PreprintPDF Available
Preprints and early-stage research may not have been peer reviewed yet.

Abstract and Figures

The maritime industry is experiencing a technological revolution that affects shipbuilding, operation of both seagoing and inland vessels, cargo management, and working practices in harbors. This ongoing transformation is driven by the ambition to make the ecosystem more sustainable and cost-efficient. Digitalization and automation help achieve these goals by transforming shipping and cruising into a much more cost- and energy-efficient, and decarbonized industry segment. The key enablers in these processes are always-available connectivity and content delivery services, which can not only aid shipping companies in improving their operational efficiency and reducing carbon emissions but also contribute to enhanced crew welfare and passenger experience. Due to recent advancements in integrating high-capacity and ultra-reliable terrestrial and non-terrestrial networking technologies, ubiquitous maritime connectivity is becoming a reality. To cope with the increased complexity of managing these integrated systems, this article advocates the use of artificial intelligence and machine learning-based approaches to meet the service requirements and energy efficiency targets in various maritime communications scenarios.
Content may be subject to copyright.
AI-Aided Integrated Terrestrial and Non-Terrestrial
6G Solutions for Sustainable Maritime Networking
Salwa Saafi, Olga Vikhrova, G´
abor Fodor, Jiri Hosek, and Sergey Andreev
This article has been accepted for publication in a future issue of IEEE Network, but has not been fully edited. Content may change prior to final publication.
©2022 IEEE. Personal use of this material is permitted. Permission from IEEE must be obtained for all other uses, in any current or future media, including
reprinting/republishing this material for advertising or promotional purposes, creating new collective works, for resale or redistribution to servers or lists, or
reuse of any copyrighted component of this work in other works.
Abstract—The maritime industry is experiencing a techno-
logical revolution that affects shipbuilding, operation of both
seagoing and inland vessels, cargo management, and working
practices in harbors. This ongoing transformation is driven by
the ambition to make the ecosystem more sustainable and cost-
efficient. Digitalization and automation help achieve these goals
by transforming shipping and cruising into a much more cost-
and energy-efficient, and decarbonized industry segment. The
key enablers in these processes are always-available connectivity
and content delivery services, which can not only aid ship-
ping companies in improving their operational efficiency and
reducing carbon emissions but also contribute to enhanced crew
welfare and passenger experience. Due to recent advancements
in integrating high-capacity and ultra-reliable terrestrial and
non-terrestrial networking technologies, ubiquitous maritime
connectivity is becoming a reality. To cope with the increased
complexity of managing these integrated systems, this article
advocates the use of artificial intelligence and machine learning-
based approaches to meet the service requirements and energy
efficiency targets in various maritime communications scenarios.
The maritime industry anticipates a substantial increase in
the number of operating vessels, new harbors, and routes
worldwide in response to the trade facilitation initiatives
supported by the World Trade Organization. These initiatives
aim to speed up international trade and unlock inclusive
economic development. As a fast-growing sector, this industry
is also experiencing external pressures due to environmental
concerns. Today’s greenhouse gas emissions from shipping
are estimated as 2.6% of the total global emissions, which
is the equivalent of emissions from a large country [1]. To
cope with this escalation, environmental sustainability and
digital inclusion practices become fundamental and have to
be incorporated across all maritime operations.
For sustainable industry, the maritime sector needs to go
through an extensive optimization and evolution toward fully
autonomous, globally connected, and digitalized operations
with zero-emissions [2]. The success of automation of the
maritime industry relies heavily on dynamic networking and
artificial intelligence (AI) technologies that foster the emerging
applications, such as intelligent harbors, remote on-board
maintenance, and autonomous docking. These require un-
precedentedly high data rates, large-scale connectivity between
a large number of dissimilar terminals, life-long learning and
Salwa Saafi (corresponding author, email: is with Brno
University of Technology and Tampere University. Olga Vikhrova, and Sergey
Andreev are with Tampere University, Finland. G´
abor Fodor is with KTH
Royal Institute of Technology. Jiri Hosek is with the Department of Telecom-
munications, Brno University of Technology, Czech Republic.
inference at the smart end-points, and seamless operation of
terrestrial and non-terrestrial networks.
Recent findings showed that nearly 90% of data generated
on-board never leaves the deck, which means that operators
are missing out on valuable insights and analytics for improved
logistics, cost of maintenance, and resource utilization [3].
Owing to recent advances in satellite technology, the number
of connected vessels has doubled over the last 5 years, but only
75% of vessels have on-board Internet access today. Satellite-
based backhauling remains extremely costly and inherently
limited, thus presenting serious challenges for the widespread
technology adoption to meet the growing needs of the mar-
itime industry. However, existing maritime communication
systems offer dedicated services opportunistically (e.g., in
proximity to coastal infrastructure), rather than genuinely
inter-connecting humans, vessels, and ports into a holistic
While attempting to integrate several non-terrestrial net-
works into a unified infrastructure to facilitate the demand-
ing intelligent broadband services for maritime operations,
cellular Fifth Generation (5G) systems become increasingly
convoluted and energy consuming, which risks compromis-
ing the fundamental need for sustainability [4]. By con-
trast, Sixth Generation (6G) technologies are envisaged not
only as those employing higher frequencies (e.g., millimeter-
wave (mmWave) and Terahertz bands) to achieve extreme
throughputs, but also as solutions capable of supporting AI-
aided closed-loop automation. Such systems are expected to
enable the ultimate potential of the zero-touch architecture
proposed by the European Telecommunications Standards In-
stitute (ETSI) for fully automated networks [5].
In this article, we first examine the most important of the
emerging maritime use cases, which call for a careful design
of a 6G maritime network (6G-MN) that facilitates commu-
nication and computation applications. The key components
of 6G-MNs are dynamic networking – where communication
links are created and activated on-demand – and distributed
intelligence, where learning and inference are employed at
different levels of the system. In our supportive study, we
demonstrate how machine learning (ML) can aid and outper-
form traditional model-based approaches for energy-efficient
topology management and scheduling in dynamic maritime
networks. We then identify communication- and learning-
related challenges in future AI-aided 6G-MNs.
The rest of this article is organized as follows. We first
introduce the rationale for building future maritime commu-
nication systems around cellular networks and discuss the
essential maritime use cases. We then review the challenges
arXiv:2201.06947v1 [cs.NI] 18 Jan 2022
of sustainable operation in 6G-MNs and highlight the key
role of AI for network-wide optimization in terms of topology
management and resource allocation. We also summarize the
open issues related to the integration of AI-based solutions in
6G-MNs and offer concluding remarks and future perspectives
on this work in the final section.
A. Existing Maritime Communication Systems
Targeting navigation safety and maritime environment pro-
tection, the International Maritime Organization (IMO) and
the International Telecommunication Union (ITU) coopera-
tively launched the global maritime distress and safety sys-
tem (GMDSS). The latter includes a set of terrestrial and
satellite radio technologies employed for people and vessel
rescue in distress. To further improve ship-to-ship and ship-to-
shore navigation accuracy, the IMO introduced the automatic
identification system (AIS), which complements marine radars
with tracking information for vessel collision avoidance and
better situational awareness [6].
Despite being an effective technology for navigation assis-
tance and maritime emergency services, the AIS provides low
data rate communications for the exchange of basic navigation
parameters such as speed, position, and direction [4]. This
limitation motivated the International Association of Marine
Aids to Navigation and Lighthouse Authorities (IALA) to
develop their own very high frequency data exchange system
(VDES). Building upon AIS capabilities, VDES encompasses
several communication subsystems aiming to provide higher
data rates, enhance the operating ranges by the integration
of satellite components, augment security mechanisms, and
support new maritime use cases such as e-Navigation [4]. The
concept was introduced by the IMO to harmonize maritime
navigation systems in offshore and coastal regions [6].
Although regulatory bodies continue to improve the systems
discussed above, the ongoing evolution toward a digitalized
maritime industry poses new challenges to maritime commu-
nication system design. With the increasing number of vessels,
growing level of ship autonomy, and widespread adoption
of Internet of Things (IoT) technologies, novel connectivity
solutions need to ensure cost-efficiency, scalability, and service
availability [7]. Such communication systems for modern
maritime operations have to support not only the existing e-
Navigation and GMDSS services, but also the emerging broad-
band and low-latency applications discussed in the sequel.
B. Emerging 6G-MN Use Cases
Based on the IMO and AIS specifications and our under-
standing of market trends, the emerging maritime use cases
can be grouped into six categories as illustrated in Fig. 1. This
figure also maps the use cases onto connectivity requirements
in terms of 5G service classes, namely, enhanced mobile
broadband (eMBB), ultra-reliable low-latency communica-
tions (URLLC), and massive machine-type communications
Massi ve Machin e-Type
Communication Critical
Bro adband
Vessel l ogist ics
Marit ime searc h
and r escue Harbo r logi stics
Ship borne IoT
On-bo ard
Navig ation and
fleet management
Fig. 1. Use case categories in 6G-MNs
The navigation and fleet management applications facilitate
the exchange of telematic information for enhanced situa-
tional awareness and maritime fleet management. They provide
mission-critical services for vessels of different types (i.e.,
cargo, law-enforcement, research, commercial, and leisure)
and shore-based traffic management organizations using high-
speed broadband links.
Shipborne IoT aims to improve on-board operation and
navigation by communicating vessel’s speed, fuel consump-
tion, and carbon dioxide emission information to the on-board
sensor fusion systems. Narrowband and massively deployed
sensors generate abundant machine-type raw data for sub-
sequent analysis and feature extraction. Vessel logistics is
another shipborne use case category where the crew utilizes
on-board communications for staff coordination and supply
management and in the case of internal emergencies. On top
of that, on-board infotainment provides passengers with access
to video streaming, gaming, and interactive applications. These
new shipborne use case categories comprise the requirements
of broadband, critical, and massive machine-type communica-
tions, thus making the system highly heterogeneous in terms
of traffic patterns, quality of service (QoS) requirements, and
device capabilities.
As the name implies, the use case category of maritime
search and rescue (SAR) provides medical emergency and
“Man Overboard” rescue services that entail broadband and
mission-critical applications to connect users in distress, on
vessels, and around shore-based facilities. To continue with
shore-based applications, harbor logistics offers a range of
services for planning, organization, and inspection of harbors
and industrial port operations. Ship loading/unloading coordi-
nation, asset tracking, warehouse management, short-sea, and
feeder shipping can be monitored via these services. Similarly
to vessel logistics use cases, harbor logistics may involve all
three types of communication regimes.
Solution Related concepts Relevant 6G-MN use cases
5G service classes eMBB, URLLC, mMTC Broadband, critical, and massive machine-type communications for all 6G-MN
use cases
NR NTN Satellite communication networks, unmanned
aerial systems, HAPs
Coverage extension, network access in offshore areas and NLOS scenarios
Mobile IAB Wireless backhaul, IAB-donor, IAB-nodes Capacity improvement within a vessel, coverage extension in offshore areas
MCX services MCPTT, MCData, MCVideo, off-network MCX Rescue services, response to shipborne emergencies
LTE/NR sidelink D2D communications, ProSe, UE-to-network
Ship-to-ship communications for navigation and collision avoidance, sensor
group communications for shipborne IoT, relaying for coverage extension
MEC Edge cloud servers, computation task offloading Low latency and low energy consumption in shore-based and offshore applica-
Cellular LPWA LTE-M, NB-IoT, mMTC Shipborne cellular IoT
NR RedCap Industrial sensors, surveillance cameras, wear-
Using data collected by RedCap devices in several 6G-MN use cases
5G XR AR, MR, VR XR for mission-critical maritime SAR, XR conferencing for vessel logistics
5G LAN 5G LAN-type access, enterprise network com-
On-board services including vessel logistics and infotainment
NPN Public network-integrated NPNs, standalone
Smart ports/harbors, enhanced on-board connectivity
Positioning High-accuracy positioning, RAT-dependent,
RAT-independent, hybrid solutions
Vessel location awareness for navigation and fleet management, indoor posi-
tioning services for staff management
Platooning Cooperative platoons, autonomous vessels Short-sea shipping and feeder services within harbor logistics use cases
C. Prospective Solutions for 6G-MNs
Several attempts to interconnect vessels in coastal waters
and build a bridge to the port have been successful by virtue
of cellular coverage. However, due to the limited capacity of
communication links, existing systems for maritime communi-
cations provided by, for example, Cellnex Telecom or Telenor
Maritime, fail to cover deep offshore areas and support delay-
critical and bandwidth-hungry use cases. By contrast, 5G and
beyond networks can provide a flexible and adaptive mobile
communication platform for the modernization and long-term
support of the maritime industry not only in coastal but
also in offshore areas, as confirmed by the Third Generation
Partnership Project (3GPP) in TR 22.819.
Initial 6G systems are to be primarily supported by the
existing 5G infrastructures, thus benefiting from the advance-
ment of cellular technologies that can foster the deployment of
even more agile and intelligent applications discussed above.
Enabling solutions, related cellular concepts, and relevant 6G-
MN use cases are offered in Table I.
Hybrid satellite–terrestrial networks have been utilized to
complement the high capacity of shore-based systems by the
wide-area coverage of satellite communications. In addition
to satellite access, New Radio (NR) technologies can support
other non-terrestrial access components, such as high-altitude
platforms (HAPs) as identified by 3GPP in TR 38.811. By
employing solutions for NR to support non-terrestrial network
(NTN), 6G-MNs can benefit from the 5-layer architecture
for 6G setups as proposed in [8] to extend the coverage of
terrestrial systems and provide access to maritime services in
offshore areas and non-line-of-sight (NLOS) scenarios.
In addition to NR NTN, an attractive technology for cov-
erage and capacity extension in terrestrial 5G and beyond
networks is integrated access and backhaul (IAB), which was
introduced by 3GPP TR 38.874 in Release 16. It is based on
using part of the wireless access spectrum for the backhaul
connections between remote base stations. While fixed IAB-
nodes can extend the capacity of coastal 6G-MNs, mobile IAB,
as suggested among the 3GPP Release 17 enhancements, can
be employed for dynamic backhaul solutions, thus shaping mo-
bile maritime mesh networks for both coverage and capacity
extension. This can be considered as a more flexible, scalable,
and affordable broadband access option for a wide range of
cloud services.
Further, 3GPP public safety services, including mission-
critical push-to-talk (MCPTT), mission-critical data (MC-
Data), and mission-critical video (MCVideo), can be applied
for emergency services. These services can also be provided
using the off-network mode in areas wherein cellular coverage
is temporarily unavailable or network performance is limited
in terms of capacity or latency.
NR sidelink for device-to-device (D2D) communications
and proximity services (ProSe) can help avoid collisions as
part of the navigation and fleet management use case. Sidelink-
based ship-to-ship communications allow vessels to assist
maritime coordination centers during SAR missions. User
equipment (UE)-to-network relay, as another form of D2D
wherein an indirect network connection is provided by a
relay UE, can extend the coverage of terrestrial systems for
shipborne applications in coastal waters.
In 5G networks, computation resources are moved closer
to the end devices. Heavy computation tasks and raw data
can be offloaded to proximate multi-access edge computing
(MEC) servers, thus minimizing end-to-end transmission de-
lays and energy consumption. Cellular low-power wide-area
(LPWA) technologies, namely, LTE machine-type communica-
tions (LTE-M) and narrowband IoT (NB-IoT), can be deployed
in 6G-MNs to support the use cases with mMTC requirements.
In a similar context, the need for NR devices with reduced ca-
pabilities (RedCap), such as on-board industrial sensors, video
surveillance cameras, and wearables, has been addressed by
3GPP in TR 38.875. Future 6G-MNs can benefit from the NR
RedCap-enabled services for enhanced vessel navigation and
monitoring, harbor logistics, and remote on-board assistance.
The latter example refers to one of the extended reality (XR)
use cases defined in TR 26.928, which includes augmented
reality-guided assistance in remote locations, mixed reality-
based sharing, and virtual reality-based telepresence collabo-
Non-public network (NPN) and local area network (LAN)-
type access for industrial IoT use cases has already been
considered by 3GPP for maritime scenarios in TR 22.819.
With the planned enhancements in Release 17, NPNs can
be employed in smart harbors to build private networks with
adequate QoS and security guarantees. A range of positioning
schemes, including radio access technology (RAT)-dependent
and RAT-independent solutions, are ratified in 3GPP Release
16 specifications and can be used separately or in a hybrid
manner to meet the required positioning accuracy in 6G-
MN use cases. Accurate positioning can also enable future
maritime platooning services. Aiming to improve navigation
safety and reduce fuel consumption, autonomous vessels can
move in cooperative platoons and be employed in feeder
services to manage short-sea shipping from hub ports to feeder
ports in inland waterways [9].
Recently, recognizing the high interest from maritime stake-
holders, 3GPP has officially included the work on maritime
communication (MARCOM) services over cellular systems in
its beyond Release 16 standardization efforts in TS 22.119.
An important challenge that needs to be considered in these
systems is energy efficiency as part of the sustainability goal
in maritime operations. Even though 5G NR system design
offers better bit-per-Joule energy efficiency as compared to
the previous generations of mobile technology, a typical 5G
site has nearly 70% higher energy consumption than a base
station deploying a mix of 2G, 3G, and 4G radios due to
the use of additional power-hungry components [10]. Further
cell densification together with link heterogeneity can make it
difficult to optimize 5G and beyond deployments in real time,
which calls for a more adaptive and less energy consuming
system design discussed in the following section.
A. Network Sustainability Challenges
The use cases identified in Fig. 1 are different from the
scenarios behind the operation of well-established mobile
and vehicular ad-hoc networks. For instance, they can create
more stable multi-hop network formations over longer time
spans, may need to transmit heterogeneous data over larger
communication ranges, and might also include more complex
and advanced on-board and in-port communication scenarios
similar to industrial IoT applications.
A large volume of data generated on-board the vessels in
deep offshore areas needs to be offloaded to the mainland
for efficient port operation, or distributed further within the
network to facilitate optimal maritime navigation and logistics.
Real-time data collected along the navigation routes are crucial
for autonomous shipping, as ports and vessels can access this
information upon request. This can mitigate deviations from
the optimal port operation and failures of communication links.
In contrast, infotainment and extended reality-based applica-
tions for passengers and crew members of small vessels and
large cruisers can require rapid dissemination of heavy content
from the land to numerous destinations in different parts of the
world. As these examples suggest, maritime operation targets
robust, scalable, high-performance, and adaptive mechanisms
for the orchestration of dissimilar services over dynamically
formed mobile networks.
In 6G-MNs, multi-hop communications and flexible mesh
topologies can be supported by IAB and D2D solutions.
These technologies help overcome the vulnerabilities of highly
directional and fast fading links, tolerate increased interference
levels, and utilize radio resources more efficiently. However,
due to node mobility within the unique and challenging inte-
grated maritime infrastructure, 6G-MNs need to continuously
adapt their resource allocation and scheduling policies over
dynamic topologies.
Our envisaged 6G-MN illustrated in Fig. 2 comprises a
high-performance terrestrial network segment that inherently
supports operation and maintenance of smart harbors and
industrial ports. The non-terrestrial part encompasses a dy-
namic IAB infrastructure, which ensures connectivity bridges
between vessels and the terrestrial segment. The former fa-
cilitates inland shipping and ubiquitous monitoring of near-
shore areas for improved human and marine life safety. An
important component of this system is the powerful edge and
cloud infrastructure for centralized and distributed learning to
enable proactive and resource-wise on-demand operation.
Traditional model-based approaches and optimization algo-
rithms may not be sufficient for satisfying the requirements
of the above system. They typically struggle with a lack
of timely global information about the system (e.g., channel
and buffer states, user mobility, or demand level) to provide
optimal control instructions and can thus yield impractical
computations due to excessive model dimensions. In turn,
data-driven methods are known to efficiently deal with both
model and algorithm deficits, and with learning functional
relations between different system parameters that are difficult
to model. These parameters include (i) user profile data
such as device position, mobility, transmission, and energy
consumption patterns, (ii) network configuration data encom-
passing instantaneous link capacity and resource utilization,
and node capabilities, and (iii) service data covering quality
of experience, subscription to cooperative learning of a ML
model, and capabilities for execution of offloaded tasks.
Allowing to extract valuable information from these massive
data, AI techniques can be used to predict traffic peaks and
system demands, detect anomalies or near-overload conditions,
and identify nodes or clusters with high energy and resource
consumption. This knowledge can alleviate the topology man-
agement and scheduling problems and underpin the solution
design for self-optimized and automated 6G-MNs with energy-
oriented optimization goals. In what follows, we provide ex-
amples of AI applications for a more energy-efficient maritime
system operation.
Autonomous on-board
computing and analytics
Remote management
and control
UHD surveillance for
people and animal safety
Real-time monitoring
and sensing
AI-aided dynamic IAB
Remote assistance and
critical communications
Entertainment and on-demand services
On-demand t opology m anagement
Energy-wise resource alloc ation
Sustainable AI for lifelong learning
Fig. 2. Our vision of a unified, scalable, reliable, and intelligent 6G system with integrated features for sustainable maritime operations
B. Energy-Centric Topology Management
Since both system topology and network load may evolve
over time, storage, compute, and spectrum resource alloca-
tions have to be provided on-demand and thus periodically
re-optimized. To manage on-demand topologies, joint link
activation and resource allocation problems need to be solved
repeatedly, and potentially compared with previous configura-
tions. As traditional model-based methods can be resource and
time consuming, a more agile approach is required to facilitate
the repetitive optimization problems.
As an illustrative example, we consider a low-dynamic
multi-hop wireless network deployed over 100 km2, where
access nodes are uniformly distributed across the given area
with the density of 0.5nodes per km2. Our goal is to
minimize the network energy consumption while delivering
heterogeneous traffic originating from cruisers or cargo vessels
and rerouted to different destinations (e.g., harbors or other
vessels). Therefore, we are aiming at the optimal allocation of
time, spectrum, and power along the optimal routing paths.
Given that the scale of the above optimization problem
increases with the number of nodes and potential routes,
the performance of traditional optimization approaches can
degrade dramatically. For solving the joint power allocation
and link scheduling problem in this system to minimize
its overall energy consumption, one needs to know all the
possible allocation patterns to employ linear programming
(LP). The number of patterns primarily depends on the number
of potential links, though not all of them may eventually be
used in the optimal solution [11].
Since the relation between the network flows and the set of
active links in the optimal configuration cannot be obtained
analytically, it may be learned by a data-driven method using
information about only some of the optimal configurations
(e.g., by solving optimization problem analytically for a num-
ber of system setups with fewer flows). Once the model is
trained, it can efficiently predict which links are critical for
the optimal configuration under any new traffic flows in the
system. Therefore, by using a deep neural network (DNN),
particularly its sub-category deep belief network (DBN), we
reduce the number of links involved in the routing and
scheduling decision in a given multi-hop layout, and thereby
alleviate the complexity and execution time of finding the
optimal configuration.
Hence, Fig. 3 demonstrates our system-level simulation
results for the energy efficiency of a maritime communication
system operating at 3.5 GHz with different traffic loads defined
as the maximum system capacity share. These results have
been first obtained by using the DNN-aided LP optimization
framework described in [11], and then compared to the base-
line system operation.
0.1 0.2 0.3 0.4 0.5
Maximum capacity share
Energy efficiency (Mbit/Joule)
Fig. 3. Energy-efficient resource management for IAB-aided backhaul solu-
Along these lines, other promising techniques such as long
short-term memory (LSTM) and auto-encoder can be applied
to discover the temporal and spatial correlations in traffic and
support the formulation of multi-objective optimization prob-
lems. For critical services with ultra-low latency requirements,
the challenge of routing is central. As shown in [12], a super-
vised DNN-aided traffic routing scheme causes much lower
overheads than the traditional options such as open shortest
path first (OSPF), while guaranteeing acceptable delay.
C. Energy-Efficient Scheduling
Due to large- and small-scale node mobility, as well as
interference fluctuations, the reported channel quality may
become outdated, misleading, or even lost at the network
side. Systematic inaccurate or imperfect knowledge of the
channel state may cause significant QoS and energy efficiency
degradation. Hence, mechanisms for channel quality prediction
can help combat different types of link blockage effects and
control radio interference while supporting redundant on-
demand topologies.
Deep learning (DL) methods can improve the accuracy of
channel reporting and, consequently, avoid inefficient resource
utilization in the core network and radio access parts. They are
well-suited for capturing non-linear and dynamic relationships
between the model input and output data. They also have
powerful prediction, inference, and data analysis capabilities
owing to the large amounts of data generated by the envi-
ronment and by the users. In particular, LSTM can handle
time series problems, which makes them attractive for channel
quality prediction and capable of alleviating the physical layer
imperfections [13].
The results of our system-level simulations summarized
in Fig. 4 demonstrate a significant improvement in energy
efficiency by applying LSTM for resource scheduling. We
employ an open-source interface between network simulator-
3 and Python-based AI frameworks. The latter train the
LSTM model using data generated by the simulator and then
return the data from the trained model back to the simulator
for testing. Communications over mmWave channels in a
single-cell network topology with mobile UEs (10m/s) are
assumed. Unlike in the baseline scenario (i.e., without ML),
in the LSTM-aided case the base station utilizes the predicted
channel quality information when making decisions about
scheduling and radio resource allocation [13]. Not only the
system energy efficiency can be improved with better channel
quality predictions, but also the end-to-end packet delay may
be significantly decreased.
A. From Centralized to Distributed Learning
As the adoption of AI technologies accelerates, the inte-
gration of various monitoring and control systems within a
centralized cloud can limit the scalability in such systems.
Hence, today’s predominantly cloud-centric AI solutions that
rely on training and inference in the remote cloud have to be
complemented by more energy-efficient, partially distributed,
and ultimately fully distributed learning mechanisms where
0 0.511.5 2
User density (users/m2)
Packet delay (s)
LSTM-aided scheduling
Energy efficiency (Mbit/Joule)
LSTM-aided scheduling
Fig. 4. Use of LSTM for energy-efficient real-time scheduling
numerous devices collaboratively train a part of a global
model [14].
Pervasive system intelligence is vital for the evolution of
maritime industry and its sustainable operation. In particular,
real-time decision-making vastly improves port logistics and
services. Through AI-assisted remote control, an operator can
digitally escort vessels safely to port. Smart fleet and asset
tracking features can improve load distribution in ports, which
decreases the volumes of carbon dioxide near the port areas.
XR applications for field engineers allow hands-on guidance
from offsite support teams who can follow the operator’s
on-site view. All of these require lifelong learning where
autonomous edge nodes (on-board the vessels or in dock
areas) can participate in sensor data collection, processing,
and sharing of resources for ML model training.
Conventionally, DNN algorithms are executed in the cloud
where training data are preprocessed at the edge before being
transferred to the cloud [14]. The edge/fog computing infras-
tructures are intended to accommodate the needs of multiple
DNN models that require locality and persistent training. They
also prevent the transmission of massive raw data over the
network. Federated learning is a practical training mechanism
wherein clients perform local ML training and forward their
results to an aggregator for further inference. Devices, edge
nodes, and cloud servers can be equivalently deemed as
clients. Under the risk of involving clients with poor channel
conditions or limited energy supply, ML model and client
selection remains challenging in these distributed learning-
based systems.
B. UE Capabilities in Device-Level Solutions
Mobile RedCap devices such as wearables can assume a
central role in real-time monitoring and ubiquitous sensing
of critical and highly dynamic processes in maritime envi-
ronments. Unmanned aerial vehicles help create situational
awareness for hinterland and smart fairway scenarios as well
as provide remote technical support for container handling
equipment. Enhanced with additional on-device learning and
inference capabilities, these systems can utilize real-time data
to offer deeper insights into energy-efficient maritime opera-
tions. Beam misalignment in dynamic maritime environments
may lead to significant data rate and energy efficiency degrada-
tion. DL-based proactive beam management at the device side
can help avoid this potential limitation. Whenever the line-of-
sight link is not available, reinforcement learning allows the
identification of optimal relay nodes in online fashion, even
with limited prior knowledge of the environment. However,
selecting the optimal relay nodes is non-trivial in dynamic
environments and when both in-band and out-of-band relaying
options are available.
In systems relying on device-level solutions, UE capabil-
ities are crucial when selecting suitable ML models. For
instance, if a moving IAB node is deployed on a floating
platform or on-board a vessel, it can be connected to a
source of renewable energy, while drones and wearable devices
operate on battery. Therefore, not all devices in 6G-MNs
are capable of training complex DNN due to their limited
storage, processing, or power capacity. Standalone compres-
sion techniques (such as pruning) have been optimized only
for DNN accuracy and without considering device energy
consumption. By combining multiple compression techniques,
one may derive compressed DL models with desired trade-
offs between performance and resource utilization [15]. For
instance, AdaDeep can automatically select various compres-
sion techniques to form a model according to not only device
capability constraints but also application-driven requirements.
C. Learning Delay and Network Reaction Time
In distributed learning under model or data split architecture,
the involved nodes need to periodically communicate ML
model parameters over the network. The time to synchronize
their results can grow significantly due to in-network transfer
delays. This synchronization latency can become even higher
when using ML models such as DNNs with thousands of
parameters. Although several solutions were proposed to re-
duce the DNN training times, the latter still depends on the
used data samples and approximation functions. By properly
selecting the nodes (e.g., depending on channel conditions and
energy availability) and adjusting the ML model parameters
(e.g., learning rates and number of epochs in DNNs), one can
reduce the learning delay.
Learning delay and ML convergence criteria are central
in future AI-aided 6G-MNs. In learning-aided architectures,
network reaction time can become a new key performance
indicator that tells how soon a new system configuration can
be enabled. It may be defined as the time between a parameter
change (e.g., number of active users or link quality) and
the network response time including ML (re-)training and
inference. Due to the limited communication ranges of high
frequency radios, a potentially large number of hops may be
required to connect two nodes of interest, which may cause an
increase in the network reaction time under topology changes.
The convergence of AI and 6G allows to build sustainable
AI-aided networks for maritime communications. With the
envisaged 6G-MN, the maritime industry can benefit from
the enabling effects of digitalization and virtualization in
reducing carbon dioxide emissions in ports and vessels. 6G-
MNs permit the integration of terrestrial and non-terrestrial
network segments, applications, and services in a holistic
manner to accommodate the need for large-scale, sustainable,
and on-demand system infrastructures. Due to network com-
plexity and dynamics, AI-aided solutions are indispensable for
prompt and customized reactions of the network to demand
fluctuations. In particular, deep learning techniques discussed
in this article can tackle 6G-MN optimization challenges at
different levels.
The challenges of efficient learning over 6G-MNs are
shaped by the distinctive features of the rapidly changing
maritime environment, remote operation with limited avail-
ability of energy and communication resources, and consider-
able learning delays in distributed systems. However, several
approaches discussed in this work, such as reinforcement
learning, can be further developed and employed to address
these issues. The insights offered by this article motivate
further research that can address the open questions and
challenges in intelligent 6G-MNs.
The authors gratefully acknowledge funding from Euro-
pean Union’s Horizon 2020 Research and Innovation pro-
gramme under the Marie Skłodowska-Curie grant agreement
No. 813278 (A-WEAR project). This work was also supported
by the Academy of Finland (projects Emc2-ML, RADIANT,
and IDEA-MILL). G. Fodor was partially supported by the
European Celtic project 6G-SKY with project ID C2021/1-9.
[1] International Transport Forum, “Decarbonising Maritime
Transport: Pathways to zero-carbon shipping by 2035,” 2018,
[Online]. Available:
decarbonising-maritime- transport-2035.pdf. [Accessed on: 19.01.2022].
[2] The European Council for Maritime Applied R&D, “Maritime
technology challenges 2030: New technologies and opportunities,
2021, [Online]. Available:
ecmar-brochure-maritme-technology-challenges-2030.pdf. [Accessed
on: 19.01.2022].
[3] Vodafone, “Near Shore Connectivity,” Vodafone, White Paper, 2019.
[4] Cellnex Telecom, “The digitalisation of maritime communications,
Gradiant in cooperation with Cellnex Telecom, 2019. 1st edition.
[5] ETSI, “Zero-touch network and Service Management (ZSM); Reference
Architecture,” European Telecommunications Standardization Institute,
Group Specification (GS) 002, 2019.
[6] ITU-R, “Modern maritime communications,” in ITU World Radiocom-
munication Seminar 2020 (WRS-20), 2020.
[7] IALA, “Maritime Radio Communications Plan,” International Associa-
tion of Marine Aids to Navigation and Lighthouse Authorities, White
Paper, 2017.
[8] H. Yanikomeroglu, “Integrated terrestrial/non-terrestrial 6G networks for
ubiquitous 3D super-connectivity,” in Proceedings of the 21st ACM
International Conference on Modeling, Analysis and Simulation of
Wireless and Mobile Systems, 2018, pp. 3–4.
[9] S. Thormann, A. Schirrer, and S. Jakubek, “Safe and Efficient Co-
operative Platooning,IEEE Transactions on Intelligent Transportation
Systems, pp. 1–13, 2020.
[10] O. Shurdi, L. Ruci, A. Biberaj, and G. Mesi, “5G Energy Efficiency
Overview,European Scientific Journal, ESJ, vol. 17, no. 1, p. 1, 2021.
[11] L. Liu, B. Yin, S. Zhang, X. Cao, and Y. Cheng, “Deep Learning
Meets Wireless Network Optimization: Identify Critical Links,IEEE
Transactions on Network Science and Engineering, vol. 7, no. 1, pp.
167–180, 2020.
[12] N. Kato, Z. M. Fadlullah, B. Mao, F. Tang, O. Akashi, T. Inoue, and
K. Mizutani, “The Deep Learning Vision for Heterogeneous Network
Traffic Control: Proposal, Challenges, and Future Perspective,” IEEE
Wireless Communications, vol. 24, no. 3, pp. 146–153, 2017.
[13] H. Yin, X. Guo, P. Liu, X. Hei, and Y. Gao, “Predicting Channel Quality
Indicators for 5G Downlink Scheduling in a Deep Learning Approach,”
arXiv, preprint 2008.01000, 2020.
[14] S. Hosseinalipour, C. G. Brinton, V. Aggarwal, H. Dai, and M. Chiang,
“From Federated to Fog Learning: Distributed Machine Learning over
Heterogeneous Wireless Networks,IEEE Communications Magazine,
vol. 58, no. 12, pp. 41–47, 2020.
[15] X. Wang, Y. Han, V. C. M. Leung, D. Niyato, X. Yan, and X. Chen,
“Convergence of Edge Computing and Deep Learning: A Comprehen-
sive Survey,” IEEE Communications Surveys Tutorials, vol. 22, no. 2,
pp. 869–904, 2020.
SALWA SAAFI ( is a Ph.D. student at the Department of Telecom-
munications at Brno University of Technology, Czech Republic and the Unit
of Electrical Engineering at Tampere University, Finland. She received her
engineering degree (2017) in telecommunications from the Higher School of
Communication of Tunis, Tunisia. Her research interests include cellular radio
access technologies, future wireless architectures, and wearable applications.
OLG A VIKH ROVA ( is a researcher at Tampere Uni-
versity, Finland. She received her Ph.D. (2021) in Information Engineering
from University Mediterranea of Reggio Calabria, Italy, and M.Sc. (2014)
in Information and Computer Science from Peoples’ Friendship University
of Russia (RUDN University), Russia. Her current research interests include
distributed edge learning and computing, integrated access and backhaul
networks, convergence of terrestrial and non-terrestrial networks.
GAB OR FOD OR ( received his Ph.D. degree in electrical
engineering from the Budapest University of Technology and Economics in
1998, his Docent degree from KTH Royal Institute of Technology, Stockholm,
Sweden, in 2019, and his D.Sc. degree from the Hungarian Academy of Sci-
ences in 2019. He is currently a master researcher with Ericsson Research and
an adjunct professor with KTH Royal Institute of Technology. He was a co-
recipient of the IEEE Communications Society Stephen O. Rice Prize in 2018.
He is serving as the Chair of the IEEE Communications Society Full Duplex
Emerging Technologies Initiative and as an Editor for IEEE Transactions on
Wireless Communications and IEEE Wireless Communications.
JIRI HOS EK ( received his M.S. and Ph.D. degrees in electrical
engineering from the Faculty of Electrical Engineering and Communication
at Brno University of Technology (BUT), Czech Republic, in 2007 and 2011,
respectively. He is currently an Associate Professor (2016) and Deputy Vice-
Head for R&D and International Relations at the Department of Telecom-
munications (2018), BUT. Jiri is also coordinating the WISLAB research
group (2012), where he deals mostly with industry-oriented projects in the
area of IoT and home automation services. Jiri (co-)authored more than 130
research works on networking technologies, wireless communications, quality
of service, quality of experience, and IoT applications.
SER GEY AN DR EEV ( is an associate professor of
communications engineering and Academy Research Fellow at Tampere
University, Finland. He has been a Visiting Senior Research Fellow with
King’s College London, UK (2018-20) and a Visiting Postdoc with University
of California, Los Angeles, US (2016-17). He received his Ph.D. (2012) from
TUT as well as his Specialist (2006), Cand.Sc. (2009), and Dr.Habil. (2019)
degrees from SUAI. He served as lead series editor of the IoT Series (2018-
21) for IEEE Communications Magazine and as editor for IEEE Wireless
Communications Letters (2017-19).
ResearchGate has not been able to resolve any citations for this publication.
Cooperative automotive platooning can improve safety and efficiency on the road. Look-ahead control of an entire platoon allows to reduce fuel consumption and travel time in open road scenarios, but dense traffic requires continuous adaptation of far-sighted plans. To achieve efficient individual vehicle control, these control systems need to be informed appropriately. For this purpose a novel concept for distributed model predictive control of the platoon vehicles is proposed which safely allows dense spacing and keeps communication requirements small while being robust against communication loss. A safety-extension separates safety constraints from the design of the tracking control goals and enables agreed-upon behavior in terms of temporarily limited decelerations. Driving corridors based on position errors are utilized to select suitable control modes or trigger prediction updates to following vehicles. Realistic vehicle dynamics co-simulations demonstrate the platoon safety and performance in selected scenarios, including emergency braking and maneuver tracking subject to traffic disturbances. The proposed measures are effective with realistic model errors, provide implicit collision safety and show string stability with low communication requirements.
Machine learning (ML) tasks are becoming ubiquitous in today's network applications. Federated learning has emerged recently as a technique for training ML models at the network edge by leveraging processing capabilities across the nodes that collect the data. There are several challenges with employing conventional federated learning in contemporary networks, due to the significant heterogeneity in compute and communication capabilities that exist across devices. To address this, we advocate a new learning paradigm called fog learning, which will intelligently distribute ML model training across the continuum of nodes from edge devices to cloud servers. Fog learning enhances federated learning along three major dimensions: network, heterogeneity, and proximity. It considers a multi-layer hybrid learning framework consisting of heterogeneous devices with various proximities. It accounts for the topology structures of the local networks among the heterogeneous nodes at each network layer, orchestrating them for collaborative/cooperative learning through device-to-device communications. This migrates from star network topologies used for parameter transfers in federated learning to more distributed topologies at scale. We discuss several open research directions toward realizing fog learning.
Ubiquitous sensors and smart devices from factories and communities are generating massive amounts of data, and ever-increasing computing power is driving the core of computation and services from the cloud to the edge of the network. As an important enabler broadly changing people’s lives, from face recognition to ambitious smart factories and cities, developments of artificial intelligence (especially deep learning, DL) based applications and services are thriving. However, due to efficiency and latency issues, the current cloud computing service architecture hinders the vision of “providing artificial intelligence for every person and every organization at everywhere”. Thus, unleashing DL services using resources at the network edge near the data sources has emerged as a desirable solution. Therefore, edge intelligence, aiming to facilitate the deployment of DL services by edge computing, has received significant attention. In addition, DL, as the representative technique of artificial intelligence, can be integrated into edge computing frameworks to build intelligent edge for dynamic, adaptive edge maintenance and management. With regard to mutually beneficial edge intelligence and intelligent edge, this paper introduces and discusses: 1) the application scenarios of both; 2) the practical implementation methods and enabling technologies, namely DL training and inference in the customized edge computing framework; 3) challenges and future trends of more pervasive and fine-grained intelligence. We believe that by consolidating information scattered across the communication, networking, and DL areas, this survey can help readers to understand the connections between enabling technologies while promoting further discussions on the fusion of edge intelligence and intelligent edge, i.e., Edge DL.
Conference Paper
Since the development of the 4G LTE standards around 2010, the research communities both in academia and industry have been brainstorming to predict the use cases and scenarios of 2020s, to determine the corresponding technical requirements, and to develop the enabling technologies, protocols, and network architectures towards the next-generation (5G) wireless standardization. This exploratory phase is winding down as the 5G standards are currently being developed with a scheduled completion date of late-2019; the 5G wireless networks are expected to be deployed globally throughout 2020s. As such, it is time to reinitiate a similar brainstorming endeavour followed by the technical groundwork towards the subsequent generation (6G) wireless networks of 2030s. One reasonable starting point in this new 6G discussion is to reflect on the possible shortcomings of the 5G networks to-bedeployed. 5G promises to provide connectivity for a broad range of use-cases in a variety of vertical industries; after all, this rich set of scenarios is indeed what distinguishes 5G from the previous four generations. Many of the envisioned 5G use-cases require challenging target values for one or more of the key QoS elements, such as high rate, high reliability, low latency, and high energy efficiency; we refer to the presence of such demanding links as the super-connectivity.
With the superior capability of discovering intricate structure of large data sets, deep learning has been widely applied in various areas including wireless networking. While existing deep learning applications mainly focus on data analysis, the role it can play on fundamental research issues in wireless networks is yet to be explored. With the proliferation of wireless networking infrastructure and mobile applications, wireless network optimization has seen a tremendous increase in problem size and complexity, calling for a paradigm for efficient computation. This paper presents a pioneering study on how to exploit deep learning for significant performance gain in wireless network optimization. Analysis on the flow constrained optimization problems suggests the possibility that a smaller-sized problem can be solved while sharing equally optimal solutions with the original problem, by excluding the potentially unused links from the problem formulation. To this end, we design a deep learning framework to find the latent relationship between flow information and link usage by learning from past computation experience. Numerical results demonstrate that the proposed method is capable of identifying critical links and can reduce computation cost by up to 50% without affecting optimality, thus greatly improve the efficiency of solving network optimization problems.
Recently, deep learning, an emerging machine learning technique, is garnering a lot of research attention in several computer science areas. However, to the best of our knowledge, its application to improve heterogeneous network traffic control (which is an important and challenging area by its own merit) has yet to appear because of the difficult challenge in characterizing the appropriate input and output patterns for a deep learning system to correctly reflect the highly dynamic nature of large-scale heterogeneous networks. In this vein, in this article, we propose appropriate input and output characterizations of heterogeneous network traffic and propose a supervised deep neural network system. We describe how our proposed system works and how it differs from traditional neural networks. Also, preliminary results are reported that demonstrate the encouraging performance of our proposed deep learning system compared to a benchmark routing strategy (Open Shortest Path First (OSPF)) in terms of significantly better signaling overhead, throughput, and delay.
This work is devoted to the analysis of the technical state of maritime radio electronic equipment. It is pointed out that significant changes occurred in traditional telecommunication systems. In particular, automated and adaptive telecommunication systems, highly effective interference-suppression codes, digital selective call systems are used now. Based on these systems, the authors formulated general principles for developing a basically new system, called the global sea telecommunication system to provide the safe navigation. It is pointed out that it is necessary to pay much attention to the technical servicing of all systems and, also, it is judicious to introduce the post of a captain assistant on radioelectronics.
Decarbonising Maritime Transport: Pathways to zero-carbon shipping by 2035
  • International Transport Forum
International Transport Forum, "Decarbonising Maritime Transport: Pathways to zero-carbon shipping by 2035," 2018, [Online]. Available: decarbonising-maritime-transport-2035.pdf. [Accessed on: 19.01.2022].
Near Shore Connectivity
  • Vodafone
Vodafone, "Near Shore Connectivity," Vodafone, White Paper, 2019.