝑤𝑠 𝐶𝑙
1
sin 𝜃 𝑀𝑎𝑏 = 𝜌𝑉𝑎2 𝑆 ( 𝑐 𝐶𝑚 ) (13)
2
𝐹𝑔𝑏 = 𝑚𝑔 ( cos 𝜃 sin ∅ ) (9) 𝑤𝑠 𝐶𝑛
cos 𝜃 cos ∅

𝐶𝑋 𝑏 𝑙1 (Ω12 + Ω22 − Ω23 − Ω24 )

1
𝐹𝑎𝑏 = 𝜌𝑉𝑎2 𝑆 (𝐶𝑌 ) (10) 𝑏
𝑀𝑚𝑜𝑡𝑜𝑟 = [ 𝑏𝑙2 (Ω12 + Ω22 − Ω23 − Ω24 ) ] (14)
2
𝐶𝑍 𝑑 (Ω12 + Ω22 − Ω23 − Ω24 )

𝐹𝑒𝑛𝑔 𝑥
𝑏
𝐹𝑒𝑛𝑔 = 𝑚𝑔 ( 0 ) (11) IV. RADIAL BASIS FUNCTION NEURAL
0 NETWORK

0 Figure 6 shows the RBF NN model, which mimics

𝑏 the quadplane’s dynamics. The RBF network is a
𝐹𝑚𝑜𝑡𝑜𝑟 = [ 0 ] (12)
−𝑏(Ω12 + Ω22 + Ω23 + Ω24 ) feedforward NN comprising a nonlinear hidden layer and
a linear output layer. The hidden layer utilizes basis
In Equation (11), the engine thrust component is functions as activation functions. A Gaussian function is a
assumed to work along the x-axis of the body frame, thus basis function that can be applied as an activation function
it does not produce a moment of the center of mass of the in the hidden neuron.

Figure 6 RBF NN model.

346 Mohammad Fahmi Pairan Syariful Syafiq Shamsudin Mohd Fauzi Yaakub

Equation (15) presents the Gaussian function, where 1 2

μ and σ are the center and the width of the radial basis in ŷ(t) = Bk + ∑hj=1 w1kj (exp (− ‖φ − μj ‖ )) (16)
σj 2
the hidden layer, respectively, and φ is the input vector
introduced into NN. This RBF model will be used in system identification
1
model structure selection and DE tuning PID as a
∅(𝑟) = − exp ‖𝜑 − 𝜇‖2 (15) dynamics model to calculate the cost function.
𝜎𝑗 2

The output of the RBF is a linear combination of the V. SYSTEM IDENTIFICATION

outputs of its hidden unit function. The output calculation
of the RBF structure representing a single hidden layer of In brief, system identification is a method to generate
the RBF network is given by Equation (16), where w1kj is a model of quadcopter dynamics from the obtained flight
the weight between the hidden layer and the output layer, data, which can offer improved prediction accuracy and an
μj and σj are the center and width of the Gaussian function understanding of the system's characteristics. This method
of the jth hidden layer, respectively, h is the hidden neuron, is based on RBF with the NNARX (Neural Network-Auto
Bk is the bias element for the output layer, and i, j, and k Regressive Structure with Extra Input) model structure,
denote the number of inputs, hidden neurons, and outputs and it can be directly compared to the simulation model
in the NN, respectively. [10]. The proposed system identification uses an RB-NN
model, which runs parallel to the quadplane plant, as
shown in Figure 7.

Figure 7 System identification flow based on the NNARX RBF model.

The regressor structure is applied to the RBF model, excitations to be used. According to recent research on
and it will predict the output based on training data. The hybrid UAVs, no specific input excitation is applicable to
training algorithm adjusts the output error, and the system all flight modes, and there are limited comprehensive
identification process will be continued until all of the studies on hybrid UAV system identification. Flight data
input-output pairs have been trained. The major steps of collection for quadplane UAVs is carried out in the
the system identification method include data collection, simulation environment (X-Plane Flight Simulation
model selection, estimation, and validation processes, Software), as illustrated in Figure 8. To simplify the flight
which are discussed in detail in the following sections. data acquisition experiments, the quadplane's input
excitation for the training data has been narrowed down to
5.1 Flight Data Collection only frequency-swept and doublet signals. Additionally,
Flight test data gathering is crucial for system the doublet input is decided to be the validation method for
identification as it extracts input-output data from in-flight data acquisition in both quadcopter and fixed-wing modes
quadplanes. Careful preparation of flight data is essential [21-28]. Automatic computer-generated frequency-sweep
for understanding system dynamics before applying real- inputs are applied in this study for better sweep input
time identification techniques. As a result, an appropriate excitation. They are generated using Equation (17), where
input excitation should be able to stimulate the dynamics A is the sweep amplitude, typically 10% of the maximum
of the quadplane that needs to be identified. The rotorcraft deflection limits, and θ(t) is provided by Equation (18)
and fixed-wing UAV each use different input excitations [10]. Meanwhile, the frequency progression is given
to stimulate their dynamics. Since the quadplanes have through Equation (19), where parameter K is given by
three flying modes, there are three different input Equation (20).
 Autotuning PID of Quadplane Hybrid UAV Using Differential Evolution Algorithm 347

Figure 8 Data collection process in the X-plane environment

𝛿𝑠𝑤𝑒𝑒𝑝 = 𝐴 sin[𝜃(𝑡)] (17) The training and validation of the data collection for
both flying modes are done by implementing doublet input
𝑇
𝜃(𝑡) = ∫0 𝑟𝑒𝑐 𝜔(𝑡) 𝑑𝑡 (18) by exciting the deflection of the elevator control surface,
δele, and the deflection of the aileron control surface, δail.
The procedure is outlined as follows:
𝜔 = 𝜔𝑚𝑖𝑛 + 𝐾( 𝜔𝑚𝑎𝑥 − 𝜔𝑚𝑖𝑛 ) (19)
• Step 1: The quadplane takes off and hovers 10 m above
𝐶 𝑡 the ground within one minute to validate data gathering
𝐾 = 𝐶2 [𝑒𝑥𝑝 ( 1 ⁄𝑇 ) − 1] (20) in quadcopter mode.
𝑟𝑒𝑐

• Step 2: The first doublet input is applied to δele to excite

For Equation (20), the values C1=4.0 and C2=0.0187 the dynamics model of the quadplane.
are found to be suitable for a wide range of applications
[10]. Moreover, the automated sweep should include • Step 3: The quadplane returns to steady hover
additional features like trim duration, constant frequency, conditions for 30 seconds after finishing the first
and sweep parameter values based on the frequency range doublet excitation.
and control input units [10]. Flight training data for quad • Step 4: Doublet input is repeated for the second and
mode will be collected in the velocity control mode at an third times to ensure sufficient quality data collection.
altitude of at least 10 m with wind conditions below 5 • Step 5: Repeat Step 2 to Step 4 for δail.
knots, as suggested by Tischler and Remple [10]. It should
be noted that the X-Plane flight tests have been conducted • Step 6: The quadplane is set to initial flight conditions
in calm weather conditions for roll and pitch angle. at 300 m above the ground with a 20 m/s initial forward
Frequency sweeps excite are done by exciting the speed for fixed wing mode.
deflection of the elevator control surface, δele, and the • Step 7: The first doublet input is applied to δele to excite
deflection of the aileron control surface, δail as follows: the dynamics model of the quadplane.
• Step 1: The quadplane takes off and hovers 10 m above • Step 8: The quadplane returns to steady hover
the ground within a one-minute duration. conditions for 30 seconds after finishing the first
• Step 2: First automatic frequency sweeps are applied to doublet excitation.
δele to excite the dynamics model of the quadplane in • Step 9: Doublet input is repeated for the second and
30-second intervals. third times to ensure sufficient quality data collection.
• Step 3: The quadplane returns to steady hover • Step 10: Repeat Step 7 to Step 9 for δail.
conditions for 30 seconds after the first frequency
sweeps. Table 1 tabulates the type of input excitation used for
• Step 4: Frequency sweeps are repeated for the second training and validation purposes, and the output/input
and third times to ensure sufficient quality data recorded during the flight data gathering.
collection.
• Step 5: Repeat Step 2 to Step 4 for δail.
348 Mohammad Fahmi Pairan Syariful Syafiq Shamsudin Mohd Fauzi Yaakub

Table 1 Inputs excitation for system identification of quadplane

No. Quad/FW Mode Output Input Training Validation

1 Pitch θ δele
Frequency Sweep Doublet
2 Roll ø δail

5.2 Model Structure Selection Considering the SISO case, the linear input-output
Choosing a model structure involves picking a relationship of the ARX model can be expressed as
specific configuration to aptly describe the connection Equation (21), where ny and nu are the size of the past
between the system's inputs and outputs. The choice of output and input, respectively. At the same time, e(t) is the
model structure ensures an accurate prediction of the flight disturbance. Alternatively, the general polynomial in
dynamics using a specific ARX structure, as illustrated in Equation (21) can be rewritten in a compact form, as
Figure 9. shown in Equation (22).

𝑦(𝑡) + 𝑎1 𝑦(𝑡 − 1) + ⋯ + 𝑎𝑛𝑦 𝑦(𝑡 − 𝑛𝑦 ) = 𝑏1 𝑢(𝑡) + 𝑏2 𝑢(𝑡 − 2) + ⋯ + 𝑏𝑛𝑥 𝑢(𝑡 − 𝑛𝑢 ) + 𝑒(𝑡) (21)

𝑦̂(𝑡 + 𝑘|𝑡, 𝜃) = 𝜑(𝑡 + 𝑘)𝜃 + 𝑒(𝑡) (22) this case reflect the ARX model coefficients and time-
lagged input-output data.
The disturbances are part of the system dynamics, so
they are not included in the ARX model of RBF in Figure 𝜃 = [𝑎1 𝑎1 ⋯ 𝑎𝑛𝑦 𝑏1 𝑏2 ⋯ 𝑏𝑛𝑥 ] (23)
9. Equation (23) and Equation (24) indicate the parameter
vector and time regression vector, respectively, which in

𝜑(𝑡) = [𝑦(𝑡 − 1) ⋯ 𝑦(𝑡 − 𝑛𝑦 ) 𝑢(𝑡 − 1) ⋯ 𝑢(𝑡 − 𝑛𝑢 ) ] (24)

Jacobian matrix, and αmax and αmin are the maximum and
minimum eigenvalues, respectively. Note that the CT
method introduces αmax and αmin parameters to bound the
eigenvalues of P(t) to overcome the instability and
covariance blow-up if the NN model has many parameters
[30].
−1
𝐾(𝑡) = 𝑃(𝑡 − 1)Π(𝑡)(1 + Π 𝑇 (𝑡)𝑃(𝑡 − 1)Π(𝑡)) (25)

θ̂(𝑡) = θ̂(𝑡 − 1) + 𝐾(𝑡)𝑒𝑡 (26)

𝑃̅ (𝑡) = 𝑃(𝑡 − 1) − 𝐾(𝑡)Π 𝑇 (𝑡)𝑃(𝑡 − 1) (27)

𝛼𝑚𝑎𝑥 −𝛼𝑚𝑖𝑛
𝑃(𝑡) = 𝑃̅ (𝑡) − 𝛼𝑚𝑖𝑛 𝛪 (28)
𝑡𝑟𝑎𝑐𝑒(𝑃̅ (𝑡))
Figure 9 ARX model of RBF for pitch and roll
5.4 Model Validation
There are several ways to validate the NN model,
5.3 Model Estimation such as one step-ahead prediction (OSA), mean squared
Model estimation in the SI process involves error (MSE), correlation tests, and chi-square tests [31]. In
determining NN parameters that minimize the error this study, OSA and MSE have been selected to validate
between identified systems and true models. The rGN the performance of the proposed network model. The OSA
algorithm with Constant Trace (CT) is selected for NN is a simple plot that compares actual measurement data
training due to its ability to avoid full Hessian matrix with the model prediction over the test data set. The
inversion, produce faster predictions, and adapt to prediction error calculation, such as the root mean square
dynamic changes, although the model structure is not error (RMSE) and the coefficient of determination, R2
optimized [29]. The CT algorithm equations are given by criterion can also be used to describe how well the NN
Equations (25) to (28), where P(t) represents the model matches the actual measurement data. These two
covariance matrix, θ̂(𝑡) represents the updated NN prediction error criteria are given in Equation (29) and
parameter vector, et represents prediction error, Π(t) is the Equation (30).
 Autotuning PID of Quadplane Hybrid UAV Using Differential Evolution Algorithm 349

1 of the crossover, each individual for the new generation

𝑅𝑀𝑆𝐸 = √ ∑𝑁 ̂𝑖 (𝑡) − 𝑦𝑖 (𝑡))2
𝑡=1(𝑦 (29) candidate has been finalized. Equation (33) defines the
𝑁
crossover operation, where 𝑢𝑖,𝐺+1 refers to the
∑(𝑦̂ (𝑡)−𝑦 (𝑡))2 individual candidate that is produced for the next
𝑅2 = 1 − ∑(𝑦𝑖 (𝑡)−𝑦̅̅̅(𝑡))
𝑖
2 (30)
𝑖 𝑖 generation, CR is the crossover ratio, and r𝑗 stands for
random number.
𝑦̂𝑖 (𝑡) is the predicted RBF NN output, 𝑦𝑖 (𝑡) is the
measurement output of the quadplane, N is the number of 𝑣𝑖,𝐺+1 𝑖𝑓 𝑟𝑗 ≤ 𝐶𝑅
outputs, 𝑦̅(𝑡) is the mean value of the measurement data. 𝑢𝑖,𝐺+1 = { } (33)
𝑖 𝑥𝑖,𝐺 𝑜𝑡ℎ𝑒𝑟𝑤𝑖𝑠𝑒
Following the selection of a suitable NN architecture, the
RBF model is utilized to predict the quadplane output, • Selection: The selection is based on the performance of
which is then compared with the reference input. The error each target vector, Xi,G by comparing its fitness value
is used in the calculation of the cost function in the DE with that of the trial vector, Ui,G. In this case, the vector
algorithm. with a lower fitness value is selected for the next
generation. This procedure is given by Equation (34).
VI. DIFFERENTIAL EVOLUTION
ALGORITHM 𝑈𝑖,𝐺 𝑖𝑓 𝑓 (𝑈𝑖,𝐺 ) ≤ 𝑓(𝑋𝑖,𝐺 )
𝑋𝑖,𝐺+1 = { } (34)
𝑋𝑖,𝐺 𝑜𝑡ℎ𝑒𝑟𝑤𝑖𝑠𝑒
DE is a population-based metaheuristic algorithm
that uses the population of D-dimensional individuals to These steps are repeated until the pre-set number of
optimize problems. It is composed of four major phases, iterations has been achieved.
which are initialization, mutation, crossover, and selection.
It starts with the random generation of candidate solutions.
After that, the mutation, crossover, and selection VII. DE BASED PID CONTROLLER DESIGN
procedures are conducted on the candidate solutions to
create trial vectors. Conducting these operations The DE method and RBF neural network model are
sequentially constitutes one cycle of the DE algorithm, used to autotune the quadplane UAV's attitude PID. The
which leads to the new generation. After each cycle, a new system identification mechanism and DE PID controller
generation of candidate solutions is formed, and the are built within the LabVIEW 2019 software environment.
process continues until the user-defined stopping criteria Figure 10 shows the general procedure for autotuning PID
are met. Each of the individual members of the population using DE and the RBF network model. This study uses the
is a candidate solution, and it is defined as in Equation (31), GA to determine the PID gain controller's ideal tuning
where D is the problem's dimensionality, NP is the parameters. A population of initial solutions of 𝑘𝑝 , 𝑘𝑖 ,
population size, and i takes values of 1, 2…, NP. and 𝑘𝑑 of PID gains will first be produced via the DE
algorithm. The population size and the spectrum of PID
{𝑋𝑖𝐺 = (𝑋𝑖,1
𝐺 𝐺
, 𝑋𝑖,2 𝐺
, … , 𝑋𝑖,𝐷 )} (31) improvements are the two main requirements for this step.
The intended input will then be used to run a PID with an
In brief, the different DE algorithm phases are further RBF neural network model created from system
described as follows: identification to evaluate the objective function using the
• Initialization: The algorithm is initialized by defining first PID gains. The objective function selected is the
three parameters, which are the population size (NP), quadratic cost function in Equation (35), which comes
mutation factor (F), and crossover rate (CR). Available from the PID control system. The following is the equation
candidate solutions inside the defined search space are for the objective function:
then randomly selected to make the initial generation.
• Mutation: The mutation strategy 'DE/rand/1/exp' is used 𝐽 = (𝑥 𝑇 𝑄𝑥 + 𝑢𝑇 𝑅𝑢)𝑑𝑡 (35)
to generate mutation vectors in each generation. The
mutation step is performed in a pool of at least three After obtaining the goal function, DE algorithms
individuals, with the difference between the two execute genetic operations such as selection, crossover,
randomly selected individuals multiplied by F and then and mutation to create and modify new PID gains. The DE
added to the third individual. This computation is method will iterate until the generated number reaches the
presented by Equation (32), and this is done for each maximum value, provided that the generated number is
individual in the population. smaller than the maximum. If this condition is satisfied,
the most recent PID gains are considered optimal.
𝑉𝑖𝐺 = 𝑋𝑟1
𝐺 𝐺
+ 𝐹. (𝑋𝑟2 𝐺)
− 𝑉𝑟3 (32)
VIII. RESULTS AND DISCUSSION
• Crossover: The mutation-induced individual from the
previous phase undergoes crossover with the existing The data collection experiment is done on X-Plane 9
one. The crossover step produces a random parameter if environment, as shown in Figure 11.
candidates exceed the crossover ratio threshold or is
selected from the existing one if it does not. At the end
350 Mohammad Fahmi Pairan Syariful Syafiq Shamsudin Mohd Fauzi Yaakub

Figure 10 Overview process of PID autotuning using DE with RBF network model.

Figure 11 Flight test data collection in X-Plane

 Autotuning PID of Quadplane Hybrid UAV Using Differential Evolution Algorithm 351

Figure 12 Frequency sweep in δele input axis

Figure 13 Frequency sweep in δail input axis

The frequency sweep inputs are applied to the Table 2 Amplitude and frequency value for frequency
elevator and aileron as shown in Figure 12 and Figure 13, swept parameter.
following previously discussed guidelines while
maintaining 70% full throttle. The frequency ranges of the
No. Mode Attitude A 𝛚𝐦𝐢𝐧 𝛚𝐦𝐚𝐱
sweep inputs of the quadplane’s pitch, roll, aileron, and
elevator input are recorded approximately between 0.3 1 Pitch 0.1
rad/s and 12 rad/s. Table 2 presents the frequency swept Quad
parameter used in flight data collection. 2 Roll 0.05
Five fixed hidden neurons of the RBF with two past 0.3 12
outputs and one past input NNARX structure are chosen 3 Pitch
for the real-time identification of the quadcopter. The trial- FW 0.2
and-error method with hold-out validation is employed to 4 Roll
determine the regression vector structures and the hidden
neuron RBF. The RBF is trained using the CT algorithm,
with αmax and αmin set to 10 and 0.1, respectively. The mean
training time, RMSE, and R2 are used to compare how
well the predictions were made. Table 3 summarizes the
RBF network identification.
352 Mohammad Fahmi Pairan Syariful Syafiq Shamsudin Mohd Fauzi Yaakub

Table 3 RBF network specification The OSA prediction of roll and pitch angle responses
from the 2-1 RBF NN structure is shown in Figure 14. In
Number of hidden neurons 5 the figure, the red solid line represents the predicted output
from RBF, and the black line represents the measured
Number of past outputs 2 angle from the X-Plane simulation. It can be observed that
the NNARX RBF model is able to adapt well to dynamic
Number of past inputs 1 changes and follow a smooth path with an improved
prediction for the remaining data. Meanwhile, Figure 15
Radial basis shows the roll and pitch angle responses as predicted from
Activation function in hidden layer function the 2-1 RBF NN structure for the transition mode.
In Figure 14, the black solid line represents the
Linear measured X-Plane simulation attitude data, whereas the
Activation function in output layer function solid blue line represents the predicted output from RBF
𝛼𝑚𝑎𝑥 =10 NN. These attitude responses from the RBF NN model
CT training parameter seem to be overlaid with the measured quadplane
𝛼𝑚𝑖𝑛 =0.1 responses almost perfectly. It can be seen from the data in
Figure 14 that the 2-1 RBF trained using the CT algorithm
Total number of regressors 6 is capable of following a smooth path and adapting well to
dynamic changes with good accuracy.

20
6 Roll angle Pitch angle
Roll angle RBF 15
Pitch angle RBF
4
10
Pitch angle (Degree)
Roll angle (Degree)

2 5

0 0
15

-2 -5 10

4
5

-10 0

-4 0

-5
-2

-4 -15 -10

-6 8100 8200 8300 8400 8500 8600 8700 8800 -15

10600 10700 10800 10900

-20
0 2000 4000 6000 8000 10000 12000 14000 16000 18000 0 2000 4000 6000 8000 10000 12000 14000 16000 18000
Sampling Time Time (Samples)
(a) Roll angle (b) Pitch angle

Figure 14 RBF System identification for quadplane roll and pitch angle in quadcopter mode

(a) Roll angle (b) Pitch angle

Figure 15 RBF System identification for quadplane roll and pitch angle in transition mode.
 Autotuning PID of Quadplane Hybrid UAV Using Differential Evolution Algorithm 353

0.20 0.00
Roll Pitch
0.15 Roll RBF -0.05 Pitch RBF

0.10 -0.10

0.05 -0.15
Roll (rad)

Pitch (rad)
0.00 -0.20
0.15
-0.10

0.10
-0.05 -0.15

0.05
-0.25
-0.20

0.00
-0.10
-0.05
-0.30 -0.25

-0.30

-0.15 -0.10

-0.15
-0.35 -0.35

3350 3400 3450 3500 3550 3600 2850 2900 2950 3000 3050 3100 3150

-0.20
0 1000 2000 3000 4000 5000 -0.40
0 1000 2000 3000 4000 5000
Sampling Time
Sampling Time
(a) Roll angle (b) Pitch angle

Figure 16 RBF System identification for quadplane roll and pitch angle in fixed wing mode.

The plotted transition mode of pitch and roll rate represents the X-Plane data. It can be observed from the
responses, which are derived from the RBF model, are figure that the predicted outputs from the RBF are able to
visually depicted in Figure 15 alongside the corresponding track the actual attitude measurement of the quadplane in
actual attitude responses obtained from the X-Plane. The fixed wing mode. Moreover, RBF outputs closely match
predicted output of the RBF is depicted by the red solid the actual measured outputs of the quadplane. These
line in the graph, whereas the X-Plane data is represented results also indicate that the proposed RBF model has been
by the black dotted line. It can be observed from the figure appropriately trained to imitate the attitude dynamics of
that the predicted outputs from the RBF are able to track the quadplane in fixed-wing mode with good accuracy.
the actual attitude measurement of the quadplane in Overall, Table 4 summarizes the performance of RBF
transition mode. Additionally, it is noteworthy that the prediction in terms of training time, RMSE, and R2. Based
outputs produced by the RBF closely align with the on the results, it can be concluded that the proposed RBF
measured outputs of the quadplane. NN system identification can successfully predict the
On the other hand, the fixed wing pitch and roll rate attitude dynamics of the quadplane in all flight modes.
responses obtained from the RBF model are plotted Therefore, the complexity of switching flight modes
against actual attitude responses from the X-Plane in between the two discrete flight profiles and their dynamic
Figure 16. The blue solid line in the graph indicates the models can be ruled out.
predicted output of RBF, while the black solid line

Table 4 Performance of RBF system identification

Mean Computing
No. Mode Attitude R2 RMSE
Time (ms)
1 Pitch 0.956 0.0046
Quad 4.96
2 Roll 0.988 0.0056
3 Pitch 0.975 0.0064
Transition 4.97
4 Roll 0.957 0.0116
5 Pitch 0.996 0.0032
Fixed Wing 4.99
6 Roll 0.936 0.0080

The detailed DE algorithm parameters used in tuning PID parameter range, as shown in Table 6. Table 7 reveals
the PID controller are presented in Table 5. Note that D each PID parameter's value for each controller type. The
has been set based on the number of parameters used in the values are achieved automatically using the DE algorithm
objective function. In this case, problem dimension refers and manual tuning. Manual tuning of PID controls
to the number of PID parameters (i.e., kp, ki and kd) that involves setting the gains through a trial-and-error method.
equals to three. The boundary constraint is set based on the
354 Mohammad Fahmi Pairan Syariful Syafiq Shamsudin Mohd Fauzi Yaakub

Table 5 DE algorithm parameters Table 7 PID gain coefficients for quad, transition and
fixed wing modes.
Population Size 20
Scale Factor (F) 0.90 Tuning Attitude kp ki kd
Crossover Probability (CR) 0.95 1.50
Roll 0.08 0.00
Mutation Method Random Manual
Pitch 0.50 0.08 0.00
Iteration 50
Roll 2.39 0.32 0.34
DE
Pitch 1.80 0.22 0.01
Table 6 PID parameter range
This involves setting the integral and derivative gains
Parameter Maximum Minimum to zero, increasing the proportional gain until oscillation is
observed, and then increasing the integral gain until the
kp 5 0
offset is corrected. The derivative gain can be increased to
ki 1 0 reduce overshoot if overshoot occurs. The PID gains are
the same as those used for flight data collection and have
kd 1 0 been tested on X-Plane 9.

0.30
0.25
Setpoint Roll
0.25 Manual Roll
0.20

Pitch 0.20
0.15 Manual Pitch
Roll (rad)
Pitch (rad)

0.15
0.10
0.10

0.05
0.05

0.00
0.00

-0.05 -0.05
0 200 400 600 800 0 200 400 600 800
Time (s) Time (s)
(a) Roll angle (b) Pitch angle

Figure 17 Step response of manual tuning PID.

0.25 0.30

0.25
SetpointRoll
0.20
DE Roll
Setpoint Pitch Angle 0.20
0.15
DE Pitch
Pitch Angle (rad)

0.15
0.10
Roll (rad)

0.10

0.05 0.05

0.00 0.00

-0.05
-0.05
-0.10
-0.10
-0.15
0 200 400 0 200 400

Time (s) Time (s)

(a) Pitch (b) Roll

Figure 18 Step response of DE tuning PID.

 Autotuning PID of Quadplane Hybrid UAV Using Differential Evolution Algorithm 355

The PID controller is tested under a 0.2 rad step input services, offering a faster, more efficient, and eco-friendly
reference signal. The research compares the time domain logistic solution.
performance of manual PID and PID-DE in the attitude of
the quadplane in terms of percentage overshoot, rise time, ACKNOWLEDGMENTS
and peak time. Figure 17 and Figure 18, respectively, show
the typical step responses for manual PID and PID-DE in The authors acknowledge the support for the research
the attitude of the quadplane. DE tuning has been shown study from Universiti Tun Hussein Onn Malaysia (UTHM)
to significantly improve the roll and pitch performance of through Tier1 (Vot: Q112).
a quadplane UAV compared to manual tuning. The results
show that roll and pitch angle responses from PID-DE REFERENCES
have a lower percentage overshoot and a faster rise and
peak time compared to the manual tuning PID. Table 8 [1] Nguyen HT, Quyen TV, Nguyen CV, Le AM, Tran
summarizes the time domain performance for manual HT, Nguyen MT, “Control algorithms for UAVs: A
tuning and the PID-DE. comprehensive survey,” EAI Endorsed Transactions
on Industrial Networks and Intelligent Systems, Vol.
7, No. 23, 2020.
Table 8 Manual PID and DE PID performance. [2] Ngo TL, Hoang DT, Le THH, “Dynamics modeling
and pitching parameters identification of a novel
Performance Attitude Tuning Value hybrid UAV,” Journal of Aeronautics, Astronautics
and Aviation, Vol. 52, No. 1, 2020, pp. 25-36.
Manual 10.00 [3] Nasir REM, Mustapha MF, Mohamad F, Hamid
Pitch
DE 4.15 AHA, Noh MHM, “Low-fidelity aerodynamic
Overshoot
computational simulation on fixed wing-multirotor
(%) Manual 29.65
Roll hybrid unmanned aerial vehicle,” Journal of
DE 27.85 Aeronautics, Astronautics and Aviation, Vol. 55, No.
Manual 283 3S, 2023, pp. 439-452.
Pitch [4] Goetzendorf-Grabowski T, Mieloszyk J, Figat M,
Peak time, DE 99 Tarnowski A, Hernik B, “Lightweight UAV for
Tp (s) Manual 141 emergency medical service,” Proceeding of the 32 th
Roll Congress of the International Council of the
DE 46 Aeronautical Sciences (ICAS), 2021.
Manual 113 [5] Yu S, Kwon Y, “Development of VTOL drone for
Pitch stable transit flight,” Journal of Computer and
Rise time, DE 29
Communications, Vol. 5, No. 7, 2017, pp. 36-43.
Tr (s) Manual 38 [6] Saeed AS, Younes AB, Islam S, Dias J, Seneviratne
Roll
DE 10 L, Cai G, “A review on the platform design, dynamic
modeling and control of hybrid UAVs,” Proceeding
of International Conference on Unmanned Aircraft
Systems, 2015.
IX. CONCLUSIONS
[7] Escareno J, Stone RH, Sanchez A, Lozano R,
“Modeling and control strategy for the transition of
In conclusion, this study introduces a comprehensive
a convertible tail-sitter UAV,” Proceedings of the
approach to address the dynamics and control challenges
European Control Conference, 2007.
associated with the use of quadplane UAVs. The utilization
[8] Naldi R, Marconi L, “Optimal transition maneuvers
of DE for autotuning attitude PID controllers, combined
for a class of V/STOL aircraft,” Automatica, Vol. 47,
with the integration of a unified dynamics model derived
No. 5, 2011, pp. 870-879.
through system identification with X-Plane simulation,
[9] Morelli EA, Klein V, Aircraft System Identification:
has yielded a significant improvement in the UAV
Theory and Practice. Williamsburg, Virginia,
performance. The RBF NN system has successfully
Sunflyte Enterprises, 2016.
predicted quadplane attitude dynamics with a high R2
[10] Tischler MB, Remple RK, Aircraft and Rotorcraft
value. The DE algorithm also improves the roll and pitch
System Identification. Reston, Virginia, American
angle responses, achieving lower overshoot, faster rise,
Institute of Aeronautics and Astronautics, 2012.
and peak times, and eliminating manual PID tuning effort.
[11] Shamsudin SS, “The Development of Neural
Future work could expand the NN system identification to
Network Based System Identification and Adaptive
include MLP, RNN, and deep learning networks for better
Flight Control for an Autonomous Helicopter
prediction accuracy and convergence rate. Moreover,
System,” PhD Thesis, University of Canterbury,
autotuning PID with self-organization can reduce
2013.
computation time and improve algorithm speed. All in all,
[12] Saeed AS, Younes AB, Cai C, Cai G, “A survey of
the advancements in control algorithms and systems have
hybrid unmanned aerial vehicles,” Progress in
aided in the feasibility and viability of drone delivery
Aerospace Sciences, Vol. 98, 2018, pp. 91-105.
 356 Mohammad Fahmi Pairan Syariful Syafiq Shamsudin Mohd Fauzi Yaakub

[13] Awang Jumat N, Ogunwa TT, Abdullah EJ, Chahl J, [26] Wei W, Schwartz N, Cohen K, “Frequency-domain
Romli FI, Abdul Majid DL, “Flapping actuation system identification and simulation of a quadrotor
using temperature feedback control of coated shape controller,” Proceedings of the AIAA Modeling and
memory alloy actuators,” Microsystem Technologies, Simulation Technologies Conference, 2014.
Vol. 27, 2021, pp. 3299-3311. [27] Lee Y, Kim S, Suk J, Koo H, Kim J, “System
[14] Wang W, Yuan X, Zhu J, “PID tuning via differential identification of an unmanned aerial vehicle from
evolution for quadrotor UAVs trajectory tracking,” automated flight tests,” Proceedings of the 1st UAV
Proceedings of the IEEE Symposium Series on Conference, 2002.
Computational Intelligence, 2016. [28] Licitra G, Bürger A, Williams P, Ruiterkamp R,
[15] Abdullah EJ, Majid DL, Romli FI, Gaikwad PS, Diehl M, “Aerodynamic model identification of an
Yuan LG, Harun NF, “Active control of strain in a autonomous aircraft for airborne wind energy,”
composite plate using shape memory alloy Optimal Control Applications and Methods, Vol. 40,
actuators,” International Journal of Mechanics and No. 3, 2019, pp. 422-447.
Materials in Design, Vol. 11, 2016, pp. 25-39. [29] Shamsudin SS, Chen X, “Recursive Gauss-Newton
[16] Abdollahi A, Tabar AF, Khodadadi H, “Optimal based training algorithm for neural network
controller design for quadrotor by genetic algorithm modelling of an unmanned rotorcraft dynamics,”
with the aim of optimizing the response and control International Journal of Intelligent Systems
input signals,” Cumhuriyet University Faculty of Technologies and Applications, Vol. 13, 2014, pp.
Science Journal, Vol. 36, No. 3, 2015, pp. 135-147. 56-80.
[17] Mohammed MJ, Rashid MT, Ali AA, “Design [30] Shamsudin SS, Chen X, “Identification of an
optimal PID controller for quad rotor system,” unmanned helicopter system using optimised neural
International Journal of Computer Applications, Vol. network structure,” International Journal of
106, No. 3, 2014, pp. 15-20. Modelling, Identification and Control, Vol. 17, No.
[18] Gün A, “Attitude control of a quadrotor using PID 3, 2012, pp. 223-241.
controller based on differential evolution algorithm,” [31] Mashor MY, “Performance comparison between
Expert Systems with Applications, Vol. 229, 2023, HMLP, MLP and RBF networks with application to
120518. on-line system identification.” Proceedings of IEEE
[19] Saad MS, Jamaluddin H, Darus IZM, Conference on Cybernetics and Intelligent Systems,
“Implementation of PID controller tuning using 2004.
differential evolution and genetic algorithms,”
International Journal of Innovative Computing,
Information and Control, Vol. 8, No. 11, 2012, pp.
7761-7779.
[20] Trias A, Yahya A, Syahrullail S, Baharom A, Nor
Liyana SH, “PID controller tuning by differential
evolution algorithm on edm servo control system,”
Applied Mechanics and Materials, Vol. 284, 2013,
pp. 2266-2270.
[21] Cai G, Al Mehairi H, Al-Hosani H, Dias J,
Seneviratne L, “Frequency-domain flight dynamics
model identification of MAVs-miniature quad-rotor
aerial vehicles,” Proceedings of the IEEE/RSJ
International Conference on Intelligent Robots and
Systems, 2014.
[22] Adiprawita W, Ahmad AS, Sembiring J,
“Automated flight test and system identification for
rotary wing small aerial platform using frequency
responses analysis,” Journal of Bionic Engineering,
Vol. 4, No. 4, 2007, pp. 237-244.
[23] Agha M, “System Identification of a Circulation
Control Unmanned Aerial Vehicle,” PhD Thesis,
University of Denver, 2017.
[24] Bhandari S, Navarro P, Ruiz A, “Flight testing, data
collection, and system identification of a multicopter
UAV,” Proceedings of the AIAA Modeling and
Simulation Technologies Conference, 2017.
[25] Dorobantu A, Murch A, Mettler B, Balas G, “System
identification for small, low-cost, fixed-wing
unmanned aircraft,” Journal of Aircraft, Vol. 50, No.
4, 2013, pp. 1117-1130.

View publication stats