Adaptive proportional integral derivative deep feedforward network for quadrotor trajectory-tracking flight control

International Journal of Electrical and Computer Engineering (IJECE)
Vol. 12, No. 4, August 2022, pp. 3607~3619
ISSN: 2088-8708, DOI: 10.11591/ijece.v12i4.pp3607-3619  3607
Journal homepage: http://ijece.iaescore.com
Adaptive proportional integral derivative deep feedforward
network for quadrotor trajectory-tracking flight control
El Ayachi Chater1
, Halima Housny2
, Hassan El Fadil2
1
LASTIMI Laboratory, Higher School of Technology, Mohammed V University, Rabat, Morocco
2
ISA Laboratory, National School of Applied Sciences, Ibn Tofail University, Kenitra, Morocco
Article Info ABSTRACT
Article history:
Received Apr 29, 2021
Revised Mar 21, 2022
Accepted Apr 2, 2022
When the controlled system is subject to parameter variations and external
disturbances, a fixed-parameter proportional integral derivative (PID)
controller cannot ensure its stabilization. In this case, its control requires
online parameter adjustment. Specifically, as the quadrotor is a multi-input
multi-output, nonlinear, and underactuated system, robust control is
necessary to ensure efficient trajectory tracking flights. In this paper, an
adaptive proportional integral derivative (APID) controller is proposed to
control the quadrotor systems. This APID-based control strategy uses a two
hidden layer deep feedforward network (DFN), where the one-step secant
algorithm is chosen for initializing the DFN parameters. All the design steps
of the proposed adaptive controller are described. The multidimensional
particle swarm optimization (PSO) algorithm is used for tuning the DFN
parameters. Then, using two simulation efficiency tests, a comparison
between the proposed PSO-based APID-DFN, the (non-optimized)
APID-DFN, the feedforward network APID, and the fixed-parameter PID
controllers proves much efficiency of the proposed adaptive controller. The
results illustrate that the proposed PSO-based APID-DFN controller can
ensure good quadrotor system stabilization and achieve minimum overshoot
and faster convergence speed for all quadrotor motions. Thus, the proposed
control strategy could be considered an additional intelligent method-based
adaptive control for unmanned aerial vehicles.
Keywords:
Adaptive proportional integral
derivative
Deep neural network
Feedforward network
Multidimensional particle
swarm optimization
Quadrotor system
This is an open access article under the CC BY-SA license.
Corresponding Author:
El Ayachi Chater
LASTIMI Lab., Higher School of Technology, Mohammed V University
Rabat, Morocco
Email: elayachi.chater@est.um5.ac.ma
1. INTRODUCTION
Over the last years, the research control community has shown an increasing interest in flying
vehicles without onboard human pilots known as unmanned aerial vehicles (UAVs). UAVs’ civilian
applications have increased in diverse fields due to their low cost. Their application ranges from homeland
security, disaster relief, and weather forecasting to power line inspection and precision agriculture [1]–[3].
Besides, as in the outdoors environment, the aerial vehicles are exposed to adverse atmospheric conditions,
reliable and robust control strategies are necessary.
Many papers deal with the control problem of the multi-rotor UAV systems. Some of these works
propose the development of linear controllers like the proportional integral derivative (PID) and
linear-quadratic regulators (LQR) [4], [5]. Besides, other works suggest the development of nonlinear
methods to ensure UAV system stability. Among the latter, we can list the backstepping control approach [6],
[7] and the sliding mode approaches [8], [9]. In addition, intelligent control strategies have been vastly used

 ISSN: 2088-8708
Int J Elec & Comp Eng, Vol. 12, No. 4, August 2022: 3607-3619
3608
to stabilize the quadrotor system, such as fuzzy logic control [10]–[12], artificial neural network (ANN) [13],
and adaptive fuzzy inference system [14], [15]. Specifically, due to their simplicity and high reliability, the
proportional integral derivative (PID) controllers have been widely used in industrial systems. However, the
conventional PID controllers are not always satisfying in nonlinear systems such as the highly coupled
quadrotor system.
Several research works are still proposed for stabilizing and controlling the quadrotor system flight.
For instance, a real-time experimental test is implemented to evaluate the PID controller for the quadrotor
system. First, the PID controller parameters were selected to have the desired energy consumption [16].
Then, a PID flight controller for a small aerial vehicle was presented in [17], where the proposed algorithm
was deployed using Bluetooth® low energy connection via a personal area network (PAN). Besides, a design
and implementation solution that utilizes low-cost components with a PID control for a quadrotor system was
proposed with acceptable experimental test results [4].
Indeed, the most crucial step for the PID controller design is finding appropriate gains that permit
the stabilization of the nonlinear system despite system parameter variation and external disturbances. Thus,
several methods are utilized for tuning the PID control gains. For instance, the PID parameters could be
determined using the classical Zeigler-Nichols (ZN) method with tests for attitude stabilization of a quadrotor
system. However, despite the easy implementation of the ZN method, it could generally generate an
unacceptable overshoot. Therefore, optimization techniques are usually used to (offline) tune the PID
controller gains according to a fitness function. Among these techniques, particle swarm optimization (PSO)
[18], genetic algorithm (GA) [19], and ant colony optimization (ACO) are usually utilized [20]. These
techniques are shown to be better than the ZN method. However, they only provide constant control
parameters, which could lead to small control quality in the presence of external disturbances. To overcome
this drawback, the online tuning of the PID control parameters is more effective. Mainly, combining the PID
control strategy with intelligent techniques improves the quadrotor control performances. Thus, several
recent published works have incorporated intelligent algorithms to ensure online PID parameter adjustment,
such as fuzzy logic control (FLC) [21], [22] and neural networks (NN) [23]–[25].
Artificial neural networks (ANNs) have been widely studied due to their system identification and
control design advantages. The deep neural network (DNN) is primarily defined as an exciting area of neural
network (NN), which has gained popularity in recent years. The architecture of DNN is based on a
multi-layer NN that contains multiple hidden layers. A typical model of DNN is the deep feedforward
network (DFN). This self-learning algorithm allows filtering information through multiple hidden layers in
forwarding directions using the ANN algorithm. Significantly, the DNN has enabled significant progress in
sound and image processing applications, including feature detection, facial recognition, object
identifications, computer vision, and text classification [26]–[28]. Besides, potential applications of DNN are
numerous in control system engineering [29]. NN and similar approaches such as DNN and DFN can provide
better results when used to online tune the PID controller parameters [13], [23]–[25] for controlling the
quadrotor system. For instance, an ANN was utilized to adjust the PID parameters [23], where a comparative
study of three types of training methods (Bayesian regularization, Levenberg-Marquardt, and scaled
conjugate gradient) was performed to minimize the mean square error (MSE) when controlling the roll, pitch,
yaw, and altitude of a quadrotor system. Then, as the proposed adaptive ANN-based PID can improve the
quadrotor tracking performances, a decentralized PID neural network (PIDNN) control scheme was proposed
to stabilize a quadrotor’s attitude in the presence of the Dryden model of wind disturbance. Simultaneously, a
conventional PID controller was utilized in the outer loop to generate the inner-loop reference path [24].
Then, it was shown that this controller could reject the external disturbances with good stability. An adaptive
PID controller was also proposed to stabilize a quadrotor system’s attitude and position with unknown
variable payloads. It was shown that the neural network could deal with the unknown variable payload by
online tuning of the PID control parameters [25]. A comparative analytical method has also shown the
advantages of ANNs over a UAV attitude controller’s PID conventional control method [13].
Inspired by the above works and considering their result limitations, this paper aims to present a
new control strategy based on an adaptive PID deep feedforward network (APIDDFN) for quadrotor UAVs.
Specifically, the proposed learning control solution uses: i) deep feedforward network composed of two
hidden layers, ii) the one-step secant (OSS) method selects the deep feedforward network’s initial weights
and bias, iii) a multidimensional particle swarm optimization (PSO) algorithm to adjust the elaborated deep
feedforward network (DFN), and iv) a quadrotor (model-based) simulation setup in Dryden wind turbulence
and assuming a payload mass change during the vehicle motions. It is worth noticing that the system model is
always highly simplified, the relation between the body-fixed and earth-fixed velocities is omitted, and
external disturbances are ignored in almost previous works related to controlling the quadrotor UAVs. No
simplifications are used in this work.

Int J Elec & Comp Eng ISSN: 2088-8708 
Adaptive proportional integral derivative deep feedforward network for … (El Ayachi Chater)
3609
Our new result may be summarized as follows: using the DFN learning algorithm, a performance
evaluation obtained through analyses and a model-based simulation setup shows that combining a PID
controller with the DFN algorithm allows designing an efficient adaptive controller that can stabilize the
aerial vehicle Euler angles and guarantee a perfect trajectory tracking for arbitrary linear position reference
trajectory, despite payload mass variations and external wind disturbances. Thus, the proposed control
strategy could be considered an additional controller for nonlinear highly coupled quadrotor UAVs.
This paper is organized as follows: the underactuated quadrotor dynamics, the Dryden wind model,
and the APIDDFN control design are described in section 2. Then, section 3 presents the simulation results
and discussion. Finally, section 4 concludes the paper.
2. RESEARCH METHOD
In this section, the APIDDFN controller design is presented for the quadrotor system in the presence
of the Dryden wind disturbance and payload mass change. For clarity, the quadrotor dynamics and the
Dryden wind turbulence models are first presented. Then, the whole control design steps are described in
detail.
2.1. Quadrotor dynamic modelling
A quadrotor is equipped with four rotors: two rotors (1 and 3) rotate in a counterclockwise direction,
while the other rotors (2 and 4) rotate in a clockwise direction. This four-rotor actuation can generate six
motions: three linear translations (along the 𝑥, 𝑦, and 𝑧 axes) and three angular rotations, namely the 𝜑-roll,
𝜃-pitch, and 𝜓-yaw Euler angles. Thus, the quadrotor is an underactuated highly coupled system with its six
degrees of freedom (6DOF) and only four input actuators.
Now, to establish the quadrotor model, two coordinate systems should be considered: an earth-fixed
coordinate frame (𝐹𝑒 = (𝑂, 𝑋𝑒, 𝑌𝑒, 𝑍𝑒)), and a body-fixed coordinate frame (𝐹𝑏 = (𝐺, 𝑋𝑏, 𝑌𝑏, 𝑍𝑏)), whose
origin is the vehicle center of gravity (CoG). For clarity, Figure 1 shows the quadrotor X structure. The
quadrotor system model may be summarized as (1) [30]:
{
𝑥̈ = (𝑐𝜑𝑠𝜃𝑐𝜓 + 𝑠𝜑𝑠𝜓)
𝑇
𝑚
+
1
𝑚
𝑓𝑤𝑥
𝑦̈ = (𝑐𝜑𝑠𝜃𝑠𝜓 − 𝑠𝜑𝑐𝜓)
𝑇
𝑚
+
1
𝑚
𝑓𝑤𝑦
𝑧̈ = 𝑐𝜑𝑐𝜃
𝑇
𝑚
− 𝑔 +
1
𝑚
𝑓𝑤𝑧
𝜑̈ =
𝜏𝜑
𝑗𝑥
+
𝑗𝑦−𝑗𝑧
𝑗𝑥
𝜃̇𝜓̇
𝜃̈ =
𝜏𝜃
𝑗𝑦
+
𝑗𝑧−𝑗𝑥
𝑗𝑦
𝜑̇ 𝜓̇
𝜓̈ =
𝜏𝜓
𝑗𝑧
+
𝑗𝑥−𝑗𝑦
𝑗𝑧
𝜑̇ 𝜃̇
(1)
where 𝑥, 𝑦 and 𝑧 denote the cartesian three-space position, 𝜑, 𝜃, and 𝜓 denote the aerial vehicle orientation,
𝑇 is the total thrust, 𝑚 denotes the vehicle mass, 𝑔 is the gravitational acceleration, 𝑗𝑥, 𝑗𝑦, and 𝑗𝑧 denote the
moments of inertia along the three axes 𝑥, 𝑦, and 𝑧, respectively, 𝜏𝜑, 𝜏𝜃, and 𝜏𝜓 denote the propeller torques,
and 𝑓
𝑤𝑥, 𝑓
𝑤𝑦, and 𝑓
𝑤𝑧 denotes the aerodynamic force. For simplicity, we adopted the following notations:
𝑐𝛽 : = 𝑐𝑜𝑠 (𝛽) , and 𝑠𝛽 ∶= 𝑠𝑖𝑛 (𝛽) , for 𝛽 ∈ [
− 𝜋
2
,
𝜋
2
].
2.2. The Dryden wind turbulence
The turbulence process is stochastic and is expressed by velocity spectra [31]. In this work, we
apply the most frequently used model called the Dryden turbulence model [32]. Without loss of generality,
we admit that the turbulence field is fixed in time and space. However, we assume using a random,
homogenous, and isotropic turbulent wind. The Dryden turbulence model is characterized by its power
spectral densities (PSD). These power spectral densities correspond to the longitudinal, lateral, and vertical
linear velocity components for the MIL-F-8785C model [31] and are defined by (2):
𝛷𝑥(𝑤) =
2𝜎𝑥
2𝐿𝑥
𝜋𝑉
1
1+(𝐿𝑥
𝑤
𝑉
)
2 , 𝛷𝑦(𝑤) =
𝜎𝑦
2𝐿𝑦
𝜋𝑉
1+3(𝐿𝑦
𝑤
𝑉
)
2
[1+(𝐿𝑦
𝑤
𝑉
)
2
]
2 , 𝛷𝑧(𝑤) =
𝜎𝑧
2𝐿𝑧
𝜋𝑉
1+3(𝐿𝑧
𝑤
𝑉
)
2
[1+(𝐿𝑧
𝑤
𝑉
)
2
]
2 (2)
where the turbulence intensities (𝜎𝑥, 𝜎𝑦 and 𝜎𝑧) correspond to the turbulence magnitude in the longitudinal,
lateral, and vertical axes, respectively. The turbulence scale length (𝐿𝑥, 𝐿𝑦 and 𝐿𝑧) represent the turbulence

 ISSN: 2088-8708
3610
field length in the longitudinal, lateral, and vertical axis, respectively. 𝑤 is the time-frequency, and 𝑉 is the
air velocity.
Implementing the Dryden wind model consists of passing the band-limited Gaussian white noise
signal through a shaping filter function to generate an output signal with spectral properties. The turbulence
signal must have the same frequency spectrum as the PSD of the Dryden wind. For this reason, we should
use the following filters defined by (3) [33]:
𝐻𝑥(𝑠) = 𝜎𝑥√
2𝐿𝑥
𝜋𝑉
.
1
1+
𝐿𝑥
𝑉
𝑠
, 𝐻𝑦(𝑠) = 𝜎𝑦√
𝐿𝑦
𝜋𝑉
.
1+
√3𝐿𝑦
𝑉
𝑠
(1+
𝐿𝑦
𝑉
𝑠)
2 , 𝐻𝑧(𝑠) = 𝜎𝑧√
𝐿𝑧
𝜋𝑉
.
1+
√3𝐿𝑧
𝑉
𝑠
(1+
𝐿𝑧
𝑉
𝑠)
2 (3)
Continuous Dryden filter is a low-pass filter, where any frequency higher than the cutoff frequency
is eliminated. The filter cutoff frequency is given by the ratio of turbulence scale length to the airspeed. Then,
according to [31], at low altitude, the turbulence scale lengths and intensities are defined by the following
equation where 𝑊20 denotes the wind speed at the height of 6 meters, and ℎ denotes the quadrotor altitude.
𝐿𝑧 = ℎ, 𝐿𝑥 = 𝐿𝑦 =
ℎ
(0.177+0.000823ℎ)1.2 , 𝜎𝑧 = 0.1𝑊20, 𝜎𝑥 = 𝜎𝑦 =
0.1𝑊20
(0.177+0.000823ℎ)0.4 (4)
Figure 1. Quadrotor system X-structure with its coordinate frames
2.3. Adaptive PID deep feedforward network design
In this work, the control of the quadrotor system is based on an APIDDFN. This controller is
designed by combining the conventional PID control strategy with the DFN algorithm. The DFN algorithm is
the most common kind of neural network architecture mapping input to output through hidden nodes using a
sequence of layered transformations that connect the neurons of each layer to those of the next layer. For
clarity, algorithm 1 gives the main steps of the entire controller design process. Then, a description of this
algorithm is given below.
Algorithm 1. PSO-APIDDFN algorithm steps for controlling the quadrotor system
1 Collect the quadrotor control system dataset
2 Set the deep feedforward network (DFN) architecture
3 Initialize the DFN parameters (weights and bias)
4 Design the adaptive DFN controller
5 Tune the DFN parameters using the multidimensional PSO algorithm
The first step in the APIDDFN design is collecting the training data sets necessary to carry out the
learning process. Indeed, we use the PID control strategy to collect the data set required to design the deep
feedforward networks (DFNs), namely 𝐷𝐹𝑁𝑥, 𝐷𝐹𝑁𝑦, 𝐷𝐹𝑁𝑧, 𝐷𝐹𝑁𝜑, 𝐷𝐹𝑁𝜃 and 𝐷𝐹𝑁𝜓. The data sets
collected from the quadrotor system response are composed of 2-dimensional vector [𝑒 𝑈 ]𝑇
where 𝑒 is the
controller error that denotes the difference between the actual and desired output value, and 𝑈 is the PID
controller output for the corresponding system state.

3611
One of the essential features that affect the resultant network in the learning task is the network
architecture, as the hidden layer number and neuron number in each layer are essential in the network design.
One hidden layer and a few neurons number may be insufficient. However, multi-layer architecture with
many neurons may imply a superior computation time and may not be more efficient for solving complex
problems.
In the proposed APIDDFN, each DFN has four layers: one input layer, two hidden layers, and one
output layer. The input layer has two neurons corresponding to error signal 𝑒 and the feedback value from the
current system output. The output layer has three (two) neurons that consist of the 𝑘𝑝, 𝑘𝑖 and 𝑘𝑑 (𝑘𝑝 and 𝑘𝑑)
PID (PD) gains. The structure of the proposed deep feedforward network is shown in Figure 2. Roughly, each
connection has weight and bias values that are represented by matrices of numbers, where 𝐼𝑛 = [𝐼𝑛1 𝐼𝑛2]𝑇
denotes the network input vector, and 𝑊𝑘
(𝑏𝑘
), for 𝑘 = {1,2,3}, denotes the matrix value that represents the
weight (bias) of the connection between consecutive layers. The mathematical equations of the DFN
algorithm can be described as follows:
- Layer 0 (input layer):
ℎ0
= 𝐼𝑛 (5)
where 𝐼𝑛 is the network input vector. In this paper, as the input layer comprises two neurons, the input vector
dimension is (2x1).
- Layer 1 to layer 2 (hidden layers):
ℎ𝑖
= 𝛷(𝑏𝑖
+ 𝑊𝑖
ℎ𝑖−1) (6)
For each hidden layer 𝑖, 𝑖 ∈ {1,2}, 𝑊𝑖
is the layer weight that represents the strength of the
connection between the neurons of the hidden layer 𝑖 and the previous layer (𝑖 − 1), 𝑏𝑖
is the neuron bias
added to produce the net input, and 𝛷 is the hidden layer activation function. Noting that, 𝑊𝑖
is a (𝑟x𝑐)
matrix and 𝑏𝑖
is a (𝑟x1) vector, where 𝑟 is the current layer neuron number, and 𝑐 is the previous layer
neuron number.
In this work, 𝑊1
is a (3x2) matrix, 𝑏1
and 𝑏2
are a (3x1) vectors, 𝑊2
is a (3x3) matrix. The
activation function applied to each hidden layer is the sigmoid function, which is defined by (7):
𝛷(𝑧) =
1
1+𝑒−𝑧 (7)
- Layer 3 (output layer):
𝑦𝑝 = 𝛹(𝑏3
+ 𝑊3
ℎ2) (8)
where 𝛹 is the output layer activation function and 𝑦𝑝 = [𝑜𝑢𝑡1 𝑜𝑢𝑡2 𝑜𝑢𝑡3]𝑇
denotes the network output
vector. In this work, the output vector dimension is (3(2)x1), 𝑊3
is an (3(2)x3) matrix, 𝑏3
is an (3(2)x1)
vector and the linear activation (i.e., no activation function) is chosen as activation function.
The next step consists of initializing the deep feedforward network weights and biases. The initial
training of DFN parameters is the process that permits finding the weights and bias values for all network
layers to map the input data to the associated output. This training operation needs running several iterations
making minor changes to the DFN parameters until a minimum of a cost function is reached. Many
algorithms can be used, such as Levenberg-Marquardt (LM) and the backpropagation algorithms that use
Jacobian derivatives, known as the fastest algorithm [34]. Roughly, the backpropagation training function
could use the gradient derivatives, such as gradient descent (GD), the Broyden, Fletcher, Goldfarb, and
Shanno (BFGS) quasi-newton method, or the one-step secant (OSS) algorithm. In this work, after several
tests, we have chosen the one-step secant (OSS) algorithm as it was shown to ensure satisfying results using
less memory and computational time compared with the BFGS algorithm [35].
In this paper, the APIDDFN is carried out to stabilize the quadrotor system for trajectory tracking
objective. The APIDDFN control method comprises the conventional PID control strategy and the deep
feedforward network (DFN) algorithm. Especially for the quadrotor system, six APIDDFNs controllers
denoted as 𝐴𝑃𝐷𝐷𝐹𝑁𝑥, 𝐴𝑃𝐷𝐷𝐹𝑁𝑦, 𝐴𝑃𝐼𝐷𝐷𝐹𝑁𝑧, 𝐴𝑃𝐼𝐷𝐷𝐹𝑁𝜑, 𝐴𝑃𝐼𝐷𝐷𝐹𝑁𝜃 and 𝐴𝑃𝐼𝐷𝐷𝐹𝑁𝜓, are designed.
Thus, six DFNs controllers are used to online adjusting the gains of the PID (PD) controllers for 𝑧, 𝜑, 𝜃, and
𝜓 (𝑥 and 𝑦). Figure 3 shows the structure of the 𝐴𝑃𝐼𝐷𝐷𝐹𝑁𝑧 for the 𝑧 system-state, where 𝑈1 is the controller
output, 𝑧 is the current system output, and 𝑒𝑧 is the difference between the desired trajectory 𝑧𝑑𝑒𝑠 and its
actual value 𝑧. For clarity, a global control system schema is provided in Figure 4.

 ISSN: 2088-8708
3612
Figure 2. Structure of the designed deep feedforward network
Figure 3. Structure of the APIDDFNz controller designed for z-state
Figure 4. Global control system schema
Besides, once the network architecture has been chosen, the weights 𝑊𝑘
and biases 𝑏𝑘
should be
trained in each layer, for 𝑘 ∈ {1,2,3}. Appropriate values are obtained for these parameters by minimizing a
cost function that computes the difference between the predicted output obtained and the actual output value.
Then, without a loss of generality, we choose the integral of time-weighted absolute error (ITAE) as a
performance criterion in the considered control problem. Thus, letting 𝑒(𝑡) denote the error between the
target value of the output and the actual output value, the ITAE performance criterion is defined by (9):
𝐼𝑇𝐴𝐸 = ∫ 𝑡|𝑒(𝑡)|𝑑𝑡
∞
0
(9)
The next step consists of adjusting the weights and bias to better dynamic behavior for the
closed-loop quadrotor system until the chosen cost function reaches a minimal value. Thus, we use the
multidimensional PSO algorithm to find the 33 (29) DFN parameters’ optimum values. Indeed, PSO is a
population-based stochastic optimization algorithm that has experienced many enhancements [36]. Thus,
PSO has been extensively utilized to solve optimization problems, where it has proven its effectiveness in

3613
different research areas [37]. To describe the specific DFN parameter tuning process, algorithm 2 gives the
multidimensional PSO algorithm main steps. Each step of algorithm 2 may be described as follows:
First, the PSO algorithm parameters should be selected. Practically, to tune the optimum value of the
33 (29) DFN parameters, the population number 𝑛 is set to 200, and the maximum iteration is set to 5.
However, the right choice of the search space parameter (𝑠𝑠), which defines the search space limits during
the tuning process, allows computing the parameters’ maximum possible values to be tuned. For this reason,
we choose a different search space parameter for each system state. Then, the PSO algorithm is initialized
with random particles that correspond to the DFN weights and bias parameters’ possible values. In steps 3
and 4, the particles are evaluated to compute the local best fitness 𝜒1(𝑡), which is known as the cognitive
component, and the global best fitness 𝜒2(𝑡), which is known as the social component. In step 5, the PSO
algorithm computes the values of actual position 𝜒(𝑡) and actual velocity 𝑣(𝑡) of the particles, which affects
the movement of the particles according to (10):
{
𝑣(𝑡 + 1) = 𝛾 𝑣(𝑡) + 𝑐1𝑟1(𝜒1(𝑡) − 𝜒(𝑡)) + 𝑐2𝑟2(𝜒2(𝑡) − 𝜒(𝑡))
𝜒(𝑡 + 1) = 𝜒(𝑡) + 𝑣(𝑡 + 1)
(10)
where 𝛾 is the inertia weight used to adjust the impact of the previous velocity value upon the current one.
(𝑐1, 𝑐2) numbers are the cognitive and social constants, and (𝑟1,𝑟2) numbers are arbitrarily chosen in [0,1].
Afterward, if the maximum iteration number is not yet reached, the algorithm returns to step 2. Otherwise,
algorithm 2 provides the optimum values of DFN parameters.
Algorithm 2: Multidimensional PSO algorithm for the DFN
1
2
3
4
5
Begin
Set the PSO parameters (population number, search space, maximum of iteration …)
Loop:
Initialize random particles (DFN weights and bias)
Evaluate the local best fitness for each particle
Evaluate the global best fitness
Update the velocity and position of particles
Until maximum iteration is reached
3. RESULTS AND DISCUSSION
The completely dynamic control system setup is implemented using the MATLAB simulation tool
to test and show the proposed controller’s validity for the quadrotor aerial vehicle. First, the feedforward
network (FN) and the DFN structures adjust the PID/PD controller gains for each quadrotor system state.
Then, the elaborated final DFN weights and biases are optimized using the PSO algorithm. Besides, a
comparison between all the above controllers is provided to show the controller’s effectiveness. Finally, to
highlight the proposed PSO-based APIDDFN controller validity, an efficiency test is performed. The
quadrotor aerial vehicle’s payload is assumed to be changing during its motions and subjected to the Dryden
turbulence wind.
The quadrotor parameters are chosen similar to those used in [38]. Then, the proposed controller is
conducted to stabilize the quadrotor attitude and ensure the following 𝑥, 𝑦, and 𝑧 trajectory-tracking:
𝑥𝑑𝑒𝑠 = {
1 𝑡 ∈ [25𝑠 55𝑠]
0 𝑜𝑡ℎ𝑒𝑟𝑤𝑖𝑠𝑒
, 𝑦𝑑𝑒𝑠 = {
1 𝑡 ∈ [45𝑠 75𝑠]
, 𝑧𝑑𝑒𝑠 = {
1 𝑡 ∈ [5𝑠 95𝑠]
(11)
Then, we aim to show the proposed APIDDFN controller effectiveness compared to three other control
methods implemented to stabilize the quadrotor system using the following simulation scenario.
3.1. Quadrotor control by the conventional PID controller
The first control method consists of controlling the quadrotor system by a fixed-parameter PID/PD
controller. Table 1 summarizes the PID/PD gains obtained for each system state using the MATLAB PID
Tuner. Thus, to design the FN and DFN systems, the training dataset is collected from error and output
values when the above-tuned PID controller controls the quadrotor system.
3.2. Quadrotor control by the adaptive PID feedforward network
The second method consists of controlling the quadrotor by the APIDDFN, where six APIDFN
controllers are implemented. Each APIDFN controller comprises a PID and a feedforward network (FN)
algorithm containing three layers: one input layer, one hidden layer, and one output layer. The one-step

 ISSN: 2088-8708
3614
secant (OSS) algorithm is used to initialize the FN weight and bias. Table 2 summarizes the 21(17) FN
parameters obtained for 𝑧, 𝜑, 𝜃, and 𝜓 (𝑥 and 𝑦) quadrotor system states.
Table 1. PID/PD controller initial gains
𝒙 𝒚 𝒛 𝝋 𝜽 𝝍
𝑲𝒑 0.1 0.1 7.5 0.2339 0.2339 0.00274
𝑲𝒊 -- -- 3.075 0.0929 0.0929 0.0001
𝑲𝒅 0.2 0.2 5.5 0.08638 0.08638 0.1
Table 2. FN system parameters
W1
b1
W2
b2
x [
−1.7715 −1.2277
−1.0460 2.1203
1.8570 1.5275
] [
2.7265
0.0176
2.4471
] [
−0.1914 −0.0024 −0.0470
−0.5157 −0.0110 −0.0492
] [
0.2428
0.5652
]
y [
−0.1625 −2.1594
−0.8279 2.1314
1.0827 −1.4031
] [
−2.4831
0.1042
2.7171
] [
0.0856 0.0326 0.2804
−0.1094 −0.0268 −0.0232
] [
−0.1916
−0.0954
]
z [
−1.8357 −1.3429
−1.4657 −1.9203
2.2023 0.2091
] [
2.6015
−0.0144
2.6620
] [
−0.1721 0.0042 −0.3515
0.4011 −0.0161 −0.1328
0.1481 −0.0101 −0.1933
] [
0.5038
−0.3017
0.0980
]
𝜑 [
1.9734 −1.4623
1.8060 −1.3916
2.3032 −0.9046
] [
−2.3548
−0.0444
2.3236
] [
0.0567 0.0694 −0.5032
−0.3739 0.1919 −0.9850
−0.8101 0.1823 −0.3795
] [
0.5841
0.5969
−0.4967
]
𝜃 [
1.3642 1.8727
0.8303 2.2216
−1.5594 −1.7780
] [
−2.5868
0.0926
−2.4978
] [
−0.8604 0.1266 0.3287
0.1905 0.0896 0.7230
0.6688 −0.1112 −0.8360
] [
−0.5630
0.8973
−0.1628
]
𝜓 [
1.8379 0.9877
2.1095 −0.9542
−1.6207 1.7556
] [
−2.7657
0.1496
−2.3730
] [
−0.1470 −0.0861 −0.2114
0.0432 −0.0209 −0.0446
−0.0323 0.0070 0.0184
] [
−0.2902
0.1077
0.0518
]
3.3. Quadrotor control by the adaptive PID deep feedforward network
The third method consists of controlling the quadrotor system using the APIDDFN that comprises
one additional hidden layer than the APIDFN. The OSS algorithm is also used as a training method for
initializing the DFN parameters. Table 3 provides the 33 (29) DFN parameters for 𝑧, 𝜑, 𝜃, and 𝜓 (𝑥 and 𝑦)
quadrotor system state.
Table 3. DFN system parameters
W1
b1
W2
b2
W3
b3
x [
-1.2635 -1.8703
-1.4752 -1.8558
1.3386 1.7651
] [
2.5904
0.0019
2.6482
] [
-1.0528 -1.0635 1.0501
1.4393 -0.8603 -1.0094
-0.2118 1.2808 1.3327
] [
2.1729
-0.0216
-2.1444
] [
-0.6447 -0.8081 -0.9887
-0.3314 -0.1791 -0.2768
] [
0.2324
0.2182
]
y [
2.2750 0.8316
1.4493 1.6726
0.9982 -1.9617
] [
-2.4105
-0.1869
2.7044
] [
1.5614 -1.2710 0.0251
-1.7011 0.3893 1.2190
-0.1804 -1.1829 0.8886
] [
-2.2945
0.1096
2.8061
] [
-0.1279 -0.1177 -0.3330
0.8719 0.6333 0.2293
] [
0.3245
-0.0110
]
z [
-1.9363 -1.4086
-1.8222 1.4957
-0.1388 2.1700
] [
2.4772
0.0698
-2.6938
] [
0.5245 1.2985 1.4103
0.8245 -0.9529 -1.5775
0.6241 1.4304 0.9536
] [
-2.0345
0.1698
2.2061
] [
0.4751 0.1462 -0.1361
0.3874 0.3852 0.0725
0.7497 0.3510 -0.0788
] [
0.4341
-0.0046
0.5572
]
𝜑 [
-2.3148 -0.7212
0.4887 -2.3122
-2.2385 -0.9351
] [
2.4183
-0.2637
-2.4162
] [
-0.3337 -1.5494 0.0845
0.9354 -1.3129 -1.2057
-0.0742 -1.9902 -0.5102
] [
2.4315
0.0086
1.9356
] [
-0.4162 -0.1092 0.5896
-0.2739 0.0601 0.0690
0.3692 -0.3557 -0.3936
] [
-0.0923
0.0684
0.4138
]
𝜃 [
-1.0699 2.0483
-2.2879 -0.7388
1.6393 -1.7824
] [
2.5645
0.0132
2.4319
] [
1.4605 0.0494 -1.4034
1.3389 -0.8714 -1.0487
-1.0406 0.0780 -1.7284
] [
-1.9564
0.0249
-2.0190
] [
0.0938 0.3084 0.0920
0.5909 -0.4233 0.2745
0.6534 -0.3072 -0.5427
] [
0.0978
0.8989
0.2059
]
𝜓 [
1.3434 -1.6871
0.9313 2.2096
0.9701 3.3114
] [
-2.7498
0.0267
1.2526
] [
1.9468 -0.0490 -0.4695
0.9199 0.0671 1.2957
0.8037 -0.3096 -1.9388
] [
-1.9837
-1.3488
1.6927
] [
0.4094 -1.3674 -0.2009
-0.0300 0.1989 0.0365
0.0016 0.7188 0.1032
] [
-0.7306
0.1025
0.6457
]
3.4. Quadrotor control by the PSO-based adaptive PID deep feedforward network
The fourth method consists of controlling the quadrotor system using the APIDDFN controller. In
this case, the control parameters are optimized using the PSO algorithm (PSO-APIDDFN). For clarity,
Table 4 shows the new 33(29) PSO-tuned DFN parameters for each quadrotor system state.

3615
Table 4. DFN system parameters tuned by the multidimensional PSO algorithm
𝑊1
𝑏1
𝑊2
𝑏2
𝑊3
𝑏3
x [
0.3320 4.0000
−0.4583 3.4840
1.2157 1.5322
] [
3.6509
−0.3025
2.2516
] [
0.2276 2.6606 −2.0000
−0.7656 −0.7084 −1.6188
2.2548 0.2497 0.7498
] [
−2.1355
−0.0816
1.6769
] [
−0.5444 0.0178 0.1341
−1.2507 0.0271 −1.0000
] [
1.4377
1.0806
]
y [
0.0021 1.6944
2.8716 −1.4658
−0.0741 1.3784
] [
−3.5567
−0.8757
−2.7979
] [
0.5380 1.1247 2.0000
−2.7043 1.3204 0.9411
0.0401 −0.0941 0.3346
] [
−1.0000
−0.6482
2.0174
] [
0.8007 1.3752 0.9639
−0.4098 0.0989 1.0963
] [
2.0000
−0.2658
]
z [
0 1.2998
−0.8544 3.0684
1.4629 1.1332
] [
2.5406
1.1023
4.0000
] [
−0.7441 1.1434 1.8354
0.7360 −1.0000 −0.3674
−0.3268 −2.1546 −1.4486
] [
−1.9879
0.9285
−2.8496
] [
1.0000 1.4376 −1.1577
0.1954 −0.9893 −0.5038
−2.0000 0.7333 −1.7944
] [
0.8648
1.4695
−0.2768
]
𝜑 [
−2.0000 −1.7835
−1.5054 −2.4023
−0.4962 0.4424
] [
4.0000
−0.3487
−3.1380
] [
0 −0.5467 −0.3974
0.6310 −0.6741 −1.6522
−3.0000 −0.8071 1.0936
] [
2.9705
1.0000
−2.9764
] [
0.5489 −0.1122 −0.1534
−0.1632 0.8813 0.2801
2.0000 2.0000 −1.0000
] [
1.1646
1.4386
1.0000
]
𝜃 [
0 −0.5770
−0.3557 3.6338
−1.2773 −2.0000
] [
−2.3539
0.0436
−3.8555
] [
2.0000 1.8656 −0.7121
−0.2626 −1.3522 −0.7005
−1.5814 −1.3641 −3.0000
] [
−3.0216
1.0000
0
] [
0.9232 0.0311 −0.4625
2.0000 −0.3400 0.0296
−0.7128 2.0000 2.0000
] [
−1.0655
1.0000
2.0000
]
𝜓 [
1.9006 −0.6465
−2.3708 1.0607
−1.2146 1.5179
] [
−2.3242
−0.3009
−1.0000
] [
1.8107 1.3353 −1.5015
1.6569 1.6522 −0.9362
−0.7715 1.0510 0
] [
−0.9729
−0.0321
2.8002
] [
−0.6805 −1.2258 0.4413
0.9313 −0.7401 −0.4986
−0.2053 0.1624 0.9461
] [
1.2186
0.6912
2.0000
]
3.5. Controllers’ performance comparison
It is worth noticing that the quadrotor linear acceleration involves the total thrust force and the Euler
angle states. Then, it is challenging to stabilize this system in an open-loop scheme [39]. For this reason, the
quadrotor system should be stabilized using a suitable control strategy. Besides using the PSO algorithm, the
proposed control method allows better system performance in the closed-loop.
The time response for linear positions is shown in Figure 5. The performance characteristics
comparison such as percent of overshoot (𝑀𝑝), rise time (𝑇𝑟), settling time (𝑇𝑠), and 𝐼𝑇𝐴𝐸 criterion between
these four control strategies is summarized in Table 5. It can be seen from Figure 5 and Table 5 that the
APIDDFN is slightly better than the APIDFN. Thus, it is evident that using additional hidden layers has
improved the control system’s performance. Besides, except for the rise time of the 𝑧 and 𝑥 system states, the
PSO-APIDDFN controller shows better performances, a faster convergence speed, a minimum overshoot,
and a minor 𝐼𝑇𝐴𝐸 performance criterion compared to the other controllers
Then, Figure 6 shows the 𝜑, 𝜃, and 𝜓 quadrotor attitudes for the PID, APIDFN, APIDDFN, and
PSO-APIDDFN controllers. It is shown that the four control methods can ensure the stabilization of the
quadrotor attitude. Roughly, 𝜑 and 𝜃 angle values do not exceed 4 degrees for all these control strategies as
shown in Figures 6(a) and 6(b). Besides, the 𝜓-yaw angle value is almost null, especially for the PSO-
APIDDFN controller as shown in Figure 6(c).
Table 5. Control system performance comparison
𝑥 𝑦 𝑧
𝑀𝑝(%) 𝑇𝑟(𝑠) 𝑇𝑠(𝑠) 𝐼𝑇𝐴𝐸 𝑀𝑝(%) 𝑇𝑟(𝑠) 𝑇𝑠(𝑠) 𝐼𝑇𝐴𝐸 𝑀𝑝(%) 𝑇𝑟(𝑠) 𝑇𝑠(𝑠) 𝐼𝑇𝐴𝐸
PID 9.432 1.763 7.376 0.491 9.432 1.763 7.376 0.722 9.975 0.6103 3.613 0.144
APIDFN 8.719 1.770 7.292 0.481 8.845 1.780 7.088 0.699 9.916 0.6102 3.609 0.143
APIDDFN 7.890 1.702 6.355 0.395 8.671 1.776 7.277 0.709 9.701 0.61 3.574 0.144
PSO-APIDDFN 1.026 1.751 2.968 0.075 1.134 1.729 2.929 0.111 1.529 0.6923 2.144 0.0523
3.6. Efficiency test
The simulation tests are also performed to illustrate the above controller efficiency when the
quadrotor system is subject to external disturbances. Roughly, two types of disturbances are considered: a
quadrotor mass change and external wind. First, Figure 7 shows the case of 20% mass increasing between
10 s and 30 s, and decreasing, between 60 s and 80 s as shown in Figure 7(a), and its effect on the vehicle
z-altitude as shown in Figure 7(b). The total mass change caused a slight overshoot in the PSO-APIDDFN
(about 5%) than the other controllers (about 16.6%). Thus, the proposed controller can compensate for the
mass change by changing the PID/PD gains. Then, Figure 8 illustrates the z-altitude PID gains variation in
the case of APIDFN as shown in Figure 8(a), APIDDFN as shown in Figure 8(b), and PSO-APIDDFN as
shown in Figure 8(c) controllers when the quadrotor system is subject to a mass change. These subfigures
show the adaptive behavior of the three controllers. However, unlike the PID controller gains that remain at
the same value in all simulation time, the gains elaborated in the PSO-APIDDFN controller change according
to the behavior of the quadrotor system outputs.

 ISSN: 2088-8708
3616
Secondly, we simulate the effect of the external wind on the quadrotor system translational motion.
This disturbance is conceived using the above Dryden model, where the wind speed is 20 m/s. Figure 9
shows the trajectory tracking of the quadrotor system under the wind. The above results show that the
PSO-APIDDFN controller has an excellent wind disturbance-rejection compared to the other control
strategies, for x-state as shown in Figure 9(a), y-state as shown in Figure 9(b), and z-state as shown in
Figure 9(c).
Figure 5. 𝑥, 𝑦, and 𝑧 linear positions time responses
(a) (b) (c)
Figure 6. Euler angles stabilization a) 𝜑-roll, b) 𝜃-pitch, and c) 𝜓-yaw

3617
(a) (b)
Figure 7. Altitude response in case of quadrotor mass change (a) altitude response and (b) mass change
(a) (b) (c)
Figure 8. Variation of altitude PID gains in case of (a) APIDFN, (b) APIDDFN, and (c) PSO-APIDFN
(a) (b) (c)
Figure 9. Effect of the wind on the translational motion: (a) x-position, (b) y-position, and (c) z-altitude
4. CONCLUSION
In this paper, an adaptive proportional integral derivative deep feedforward network (APIDDFN)
controller is proposed for stabilizing the trajectory tracking control of an underactuated quadrotor system.
The deep feedforward neural network (DFN) is used to online adjusting the PID controller parameters. A
one-step secant algorithm is utilized to initialize the DFN system weights and biases. Besides, a
multidimensional particle swarm optimization (PSO) algorithm is utilized to optimize the DFN parameters.
The simulation results illustrate that the proposed PSO-APIDDFN controller can achieve faster convergence

 ISSN: 2088-8708
3618
speed and show minimum overshoot than the non-optimized adaptive PID feedforward network and adaptive
PID deep feedforward network. These results demonstrate the efficiency of the proposed method in designing
an intelligent controller for stabilizing the quadrotor UAV. Future works are related to integrating an
observer in the control system to estimate unmeasured translational and rotational quadrotor system states
and rates.
REFERENCES
[1] W. Gu, K. P. Valavanis, M. J. Rutherford, and A. Rizzo, “UAV model-based flight control with artificial neural networks: a
survey,” Journal of Intelligent and Robotic Systems, vol. 100, no. 3–4, pp. 1469–1491, Dec. 2020, doi: 10.1007/s10846-020-
01227-8.
[2] S. Y. Choi and D. Cha, “Unmanned aerial vehicles using machine learning for autonomous flight; state-of-the-art,” Advanced
Robotics, vol. 33, no. 6, pp. 265–277, Mar. 2019, doi: 10.1080/01691864.2019.1586760.
[3] A. Carrio, C. Sampedro, A. Rodriguez-Ramos, and P. Campoy, “A review of deep learning methods and applications for
unmanned aerial vehicles,” Journal of Sensors, vol. 2017, pp. 1–13, 2017, doi: 10.1155/2017/3296874.
[4] A. R. Al Tahtawi and M. Yusuf, “Low-cost quadrotor hardware design with pid control system as flight controller,” Telkomnika
(Telecommunication Computing Electronics and Control), vol. 17, no. 4, pp. 1923–1930, Aug. 2019, doi:
10.12928/TELKOMNIKA.v17i4.9529.
[5] Y. Al Younes, A. Drak, H. Noura, A. Rabhi, and A. El Hajjaji, “Robust model-free control applied to a quadrotor UAV,” Journal
of Intelligent and Robotic Systems, vol. 84, no. 1–4, pp. 37–52, Dec. 2016, doi: 10.1007/s10846-016-0351-2.
[6] M. A. Basri and A. Noordin, “Optimal backstepping control of quadrotor UAV using gravitational search optimization
algorithm,” Bulletin of Electrical Engineering and Informatics (BEEI), vol. 9, no. 5, pp. 1819–1826, Oct. 2020, doi:
10.11591/eei.v9i5.2159.
[7] E. Chater, H. Housny, and H. El Fadil, “Robust control design for quadrotor trajectory path tracking,” in 2019 8th International
Conference on Systems and Control (ICSC), Oct. 2019, pp. 21–26, doi: 10.1109/ICSC47195.2019.8950509.
[8] A. Noordin, M. A. M. Basri, and Z. Mohamed, “Sliding mode control for altitude and attitude stabilization of quadrotor UAV
with external disturbance,” Indonesian Journal of Electrical Engineering and Informatics (IJEEI), vol. 7, no. 2, pp. 203–210, May
2019, doi: 10.52549/ijeei.v7i2.1149.
[9] E. A. Chater, H. Housny, and H. El Fadil, “Robust sliding mode control for quadrotor UAV,” in 2020 IEEE 2nd International
Conference on Electronics, Control, Optimization and Computer Science (ICECOCS), Dec. 2020, pp. 1–6, doi:
10.1109/ICECOCS50124.2020.9314584.
[10] N. I. Vitzilaios and N. C. Tsourveloudis, “An experimental test bed for small unmanned helicopters,” Journal of Intelligent and
Robotic Systems, vol. 54, no. 5, pp. 769–794, May 2009, doi: 10.1007/s10846-008-9284-8.
[11] N. Maharani Raharja, E. Firmansyah, A. Imam Cahyadi, and I. Iswanto, “Hovering control of quadrotor based on fuzzy logic,”
International Journal of Power Electronics and Drive Systems (IJPEDS), vol. 8, no. 1, pp. 492–504, Mar. 2017, doi:
10.11591/ijpeds.v8.i1.pp492-504.
[12] I. Iswanto, O. Wahyunggoro, and A. Imam Cahyadi, “Path planning based on fuzzy decision trees and potential field,”
International Journal of Electrical and Computer Engineering (IJECE), vol. 6, no. 1, pp. 212–222, Feb. 2016, doi:
10.11591/ijece.v6i1.8606.
[13] J. Muliadi and B. Kusumoputro, “Neural network control system of UAV altitude dynamics and its comparison with the PID
control system,” Journal of Advanced Transportation, vol. 2018, pp. 1–18, 2018, doi: 10.1155/2018/3823201.
[14] H. Housny, E. A. Chater, and H. El Fadil, “Multi closed-loop adaptive neuro-fuzzy inference system for quadrotor position
control,” Advances in Science, Technology and Engineering Systems Journal, vol. 5, no. 5, pp. 526–535, 2020, doi:
10.25046/aj050565.
[15] H. Housny, E. A. Chater, and H. El Fadil, “Observer-based enhanced ANFIS control for a quadrotor UAV,” International Review
on Modelling and Simulations (IREMOS), vol. 14, no. 1, Feb. 2021, doi: 10.15866/iremos.v14i1.18991.
[16] P. E. I. Pounds, D. R. Bersak, and A. M. Dollar, “Stability of small-scale UAV helicopters and quadrotors with added payload
mass under PID control,” Autonomous Robots, vol. 33, no. 1–2, pp. 129–142, Aug. 2012, doi: 10.1007/s10514-012-9280-5.
[17] A. Noordin, M. A. M. Basri, and Z. Mohamed, “Simulation and experimental study on PID control of a quadrotor MAV with
perturbation,” Bulletin of Electrical Engineering and Informatics (BEEI), vol. 9, no. 5, pp. 1811–1818, Oct. 2020, doi:
10.11591/eei.v9i5.2158.
[18] H.-K. Tran and J.-S. Chiou, “PSO-based algorithm applied to quadcopter micro air vehicle controller design,” Micromachines,
vol. 7, no. 9, Sep. 2016, doi: 10.3390/mi7090168.
[19] S.-E.-I. Hasseni and L. Abdou, “Decentralized PID control by using GA optimization applied to a quadrotor,” Journal of
Automation, Mobile Robotics and Intelligent Systems, vol. 12, no. 2, pp. 33–44, Jun. 2018, doi: 10.14313/JAMRIS_2-2018/9.
[20] T. K. Priyambodo, A. Dharmawan, O. A. Dhewa, and N. A. S. Putro, “Optimizing control based on fine tune PID using ant
colony logic for vertical moving control of UAV system,” AIP Conference Proceedings, 2016, doi: 10.1063/1.4958613.
[21] H. Housny, E. A. Chater, and H. El Fadil, “Fuzzy PID control tuning design using particle swarm optimization algorithm for a
quadrotor,” in 2019 5th International Conference on Optimization and Applications (ICOA), Apr. 2019, pp. 1–6, doi:
10.1109/ICOA.2019.8727702.
[22] A. Benbouali, F. Chabni, R. Taleb, and N. Mansour, “Flight parameters improvement for an unmanned aerial vehicle using a
lookup table based fuzzy PID controller,” Indonesian Journal of Electrical Engineering and Computer Science (IJEECS), vol. 23,
no. 1, pp. 171–178, Jul. 2021, doi: 10.11591/ijeecs.v23.i1.pp171-178.
[23] S. Bari, S. S. Zehra Hamdani, H. U. Khan, M. ur Rehman, and H. Khan, “Artificial neural network based self-tuned PID
controller for flight control of quadcopter,” in 2019 International Conference on Engineering and Emerging Technologies
(ICEET), Feb. 2019, pp. 1–5, doi: 10.1109/CEET1.2019.8711864.
[24] Y. Chen, Y. He, and M. Zhou, “Decentralized PID neural network control for a quadrotor helicopter subjected to wind
disturbance,” Journal of Central South University, vol. 22, no. 1, pp. 168–179, Jan. 2015, doi: 10.1007/s11771-015-2507-9.
[25] Y.-F. Teng, B. Hu, Z.-W. Liu, J. Huang, and Z.-H. Guan, “Adaptive neural network control for quadrotor unmanned aerial
vehicles,” in 2017 11th Asian Control Conference (ASCC), Dec. 2017, pp. 988–992, doi: 10.1109/ASCC.2017.8287305.
[26] N. M. Ibrahim, S. T. F. Al-Janabi, and B. Al-Khateeb, “Electricity-theft detection in smart grids based on deep learning,” Bulletin
of Electrical Engineering and Informatics, vol. 10, no. 4, pp. 2285–2292, Aug. 2021, doi: 10.11591/eei.v10i4.2875.

3619
[27] Y. Zhang, P. Phillips, S. Wang, G. Ji, J. Yang, and J. Wu, “Fruit classification by biogeography-based optimization and
feedforward neural network,” Expert Systems, vol. 33, no. 3, pp. 239–253, Jun. 2016, doi: 10.1111/exsy.12146.
[28] B.-S. Oh, K.-A. Toh, A. B. J. Teoh, and Z. Lin, “An analytic Gabor feedforward network for single-sample and pose-invariant
face recognition,” IEEE Trans. on Image Processing, vol. 27, no. 6, pp. 2791–2805, Jun. 2018, doi: 10.1109/TIP.2018.2809040.
[29] S. Edhah, S. Mohamed, A. Rehan, M. AlDhaheri, A. AlKhaja, and Y. Zweiri, “Deep learning based neural network controller for
quad copter: application to hovering mode,” in 2019 International Conference on Electrical and Computing Technologies and
Applications (ICECTA), Nov. 2019, pp. 1–5, doi: 10.1109/ICECTA48151.2019.8959776.
[30] H. Housny, E. Chater, and H. El Fadil, “Multi-closed-loop design for quadrotor path-tracking control,” in 2019 8th International
Conference on Systems and Control (ICSC), Oct. 2019, pp. 27–32, doi: 10.1109/ICSC47195.2019.8950659.
[31] D. Lee, J. Sim, K. Han, C. Kim, and G. O. Koh, “Angle-of-attack command longitudinal control for supersonic advanced trainer
aircraft,” Int. J. of Aeronautical and Space Sciences, vol. 22, no. 1, pp. 120–128, Feb. 2021, doi: 10.1007/s42405-020-00279-2.
[32] J. Chen, L. Wang, J. Fu, and Z. Yang, “Engineering comprehensive model of complex wind fields for flight simulation,”
Aerospace, vol. 8, no. 6, May 2021, doi: 10.3390/aerospace8060145.
[33] H. Ji, R. Chen, and P. Li, “Analysis of helicopter handling quality in turbulence with recursive Von Kármán model,” Journal of
Aircraft, vol. 54, no. 5, pp. 1631–1639, Sep. 2017, doi: 10.2514/1.C034189.
[34] A. A. Suratgar, M. B. Tavakoli, and A. Hoseinabadi, “Modified Levenberg-Marquardt method for neural networks training,”
World Acad Sci Eng Technol, vol. 6, no. 1, pp. 46–48, 2005.
[35] M. Brunato and R. Battiti, “A telescopic binary learning machine for training neural networks,” IEEE Transactions on Neural
Networks and Learning Systems, vol. 28, no. 3, pp. 665–677, Mar. 2017, doi: 10.1109/TNNLS.2016.2537300.
[36] D. Wang, D. Tan, and L. Liu, “Particle swarm optimization algorithm: an overview,” Soft Computing, vol. 22, no. 2, pp. 387–408,
Jan. 2018, doi: 10.1007/s00500-016-2474-6.
[37] Y. Zhang, S. Wang, and G. Ji, “A comprehensive survey on particle swarm optimization algorithm and its applications,”
Mathematical Problems in Engineering, vol. 2015, pp. 1–38, 2015, doi: 10.1155/2015/931256.
[38] S. Bouabdallah, “Design and control of quadrotors with application to autonomous flying,” Lausanne, EPFL, 2007, 10.5075/epfl-
thesis-3727.
[39] K. P. Valavanis, Ed., Advances in unmanned aerial vehicles. Dordrecht: Springer Netherlands, 2007.
BIOGRAPHIES OF AUTHORS
El Ayachi Chater received a B.S. degree in Electronics from the ENSET,
Hassan II University, Mohammedia, Morocco, in 1992, and the Agrégation degree in
Electrical Engineering from the ENSET, Mohammed V University, Rabat, Morocco, in
1999. He received an M.S. and a Ph.D. Degrees in Control Engineering from the EMI,
Mohammed V University, Rabat, Morocco, in 2002 and 2013, respectively. From 2016, he
was an assistant professor, Physics department, ENS, Mohammed V University, Rabat,
Morocco. Then, from 2019, he has been teaching control systems and industrial computing,
electrical engineering department, EST, Salé, Morocco. His research interest includes
nonlinear and intelligent control of electromechanical and mechatronic systems. He has
published several journal/conference papers on these topics. He can be contacted at email:
elayachi.chater@est.um5.ac.ma.
Halima Housny received a B.S. degree in EEA from the FST, Hassan II
University, Mohammedia, Morocco, in 1998. She received an M.S. degree in control
engineering from the EMI, Mohammed V University, Rabat, Morocco, in 2003. Then, she
received a Ph.D. degree in Electrical Engineering with ISA Lab., ENSA, Ibn Tofail
university, Kenitra, Morocco, in 2021. Since 2003, she was teaching computer sciences in
high schools, Salé, Morocco. Her research interest is the optimization and intelligent
control strategies for unmanned aerial vehicles. She has already published several papers
on this topic. She can be contacted at email: housny.hal@gmail.com.
Hassan El Fadil received a B.S. degree in Electronics, the Agrégation degree
in Electrical Engineering, from the ENSET, Mohammed V University, Rabat, Morocco,
in 1994 and 1999, respectively. He received an M.S. degree and a Ph.D. degree in
Automatic Control from the EMI, Mohammed V University, Rabat, Morocco, in 2002 and
2008, respectively. He is currently Professor at the ENSA, Ibn Tofail University, Kenitra,
Morocco. His research interest includes nonlinear and adaptive control, power converters
and electric motors control, renewable energy, distributed energy resources, smart grid,
and electric vehicles. He has published over 100 journals/conference papers on these
topics. Hassan El Fadil is with ISA Lab., ENSA, Ibn Tofail University, Kenitra, Morocco.
He can be contacted at email: elfadilhassan@yahoo.fr.

Adaptive proportional integral derivative deep feedforward network for quadrotor trajectory-tracking flight control

More Related Content

Similar to Adaptive proportional integral derivative deep feedforward network for quadrotor trajectory-tracking flight control (20)

More from IJECEIAES (20)

Recently uploaded (20)

Adaptive proportional integral derivative deep feedforward network for quadrotor trajectory-tracking flight control