Citation:
Pinto Neto, E.C.;
Baum, D.M.; Almeida, J.R.d., Jr.;
Camargo, J.B., Jr.; Cugnasca, P.S.
Deep Learning in Air Traffic
Management (ATM): Applications,
Opportunities, and Open Challenges.
Aerospace
2023
,
10
, 358.
https://
doi.org/10.3390/aerospace10040358
Academic Editor:
Álvaro
Rodríguez-Sanz
Received:
3 February 2023
Revised:
23 March 2023
Accepted:
23 March 2023
Published:
4 April 2023
Copyright:
©
2023
by
the
authors.
Licensee
MDPI,
Basel,
Switzerland.
This
article
is
an
open
access
article
distributed
under
the
terms
and
conditions
of
the
Creative
Commons
Attribution
(CC
BY)
license
(https://
creativecommons.org/licenses/by/
4.0/).
aerospace
Review
Deep Learning in Air Traffic Management (ATM): A Survey
on Applications, Opportunities, and Open Challenges
Euclides Carlos Pinto Neto, Derick Moreira Baum, Jorge Rady de Almeida, Jr., João Batista Camargo, Jr.
and Paulo Sergio Cugnasca *
Safety Analysis Group (GAS-POLI), University of São Paulo (USP), São Paulo 05508-090, Brazil;
euclidescpn@usp.br (E.C.P.N.); derick.baum@usp.br (D.M.B.); jorgerady@usp.br (J.R.d.A.J.);
joaocamargo@usp.br (J.B.C.J.)
*
Correspondence:
cugnasca@usp.br
Abstract:
Currently, the increasing number of daily flights emphasizes the importance of air trans-
portation.
Furthermore, Air Traffic Management (ATM) enables air carriers to operate safely and
efficiently through the multiple services provided.
Advanced analytic solutions have demonstrated
the potential to solve complex problems in several domains, and Deep Learning (DL) has attracted
attention due to its impressive results and disruptive capabilities.
The adoption of DL models in ATM
solutions enables new cognitive services that have never been considered before. The main goal of this
research is to present a comprehensive review of state-of-the-art Deep Learning (DL) solutions for Air
Traffic Management (ATM). This review focuses on describing applications, identifying opportunities,
and highlighting open challenges to foster the evolution of ATM systems.
To accomplish this, we
discuss the fundamental topics of DL and ATM and categorize the contributions based on different
approaches.
First, works are grouped based on the DL approach adopted.
Then, future directions are
identified based on the ATM solution area.
Finally, open challenges are listed for both DL applications
and ATM solutions.
This article aims to support the community by identifying research problems to
be faced in the future.
Keywords:
deep learning; air traffic management; survey; convolutional neural networks (CNN);
generative adversarial networks (GAN); recurrent neural network (RNN); autoenconder
1.
Introduction
Currently, the increasing number of daily flights emphasizes the importance of the air
transportation system.
Furthermore, Air Traffic Management (ATM) enables air carriers
to operate safely and efficiently through the multiple services provided, e.g., Air Traffic
Flow Management (ATFM), Airspace Management (ASM), and Flight Information Services
(FIS) [
1
,
2
].
Air Traffic Control (ATC) technologies have been improved over the decades,
and
new
technologies
are
needed
to
further
enhance
existing
procedures
and
support
future demands.
Moreover,
advanced
analytic
solutions
have
demonstrated
their
potential
to
solve
complex problems in several domains [
3
].
In this context, Deep Learning (DL) has attracted
attention
due
to
its
impressive
results
and
disruptive
capabilities
[
4
].
The
evolution
of
computational power has enabled DL to be used in several contexts, and the current massive
amount of data produced by existing systems empowers new applications.
New DL-based
solutions are currently under development in various domains, and efforts have been made
toward developing DL-based ATM solutions.
Although several ATM solutions rely on deterministic methods, adopting DL models
enables
new
cognitive
services
never
considered
before.
Solutions
vary
from
aircraft
performance (e.g., flight state, parameters, and trajectory optimization) to human factors
(e.g., fatigue assessment).
In all these cases, DL can be used and further improve current
operations in terms of airspace safety and efficiency.
Conversely, several challenges must
Aerospace
2023
,
10
, 358.
https://doi.org/10.3390/aerospace10040358
https://www.mdpi.com/journal/aerospace
Aerospace
2023
,
10
, 358
2 of 30
be overcome to integrate DL solutions in airspace operations.
Furthermore, it is not simple
to clearly define the challenges faced by state-of-the-art strategies and open challenges.
Finally, it is also complex to relate different Deep Learning (DL) architectures with ATM
problems to simplify identifying research gaps.
The main goal of this research is to present a comprehensive review of state-of-the-art
Deep
Learning
(DL)
solutions
for
Air
Traffic
Management
(ATM).
This
review
focuses
on describing applications, identifying opportunities, and highlighting open challenges
to foster the evolution of ATM systems.
To accomplish this, we discuss the fundamental
topics
of
DL
and
ATM
and
categorize
the
contributions
based
on
different
approaches.
First, works are grouped based on the DL approach adopted.
Then, future directions are
identified based on the ATM solution area.
Finally, open challenges are listed for both DL
applications and ATM solutions.
This article aims to support the community by identifying
research problems to be faced in the future.
Therefore, the main contributions of this research are:
•
A
comprehensive
review
of
state-of-the-art
Deep
Learning
(DL)
solutions
for
Air
Traffic Management (ATM);
•
Future directions based on insights of single contributions and ATM solutions groups;
•
An extensive list of open challenges in the context of Deep Learning (DL) applications
in ATM;
•
An extensive list of open challenges from the ATM solutions standpoint.
The paper is organized as follows:
Section
2
discusses aspects of Deep Learning (DL)
and Air Traffic Management (ATM). Secondly, Section
3
reviews DL application in ATM
solutions.
Then, Section
3.6
presents several insights on the review and future directions
for contributions belonging to different categories.
Finally, Sections
4
and
5
present several
open challenges and the conclusions of this research, respectively.
2.
Background
The works reviewed in this research involve several concepts related to Air Traffic
Management (ATM) and Deep Learning (DL). This Section describes fundamental aspects
to a better understanding of the contributions reviewed.
2.1.
Deep Learning (DL)
Currently, the complexity of systems and applications in several domains is increasing.
Applications need to embed some advanced reasoning to accomplish the tasks they have
been designed to.
Conversely, this is complex for several reasons.
In the past few years,
Deep
Learning
(DL)
[
5
,
6
]
applications
have
been
exceptionally
successful
in
multiple
challenging tasks,
and more attention is attracted once the intelligent decision emerges
from patterns hidden in large multi-dimensional datasets.
Deep Learning (DL) can be represented by a neural network with a large number of
layers
and
parameters
using
a
cascade
of
multiple
layers
of
nonlinear
processing
units
for feature extraction and transformation [
7
,
8
].
Several domains have benefited from DL
applications, e.g., healthcare [
9
,
10
], transportation [
11
,
12
], and manufacturing [
13
].
Moreover, several architectures have been proposed with different goals.
More archi-
tectures are under development to solve specific problems.
This research considers five
of the most popular architectures:
Deep Neural Networks (DNN), Convolutional Neural
Networks (CNNs), Recurrent Neural Networks (RNN), Generative Adversarial Networks
(GANs), and Autoencoders (AE).
•
Feed-Forward Networks:
Feed-Forward Neural networks (also referred to as Deep
Neural
Networks—DNN—in
this
article)
consists
of
neurons
ordered
into
layers.
The
first
layer,
called
the
input
layer,
the
last
layer,
called
the
output
layer,
and
the
hidden
layers
[
14
].
Neurons
can
be
considered
processing
units
connected
to
synaptic
weights.
These
neurons
produce
an
output
using
an
activation
function,
which is sent to the following layer [
15
].
These networks are usually trained using the
back-propagation algorithm (used to compute gradients) and the Stochastic Gradient
Aerospace
2023
,
10
, 358
3 of 30
Descent
(SGD)
algorithm
to
optimize
the
weights
(using
the
gradient
computed
previously).
Figure
1
illustrates a simple DNN and highlights the input, hidden, and
output layers.
The number of nodes and hidden layers can change depending on the
problem faced, similar to the input vector.
This architecture has been widely used and
has presented tremendous success in several initiatives.
Figure 1.
Example of Deep Neural Network (DNN) [
14
,
15
].
•
Convolutional Neural Networks (CNNs):
Convolutional Neural Networks (CNNs)
are a category of Deep Learning (DL) models designed to process data in a grid-like
topology (e.g., time-series and image data).
CNNs are usually composed of three types
of layers: convolutional, pooling, and fully connected layers [
16
,
17
]. The convolutional
layers are responsible for extracting important features.
The pooling layers reduce
the resolution of features, making them robust against noise and distortion.
Finally,
the
fully-connected
layers
produce
class
scores
from
the
activations
[
18
].
Figure
2
illustrates a simple CNN model.
Figure 2.
Example of Convolutional Neural Networks (CNNs) [
16
,
18
].
•
Recurrent Neural Networks (RNN):
Recurrent Neural Networks (RNNs) represent a
neural network architecture used to detect patterns in sequences (e.g., images, text,
or numerical time series) [
19
].
Important RNN features are the feedback connection
and memory, which enable activations to flow in a loop and temporal processing [
20
].
Figure
3
illustrates a simple example of an RNN.
Aerospace
2023
,
10
, 358
4 of 30
Figure 3.
Example of Recurrent Neural Network (RNN) [
19
,
20
].
•
Generative
Adversarial
Networks
(GANs):
This
architecture
is
based
on
the
com-
petition between a generation and a discriminator.
In this sense, the generator uses
random noise to produce fake data while the discriminator tries to distinguish real data
from fake data.
When the generator can produce data that cannot be appropriately
classified as fake by the discriminator, the model can produce realistic data [
21
,
22
].
Figure
4
illustrates a simple GAN architecture.
Figure 4.
Example of Generative Adversarial Network (GAN) [
21
,
22
].
•
Autoencoders (AE):
This specific type of neural network was developed to encode
inputs into a compressed and meaningful representation.
After this reduced version
of the provided features is produced, the model decodes it back, aiming to produce
an
output
as
close
as
possible
to
the
input
[
23
,
24
].
Figure
5
illustrates
a
simple
AE architecture.
Figure 5.
Example of Autoenconder (AE) [
23
,
24
].
2.2.
Air Traffic Management (ATM)
Currently, the air transportation system connects countries and plays a major role in
society.
Daily operations are safe and efficient due to the many advances in technology
and regulations.
In fact, new technologies are required to further improve the National
Airspace System (NAS) operation.
Air Traffic Management (ATM) can be defined as “the dynamic, integrated manage-
ment of air traffic and airspace including air traffic services,
airspace management and
Aerospace
2023
,
10
, 358
5 of 30
air traffic flow management [...]
in collaboration with all parties and involving airborne
and ground-based functions” [
25
].
To maintain safety and efficiency levels, the Air Traffic
Control
(ATC)
guarantees
smooth
airspace
flow
through
ATC
services.
The
airspace
is
divided into regions [
1
] and served by Air Traffic Controllers (ATCos) via Air Traffic Flow
Management (ATFM), Airspace Management (ASM), and Air Traffic Services (ATS), as
illustrated in Figure
6
.
These services comprise flight information service, alerting service,
air traffic advisory service, area control service, approach control service, and aerodrome
control service [
1
,
2
].
Figure 6.
Overview of ATM components [
1
,
26
–
28
].
However,
ATC
faces
several
challenges.
For
example,
communication
[
29
],
effi-
ciency [
30
], safety [
31
], excessive ATCo workload [
32
], and congestion [
33
].
Several stake-
holders
have
been
working
on
new
technologies
to
modernize
and
include
new
ATC
capabilities in the National Airspace System (NAS). Conversely, the increasing number of
flights demonstrates that further technological development is required.
These technologies vary from application and the service targeted.
Furthermore, four
critical ATM application-oriented research areas highlighted by SESAR [
34
] are:
•
ATM Operations, Architecture, Performance, and Validation (OAPV):
focuses on so-
lutions
to
enhance
and
enable
trajectory-based
operations,
considering
technologies
related to aircraft trajectory. It may include trajectory planning [
35
,
36
], prediction [
37
,
38
],
generation [
39
,
40
], optimization [
41
,
42
], and clustering [
43
,
44
];
•
Enabling Aviation Infrastructure (EAI):
includes technologies to enable more flexible
architectures.
Involves
ground
and
airborne
systems
that
can
be
useful
for
ATM.
For example, aircraft health prediction [
45
], optimization [
46
], and management [
47
];
•
High-Performing Airport Operations (HPAO):
targets emerging technologies to im-
prove situational awareness for tower controllers.
For example, it may include tempo-
ral aspects that affect the airspace operations, such are departure delay
prediction [
48
,
49
]
and arrival delay prediction [
50
];
•
Advanced
Air
Traffic
Services
(AATS):
involves
tools
to
improve
departure
and
arrival processes, separation management, air and ground safety, and systems to sup-
port flight planning.
This area refers to solutions considering the interaction between
humans and computers and may include augmentation solutions [
51
,
52
] and behav-
ioral technologies [
53
].
Moreover, it also considers airspace complexity
solutions [
54
]
,
e.g., initiatives
related to complexity estimation [
55
] and reduction
[
56
,
57
]
are exam-
ples of solutions in this portfolio.
Therefore, this research focuses on reviewing DL efforts in these five ATM solutions
categories.
Moreover, we also present an in-depth discussion on future efforts and open
challenges for each group.
In the past few years, several ATM solutions have been proposed in the literature. These pro-
posals are based on several methods and techniques, e.g., deterministic
optimization [
58
]
, stochas-
tic
solvers
[
59
],
and
data-driven
strategies
[
60
].
Furthermore,
Artificial
Intelligence
(AI)
and
Aerospace
2023
,
10
, 358
6 of 30
Deep
Learning
(DL)
present
state-of-the-art
performance
in
challenging
tasks
(e.g.,
image
classification [
61
])
while solving scientific problems in multiple areas. Therefore, there has been
an effort towards leveraging AI and DL capabilities to foster the development of next-gen ATM
systems.
Figure
7
shows
the
evolution
of
results
count
provided
by
Google
Scholar
using
“ATM
+
Artificial
Intelligence” and “ATM + Deep Learning” as search strings.
This shows an
increase in the interest of AI and DL in ATM solutions.
Figure 7.
Evolution of AI and DL search results in ATM in the past few years.
In this sense, the development of new solutions relies on the understanding of the
existing body of knowledge and research gaps.
In this research, we focus on reviewing
multiple
DL-based
initiatives
in
ATM
to
support
the
development
of
future
solutions.
Conversely,
we focus on the identification of research opportunities while highlighting
how DL methods are used.
The statistical evaluation of the performance of such methods
compared to baseline solutions is in the scope of future works.
3.
Literature Review
This section presents all works reviewed categorized into five classes, matching the
architectures previously discussed:
Deep Neural Networks (DNN), Convolutional Neural
Networks (CNNs), Recurrent Neural Networks (RNN), Generative Adversarial Networks
(GANs), and Autoencoders (AE). Once the review of all works is presented, we compare
the contributions in Table
1
.
Each work is classified regarding the following attributes:
•
Year:
Describes the year in which the article was published;
•
ATM Area:
Categorizes the article into one of the four ATM solutions areas previously
described, i.e., OAPV, EAI, HPAO, and AATS;
•
ATM System:
Indicates if the solution focuses on Air Traffic Services (ATS), Airspace
Management (ASM), or Air Traffic Flow Management (ATFM);
•
Flight solution:
Indicates if the solution is directly applicable to one (S) or multiple
(M) aircraft;
•
Deep
Learning
(DL)
Application:
Refers to the aspects of the Deep Learning (DL)
application, indicating if the authors presented details on the architecture (Arc), vali-
dation (Val), and deployment (Dp);
•
Airspace Key Performance Indicator (KPI):
Indicates the main target of the proposed
method regarding airspace operations.
Works contributions are classified into safety
Aerospace
2023
,
10
, 358
7 of 30
(Sft), efficiency (Ef), and sustainability (Sus).
Although some initiatives overlap multi-
ple KPIs, we intend to identify the primary focus;
•
Air Traffic Controller (ATCo):
Indicates if the proposed solution is intended to sup-
port the operation of ATC professionals.
This attribute identifies if the solution pro-
posed considers human factors (HF—e.g., mental workload and fatigue identification)
and augmentation capabilities (Aug—e.g.,
indicates if the solution indents to help
professionals in the task).
3.1.
Applications of Deep Neural Networks (DNN) Networks in ATM
The authors in [
62
] propose an Air Traffic Control (ATC) model to guide an arbitrary
number of aircraft across three-dimensional, unstructured airspace safely.
This challenging
problem relies on the complex set of tasks performed by ATC due to the increasing number
of
aircraft.
Therefore,
the
authors
emphasize
that
autonomous
ATC
functionalities
are
necessary to support future operations.
To accomplish this,
graph-based deep learning
approaches are used to handle the input ordering of aircraft and the varying number of
aircraft.
In the 24 h simulation experiment, the proposed method managed the airspace by
avoiding 100% of potential collisions and preventing 89.8% of potential conflicts. In addition
to
this
effort,
the
authors
in
[
63
]
focus
on
defining
which
variables
determine
airspace
complexity based on machine learning models.
In this sense, DL interpretability can play
an important role in future works [
64
,
65
].
Wang et al. [
66
] apply cutting-edge DL techniques to predict flight departure demand in
a strategic time horizon. This effort is intended to support MITRE’s Pacer program to improve
operators’
situation
awareness
of
the
potential
for
departure
delays
during
busy
periods.
To accomplish this, the authors leverage better data sources (i.e., Aviation System Performance
Metrics (ASPM) and System Wide Information Management (SWIM)) and robust forecasting
algorithms.
The
authors
trained
forecasting
models
with
DL
techniques
of
sequence
to
sequence (seq2seq) and seq2seq with attention and showed through field demonstrations that
the Mean Squared Error (mse) can be reduced using the proposed strategy.
The authors in [
67
] investigate the effectiveness of the Hybrid Deep Learning (HDL)
in the departure delay severity prediction for ten major airports in the U.S. that experience
high ground and air congestion.
In fact, the effectiveness of airports and airlines greatly
relies
on
punctuality,
and
HDL
models
have
demonstrated
promising
results
in
many
complex problems.
This motivated the authors to propose a strategy to analyze structured
air traffic data as a combination of a Feed-Forward Artificial Neural Network model, and a
gradient boosted tree model (XGBoost).
The proposed strategy achieves a rise of 22.95% in
accuracy when compared to a pure neural network model.
In [
68
], the authors investigate the feasibility of machine learning methods for cost
reduction and service quality improvement in low-cost airlines (LCAs) based on the use
of
predictive
modeling
approaches
and
real
airline
datasets.
Two
major
problems
are
faced,
i.e.,
fuel consumption prediction and flight delay prediction.
To accomplish this,
the authors use different methods,
such as Random Forest,
XGBoost,
and Deep Neural
Network.
The experiments conducted showed that the proposed approach predicts fuel
consumption
and
delays
with
high
accuracy.
Finally,
the
authors
conclude
that
these
models are effective for the investigated airports using the information available one day
before the flight.
Bala
et
al.
[
69
]
evaluate
the
performances
of
Deep
Feed-Forward
Neural
Network,
Neural Network, and Support Vector Machine models on a binary classification problem
using flight on-time data records from the United State Bureau of Transportation Statistics.
As
previously
discussed,
flight
delays
impact
airport
and
airline
operations,
resulting
in significant economic losses.
This motivates the accurate prediction of such factors to
enable
informed
decision-making
in
the
aviation
industry.
Conversely,
it
depends
on
several aspects, e.g., the air transportation system complexity and airport infrastructure.
The
experiments
showed
that
the
proposed
strategy
can
be
used
to
tackle
the
problem
highlighted.
The
authors
emphasized
the
contribution
of
this
initiative
to
the
aviation
Aerospace
2023
,
10
, 358
8 of 30
industry
and
the
air
transportation
unit
concerning
improving
passengers’
experience
through better flight delay decision support systems.
In [
70
], the authors develop a flight delay prediction system based on domestic flights
inside
the
United
States
of
America.
The
impact
of
delays
on
the
airline
business
is
significant.
Although there is an interest in increasing the predictability of such events,
this remains a major challenge currently.
In order to tackle this problem, the authors use
machine learning and Deep Learning methods.
The case studies demonstrated that the
models learn the cause of flight delays and cancellations and associate them with the link
between departure and arrival delays.
Similarly, the authors in [
71
] propose a predictive
solution for fight delays based on Deep Learning (DL) and on the Levenberg
-
-Marquardt
(LM)
algorithm.
The
results
obtained
from
the
experiments
showed
that
the
proposed
model efficiently predicts delays.
Mas-Pujol et al. [
72
] propose two Deep Learning models to mimic the current proce-
dure’s behavior to help specialized ATCos detect the imbalances that will require regulation.
The flight allocation in the current ATC system is required to be time-efficient, cost-efficient,
and safe through the Demand–Capacity Balancing process.
This process entails analyzing
corrective actions in the form of regulations in areas with high demand to avoid overload.
However, this procedure is complex, time-consuming, and based on ATCo’s experience.
To tackle this problem, the authors use a CNN and an RNN to demonstrate that regulation
can be predicted with over 80% accuracy for the context considered.
Chakrabarty [
73
] focuses on the flight arrival delay prediction for flights operated by
American Airlines, predicting possible arrival delays of the flight using machine learning
approaches.
In fact, flight delays result in airline companies operating commercial flights
incurring losses, and new methods to avoid them are needed.
The experiments conducted
showed that the Gradient Boosting Classifier model achieves maximum accuracy of 85.73%.
Finally, the author indicates that using Machine Learning-Deep Learning Hybrid Models
tuned with Grid Search to achieve better model performance is in the scope of future works.
The authors in [
74
] examine requirements to be deployed different techniques opera-
tionally in an ATM system, exploring aspects of such as verification, regulatory certification,
and end-user acceptance.
The success of AI solutions motivates their application in aviation
systems.
In this research, the authors consider a novel cognitive Human–Machine Interface
(HMI) configured through machine learning.
The authors highlight that the increasing
levels
of
automation
and
autonomy
are
expected
to
include
certification
requirements,
and a discussion is conducted regarding how ATM systems can be accommodated into
the existing certification framework for aviation systems.
This research brings important
insights into the application of how ML and DL can be part of ATM systems and presents
future directions that converge with DL research topics (e.g., explainability).
In [
75
], the authors propose a DL method to construct an aircraft network and utilize
the complexity indices to characterize it.
As the basic unit of the airspace system is the Air
Traffic Control Sector (ATCS) is the basic unit of the airspace system, identifying congestion
in such areas enables decision support for strategic planning and daily operations.
The ex-
isting approaches focus on the static structure and the dynamic operational features, which
has motivated the authors to develop a more flexible strategy.
In this sense, congestion
identification becomes the complexity of the aircraft network and is detected by a Deep
Active Learning (DAL) model.
The experiments showed that this approach outperforms
existing mainstream methods in the four objective evaluation indices.
Facing air transportation delays as an outcome of local airport dynamics and the global
propagation process, the authors propose a method to assess airport identifiability in [
76
].
The focus of this research is to demonstrate how DL models can recognize airports with
high precision and that delays are more dependent on each airport’s characteristics than
the global air transportation system’s effects.
As a natural result, identifiability is higher
for large and highly connected airports.
Finally, the authors highlight that the proposed
approach is superior to the mainstream approaches in multiple aspects.
Aerospace
2023
,
10
, 358
9 of 30
Boggavarapu
et
al.
[
77
]
present
a
delay
estimation
DL
model
trained
using
the
air
traffic and weather data obtained from the U.S. Bureau of Transportation Statistics and
NOAA—National
Oceanic
and
atmospheric
administration.
Facing
flight
delays
as
a
major challenge in the air transportation system, the authors acknowledge that delays are
influenced by several factors.
In fact, this complexity hardens the accurate prediction of
delays in different scenarios.
To tackle this problem, a Gated Recurrent Unit (GRU) network
is
adopted
due
to
the
recurrent
and
time-series
nature
of
the
dataset.
The
experiments
performed showed that the proposed approach is effective in estimating departure delays
based on a case study focusing on the Chicago airport.
In
[
78
],
the
authors
address
to
provide
an
overview
of
the
state
of
the
art
for
ap-
plying DL to the aircraft design, dynamics, and control field.
Several DL solutions have
been
proposed
in
this
context,
focusing
on
an
information-rich,
data-driven
approach.
Two main groups are considered:
own-ship aircraft modeling, including proposals that
have been/can be implemented online for the aircraft design/dynamics/control, and other
airplane research works, DL-based solutions for offline monitoring of the aircraft operation.
The
authors
describe
several
efforts
throughout
the
paper
and
point
out
several
open
challenges to be addressed in future works.
The authors in [
79
] present a DL-based approach to augment the job of both ground
controllers and pilots.
The current challenges faced in the aviation industry (e.g., profes-
sional shortage) represent a concern for aeronautic enterprises and regulators, given the
increasing number of annual flights.
In this context, the Single Pilot Operations concept
relies on automation in several layers of the air transportation system.
The authors use
Meteorological Terminal Air Reports to create a model based capable of determining the
approach trajectory of an aircraft thirty minutes before the landing time.
The experiments
performed were conducted on aircraft trajectories from Toulouse to Seville, demonstrating
that the proposed strategy achieves over 90% accuracy in the prediction task.
In [
80
], the authors’ Machine Learning (ML) and Artificial Intelligence (AI) methods
are proposed to control and predict the state of air traffic.
ATC plays a pivotal role in society
and acts in safety-critical scenarios.
Therefore, there is a mental workload experienced by
Air Traffic Controllers (ATCos) that needs to be maintained at low levels.
The proposed
strategy, based on different statistical methods (e.g., neural network), presents high accuracy
prediction compared to other statistical algorithms with over 95% accuracy.
In fact, this
contribution emphasizes that DL techniques can be used to build up multiple prediction
services to support the ATC operation.
The
authors
in
[
81
]
present
a
survey
and
a
DL-based
model
of
real-time
aircraft
tracking systems.
This problem is a current challenging issue in the literature for several
reasons, e.g., the need for an accurate and complete data transfer from aviation to ground
systems. Conversely, aircraft tracking becomes difficult due to data loss caused by telemetry
or data acquisition.
In this context, the authors present a survey of aircraft tracking systems,
categorizing works into three classes (i.e., mathematical, machine learning-based, and Deep
Learning-based).
After that, a real-time Deep Learning-based Aircraft Tracking (DeepAT)
system that enables real-time tracking of an aircraft is introduced.
DeepAT offers promising
results in the experiments performed to prevent data loss in different applications.
3.2.
Applications of Convolutional Neural Networks (CNN) in ATM
In [
82
], the authors propose an advanced Bayesian Deep Learning method for aircraft
trajectory prediction considering weather impacts.
Trajectory prediction is a challenge, but
a required aspect of the next-generation National Air Transportation System (NATS), and
reliable prediction models must consider uncertainties from various sources.
Then, the
authors introduce a deterministic trajectory prediction model with classical deep learning
methods to handle both spatial and temporal information.
Moreover, multiple layers are
used, e.g., CNN, RNN, and fully connected layers.
The experiments showed a significant
reduction in prediction variance compared to existing methods.
Aerospace
2023
,
10
, 358
10 of 30
The
authors
in
[
83
]
propose
a
Hybrid
Deep
Neural
Network
(HDNN)
for
active
hazard
identification
of
Auxiliary
Power
Units
(APU)
in
civil
aircraft.
This
method
is
based on a multi-time window Convolutional Neural Network–Bidirectional Long Short-
Term Memory (CNN-BiLSTM) neural network, and integrates three models with different
time window sizes in parallel.
Hence, this combination can automatically extract features
to
represent
the
system
state
and
learn
the
time-based
patterns
in
the
time
series
data.
The authors also state that this strategy has the most stable identification performance for
data with imbalanced samples in comparison to others present in the literature.
In [
84
], the authors introduce an automatic image-based aircraft defect detection using
a pre-trained Convolutional Neural Network (CNN) for feature extraction and the Support
Vector Machine (SVM) method with a linear kernel in the classification step. Throughout the
paper,
the
author
highlights
that
this
initiative
intends
to
support
regular
maintenance
using visual and nondestructive Inspection (NDI) and a detailed description of the images
used and how the classification process works.
The experiments showed that the proposed
method is accurate (96%) and presents high performance even in simple hardware.
The authors in [
85
] present a generalizable efficient tree-based matching algorithm to
build feature maps from meteorological datasets (i.e.,
wind,
temperature,
and convective
weather). This effort is focused on aviation efficiency and targets reliable 4D aircraft trajectory
prediction. In this sense, the authors propose an end-to-end convolutional recurrent neural
network that consists of a Long Short-Term Memory (LSTM) encoder network and a mixture
density LSTM decoder network.
Then, to enable high-dimension weather representation
learning, the authors include convolutional layers into the pipeline. The case studies showed
that the learned filters successfully locate convective weather and generalize the weather-
related features using real operational data (flights from IAH to BOS).
Xie
et
al.
[
86
]
propose
an
end-to-end
Sector
Operation
Complexity
(SOC)
learning
framework based on deep CNN. This study is motivated by the lack of approaches that
do not rely on hand-crafted factors.
Indeed, these factors require specialized background
and might limit the evaluation performance of the model.
Then, the authors propose a
Multichannel
Traffic
Scenario
Image
(MTSI)
to
represent
the
overall
air
traffic
scenario
by splitting the airspace into a two-dimension grid map,
extracting high-level features,
and learning the SOC pattern with the support of CNN. The experiments showed that the
proposed strategy can effectively extract traffic complexity information from MTSIs.
The
authors
in
[
87
]
introduce
two
flight
delay
prediction
models
based
on
CNN
employing fusion of meteorological data to predict flight delays.
The first model is the
Dual-channel Convolutional Neural Network (DCNN) based on the ResNet, whereas the
second refers to the Squeeze and Excitation-Densely Connected Convolutional Network
(SE-DenseNet), which is a combination of SENet and DenseNet.
The main idea is to rely
on flight and meteorological data fusion with efficient feature recalibration procedures.
The experiments demonstrated that the accuracy of the model can be enhanced by the pro-
posed strategy, and the two networks introduced in this paper can improve the prediction
process, reaching 92.1% and 93.19%.
The authors in [
88
] focus on generating individual-sensitive resolution advisories for
air traffic conflicts.
The authors’ goal is to increase the acceptance of workload-alleviating
automation in air traffic control by adapting advisories to different controller strategies.
In fact, this personalization is reached using a CNN model trained on individual controller
data.
A
human-in-the-loop
experiment
was
performed
to
generate
datasets
of
conflict
geometries and controller resolutions, and the results demonstrated that this strategy can
predict command type, direction, and magnitude.
An unfolding contribution is presented
in [
89
], focusing on performing an exploratory investigation into conformal and individual-
sensitive automation for Air Traffic Control (ATC) based on CNNs.
There have been several
challenges
regarding
ATC
augmentation
and
automation
related
to
a
lack
of
trust
and
acceptance.
The authors investigate automation from the personalization standpoint to
individual controllers.
This approach relies on combining visual features and a tailored
CNN trained on individual controller data collection from human-in-the-loop simulations.
Aerospace
2023
,
10
, 358
11 of 30
The
results
achieved
showed
that
the
proposed
approach
can
predict
command
type,
direction, and magnitude.
In
[
90
],
the
authors
present
a
hybrid
RNN-CNN
cascade
architecture
to
predict
C-
ATC
capacity
regulations
for
en-route
traffic.
The
authors
investigate
three
different
Air
Traffic Management (ATM) frameworks to improve the cost-efficiency for flow and network
Management considering facing the detection of regulations. To accomplish this, two Deep
Learning models are combined, creating a different hybrid model. This combination comprises
a Recurrent Neural Network (RNN) and a Convolutional Neural Network (CNN) to extract
the overall airspace characteristics and process artificial images of the airspace configuration.
The experiments used historical data from two of the most regulated European regions and
showed that a cascade architecture presents average accuracy of 88.45%.
The authors in [
91
] propose a neural network structure combining CNN and LSTM
to classify hypersonic aircraft flight trajectories.
In the past few years, there has been an
increase
in
interest
in
supersonic
and
hypersonic
due
to
their
high
speed
and
complex
maneuvering mode.
Thus, the classification process compensates for their shortcomings of
insensitivity to temporal and spatial characteristics and can effectively classify two kinds of
hypersonic glide vehicles.
The classification experiments demonstrated that the proposed
model has good performance under the condition of introducing observation noise.
A CNN-based multi-feature predictive model (MF-CNN) is proposed in [
92
] to pre-
dict network-scale traffic flow with multiple spatiotemporal features and external factors.
In today’s air traffic system, traffic flow prediction is paramount for many applications,
e.g., traffic control and route guidance.
Conversely, it is not simple to identify important
features using traditional data-driven traffic flow prediction models (e.g., periodicity and
weather).
Therefore, the authors classify traffic features into temporal continuity as short-
term features and daily periodicity, and weekly periodicity as long-term features.
After
that, they are mapped into three 2D spaces.
In this process, CNNs learn high-level spa-
tiotemporal features and provide them to the logistic regression layer for final prediction.
The experiments showed that the MF-CNN model improves the predictive performance
compared to the five baseline models.
In [
93
], the authors expand previous work on thunderstorm forecasting [
94
] by ap-
plying
CNNs
to
exploit
the
spatial
characteristics
embedded
in
weather
data.
In
fact,
thunderstorms can disrupt Air Traffic Management (ATM) procedures, causing a complex
state of operation within the airspace system.
Currently, it is still a challenge to have precise
forecasts, hardening strategic planning.
Then, the authors focus on deep learning as it has
provided
promising
results
in
different
scenarios.
The
learning
task
is
formulated
as
a
binary-classification problem based on satellite data.
The experiments compared different
Deep Learning (DL) architectures, e.g., a fully Convolutional Neural Network (FCN), a
CNN-based encoder–decoder, a UNet, a pyramid-scene parsing network (PSPNet), and
Multi-Layer Perceptron (MLP). The results indicate that CNN-based architectures improve
the performance of point-prediction models and can be used to increase the prediction lead
time of thunderstorms.
In
[
95
],
the
authors
present
a
novel
processing
paradigm
to
integrate
multilingual
speech recognition for robust speech recognition in Air Traffic Control (ATC). This refers
to a single framework using an Acoustic Model (AM), a Pronunciation Model (PM), and
a
Language
Model
(LM).
The
process
works
as
follows:
the
AM
converts
ATC
speech
into phoneme-based text sequences, the PM translates these sequences into a word-based
sequence, and the LM corrects both phonemes- and word-based errors in the decoding
results.
The
AM
includes
a
CNN
and
an
RNN
considering
the
spatial
and
temporal
dependences of the speech features.
The authors used large amounts of real Chinese and
English ATC recordings and achieved a 3.95% label error rate.
Rahman et al. [
96
] converted trajectory data into images, which size does not depend on
the number of planes, and developed a multi-label conflict resolution model called ACRnet.
This model uses a CNN to classify the obtained images. To solve aircraft conflicts, Air Traffic
Controllers (ATCos) interact with flight crews observing several parameters (e.g., positioning,
Aerospace
2023
,
10
, 358
12 of 30
speed,
direction,
and
weather),
hardening
the
task
with
the
currently
congested
airspace.
In this sense, supportive systems can help ATCos in their tasks. The experiments conducted
demonstrated that ACRnet achieves high accuracy for two aircraft and three aircraft.
Liu et al. [
97
] introduce a recurrent 3D CNN (R-3DCNN) to consider the spatial and
temporal air traffic transitions comprehensively for Air Traffic Flow (ATF) prediction. The au-
thors employ a new data representation called Traffic Situation Graphics (TSG)—generated
by splitting the 3D earth space with fixed grid maps and flight levels—to illustrate traffic
flow situations in a single instant. Then, the 3D CNN and LSTM extract high-level features
(spatial and temporal) from a TSG sequence, assuming that inputs are determined by com-
bining the traffic situations on different flight levels with areas affected by other real-time
factors.
The evaluation demonstrated that the proposed strategy can obtain accurate and
stable prediction results of ATF prediction with distribution on different flight levels.
The authors in [
98
] present an effort to identify flight states based on CNN. To this end,
a novel one-dimension CNN is introduced to automatically extract useful features from
the
structural
vibration
of
a
recently
fabricated
self-sensing
wing
through
wind-tunnel
experiments.
In fact, it is challenging to identify the flight state from the complex vibration
signals
with
high
accuracy.
To
accomplish
this,
the
authors
decomposed
the
obtained
signals into various sub-signals with different frequency bands and formed the best possible
combination for multichannel inputs of the CNN. The two case studies showed that the
proposed approach can achieve high identification accuracy and robustness,
providing
new perspectives on self-awareness toward the next generation of intelligent air vehicles.
3.3.
Applications of Recurrent Neural Networks (RNN) in ATM
The authors in [
99
] propose a short-term wind speed prediction framework for bridge
traffic
control
under
strong
winds.
The
goal
is
to
improve
the
estimation
accuracy
for
the timeframe of traffic control during a typhoon.
The authors use a hybrid modeling of
wind speed at the bridge and a Time-Shifted Data Correction (TSDC) method.
The hybrid
modeling considers two available data types (i.e., structural health monitoring and regional
specialized meteorological center—RSMC) using training features based on the maximum
sustained winds of a typhoon.
Throughout the paper, a graphical and in-depth description
of all steps is presented and, as a numerical example, typhoons from 2020 were used as
test data to demonstrate the improvement in prediction performance via the use of hybrid
modeling and the TSDC method.
This initiative describes an approach that can be used in
several aspects of Air Traffic Management (ATM) in future works.
The
authors
in
[
100
]
introduce
a
four-dimensional
flight
trajectory
prediction
model
based on a Long Short-Term Memory (LSTM) network to maintain the long-term features
and manage to predict accurate trajectories. Several factors contribute to stable and safe air
traffic,
especially during the climbing and descending phases.
In this sense,
technologies
and equipment must present precise information in each flight phase to ensure fluidity and
safety. Conversely, strong external interference or blind zones present potential risks to these
operations.
Throughout the paper, several aspects are described in detail, and the authors
adopt a clear and graphical approach to introduce their contributions. Finally, the experiments
conducted showed that the proposed system is able to provide timely decision support.
Ref. [
101
] proposes a combination of convolutional layers into Long Short-Time Mem-
ory (LSTM) cells to predict the aircraft trajectory based on the weather condition and flight
plan.
Convective weather avoidance is vital in safe operations,
and it is also a primary
objective
of
the
next-generation
air
traffic
management
system.
Therefore,
the
authors
use history flight track data, the last on-file flight plan, and the time-dependent convec-
tive weather map over the period from 1 November 2018 to 5 February 2019.
The flights
investigated had JFK-LAX as the city pair.
The experiments conducted showed that the
proposed
approach
can
reduce
the
deviation
compared
to
the
last
on-file
flight
plan
in
47.0% of the predicted flight tracks.
Similarly, the authors in [
102
] propose a strategy to
predict Air Traffic Flow and Capacity Management (ATFCM) weather regulations using a
time-distributed Recurrent Neural Network.
Aerospace
2023
,
10
, 358
13 of 30
In
[
103
],
the
authors
propose
a
trajectory
prediction
model
based
on
a
dual-self-
attentive (DSA)-temporal convolutional network (TCN)-bidirectional gated recurrent unit
(BiGRU)
neural
network.
The
main
idea
is
that
TCN
provides
highly
stable
training,
high parallelism, and a flexible perceptual domain, whereas the self-attentive mechanism
can focus on features that contribute the most to the output.
Then, the BiGRU network
discovers connections between features and outputs of the trajectory sequence, optimized
by a Bayesian algorithm.
Experiments demonstrated that the DSA-TCN-BiGRU model
based on Bayesian hyperparameter optimization outperforms other models present in the
literature.
Similarly, Shi et al. [
104
] propose a flight trajectory prediction model based on a
Long Short-Term Memory (LSTM) network.
This approach can accurately predict flight
trajectories in both 3D and 4D spaces.
The authors also point out that multi-modal data
(e.g., audio and video) can be included in the process in future works.
The authors propose a 4D trajectory prediction hybrid architecture based on a combi-
nation of Convolutional Neural Network (CNN) and Long Short-Term Memory (LSTM)
in
[
105
].
Considering
a
4D
trajectory
as
a
multi-dimensional
time
series
with
plentiful
spatial-temporal
features
with
a
high
degree
of
complexity
and
uncertainty,
providing
accurate
solutions
is
complex.
In
the
training
process,
the
authors
use
real
Automatic
Dependent Surveillance-Broadcast (ADS-B) historical trajectory data and compare the pro-
posed method with a single LSTM model and Feed-Forward (FF) model on the same data
set.
Experiments showed that the trajectory prediction accuracy of the proposed strategy is
superior to a single model.
In [
106
], the authors use an improved GRU network to study the time series of traffic
flows.
The LSTM short-term traffic flow prediction based on the flow series is investigated
as a more complex version of the GRU model in this case.
In this sense, an improved GRU
with bidirectional positive and negative feedback (called the Bi-GRU) is used to complete
the short-term traffic flow prediction leveraged by the Rectified Adaptive (RAdam) model in
the optimization process.
Finally, the experiment conducted demonstrated the effectiveness
of the proposed method regarding short-term traffic flow prediction.
This contribution can
be further extended and applied to air transportation in different flight phases and regions.
The authors in [
107
] propose a model based on the Social Long Short-Term Memory
(S-LSTM)
network
to
predict
multi-aircraft
trajectory.
Facing
trajectory
prediction
as
a
paramount feature in future operations, several efforts have been made to enhance such
estimations.
Therefore, the authors focus on building a model for each aircraft and use
a pooling layer to integrate the hidden states of the associated aircraft.
The experiments
considered aircraft trajectories in the Northern California terminal area as the experimen-
tal data and showed that the proposed S-LSTM model presents fewer prediction errors
compared with the mainstream trajectory prediction models.
Moreover, the authors also
emphasize
the
importance
of
considering
aircraft
interaction
as
a
factor
in
predicting
trajectories accurately.
In [
108
], the authors introduce the Airport Traffic Flow Prediction Network (ATFPNet),
a DL-based framework to capture spatial-temporal dependencies of the historical airport
traffic
flow
in
multiple-step
situational
arrival
flow
prediction.
To
accomplish
this,
the
authors
rely
on
a
special
semantic
graph
built
on
the
flight
schedule
to
represent
the
airport network.
Moreover, the graph convolution operator and the GRU are combined to
extract transition patterns of airport traffic flow.
The experiments used a real-world airport
traffic flow dataset and showed that the ATFPNet outperforms other baselines on different
prediction horizons, achieving up to a 17% MAE improvement.
Zhao
et
al.
[
109
]
introduce
a
Deep
Long
Short-Term
Memory
(D-LSTM)
model
for
aircraft trajectory prediction tailored to complex flight environments.
In Trajectory-Based
Operations (TBO), several predictive capabilities are necessary for safe and efficient opera-
tions.
Conversely, trajectory prediction is a major challenge and airspace complexity can
compromise estimation accuracy.
In fact, the current state-of-the-art forecasting methods
are difficult to be applied in actual operation and management.
Therefore, the authors use
multi-dimensional features of aircraft trajectory, testing the proposed strategy with real
Aerospace
2023
,
10
, 358
14 of 30
flight data (ADS-B). The experiments showed that this strategy presents prediction accuracy
more than existing methods in different flight phases.
In addition to this contribution, the
authors in [
110
] use Recurrent Neural Networks (RNN) to predict air traffic density in ATC
sectors, while the authors in [
111
] propose a strategy to predict air traffic congested areas
using LSTM networks.
Ref. [
112
] studies the problem of estimating aircraft onboard parameters using ground
surveillance available parameters. The authors adopt an LSTM model and Flight Data Records
to estimate target parameters upon three aspects: fuel flow rate, flap configuration, and landing
gear position. In fact, this effort fosters ATM awareness regarding aircraft behaviors, enabling
the evaluation of system performance in terms of safety and efficiency.
Future works are
also pointed out by the authors, e.g., enhanced flap and landing gear setting prediction with
airspeed information. The insights of this paper highlights that predictive capabilities related
to aircraft performance can support informed decisions by ATC.
3.4.
Applications of Generative Adversarial Networks (GANs) in ATM
Wu et al. [
113
] propose a long-term 4D trajectory prediction model based on Generative
Adversarial Network (GAN). The authors use three deep generative models based on one-
dimensional convolution neural network (Conv1D-GAN), a two-dimensional convolution
neural
network
(Conv2D-GAN),
and
long
short-term
memory
neural
network
(LSTM-
GAN). Although 4D trajectory prediction is a capability required for future ATM systems,
it represents a complex problem due to several factors.
Using this model combination, this
paper trains and tests a predictive solution using historical 4D trajectory data from Beijing to
Chengdu, China.
The experiments conducted showed that the proposed strategy presents
promising results and that the Conv1D-GAN is the most suitable generative adversarial
network for long-term aircraft trajectory prediction.
In [
114
], a Conditional Generative Adversarial Network (CGAN) approach is proposed
for
weather-related
aircraft
trajectory
prediction
problems.
Furthermore,
the
generator
network
focuses
on
weather
feature
extraction
and
includes
two
convolutional
layers.
Then, the features are provided to a single-layer long short-term network to output the
generated trajectory.
The discriminator network tries to discriminate the inputs from the
ground truth dataset and the generated trajectory.
The experiments were conducted based
on the data obtained from Sherlock Data Warehouse (SDW), and the results suggest that
the proposed strategy outperforms other proposals present in the literature.
Aksoy et al. [
115
] present a hybrid methodology to generalize the flight trajectories
and decide whether they are abnormal or not.
The first approach relies on considering
the
time-based
features
of
the
trajectories.
This
is
composed
of
LSTM
autoencoders
to
rapidly predict the class of the flight, inherently considering the time-based features of a
flight trajectory.
The second approach on a more pattern strategy through a Generative
Adversarial Network (GAN), which generates realistic samples.
Flight trajectories are dif-
ferent even when following patterns that are flown previously and optimized for different
conditions.
These patterns can be influenced by several factors (e.g., airspace utilization,
controllers’ cognitive complexity, weather, and NOTAMs).
The obtained results showed
that this approach can classify anomalies in trajectories.
The authors in [
116
] consider multiple operational aircraft taxi-speed factors (e.g., sur-
rounding traffic on the ground and target take-off time) and adopt the Generative Adversarial
Imitation Learning (GAIL) algorithm for modeling. The main goal is to enable the model to
learn and reproduce the ground movement patterns in a real-world dataset under different
circumstances. The contributions of this research are very valuable to the ATM community
since it is difficult to predict the spatio-temporal component of aircraft-taxi trajectory.
Fur-
thermore, this initiative supports the ATM decision-making process. The proposed strategy
outperforms all the baseline models by a significant margin. For example, it achieves up to
97.1% for arrivals and 88.3% for departures concerning Spatial Completion (SC).
In
[
117
],
the
authors
develop
a
method
for
using
Generative
Adversarial
Networks
(GANs) to generate condition monitoring data of aircraft engines.
To accomplish this,
an
Aerospace
2023
,
10
, 358
15 of 30
algorithm for generating monitoring data to extend a sample of aircraft engine condition
monitoring data is proposed. Monitoring costs and difficulties represent an obstacle in today’s
operations since little condition monitoring data exists. This hardens data-driven approaches
to
aircraft
engine
maintenance.
The
experiments
conducted
showed
that,
based
on
the
condition monitoring parameters recorded for a CF6-80C2A5 engine, the proposed strategy
was able to generate data within a reasonable range. A complementary approach regarding
data
generation
is
presented
in
[
118
],
in
which
the
authors
evaluated
the
performance
of
synthetically generated snow radar images based on modified cycle-consistent adversarial
networks. In fact, this approach can be used in different ATM applications.
Hu
et
al.
[
119
]
propose
a
short-term
aircraft
Trajectory
Prediction
(TP)
framework
called TPGAN. The primary goal of this effort is to predict multi-horizon trajectories in a
single step using the Conditional Generative Adversarial Network (CGAN). Regarding the
trajectory
prediction
capabilities,
existing
works
usually
perform
the
multi-horizon
TP
task
iteratively,
suffering
from
error
accumulation
problems.
The
authors
employ
the
generator to output the predictions, whereas the discriminator learns the discriminative
features
between
ground
truth
and
predictions
and
applies
the
generative
adversarial
training strategy to optimize the proposed framework.
The experiments used a dataset
collected
from
real-world
ATC
systems
and
demonstrated
that
the
proposed
strategy
achieves significant performance improvements compared to an LSTM-based baseline.
Guo et al. [
120
] present a novel anomaly detection model based on Improved Gen-
erative Adversarial Networks and long short-term memory networks (IGAN-LSTM) to
detect anomalies in ADS-B Systems.
To accomplish this,
IGAN enhances the generator
architecture by upgrading the commonly used encoder architecture to encoder-decoder-
encoder architecture.
Then, encoding losses are used to determine whether a data sample
is anomalous or not, and LSTM networks are used to model ADS-B data with temporal
dependence.
The experiments performed showed that the proposed approach outperforms
other baseline methods from the literature.
Huang et al. [
121
] propose a novel method called Improved Wasserstein Skip-Connection
GAN
(IWGAN).
The
main
goal
is
to
integrate
the
Wasserstein-GAN
(WGAN)
and
SkipGANomaly models to distinguish normal and abnormal images, which is called the Im-
proved Wasserstein Skip-Connection GAN (IWGAN). The challenge of not having relatively
sufficient datasets in the airport field hardens the training of DL models in different ATM
applications.
Since GANs can learn the latent vector space of all images, the authors adopt a
GAN variant with autoencoders to create a hybrid model for detecting anomalies and hazards
in the airport environment.
The experiments showed that the proposed model is efficient
in solving the problem faced in this research. In addition to this initiative, Zhang et al. [
122
]
propose an image-based aircraft type recognition approach based on Conditional Generative
Adversarial Networks (GANs).
The authors in [
123
] focus on evaluating aircraft trajectory generation methods and
propose a common baseline to compare literature and new methods to generate air traffic
trajectories.
In fact, state-of-the-art methods to generate individual trajectories can lack
realism concerning common situations implemented by ATCos.
Conversely, data-driven
approaches excel at imitating operational practice but may not be simple to implement
due to aircraft performance limitations.
Therefore, the authors present an extensive set of
metrics to evaluate the quality of generated trajectories and point out as future directions
the use of this framework to objectively assess trajectory generation performance (e.g., using
GANs).
An effort correlated with this approach is presented in [
124
], in which the authors
explore the use of generative data models to learn real approach flight path probability
distributions through the use of GANs.
Lang
et
al.
[
125
]
propose
a
fault
prediction
of
aircraft
engine
based
on
data
aug-
mentation technology.
Predicting the Remaining Useful Life (RUL) of the equipment by
constructing models using historical data is challenging since the data is difficult to obtain.
One possible way to solve this problem is to develop models based on data augmentation.
To accomplish this, the authors use a GAN to study the distribution of the original dataset
Aerospace
2023
,
10
, 358
16 of 30
and to generate a new training set.
The original and the new datasets are combined to train
a Convolutional Neural Network and Long Short-Term Memory Network (CNN-LSTM)
prediction model. The case studies demonstrate that the proposed GAN-CNN-LSTM model
can effectively predict the RUL compared with the existing methods.
3.5.
Applications of Autoencoders in ATM
In
[
126
],
the
authors
propose
Deep
Learning
(DL)
techniques
to
model
Air
Traffic
Controllers’ (ATCos) reactions in resolving conflicts.
The authors focus on the Air Traffic
Controllers’ (ATCos) reaction prediction problem for Conflict Detection and Resolution
(CD&R). DL methods that can model ATCOs’ timely reactions are presented and evaluated
in real-world data sets.
Throughout the paper, the authors describe all contributions in a
graphical and in-depth manner.
The experiments conducted showed that the proposed
approach presents very high accuracy in such a detection problem.
Finally, multiple future
directions are pointed out in this research.
In fact, the authors in [
127
] present a correlated
contribution but propose a method to analyze flight trajectories, detect unusual flight behav-
iors, and infer ATC actions.
The authors in [
128
] focus on evaluating pilots’ fatigue status
using the Deep Contractive Autoencoder Network.
The authors propose a fatigue evalua-
tion index to compute the power spectrum of relative rhythms from electroencephalogram
(EEG) signals.
A novel multivariate anomaly detection model called Contextual Auto-Encoder (CAE)
is
proposed
in
[
129
].
The
main
goal
is
to
use
the
baseline
of
a
regular
LSTM-based
au-
toencoder combined with several decoders to gather data on a specific flight phase (e.g.,
climbing, cruising, or descending) in training.
Although ADS-B supports the tracking of
the high number of aircraft in the air, it also introduces cybersecurity concerns that must
be mitigated.
To tackle this problem, a dataset was created using real-life anomalies and
realistically crafted trajectory modifications, with which the CAE, alongside three anomaly
detection models from the literature, were evaluated.
Experimental results showed that
CAE achieves better results in both accuracy and speed of detection.
In addition to this con-
tribution, the authors in [
130
] focus on focus on detecting anomalies in real-time for flight
testing.
To this end, the authors propose an approach based on fine-tuned autoencoder to
extract generic underlying features, followed by a stacked LSTM.
Ref. [
131
] proposes a novel engine fault detection method based on original Aircraft
Communications Addressing and Reporting System (ACARS) data is proposed.
The au-
thors
divide
all
variables
into
separated
groups
according
to
their
correlations
and
use
an improved convolutional denoising autoencoder to extract the features of each group.
All extracted features are then fused to form feature vectors to enable fault sample iden-
tification.
The evaluation process showed that this method is efficient in fault detection
and robustness while maintaining low computational and time costs.
Furthermore, Fer-
nandez
et
al.
[
132
]
performed
descriptive
and
predictive
analyses
to
detect
anomalies,
i.e., Flight Data Monitoring (FDM) unknown hazards, during the approach phase.
Corrado et al. [
133
] introduce a novel framework based on DL methods using autoen-
coders to identify anomalies in terminal airspace operations. The primary goal is to leverage
historical aircraft trajectory data combined with weather and traffic metrics to build an
anomaly detection capability.
To accomplish this, data from multiple sources (
e.g., aircraft
trajectory, weather, traffic/congestion) is used to train the models and demonstrated on
six months of arriving flight data collected for San Francisco International Airport that the
proposed strategy has the potential to aid air traffic controllers.
In [
134
], a new deep stacked autoencoders networks method is proposed to predict
flight delay based on the relationship of time and space.
In fact, the stacked autoencoder
approach
derives
the
characteristics
of
flight
delay
information
from
massive
data
and
optimizes all the networks’ parameters with the backpropagation method.
Throughout the
article, the authors describe the contributions and the results achieved in detail (e.g., for
different periods in the future).
Results demonstrate that the prediction accuracy with deep
Aerospace
2023
,
10
, 358
17 of 30
stacked autoencoders is above 90%.
Finally, several future works can be identified from
this initiative based on the use of autoencoders to improve prediction performance.
Wu
et
al.
[
135
]
design
a
feature
association
algorithm
to
solve
the
ATM
systems
security
situation
awareness
via
a
Deep-Related
Sparse
Autoencoder
(DRSAE)
model.
In safe and efficient operations, it is pivotal that security situational awareness information
is
provided
to
the
air
traffic
management
(ATM)
system
with
an
integrated
air-ground
structure.
However, there are problems that can impact the process of situation awareness
feature extraction.
In this sense, the authors design and compare the DRSAE model with
other feature extraction models (e.g., Principal Component Analysis—PCA, Autoencoder—
AE, and Sparse Autoencoder—SAE). The results obtained showed that the DRSAE model
is robust in feature extraction of the ATM system, presenting strong expression ability.
The authors in [
136
] propose a Convolutional Variational Auto-Encoder (CVAE) for
anomaly
detection
in
high-dimensional
time-series
flight
data.
Currently,
the
current
approach for identifying vulnerabilities in NAS operations leverages domain expertise,
which works well when the system has a well-defined operating condition.
Conversely,
highly complex scenarios can be faced in the airspace, and state-of-the-art machine learning
models
usually
rely
on
supervised
learning.
In
many
cases,
labeling
data
requires
spe-
cialized expertise that is time-consuming and, therefore, largely impractical.
Motivated
by these challenges, the authors validate the proposed approach on Yahoo’s benchmark
data and on a case study of identifying anomalies in commercial flights’ take-offs.
The re-
sults showed that CVAE outperforms both classic and deep learning-based approaches in
detecting anomalies.
Zeng
et
al.
[
137
]
and
Olive
et
al.
[
138
]
focus
on
the
use
of
autoencoders
to
cluster
trajectories.
The
former
work
proposes
a
trajectory
clustering
method
based
on
Deep
Autoencoder (DAE) and Gaussian mixture model (GMM) to identify the prevailing traf-
fic flow patterns in the terminal airspace,
whereas the latter explores the application of
deep
trajectory
clustering
based
on
autoencoders
to
the
problem
of
flow
identification.
Both efforts face a challenging problem for ATM systems since trajectories might change
substantially if unexpected events happen (e.g., weather-related events).
In addition to this
effort, the authors in [
139
] propose a framework for predicting air traffic situations as a
sequence of images using an autoencoder with convolutional Long Short-Term Memory
(ConvLSTM).
Table 1.
Comparison of works that use Deep Learning (DL) in ATM solutions.
ATC System
Flight
Solution
DL Application
Airspace
KPI
ATCo
Model
Paper
Year
ATM
Area
ATS
ASM
ATFM
S
M
Arc
Val
Dp
Sft
Ef
Sus
HF
Aug
Malekzadeh et al.
[
84
]
2017
EAI
Liu et al. [
85
]
2018
OAPV
Yang et al. [
92
]
2019
AATS
Liu et al. [
97
]
2019
AATS
Chen et al. [
98
]
2019
EAI
Van et al. [
88
]
2019
AATS
Qu et al. [
87
]
2020
HPAO
Van et al. [
89
]
2020
AATS
Lin et al. [
95
]
2020
AATS
Pang et al. [
82
]
2021
OAPV
Zeng et al. [
91
]
2021
OAPV
Xie et al. [
86
]
2021
AATS
Di et al. [
83
]
2022
EAI
Mas et al. [
90
]
2022
AATS
Jardines et al. [
93
]
2022
AATS
CNN
Rahman et al. [
96
]
2022
AATS
Aerospace
2023
,
10
, 358
18 of 30
Table 1.
Cont.
ATC System
Flight
Solution
DL Application
Airspace
KPI
ATCo
Model
Paper
Year
ATM
Area
ATS
ASM
ATFM
S
M
Arc
Val
Dp
Sft
Ef
Sus
HF
Aug
Horiguchi et al.
[
68
]
2017
HPAO
Kistan et al. [
74
]
2018
AATS
Boggavarapu
et al. [
77
]
2019
HPAO
Chakrabarty et al. [
73
]
2019
HPAO
Mollinga et al. [
62
]
2020
AATS
Wang et al. [
66
]
2020
AATS
Mas et al. [
72
]
2020
AATS
Yazdi et al. [
71
]
2020
HPAO
Jimenez et al. [
79
]
2020
OAPV
Cheevachaipimol
et al. [
67
]
2021
HPAO
Bala et al. [
69
]
2021
HPAO
Dong et al. [
78
]
2021
EAI
Gholami et al. [
70
]
2022
HPAO
Tan et al. [
75
]
2022
AATS
Ivanoska et al. [
76
]
2022
HPAO
Sangeetha et al. [
80
]
2022
AATS
Çakıcı et al. [
81
]
2022
OAPV
DNN
Perez et al. [
63
]
2022
AATS
Shi et al.
[
104
]
2018
OAPV
Pang et al. [
101
]
2019
OAPV
Zhao et al. [
109
]
2019
OAPV
Shi et al. [
100
]
2020
OAPV
Ma et al. [
105
]
2020
OAPV
Jarry et al. [
112
]
2020
EAI
Shi et al. [
111
]
2021
AATS
Shu et al. [
106
]
2021
AATS
Xu et al. [
107
]
2021
OAPV
Yan et al. [
108
]
2021
AATS
Mas et al. [
102
]
2021
AATS
Lim et al. [
99
]
2022
AATS
Huang et al. [
103
]
2022
OAPV
RNN
Asirvadam et al. [
110
]
2022
AATS
Guo et al. [
120
]
2021
OAPV
Olive et al. [
123
]
2021
OAPV
Lang et al. [
125
]
2021
EAI
Jarry et al. [
124
]
2021
OAPV
Wu et al. [
113
]
2022
OAPV
Hu et al. [
119
]
2022
OAPV
GAN
Huang et al. [
121
]
2023
AATS
Aerospace
2023
,
10
, 358
19 of 30
Table 1.
Cont.
ATC System
Flight
Solution
DL Application
Airspace
KPI
ATCo
Model
Paper
Year
ATM
Area
ATS
ASM
ATFM
S
M
Arc
Val
Dp
Sft
Ef
Sus
HF
Aug
Olive et al.
[
127
]
2018
AATS
Chen et al. [
134
]
2018
HPAO
Xuyun et al. [
131
]
2019
EAI
Fernandez et al. [
132
]
2019
EAI
Que et al. [
130
]
2019
EAI
Wu et al. [
128
]
2019
AATS
Memarzadeh
et al. [
136
]
2020
Aircraft
Olive et al. [
138
]
2020
OAPV
Corrado et al. [
133
]
2021
AATS
Zeng et al. [
137
]
2021
OAPV
Kim et al. [
139
]
2021
AATS
Bastas et al. [
126
]
2022
AATS
Chevrot et al. [
129
]
2022
OAPV
AE
Wu et al. [
135
]
2022
AATS
3.6.
Further Insights:
Opportunities
This section presents an analysis of the works reviewed.
First,
a discussion on the
publications and their respective keywords is conducted.
Then, a granular analysis of the
future direction of different ATM solutions is presented.
3.7.
Paper Count and Keywords
In the past few years, several papers have been published in the context of DL and
ATM. This research focuses on presenting state-of-the-art solutions published in the past
few years.
Figures
8
and
9
illustrate the paper count by year and by ATM area.
Most of
the
papers
analyzed
have
been
published
in
the
past
few
years
in
high-impact
venues.
In fact, the trend in the data highlights that we can expect new solutions in the next few
years.
Moreover, AATS and OAPV solutions are the topics of most of the papers published.
However, EAI and HPAO solutions are also present.
Although there are clear categories for these solutions, they focus on different problems.
Thus, Figure
10
illustrates the word cloud produced by the combination of keywords of
all papers analyzed.
Terms such as “Deep Learning”, “prediction”, and “trajectory” are
clearly present in several papers.
Some other words such as “speech” and “DCNN” are
only present in a few articles.
Figure 8.
Paper count by year.
Aerospace
2023
,
10
, 358
20 of 30
Figure 9.
Paper count by ATM solution areas.
Figure 10.
Word cloud of keywords.
3.8.
ATM Solutions
The
papers
reviewed
have
been
categorized
according
to
their
main
contribution.
Here, we group works by ATM areas and present their future direction, either highlighted
by the authors or gathered from the text.
These insights are listed in Table
2
.
Several future directions are identified for OAPV, including the optimization of models,
including more features, temporal dependencies, and applications in different domains.
These are the direct unfolding of the works analyzed, and these directions highlight that
robust and accurate trajectory-based solutions can benefit from these initiatives.
Regarding AATS, several future directions are identified.
Conversely, some aspects
are pointed out by different authors, e.g., the consideration of the airspace state (e.g., sector
density) to optimize model performance and the use of realistic scenarios to simplify future
deployments.
The efforts made toward fatigue and workload detection and estimation are
applicable to several ATM systems.
Data diversity is also highlighted as a primary goal for
future works.
Finally, disruptive events are also suggested for future endeavors.
Moreover, important insights are presented in the reviewed HPAO papers, and the use
of more data and further optimization of models are described in most of them as future
directions. Some mention their applications in different domains (e.g., aircraft maintenance),
and there is a clear demand for contributions in rare and disruptive conditions (e.g., weather).
Aerospace
2023
,
10
, 358
21 of 30
Finally, EAI works emphasize the need for more data, cross-domain applications, and
further development of the proposed models.
The consideration of complexity factors (e.g.,
sector density) is also a concern.
Table 2.
Future directions for different ATM areas.
ATM
Area
Paper
Year
Model
Future Directions
Yang et al.
[
92
]
2019
CNN
Flow prediction in severe and rare weather conditions
Liu et al. [
97
]
2019
CNN
Application of proposal in a real scenario
Van et al. [
88
]
2019
CNN
Automatic generation of training data
consisting of solution space diagram (SSD)
images and conflict resolutions
Xie et al. [
86
]
2021
CNN
Application of visual-based techniques in other ATM solutions
Jardines et al. [
93
]
2022
CNN
Investigate temporal relationships in weather data
Rahman et al. [
96
]
2022
CNN
Inclusion of airspace information to improve conflict resolution
Wang et al. [
66
]
2020
DNN
Application of different DL methods to improve performance
Tan et al. [
75
]
2022
DNN
Use of Deep Active Learning in other ATM solutions
Perez et al. [
63
]
2022
DNN
Application of proposal in different sector types
Shi et al. [
111
]
2021
RNN
Quantification of the uncertainty in the predictions
Shu et al. [
106
]
2021
RNN
Consideration of special events (e.g., weather and
large activities)
Yan et al. [
108
]
2021
RNN
Use of other factors to improve prediction accuracy
(e.g., weather information, ATC information,
the influence of international flights, and dynamic
traffic movements on the network.
Mas et al. [
102
]
2021
RNN
Implementation of visual framework to apply theoretical
regulations and create feedback to re-train the existing model.
Lim et al. [
99
]
2022
RNN
Analysis of severe weather impacts on airports
Asirvadam
et al. [
110
]
2022
RNN
Airspace optimization considering workload,
weather and unplanned traffic
Zhang et al. [
122
]
2018
GAN
Unsupervised classification method to remove the need for
data labeled with type information
Rahnemoonfar
et al. [
118
]
2020
GAN
Simulation of other ATM systems (e.g., audio)
Huang et al. [
121
]
2023
GAN
Use of attention model mechanisms
Corrado et al. [
133
]
2021
AE
Hyperparameter Optimization
Kim et al. [
139
]
2021
AE
Inclusion of altitude in the proposed method
Wu et al. [
135
]
2022
AE
Use of more advanced classifiers
Van et al. [
89
]
2020
CNN
Model Optimization
Lin et al. [
95
]
2020
CNN
Increase data diversity
Mas et al. [
90
]
2022
CNN
Inclusion of additional input features to improve performance
Mollinga et al. [
62
]
2020
DNN
Inclusion of stochastic variables like weather,
addition of waypoints, and change the simulation approach
Mas et al. [
72
]
2020
DNN
Use of an hybrid DL model
Sangeetha
et al. [
80
]
2022
DNN
Consideration of rare events (e.g., weather-based events)
Pham et al. [
116
]
2021
GAN
Implementation of a multi-agent environment
Olive et al. [
127
]
2018
AE
Consideration of more evolved structures of networks
Wu et al. [
128
]
2019
AE
Application of similar strategies considering ATC professionals
AATS
Bastas et al. [
126
]
2022
AE
Improvements of predictions regarding
low-level ATCOs’ conflict resolution actions
Liu et al.
[
85
]
2018
CNN
Extension of proposed algorithm to more features (e.g.,
ATM initiatives)
Pang et al. [
82
]
2021
CNN
Consideration of rare events (e.g., weather-based events)
Zeng et al. [
91
]
2021
CNN
Approaches to handle loss of
information from data normalization.
Jimenez et al. [
79
]
2020
DNN
Use of a more extensive dataset
Çakıcı et al. [
81
]
2022
DNN
Data sharing approaches considering aircraft and ATC
Shi et al. [
104
]
2018
RNN
Use of multi-modal data, including images, audios and videos
OAPV
Pang et al. [
101
]
2019
RNN
Consideration of rare events (e.g., weather-based events)
Aerospace
2023
,
10
, 358
22 of 30
Table 2.
Cont.
ATM
Area
Paper
Year
Model
Future Directions
Zhao et al. [
109
]
2019
RNN
Application of D-LSTM to trajectory
information prediction in high density airspace
Shi et al. [
100
]
2020
RNN
Inclusion of of multi-modal data
OAPV
Ma et al. [
105
]
2020
RNN
Models for long-term 4D trajectory prediction
Xu et al. [
107
]
2021
RNN
Integration of meteorological conditions to achieve
more accurate and stable trajectory prediction
Huang et al. [
103
]
2022
RNN
Models that use the combination of weather, control, and
other uncertainties.
Pang et al. [
114
]
2020
GAN
Development of models to improve the prediction performance
Aksoy et al. [
115
]
2021
GAN
Consideration of rare events (e.g., weather-based events)
Guo et al. [
120
]
2021
GAN
Inclusion of new features to to further
improve the performance of the proposed model
Olive et al. [
123
]
2021
GAN
Application of the proposed method to
compare data-driven trajectory generation models
Jarry et al. [
124
]
2021
GAN
Analysis of tailored network architectures and learning
Wu et al. [
113
]
2022
GAN
Application of the propose method for short-term
trajectory prediction and air traffic state estimation.
Hu et al. [
119
]
2022
GAN
To adopt this approach in other ATM solutions
Olive et al. [
138
]
2020
AE
Impact assess of clustering losses on the performance of
reconstruction-based anomaly detection methods.
Zeng et al. [
137
]
2021
AE
Use of the proposed model to assist trajectory
prediction solutions
OAPV
Chevrot et al. [
129
]
2022
AE
To adopt this approach in other domains
Malekzadeh et al.
[
84
]
2017
CNN
Use of other DNN architecture for this
application
Chen et al. [
98
]
2019
CNN
Consideration of rare events (e.g., weather-based events)
Di et al. [
83
]
2022
CNN
Consideration of rare events (e.g., weather-based events)
Dong et al. [
78
]
2021
DNN
Data Collection, Labeling, and transfer Learning
Jarry et al. [
112
]
2020
RNN
Enhancing flap and landing
gear setting prediction with airspeed information
Fu et al. [
117
]
2019
GAN
To adopt this approach in other ATM solutions
Lang et al. [
125
]
2021
GAN
To adopt this approach in other ATM solutions
Xuyun et al. [
131
]
2019
AE
Use of more com prehensive fault cases to locate the
fault source
Fernandez
et al. [
132
]
2019
AE
Consideration of other airspace aspects (e.g., sector desity)
EAI
Que et al. [
130
]
2019
AE
Automate rapid development of efficient
anomaly detection on FPGAs for various applications
Qu et al.
[
87
]
2020
CNN
Use of new models to improve results
Horiguchi
et al. [
68
]
2017
DNN
Inclusion of reservation data in the analysis
Boggavarapu
et al. [
77
]
2019
DNN
Use of more airport data
Chakrabarty
et al. [
73
]
2019
DNN
Use of advanced preprocessing and sampling techniques
Yazdi et al. [
71
]
2020
DNN
Consideration of rare events (e.g., weather-based events)
Cheevachaipimol
et al. [
67
]
2021
DNN
Use of other methods to handle imbalanced data
Bala et al. [
69
]
2021
DNN
Use of DNN in aircraft maintenance
Gholami et al. [
70
]
2022
DNN
Consideration of rare events (e.g., weather-based events)
Ivanoska et al. [
76
]
2022
DNN
Consideration of other aspects in flights (e.g., aircraft usage)
HPAO
Chen et al. [
134
]
2018
AE
Consideration of rare events (e.g., weather-based events)
4.
Open Challenges
This Section discusses several possible directions for future works based on the insights
of this review.
These open challenges are presented from two perspectives:
Deep Learning
(DL) applications and ATM solutions.
4.1.
Deep Learning (DL) Applications
•
Interpretability:
Understanding the decision made by the DL models is paramount
for
safe
and
efficient
operations.
Hence,
the
proposal
of
strategies
to
explain
the
Aerospace
2023
,
10
, 358
23 of 30
decision made in the full ATM system spectrum is a challenging but necessary step
for future operations;
•
Sustainability:
From all works analyzed in this research, none of them are focused on
sustainability services for future ATM systems.
The use of DL techniques to foster the
development of a new strategy is critical for future operations;
•
Cybersecurity:
Data are shared throughout the airspace systems in today’s operations,
and future applications will require even more data. In this sense, it is critical that new
solutions use advanced techniques (e.g., DL) to detect and mitigate cybersecurity threats;
•
Urban Air Mobility (UAM):
Several of the points discussed in this research are also
applicable to new transportation paradigms, e.g., UAM. For example, DL can enable
the
development
of
trajectory-based
solutions
tailored
to
the
UAM
environment.
In fact, lessons learned in the National Airspace System (NAS) can provide insightful
directions for UAM applications [
140
,
141
];
•
Deployment:
Few of the works reviewed in this research focus on the deployment of such
solutions. Besides the complexity that developing a DL application involves, deployment
is also challenging and requires coordination with several stakeholders. Indeed, solutions
to simplify the deployment of such methods are part of the open challenges.
4.2.
ATM Solutions
Figure
11
illustrates the open challenges in ATM solutions and directions to apply ad-
vanced methods to enhance the airspace operation.
All challenges identified are discussed
in detail based on their respective classification.
Figure 11.
Open challenges:
ATM solutions.
4.2.1.
Advanced Air Traffic Services (AATS)
•
Temporal effects:
Throughout the daily operations, the airspace state changes several
times.
However,
there
is
an
intrinsic
temporal
dependence
in
the
evolution
of
the
airspace state.
In this sense, initiatives can focus on identifying temporal connections
to improve the complexity prediction capabilities;
•
Integrated Analysis:
The complex airspace ecosystem entails various systems operat-
ing simultaneously.
In this sense, an open challenge refers to using global resources
and information to understand how complexity is impacted and, ultimately, develop
more accurate complexity-based solutions;
•
Disruption Management:
The consideration of rare but disruptive events is vital in
the assessment and development of new complexity-based solutions.
Therefore, new
capabilities for predicting rare events are part of the future ATM solutions portfolio.
•
Data Collection:
Building up datasets for training models is a complex challenge due
to several factors.
The development of new techniques to easily collect data without
compromising the operation will produce valuable resources for new applications;
Aerospace
2023
,
10
, 358
24 of 30
•
Integrated
Augmentation:
Although
some
initiatives
focus
on
supporting
profes-
sionals in the ATM system, they are commonly separated and not part of the same
portfolio.
The
development
of
a
scalable
and
integrated
approach
to
be
used
as
a
baseline across the ATM system is an open challenge;
•
Training Frameworks:
The inclusion of new technologies into the National Airspace
System (NAS) requires several phases of testing and certification.
In this sense, propos-
ing new training approaches to simplify the use of these new technologies is necessary.
4.2.2.
High-Performing Airport Operations (HPAO)
•
Airport
Data
Collection:
Although
following
the
same
rules,
airports
operate
dif-
ferently
due
to
several
factors
(e.g.,
size,
number
of
gates,
and
number
of
flights).
Collecting
internal
data
to
improve
the
airport
performance
is
a
pillar
for
future
solutions, and the development of a new data collection strategy is essential;
•
Data Sampling:
The current difficulty faced in collecting the data demands methods
to generate realistic samples.
In this case, there is a need for new sampling methods
that cover the characteristics of the airport operation;
•
Disruption Management:
The impacts of disruption in delays is difficult to predict
due to the minimal number of occurrences in history.
Future solutions are required to
handle this imbalanced environment and accurately predict such events.
4.2.3.
ATM Operations, Architecture, Performance, and Validation (OAPV)
•
New
Aircraft
Concepts:
There
are
some
companies
working
on
the
production
of
supersonic aircraft that will be integrated into the NAS in the near future.
Then, new
trajectory prediction services are required to attend to the new flight configurations
and capabilities (both in terms of performance and regulations);
•
Multi-Modal
Analysis:
Trajectory-based
solutions
are
integrated
into
a
complex
ecosystem composed of several subsystems.
In this sense, using data from different
sources and configurations (e.g., audio and video) represents another open challenge;
•
Disruption Management:
Rare events are difficult to predict accurately.
The develop-
ment of new methods capable of estimating when disruptive events happen and how
they affect the aircraft trajectory needs investigation.
4.2.4.
Enabling Aviation Infrastructure (EAI)
•
Data Collection:
For several reasons, collecting data from aircraft (e.g., engines) is a
complex task.
Therefore, the development of new software and hardware technologies
for data collection will improve the results obtained by the existing and future ATM
solutions;
•
Integrated
Health
Analysis:
The
complex
ecosystem
composed
of
several
subsys-
tems that surround the aircraft can provide information for in-flight decision-making.
In this sense, integrating the in-flight solutions with the ecosystem can yield valuable
resources and is in the scope of future works;
•
New Aircraft Concepts:
As new aircraft operate in the NAS (e.g., supersonic aircraft),
flight parameters are expected to differ.
Then, adjusting the existing solutions for such
an environment is pivotal for efficient operations.
4.2.5.
Integrated Solutions (IS)
•
Collaborative Decision:
Considering that local actions can change the global airspace
mesh,
solutions
to
enable
safe,
rapid,
and
efficient
collaborative
decision-making
represent a significant advancement in today’s technologies.
However, reaching this
flawless collaboration is not simple, as it represents a research field in ATM systems.
•
DL Orchestration:
The use of DL solutions in several areas of ATC can improve efficiency.
However, enabling these separated entities to communicate and share resources (e.g.,
parameters and outputs) can provide a smooth integration experience.
However, the
orchestration of such systems is complex and represents an open challenge.
Aerospace
2023
,
10
, 358
25 of 30
•
Knowledge sharing:
Data belonging to different organizations is not always shared
due
to
several
reasons
(e.g.,
privacy).
Then,
new
privacy-preserving
techniques
(e.g., privacy-preserving transfer learning) can overcome this obstacle and enable DL
applications to be more accurate without compromising privacy.
5.
Conclusions
Currently, the increasing number of daily flights offers new travel connections and
simplifies global transportation.
In this context, Air Traffic Management (ATM) enables air
carriers to operate safely and efficiently through the multiple services provided.
The timely
success of advanced analytic solutions has demonstrated their potential to solve complex
problems in several domains, including Air Traffic Management (ATM). Although there
are several contributions in the literature, it is not simple to define the challenges faced by
state-of-the-art strategies and open challenges due to a lack of comprehensive and extensive
analysis of such contributions.
Therefore, this research presented a comprehensive review of state-of-the-art Deep
Learning (DL) solutions for Air Traffic Management (ATM). Several topics were discussed,
focusing
on
applications,
opportunities,
and
open
challenges
to
foster
the
evolution
of
ATM systems.
Several areas of ATM applications were considered, and the state-of-the-art
solutions were classified, analyzed, and compared from different perspectives.
Finally, an
extensive discussion on the open challenges was conducted to highlight the current DL-
based solutions demanded in current and future ATM contexts.
This article provided the
reader with a clear picture of the current DL-ATM landscape and the main directions for
future works.
Finally, there are several aspects to be considered in future directions, e.g., an
analysis of solutions based on other Artificial Intelligence (AI) methods (e.g., optimization)
and solutions for specific flight missions (e.g., Search and Rescue—SAR).
Author Contributions:
Conceptualization, E.C.P.N., D.M.B., J.R.d.A.J., J.B.C.J. and P.S.C.; Method-
ology, E.C.P.N. and D.M.B.; Software, E.C.P.N.; Validation, E.C.P.N., D.M.B., J.R.d.A.J., J.B.C.J. and
P.S.C.; Formal Analysis, E.C.P.N. and P.S.C.; Investigation, E.C.P.N., D.M.B., J.R.d.A.J., J.B.C.J. and
P.S.C.;
Resources,
E.C.P.N.,
D.M.B.,
J.R.d.A.J.,
J.B.C.J. and P.S.C.;
Data Curation,
E.C.P.N.,
D.M.B.,
J.R.d.A.J., J.B.C.J. and P.S.C.; Writing—Original Draft—Preparation, E.C.P.N.; Writing—Review &
Editing, E.C.P.N., D.M.B., J.R.d.A.J., J.B.C.J. and P.S.C.; Visualization, E.C.P.N.; Supervision, P.S.C.;
Project Administration, P.S.C.; Funding Acquisition, E.C.P.N., D.M.B., J.R.d.A.J., J.B.C.J. and P.S.C. All
authors have read and agreed to the published version of the manuscript.
Funding:
This research received no external funding.
Institutional Review Board Statement:
Not applicable.
Informed Consent Statement:
Not applicable.
Data Availability Statement:
Not applicable.
Conflicts of Interest:
The authors declare no conflict of interest.
References
1.
ICAO.
Air Traffic Management—DOC 4444.
2012.
Available online:
https://www.icao.int/EURNAT/Other%20Meetings%20
Seminars%20and%20Workshops/FPL%202012%20ICAO%20EUR%20Region%20Plan/Documentation%20related%20to%20
FPL%202012%20Amendment/Amendment%201%20Doc4444.EN.pdf
(accessed on 24 March 2023).
2.
ICAO.
Air Traffic Services—Annex 11.
2001.
Available online:
https://skyrise.aero/wp-content/uploads/2017/03/ICAO-
Annex-11-Air-traffic-services.pdf
(accessed on 24 March 2023).
3.
Deng, L.; Yu, D.
Deep learning:
Methods and applications.
Found.
Trends Signal Process.
2014
,
7
, 197–387.
[
CrossRef
]
4.
Chakraborty,
S.;
Tomsett,
R.;
Raghavendra,
R.;
Harborne,
D.;
Alzantot,
M.;
Cerutti,
F.;
Srivastava,
M.;
Preece,
A.;
Julier,
S.;
Rao, R.M.; et al.
Interpretability of deep learning models:
A survey of results.
In Proceedings of the 2017 IEEE Smartworld,
Ubiquitous
Intelligence
&
Computing,
Advanced
&
Trusted
Computed,
Scalable
Computing
&
Communications,
Cloud
&
Big Data Computing, Internet of People and Smart City Innovation (Smartworld/SCALCOM/UIC/ATC/CBDcom/IOP/SCI),
San Francisco, CA, USA, 4–8 August 2017; pp.
1–6.
5.
LeCun, Y.; Bengio, Y.; Hinton, G.
Deep learning.
Nature
2015
,
521
, 436–444.
[
CrossRef
] [
PubMed
]
6.
Goodfellow, I.; Bengio, Y.; Courville, A.
Deep Learning
; MIT Press:
Cambridge, MA, USA, 2016.
Aerospace
2023
,
10
, 358
26 of 30
7.
Shinde, P.P.; Shah, S. A review of machine learning and deep learning applications.
In Proceedings of the 2018 Fourth International
Conference on Computing Communication Control and Automation (ICCUBEA), Pune, India, 16–18 August 2018; pp.
1–6.
8.
Chassagnon, G.; Vakalopolou, M.; Paragios, N.; Revel, M.P.
Deep learning:
Definition and perspectives for thoracic imaging.
Eur.
Radiol.
2020
,
30
, 2021–2030.
[
CrossRef
] [
PubMed
]
9.
Chaddad, A.; Peng, J.; Xu, J.; Bouridane, A.
Survey of Explainable AI Techniques in Healthcare.
Sensors
2023
,
23
, 634.
[
CrossRef
]
[
PubMed
]
10.
Sujith, A.; Sajja, G.S.; Mahalakshmi, V.; Nuhmani, S.; Prasanalakshmi, B.
Systematic review of smart health monitoring using
deep learning and Artificial intelligence.
Neurosci.
Inform.
2022
,
2
, 100028.
[
CrossRef
]
11.
Wu, Y.; Wu, L.; Cai, H.
A deep learning approach to secure vehicle to road side unit communications in intelligent transportation
system.
Comput.
Electr.
Eng.
2023
,
105
, 108542.
[
CrossRef
]
12.
Wang, P.; Qin, J.; Li, J.; Wu, M.; Zhou, S.; Feng, L.
Optimal Transshipment Route Planning Method Based on Deep Learning for
Multimodal Transport Scenarios.
Electronics
2023
,
12
, 417.
[
CrossRef
]
13.
Li, C.; Zheng, P.; Yin, Y.; Wang, B.; Wang, L.
Deep reinforcement learning in smart manufacturing:
A review and prospects.
CIRP
J. Manuf.
Sci.
Technol.
2023
,
40
, 75–101.
[
CrossRef
]
14.
Svozil, D.; Kvasnicka, V.; Pospichal, J.
Introduction to multi-layer feed-forward neural networks.
Chemom.
Intell.
Lab.
Syst.
1997
,
39
, 43–62.
[
CrossRef
]
15.
Sazli, M.H.
A brief review of feed-forward neural networks.
Commun.
Fac.
Sci.
Univ.
Ank.
Ser.-Phys.
Sci.
Eng.
2006
,
50
, 11–17.
[
CrossRef
]
16.
O’Shea, K.; Nash, R.
An introduction to convolutional neural networks.
arXiv
2015
, arXiv:1511.08458.
17.
Wu, J.
Introduction to convolutional neural networks.
Natl.
Key Lab Nov.
Softw.
Technol.
Nanjing Univ.
China
2017
,
5
, 495.
18.
Hijazi, S.; Kumar, R.; Rowen, C.;
Using convolutional neural networks for image recognition.
J. Comput.
Commun.
2015
,
9
, 1.
19.
Schmidt, R.M.
Recurrent neural networks (rnns):
A gentle introduction and overview.
arXiv
2019
, arXiv:1912.05911.
20.
Bullinaria, J.A.
Recurrent neural networks.
Neural Comput.
Lect.
2013
,
12
, 1.
21.
Goodfellow, I.; Pouget-Abadie, J.; Mirza, M.; Xu, B.; Warde-Farley, D.; Ozair, S.; Courville, A.; Bengio, Y.
Generative adversarial
networks.
Commun.
ACM
2020
,
63
, 139–144.
[
CrossRef
]
22.
Pan, Z.; Yu, W.; Yi, X.; Khan, A.; Yuan, F.; Zheng, Y.
Recent progress on generative adversarial networks (GANs):
A survey.
IEEE
Access
2019
,
7
, 36322–36333.
[
CrossRef
]
23.
Bank, D.; Koenigstein, N.; Giryes, R.
Autoencoders.
arXiv
2020
, arXiv:2003.05991.
24.
Girin, L.; Leglaive, S.; Bie, X.; Diard, J.; Hueber, T.; Alameda-Pineda, X.
Dynamical variational autoencoders:
A comprehensive
review.
arXiv
2020
, arXiv:2008.12595.
25.
Skybrary.
Air Traffic Management (ATM). 2022.
Available online:
https://skybrary.aero/articles/air-traffic-management-atm
(accessed on 24 March 2023).
26.
Mihetec,
T.;
Jakši´c,
Z.;
Steiner,
S.
Air
Space
Management
Procedures
in
Europe.
In
Proceedings
of
the
14th
International
Conference on Transport Science ICTS; Citeseer, Portorož, Slovenia, 27 May 2011.
27.
Neto, E.C.P.; Baum, D.M.; Brinati, M.A.; Almeida, J.R.; Cugnasca, P.S.; Camargo, J.B.
A safe flow-management method for air
traffic considering the UAS presence into the non-segregated airspace.
In
Safety and Reliability–Safe Societies in a Changing World
;
CRC Press:
Boca Raton, FL, USA, 2018; pp.
2137–2145.
28.
Kistan,
T.;
Gardi,
A.;
Sabatini,
R.;
Ramasamy,
S.;
Batuwangala,
E.
An
evolutionary
outlook
of
air
traffic
flow
management
techniques.
Prog.
Aerosp.
Sci.
2017
,
88
, 15–42.
[
CrossRef
]
29.
Lin,
Y.
Spoken instruction understanding in air traffic control:
Challenge,
technique,
and application.
Aerospace
2021
,
8
, 65.
[
CrossRef
]
30.
Jackson, J.W.; Green, S.M.
Control applications and challenges in air traffic management.
In Proceedings of the 1998 American
Control Conference; ACC (IEEE Cat.
No.
98CH36207), Philadelphia, PA, USA, 26 June 1998; Volume 3, pp. 1772–1788.
31.
Strohmeier, M.; Schäfer, M.; Lenders, V.; Martinovic, I.
Realities and challenges of nextgen air traffic management:
The case of
ADS-B.
IEEE Commun.
Mag.
2014
,
52
, 111–118.
[
CrossRef
]
32.
Tobaruela, G.; Schuster, W.; Majumdar, A.; Ochieng, W.Y.; Martinez, L.; Hendrickx, P.
A method to estimate air traffic controller
mental workload based on traffic clearances.
J. Air Transp.
Manag.
2014
,
39
, 59–71.
[
CrossRef
]
33.
Roosens, P.
Congestion and Air Transport:
A challenging phenomenon.
Eur.
J. Transp.
Infrastruct.
Res.
2008
,
8
, 137–146.
34.
SESAR.
Exploring
the
Boundaries
of
Air
Traffic
Management—A
Summary
of
SESAR
Exploratory
Research
Results.
2016.
Available online:
https://op.europa.eu/en/publication-detail/-/publication/e3460b15-7499-11ed-9887-01aa75ed71a1
(accessed
on 24 March 2023).
35.
Richards, A.; How, J.P.
Aircraft trajectory planning with collision avoidance using mixed integer linear programming.
In Pro-
ceedings of the 2002 American Control Conference (IEEE Cat.
No.
CH37301), Anchorage, AK, USA, 8–10 May 2002; Volume 3,
pp. 1936–1941.
36.
Soler, M.; Olivares, A.; Staffetti, E.; Zapata, D. Framework for aircraft trajectory planning toward an efficient air traffic management.
J. Aircr.
2012
,
49
, 341–348.
[
CrossRef
]
37.
Ayhan, S.; Samet, H.
Aircraft trajectory prediction made easy with predictive analytics.
In Proceedings of the 22nd ACM SIGKDD
International Conference on Knowledge Discovery and Data Mining, San Francisco, CA, USA, 13–17 August 2016; pp.
21–30.
Aerospace
2023
,
10
, 358
27 of 30
38.
Mondoloni, S.; Rozen, N.
Aircraft trajectory prediction and synchronization for air traffic management applications.
Prog.
Aerosp.
Sci.
2020
,
119
, 100640.
[
CrossRef
]
39.
Kamgarpour, M.; Dadok, V.; Tomlin, C.
Trajectory generation for aircraft subject to dynamic weather uncertainty.
In Proceedings
of the 49th IEEE Conference on Decision and Control (CDC), Atlanta, GA, USA, 15–17 December 2010; pp. 2063–2068.
40.
Patel, R.; Goulart, P.; Serghides, V.
Real-time trajectory generation for aircraft avoidance maneuvers.
In Proceedings of the AIAA
Guidance, Navigation, and Control Conference, Chicago, IL, USA, 10–13 August 2009; p.
5623.
41.
Sridhar, B.;
Ng, H.K.;
Chen, N.Y.
Aircraft trajectory optimization and contrails avoidance in the presence of winds.
J. Guid.
Control Dyn.
2011
,
34
, 1577–1584.
[
CrossRef
]
42.
Rodionova, O.; Sbihi, M.; Delahaye, D.; Mongeau, M.
North atlantic aircraft trajectory optimization.
IEEE Trans.
Intell.
Transp.
Syst.
2014
,
15
, 2202–2212.
[
CrossRef
]
43.
Gariel, M.; Srivastava, A.N.; Feron, E.
Trajectory clustering and an application to airspace monitoring.
IEEE Trans.
Intell.
Transp.
Syst.
2011
,
12
, 1511–1524.
[
CrossRef
]
44.
Olive, X.; Morio, J.
Trajectory clustering of air traffic flows around airports.
Aerosp.
Sci.
Technol.
2019
,
84
, 776–781.
[
CrossRef
]
45.
Jianzhong, S.; Fangyuan, W.; Shungang, N.
Aircraft air conditioning system health state estimation and prediction for predictive
maintenance.
Chin.
J. Aeronaut.
2020
,
33
, 947–955.
46.
Ren, H.; Chen, X.; Chen, Y.
Reliability Based Aircraft Maintenance Optimization and Applications
; Academic Press:
Hoboken, NJ,
USA, 2017.
47.
Boller, C.
Ways and options for aircraft structural health management.
Smart Mater.
Struct.
2001
,
10
, 432.
[
CrossRef
]
48.
Sternberg, A.; Soares, J.; Carvalho, D.; Ogasawara, E.
A review on flight delay prediction.
arXiv
2017
, arXiv:1703.06118.
49.
Esmaeilzadeh, E.; Mokhtarimousavi, S.
Machine learning approach for flight departure delay prediction and analysis.
Transp.
Res.
Rec.
2020
,
2674
, 145–159.
[
CrossRef
]
50.
Mueller, E.; Chatterji, G.
Analysis of aircraft arrival and departure delay characteristics.
In Proceedings of the AIAA’s Aircraft
Technology, Integration, and Operations (ATIO) 2002 Technical Forum, Los Angeles, CA, USA, 1–3 October 2002; p.
5866.
51.
Abbass, H.A.; Tang, J.; Amin, R.; Ellejmi, M.; Kirby, S.
Augmented cognition using real-time EEG-based adaptive strategies for air
traffic control.
In
Proceedings of the Human Factors and Ergonomics Society Annual Meeting
; SAGE Publications:
Los Angeles, CA,
USA, 2014; Volume 58, pp. 230–234.
52.
National Research Council.
Flight to the Future:
Human Factors in Air Traffic Control
; National Academies Press:
Washington, DC,
USA, 1997.
53.
Aricò, P.; Borghini, G.; Di Flumeri, G.; Bonelli, S.; Golfetti, A.; Graziani, I.; Pozzi, S.; Imbert, J.P.; Granger, G.; Benhacene, R.; et al.
Human factors and neurophysiological metrics in air traffic control:
A critical review.
IEEE Rev.
Biomed.
Eng.
2017
,
10
, 250–263.
[
CrossRef
] [
PubMed
]
54.
Kopardekar, P.H.; Schwartz, A.; Magyarits, S.; Rhodes, J.
Airspace complexity measurement:
An air traffic control simulation
analysis.
Int.
J. Ind.
Eng.
Theory, Appl.
Pract.
2009
,
16
, 61–70.
55.
Zhang, K.; Liu, Y.; Wang, J.; Song, H.; Liu, D.
Tree-based airspace capacity estimation.
In Proceedings of the 2020 Integrated
Communications Navigation and Surveillance Conference (ICNS), Herndon, VA, USA, 8–10 September 2020; pp. 5C1–1.
56.
Histon, J.; Li, L.; Hansman, R.J.
Airspace structure, future ATC systems, and controller complexity reduction.
In Proceedings of
the 29th Digital Avionics Systems Conference, Salt Lake City, UT, USA, 3–7 October 2010; pp. 4–A.
57.
Juntama, P.; Chaimatanan, S.; Alam, S.; Delahaye, D.
A distributed metaheuristic approach for complexity reduction in air traffic
for strategic 4D trajectory optimization.
In Proceedings of the 2020 International Conference on Artificial Intelligence and Data
Analytics for Air Transportation (AIDA-AT), Singapore, 3–4 February 2020; pp. 1–9.
58.
Terrab, M.; Odoni, A.R.
Strategic flow management for air traffic control.
Oper.
Res.
1993
,
41
, 138–152.
[
CrossRef
]
59.
Cobano, J.A.; Alejo, D.; Heredia, G.; Ollero, A. 4D trajectory planning in ATM with an anytime stochastic approach.
In Proceedings
of the 3rd International Conference on Application and Theory of Automation in Command and Control Systems, Naples, Italy,
28–30 May 2013; pp. 1–8.
60.
Condé Rocha Murça, M.; Hansman, R.J.
Data-Driven Modeling of Air Traffic Flows for Advanced Air Traffic Management.
Ph.D.
Thesis, Massachusetts Institute of Technology, Cambridge, MA, USA, 2019.
61.
Li, Y.; Zhang, H.; Xue, X.; Jiang, Y.; Shen, Q.
Deep learning for remote sensing image classification:
A survey.
Wiley Interdiscip.
Rev.
Data Min.
Knowl.
Discov.
2018
,
8
, e1264.
[
CrossRef
]
62.
Mollinga, J.; van Hoof, H.
An autonomous free airspace en-route controller using deep reinforcement learning techniques.
arXiv
2020
, arXiv:2007.01599.
63.
Pérez Moreno, F.; Gómez Comendador, V.F.; Delgado-Aguilera Jurado, R.; Zamarreño Suárez, M.; Janisch, D.; Arnaldo Valdés,
R.M.
Determination of Air Traffic Complexity Most Influential Parameters Based on Machine Learning Models.
Symmetry
2022
,
14
, 2629.
[
CrossRef
]
64.
Carvalho, D.V.; Pereira, E.M.; Cardoso, J.S.
Machine learning interpretability:
A survey on methods and metrics.
Electronics
2019
,
8
, 832.
[
CrossRef
]
65.
Ribeiro, M.T.; Singh, S.; Guestrin, C.
Model-agnostic interpretability of machine learning.
arXiv
2016
, arXiv:1606.05386.
66.
Wang, L.; Mykityshyn, A.; Johnson, C.; Marple, B.D.
Deep learning for flight demand forecasting.
arXiv
2020
, arXiv:2011.04476.
67.
Cheevachaipimol, W.; Teinwan, B.; Chutima, P.
Flight Delay Prediction Using a Hybrid Deep Learning Method.
Eng.
J.
2021
,
25
, 99–112.
[
CrossRef
]
Aerospace
2023
,
10
, 358
28 of 30
68.
Horiguchi, Y.; Baba, Y.; Kashima, H.; Suzuki, M.; Kayahara, H.; Maeno, J.
Predicting fuel consumption and flight delays for
low-cost airlines.
In Proceedings of the AAAI Conference on Artificial Intelligence, San Francisco, CA, USA, 4–9 February 2017;
Volume 31, pp. 4686–4693.
69.
Bala Bisandu, D.; Salih Homaid, M.; Moulitsas, I.; Filippone, S.
A deep feedforward neural network and shallow architectures
effectiveness comparison:
Flight delays classification perspective.
In Proceedings of the 2021 The 5th International Conference on
Advances in Artificial Intelligence (ICAAI), Online, 20–22 November 2021; pp. 1–10.
70.
Gholami, S.; Khashe, S.
Flight Delay Prediction Using Deep Learning and Conversational Voice-Based Agents.
Am.
Acad.
Sci.
Res.
J. Eng.
Technol.
Sci.
2022
,
89
, 60–72.
71.
Yazdi, M.F.; Kamel, S.R.; Chabok, S.J.M.; Kheirabadi, M.
Flight delay prediction based on deep learning and Levenberg-Marquart
algorithm.
J. Big Data
2020
,
7
, 1–28.
[
CrossRef
]
72.
Mas Pujol, S.; Salamí San Juan, E.; Pastor Llorens, E.
A novel methodology to predict regulations using deep learning.
In Proceed-
ings of the 10th SESAR Innovation Days:
7th of December-10th of December, 2020, Online, 7–10 December 2020; pp. 1–8; Single
European Sky ATM Research (SESAR).
73.
Chakrabarty, N.
A data mining approach to flight arrival delay prediction for american airlines.
In Proceedings of the 2019 9th
Annual Information Technology, Electromechanical Engineering and Microelectronics Conference (IEMECON), Jaipur, India,
13–15 March 2019; pp. 102–107.
74.
Kistan, T.; Gardi, A.; Sabatini, R.
Machine learning and cognitive ergonomics in air traffic management:
Recent developments
and considerations for certification.
Aerospace
2018
,
5
, 103.
[
CrossRef
]
75.
Tan, X.; Sun, Y.; Zeng, W.; Quan, Z.
Congestion Recognition of the Air Traffic Control Sector Based on Deep Active Learning.
Aerospace
2022
,
9
, 302.
[
CrossRef
]
76.
Ivanoska,
I.;
Pastorino,
L.;
Zanin,
M.
Assessing
Identifiability
in
Airport
Delay
Propagation
Roles
Through
Deep
Learning
Classification.
IEEE Access
2022
,
10
, 28520–28534.
[
CrossRef
]
77.
Boggavarapu,
R.;
Agarwal,
P.;
DH,
R.K.
Aviation
Delay
Estimation
using
Deep
Learning.
In
Proceedings
of
the
2019
4th
International Conference on Information Systems and Computer Networks (ISCON), Mathura, India, 21–22 November 2019;
pp. 689–693.
78.
Dong, Y.; Tao, J.; Zhang, Y.; Lin, W.; Ai, J.
Deep learning in aircraft design, dynamics, and control:
Review and prospects.
IEEE
Trans.
Aerosp.
Electron.
Syst.
2021
,
57
, 2346–2368.
[
CrossRef
]
79.
Jiménez-Campfens, N.; Colomer, A.; Núñez, J.; Mogollón, J.M.; Rodríguez, A.L.; Naranjo, V.
Deep Learning in Aeronautics:
Air
Traffic Trajectory Classification Based on Weather Reports.
In Proceedings of the Intelligent Data Engineering and Automated
Learning–IDEAL 2020: 21st International Conference, Guimaraes, Portugal, 4–6 November 2020; pp. 148–155.
Proceedings, Part II 21.
80.
Sangeetha, V.; Andrews, S.K.; Rajavarman, V.
Air traffic control using machine learning and artificial neural network.
J. Posit.
Sch.
Psychol.
2022
,
6
, 4737–4746.
81.
Çakıcı, M.; Okay, F.Y.; Ozdemir, S.
Real-time Aircraft Tracking System:
A Survey and A Deep Learning Based Model.
In Pro-
ceedings of the 2021 International Symposium on Networks, Computers and Communications (ISNCC), Dubai, United Arab
Emirates, 31 October–2 November 2021.
82.
Pang,
Y.;
Zhao,
X.;
Yan,
H.;
Liu,
Y.
Data-driven
trajectory
prediction
with
weather
uncertainties:
A
Bayesian
deep
learning
approach.
Transp.
Res.
Part Emerg.
Technol.
2021
,
130
, 103326.
[
CrossRef
]
83.
Di, Z.; Zhuang, X.; Hongfu, Z.
A hybrid deep neural network based on multi-time window convolutional bidirectional LSTM for
civil aircraft APU hazard identification.
Chin.
J. Aeronaut.
2022
,
35
, 344–361.
84.
Malekzadeh, T.; Abdollahzadeh, M.; Nejati, H.; Cheung, N.M.
Aircraft fuselage defect detection using deep neural networks.
arXiv
2017
, arXiv:1712.09213.
85.
Liu, Y.; Hansen, M.
Predicting aircraft trajectories:
A deep generative convolutional recurrent neural networks approach.
arXiv
2018
, arXiv:1812.11670.
86.
Xie, H.; Zhang, M.; Ge, J.; Dong, X.; Chen, H.
Learning air traffic as images:
A deep convolutional neural network for airspace
operation complexity evaluation.
Complexity
2021
,
2021
, 1–16.
[
CrossRef
]
87.
Qu,
J.;
Zhao,
T.;
Ye,
M.;
Li,
J.;
Liu,
C.
Flight
delay
prediction
using
deep
convolutional
neural
network
based
on
fusion
of
meteorological data.
Neural Process.
Lett.
2020
,
52
, 1461–1484.
[
CrossRef
]
88.
van Rooijen, S.; Ellerbroek, J.; Borst, C.; van Kampen, E.
Conformal automation for air traffic control using convolutional neural
networks.
In Proceedings of the 2019 Air Traffic Management Research and Development (ATM RD) Seminar, Vienna, Austria,
17–21 June 2019.
89.
van Rooijen, S.; Ellerbroek, J.; Borst, C.; van Kampen, E.
Toward individual-sensitive automation for air traffic control using
convolutional neural networks.
J. Air Transp.
2020
,
28
, 105–113.
[
CrossRef
]
90.
Mas-Pujol, S.; Salamí, E.; Pastor, E.
Rnn-cnn hybrid model to predict c-atc capacity regulations for en-route traffic.
Aerospace
2022
,
9
, 93.
[
CrossRef
]
91.
Zeng, K.; Zhuang, X.; Xie, Y.; Xi, Z.
Hypersonic Vehicle Trajectory Classification Using Improved CNN-LSTM Model.
In Proceed-
ings of the 2021 IEEE International Conference on Unmanned Systems (ICUS), Beijing, China, 15–17 October 2021; pp. 691–696.
92.
Yang, D.; Li, S.; Peng, Z.; Wang, P.; Wang, J.; Yang, H.
MF-CNN: Traffic flow prediction using convolutional neural network and
multi-features fusion.
IEICE Trans.
Inf.
Syst.
2019
,
102
, 1526–1536.
[
CrossRef
]
Aerospace
2023
,
10
, 358
29 of 30
93.
Jardines, A.; Eivazi, H.; Zea, E.; Simarro, J.; García-Heras, J.; Soler, M.; Otero, E.; Vinuesa, R.
Thunderstorm Prediction During
Pre-Tactical Air-Traffic-Flow Management Using Convolutional Neural Networks.
Available online:
https://papers.ssrn.com/
sol3/papers.cfm?abstract_id=4236016
(accessed on 24 March 2023).
94.
Jardines, A.; Soler, M.; Cervantes, A.; García-Heras, J.; Simarro, J.
Convection indicator for pre-tactical air traffic flow management
using neural networks.
Mach.
Learn.
Appl.
2021
,
5
, 100053.
[
CrossRef
]
95.
Lin,
Y.;
Guo,
D.;
Zhang,
J.;
Chen,
Z.;
Yang,
B.
A unified framework for multilingual speech recognition in air traffic control
systems.
IEEE Trans.
Neural Netw.
Learn.
Syst.
2020
,
32
, 3608–3620.
[
CrossRef
] [
PubMed
]
96.
Rahman, M.S.; Lapasset, L.; Mothe, J.
Aircraft Conflict Resolution Using Convolutional Neural Network on Trajectory Image.
In Proceedings of the Intelligent Systems Design and Applications:
21st International Conference on Intelligent Systems Design
and Applications (ISDA 2021), Online, 13–15 December 2021; pp. 806–815.
97.
Liu, H.; Lin, Y.; Chen, Z.; Guo, D.; Zhang, J.; Jing, H.
Research on the air traffic flow prediction using a deep learning approach.
IEEE Access
2019
,
7
, 148019–148030.
[
CrossRef
]
98.
Chen, X.;
Kopsaftopoulos, F.;
Wu, Q.;
Ren, H.;
Chang, F.K.
A self-adaptive 1D convolutional neural network for flight-state
identification.
Sensors
2019
,
19
, 275.
[
CrossRef
] [
PubMed
]
99.
Lim, J.Y.; Kim, S.; Kim, H.K.; Kim, Y.K.
Long short-term memory (LSTM)-based wind speed prediction during a typhoon for
bridge traffic control.
J. Wind.
Eng.
Ind.
Aerodyn.
2022
,
220
, 104788.
[
CrossRef
]
100.
Shi, Z.; Pan, Q.; Xu, M.
LSTM-Cubic A*-based auxiliary decision support system in air traffic management.
Neurocomputing
2020
,
391
, 167–176.
[
CrossRef
]
101.
Pang,
Y.;
Xu,
N.;
Liu,
Y.
Aircraft
trajectory
prediction
using
LSTM
neural
network
with
embedded
convolutional
layer.
In
Proceedings of the Annual Conference of the PHM Society
; PHM Society:
Scottsdale, AZ, USA, 2019; Volume 11.
102.
Mas-Pujol, S.; Salamí, E.; Pastor, E.
Predict ATFCM weather regulations using a time-distributed Recurrent Neural Network.
In Proceedings of the 2021 IEEE/AIAA 40th Digital Avionics Systems Conference (DASC), San Antonio, TX, USA, 3–7 October
2021; pp. 1–8.
103.
Huang, J.; Ding, W.
Aircraft Trajectory Prediction Based on Bayesian Optimised Temporal Convolutional Network–Bidirectional
Gated Recurrent Unit Hybrid Neural Network.
Int.
J. Aerosp.
Eng.
2022
,
2022
, 2086904.
[
CrossRef
]
104.
Shi, Z.; Xu, M.; Pan, Q.; Yan, B.; Zhang, H.
LSTM-based flight trajectory prediction.
In Proceedings of the 2018 International joint
conference on neural networks (IJCNN), Rio de Janeiro, Brazil, 8–13 July 2018; pp. 1–8.
105.
Ma, L.; Tian, S.
A hybrid CNN-LSTM model for aircraft 4D trajectory prediction.
IEEE Access
2020
,
8
, 134668–134680.
[
CrossRef
]
106.
Shu, W.; Cai, K.; Xiong, N.N.
A short-term traffic flow prediction model based on an improved gate recurrent unit neural network.
IEEE Trans.
Intell.
Transp.
Syst.
2021
,
23
, 16654–16665.
[
CrossRef
]
107.
Xu,
Z.;
Zeng,
W.;
Chu,
X.;
Cao,
P.
Multi-aircraft trajectory collaborative prediction based on social long short-term memory
network.
Aerospace
2021
,
8
, 115.
[
CrossRef
]
108.
Yan, Z.; Yang, H.; Li, F.; Lin, Y.
A deep learning approach for short-term airport traffic flow prediction.
Aerospace
2021
,
9
, 11.
[
CrossRef
]
109.
Zhao, Z.; Zeng, W.; Quan, Z.; Chen, M.; Yang, Z.
Aircraft trajectory prediction using deep long short-term memory networks.
In
Proceedings
of
the
19th
COTA
International
Conference
of
Transportation
Professionals,
Nanjing,
China,
6–8
July
2019;
pp. 124–135.
110.
Asirvadam, T.V.; Rao, S.; Balachander, T.
Predicting Air Traffic Density in an Air Traffic Control Sector.
ECS Trans.
2022
,
107
, 5037.
[
CrossRef
]
111.
Shi-Garrier, L.; Delahaye, D.; Bouaynaya, N.C.
Predicting Air Traffic Congested Areas with Long Short-Term Memory Networks.
In Proceedings of the Fourteenth USA/Europe Air Traffic Management Research and Development Seminar (ATM2021), Online,
20–23 September 2021.
112.
Jarry, G.; Delahaye, D.; Feron, E.
Approach and landing aircraft on-board parameters estimation with lstm networks.
In Pro-
ceedings of the 2020 International Conference on Artificial Intelligence and Data Analytics for Air Transportation (AIDA-AT),
Singapore, 3–4 February 2020; pp. 1–6.
113.
Wu, X.; Yang, H.; Chen, H.; Hu, Q.; Hu, H.
Long-term 4D trajectory prediction using generative adversarial networks.
Transp.
C
Res.
Part Emerg.
Technol.
2022
,
136
, 103554.
[
CrossRef
]
114.
Pang, Y.; Liu, Y.
Conditional generative adversarial networks (CGAN) for aircraft trajectory prediction considering weather
effects.
In Proceedings of the AIAA Scitech 2020 Forum, Orlando, FL, USA, 6–10 January 2020; p.
1853.
115.
Aksoy, M.; Ozdemir, O.; Guner, G.; Baspinar, B.; Koyuncu, E. Flight trajectory pattern generalization and abnormal flight detection
with generative adversarial network.
In Proceedings of the AIAA Scitech 2021 Forum, Online, 19–21 January 2021; p.
0775.
116.
Pham, D.T.; Tran, T.N.; Alam, S.; Duong, V.N.
A generative adversarial imitation learning approach for realistic aircraft taxi-speed
modeling.
IEEE Trans.
Intell.
Transp.
Syst.
2021
,
23
, 2509–2522.
[
CrossRef
]
117.
Fu, Q.; Wang, H.; Zhao, J.; Yan, X.
A maintenance-prediction method for aircraft engines using generative adversarial networks.
In Proceedings of the 2019 IEEE 5th international conference on computer and communications (ICCC), Chengdu, China, 6–9
December 2019; pp. 225–229.
118.
Rahnemoonfar, M.;
Yari, M.;
Paden, J.
Radar sensor simulation with generative adversarial network.
In Proceedings of the
IGARSS 2020-2020 IEEE International Geoscience and Remote Sensing Symposium, Waikoloa, HI, USA, 26 September–2 October
2020; pp. 7001–7004.
Aerospace
2023
,
10
, 358
30 of 30
119.
Hu, Q.; Huang, G.; Shi, H.; Lin, Y.; Guo, D.
A Short-term Aircraft Trajectory Prediction Framework Using Conditional Generative
Adversarial Network.
In Proceedings of the 2022 IEEE 4th International Conference on Civil Aviation Safety and Information
Technology (ICCASIT), Dali, China, 12–14 October 2022; pp.
433–439.
120.
Guo,
X.;
Zhu,
C.;
Yang,
J.;
Xiao,
Y.
An
Anomaly
Detection
Model
for
ADS-B
Systems
Based
on
Improved
GAN
and
LSTM
Networks.
In Proceedings of the 2021 IEEE 21st International Conference on Communication Technology (ICCT), Tianjin, China,
13–16 October 2021; pp. 802–809.
121.
Huang,
K.W.;
Chen,
G.W.;
Huang,
Z.H.;
Lee,
S.H.
IWGAN: Anomaly Detection in Airport Based on Improved Wasserstein
Generative Adversarial Network.
Appl.
Sci.
2023
,
13
, 1397.
[
CrossRef
]
122.
Zhang, Y.; Sun, H.; Zuo, J.; Wang, H.; Xu, G.; Sun, X.
Aircraft type recognition in remote sensing images based on feature learning
with conditional generative adversarial networks.
Remote Sens.
2018
,
10
, 1123.
[
CrossRef
]
123.
Olive, X.; Sun, J.; Murça, M.C.R.; Krauth, T.
A framework to evaluate aircraft trajectory generation methods.
In Proceedings of the
14th USA/Europe Air Traffic Management Research and Development Seminar, Saclay, France, 29 June–2 July 2021; pp.
17–19.
124.
Jarry, G.; Couellan, N.; Delahaye, D.
On the use of generative adversarial networks for aircraft trajectory generation and atypical
approach detection.
In
Proceedings of the Air Traffic Management and Systems IV: Selected Papers of the 6th ENRI International Workshop
on ATM/CNS (EIWAC2019) 6
; Springer:
Berlin/Heidelberg, Germany, 2021; pp. 227–243.
125.
Lang, P.; Peng, K.; Cui, J.; Yang, J.; Guo, Y.
Data augmentation for fault prediction of aircraft engine with generative adversarial
networks.
In Proceedings of the 2021 CAA Symposium on Fault Detection, Supervision, and Safety for Technical Processes
(SAFEPROCESS), Chengdu, China, 17–18 December 2021; pp. 1–5.
126.
Bastas, A.; Vouros, G.
Data-driven prediction of Air Traffic Controllers reactions to resolving conflicts.
Inf.
Sci.
2022
,
613
, 763–785.
[
CrossRef
]
127.
Olive, X.; Grignard, J.; Dubot, T.; Saint-Lot, J.
Detecting controllers’ actions in past mode S data by autoencoder-based anomaly
detection.
In Proceedings of the SID 2018, 8th SESAR Innovation Days, Salzburg, Austria, 3–7 October 2018.
128.
Wu, E.Q.; Peng, X.; Zhang, C.Z.; Lin, J.; Sheng, R.S.
Pilots’ fatigue status recognition using deep contractive autoencoder network.
IEEE Trans.
Instrum.
Meas.
2019
,
68
, 3907–3919.
[
CrossRef
]
129.
Chevrot,
A.;
Vernotte,
A.;
Legeard,
B.
CAE:
Contextual
auto-encoder
for
multivariate
time-series
anomaly
detection
in
air
transportation.
Comput.
Secur.
2022
,
116
, 102652.
[
CrossRef
]
130.
Que, Z.; Liu, Y.; Guo, C.; Niu, X.; Zhu, Y.; Luk, W.
Real-time anomaly detection for flight testing using AutoEncoder and LSTM.
In Proceedings of the 2019 international conference on field-programmable technology (ICFPT), Tianjin, China, 9–13 December
2019; pp. 379–382.
131.
Xuyun, F.; Hui, L.; Zhong, S.; Lin, L.
Aircraft engine fault detection based on grouped convolutional denoising autoencoders.
Chin.
J. Aeronaut.
2019
,
32
, 296–307.
132.
Fernández, A.; Martınez, D.; Hernández, P.; Cristóbal, S.; Schwaiger, F.; Nunez, J.M.; Ruiz, J.M.
Flight data monitoring (FDM)
unknown hazards detection during approach phase using clustering techniques and AutoEncoders.
In Proceedings of the Ninth
SESAR Innovation Days, Athens, Greece, 2–6 December 2019; pp. 2–5.
133.
Corrado, S.J.; Puranik, T.G.; Pinon-Fischer, O.J.; Mavris, D.; Rose, R.; Williams, J.; Heidary, R.
Deep autoencoder for anomaly detection
in terminal airspace operations.
In Proceedings of the AIAA Aviation 2021 Forum, Online, 2–6 August 2021; p. 2405.
134.
Chen, M.; Zeng, W.; Xu, Z.; Li, J. Delay prediction based on deep stacked autoencoder networks.
In Proceedings of the Asia-Pacific
Conference on Intelligent Medical 2018 & International Conference on Transportation and Traffic Engineering 2018,
Beijing,
China, 21–23 December 2018; pp. 238–242.
135.
Wu, Z.; Bai, Z.; Zhang, L.; Wang, K.
Feature Extraction Method Based on Sparse Autoencoder for Air Traffic Management System
Security Situation Awareness.
Secur.
Commun.
Netw.
2022
,
2022
, 3757662.
[
CrossRef
]
136.
Memarzadeh,
M.;
Matthews,
B.;
Avrekh,
I.
Unsupervised
anomaly
detection
in
flight
data
using
convolutional
variational
auto-encoder.
Aerospace
2020
,
7
, 115.
[
CrossRef
]
137.
Zeng,
W.;
Xu,
Z.;
Cai,
Z.;
Chu,
X.;
Lu,
X.
Aircraft trajectory clustering in terminal airspace based on deep autoencoder and
gaussian mixture model.
Aerospace
2021
,
8
, 266.
[
CrossRef
]
138.
Olive, X.; Basora, L.; Viry, B.; Alligier, R.
Deep trajectory clustering with autoencoders.
In Proceedings of the ICRAT 2020, 9th
International Conference for Research in Air Transportation, Zilina, Slovakia, 4–5 November 2020.
139.
Kim, H.; Lee, K.
Air traffic prediction as a video prediction problem using convolutional LSTM and autoencoder.
Aerospace
2021
,
8
, 301.
[
CrossRef
]
140.
Neto, E.C.P.; Baum, D.M.; de Almeida, J.R.; Camargo, J.B.; Cugnasca, P.S.
A trajectory evaluation platform for urban air mobility
(uam).
IEEE Trans.
Intell.
Transp.
Syst.
2021
,
23
, 9136–9145.
[
CrossRef
]
141.
Bauranov, A.; Rakas, J.
Designing airspace for urban air mobility:
A review of concepts and approaches.
Prog.
Aerosp.
Sci.
2021
,
125
, 100726.
[
CrossRef
]
Disclaimer/Publisher’s
Note:
The
statements,
opinions
and
data
contained
in
all
publications
are
solely
those
of
the
individual
author(s) and contributor(s) and not of MDPI and/or the editor(s).
MDPI and/or the editor(s) disclaim responsibility for any injury to
people or property resulting from any ideas, methods, instructions or products referred to in the content.
