You are on page 1of 12

IEEE INTERNET OF THINGS JOURNAL, VOL. 9, NO.

19, 1 OCTOBER 2022 18725

Achieve Load Balancing in Multi-UAV Edge


Computing IoT Networks: A Dynamic
Entry and Exit Mechanism
Hongzhi Guo , Member, IEEE, Xiaoyi Zhou , Graduate Student Member, IEEE,
Yutao Wang , Graduate Student Member, IEEE, and Jiajia Liu , Senior Member, IEEE

Abstract—With the gradual commercialization of 5G, espe- I. I NTRODUCTION


cially the widespread application of artificial intelligence (AI)
ITH the development of artificial intelligence
technology, the Internet of Things (IoT) continues to expand
and has integrated into every aspect of our lives. While enjoy-
ing the convenience brought by IoT, we also face unprecedented
W (AI) [1], [2] and the gradual commercialization
of 5G technologies, the scale of Internet of Thing (IoT)
challenges, including ubiquitous and unpredictable demands for networks [3], [4] continues to expand and the communication
communication and computing resources. In consideration of rate between devices is gradually increasing, which enables
their flexible deployment, low cost, and easy expansion, UAV
edge computing IoT networks (UECINs), which adopt unmanned
the application of IoT technologies in many scenarios, such
aerial vehicles (UAVs) to provide fast communication and com- as smart cities, smart factories, smart agriculture, power
puting services, have emerged as a promising solution. Note that inspection, etc. While enjoying the convenience brought
there have been a number of studies focusing on UAV’s position by IoT, we also face unprecedented challenges, including
deployment and trajectory design, resource allocation in UECIN. ubiquitous and unpredictable demands for communication
However, most existing works proposed short-term service pro-
visioning systems with a fixed number of UAVs, ignoring the
and computing resources. In consideration of unmanned aerial
problem of UAVs’ limited battery power and the possible changes vehicles’ (UAVs’) flexible deployment, low cost, and easy
of ground users’ number, locations, and resource requirements. expansion, UAV edge computing IoT networks (UECINs),
To address these issues, we present a dynamic UECIN framework which adopt UAVs to provide fast communication and
with autonomous prediction characteristics, aiming to stably pro- computing services, have emerged as a promising solution.
vide mobile-edge computing services for ground users in a certain
area over a long period of time. This framework can not only
support UAV’s dynamic entry and exit according to the real-time
needs of ground users but also update their position deployment
A. Related Works
based on the distribution of ground users. As we know, we are
the first to propose UECIN with a dynamic entry and exit mech- For application scenarios in remote areas, such as smart
anism. Besides, an efficient and load-balancing task allocation agriculture [5], power inspection [6], and forest protection [7],
scheme is further given, and extensive analysis and numerical the primary issue is the insufficient infrastructure deployment,
results corroborate the feasibility and superior performance of
our framework. which may prevent the collected data from being processed
in time, causing economic property losses. The deployment
Index Terms—Entry and exit mechanism, Internet of Things of base stations (BSs) in remote areas is not accomplished
(IoT) networks, load balancing, mobile-edge computing, multi-
UAV, neural networks. overnight [8], and operators may also consider the cost and
benefit of construction. Generally, they will not choose to
build large BSs without strategic significance in remote areas.
Therefore, the opportunity for UAVs comes [9]. Their easy-
to-control features make it easier for them to freely collect
Manuscript received 3 March 2022; accepted 20 March 2022. Date of pub- and return data, and their computing and communication
lication 23 March 2022; date of current version 23 September 2022. This work capabilities allow them to serve as aerial BSs [10], [11],
was supported in part by the National Natural Science Foundation of China
under Grant 62001393; in part by the Natural Science Basic Research Program or communication relay nodes [12], [13] to provide services
of Shaanxi under Grant 2020JC-15; in part by the Fundamental Research for ground users. Many existing works have put forward the
Funds for the Central Universities under Grant D5000210817; in part by application prospects of UAVs in remote areas. For exam-
the Xi’an Unmanned System Security and Intelligent Communications ISTC
Center; and in part by the Special Funds for Central Universities Construction ple, Maddikunta et al. [14] summarized the applications of
of World-Class Universities (Disciplines) and Special Development Guidance UAVs in smart agriculture, e.g., monitoring the status of
under Grant 0639021GH0201024. (Corresponding author: Jiajia Liu.) crops, guiding farm animals, and even pollination by UAVs.
Hongzhi Guo, Yutao Wang, and Jiajia Liu are with the National Engineering
Laboratory for Integrated Aero–Space–Ground–Ocean Big Data Application Cruz et al. [15] presented a scheme for UAV to automatically
Technology, School of Cybersecurity, Northwestern Polytechnical University, detect tropical rain forest status in real time, and its recogni-
Xi’an 710072, Shaanxi, China (e-mail: liujiajia@nwpu.edu.cn). tion algorithm processing is executed on UAV. All these works
Xiaoyi Zhou is with the School of Cyber Engineering, Xidian University,
Xi’an 710071, China. show that the application of UAVs in remote areas has great
Digital Object Identifier 10.1109/JIOT.2022.3161703 prospects.
2327-4662 
c 2022 IEEE. Personal use is permitted, but republication/redistribution requires IEEE permission.
See https://www.ieee.org/publications/rights/index.html for more information.
Authorized licensed use limited to: Bibliothèque ÉTS. Downloaded on November 05,2023 at 07:17:29 UTC from IEEE Xplore. Restrictions apply.
18726 IEEE INTERNET OF THINGS JOURNAL, VOL. 9, NO. 19, 1 OCTOBER 2022

Moreover, for urban scenarios with relatively complete is limited, and it is impossible for a single UAV to fly for
infrastructure, such as smart cities [16], smart vehicular a long time. In order to ensure the Quality of Service (QoS)
networks [17], smart factories [18], etc., their infrastructure of the UAV group, it is necessary to consider the replace-
deployment is sufficient for them to complete their usual ment of UAVs. Above all, the timing of the UAVs’ entry and
tasks. However, they also face some challenges. For example, exit is important. Joining too early results in a waste of UAV
the holding of large-scale events may lead to sudden crowd resources, and joining too late may cause the UAV to fall.
gatherings, reduced communication quality, and insufficient Moreover, for a long-running UAV group, the locations and
computing resources. Mobile devices often fail to receive sig- service requirements of ground users are not static, and thus
nals at meetings or sports events. Furthermore, the occurrence the UAV group also needs entry and exit mechanisms and
of traffic accidents may cause road congestion, and the quality location deployment strategies. The entry of UAVs and a good
of communications and task processing cannot be guaranteed. deployment strategy are the guarantee of QoS, and the exit of
In addition, the quality of Non-Line-of-Sight (NLoS) commu- UAVs is a saving of resources. Finally, how to balance the
nications between BSs and devices in urban environments is load of each UAV in a time-varying UAV group is also a
not always good. In order to solve these emergencies, UAVs challenging problem [38]. For the UAV group, load balancing
may be a good choice. They are small, flexible, as well as inex- among UAVs can not only make full use of the resources of
pensive, and they can be quickly deployed to places where it is each UAV but also prevent hardware damage caused by the
needed for computing and relay work. They can also provide long-term high-load operation of some individual UAVs.
better Line-of-Sight (LoS) communications for ground users
to make up for the deficiencies of terrestrial networks [19].
Many existing works discussed the application of UAVs in C. Contributions
urban scenarios. Zhang and Ansari [20] proposed to use UAVs In order to solve the above-mentioned long-term edge com-
as computing nodes and relay nodes to solve the problem of puting problem, we propose a dynamic multi-UAV framework
limited resources in IoT networks. Khan et al. [21] discussed in this article, which includes an autonomous prediction mech-
the potential of UAV as a relay BS in cellular communica- anism and supports UAVs’ entry and exit operations. In this
tions. Qi et al. [22] developed a future smart city architecture framework, we fully consider the energy consumption of UAV
with UAVs, in which UAVs have a significant impact on the flight, CPU operation, as well as communications, and mon-
performance of the entire system. Zhao et al. [23] studied the itor the battery power of each UAV. In particular, the main
caching UAV-assisted secure transmission for scalable videos contributions of our framework are as follows.
in hyperdense networks. 1) Through the learning of backpropagation (BP) neu-
Many facts have proved that UAVs will play an impor- ral network, the framework can predict the number of
tant role in future production and life. In the early stage of ground users in a certain period of time according to
UAV research, some researchers optimized the flight path [24], the conditions such as time as well as date and then
energy consumption [25], and computing delay [26] of a sin- update the number of UAVs in advance. Next, according
gle UAV in the IoT networks. However, the resources and to the real-time resource demands of ground users, the
service capabilities of a single UAV are often limited, and framework fine-tunes the UAV group and finally deter-
it is impossible for a single UAV to carry a lot of comput- mines the number of UAVs providing the service. The
ing, communication, and battery resources [27]. Therefore, entry and exit of UAVs can be flexibly controlled to
the simultaneous service of multiple UAVs is necessary and it provide good services for ground devices while saving
can make up for the shortcomings of insufficient resources of UAV resource consumption.
a single UAV [28]. Multi-UAV systems have higher require- 2) After the number of UAVs is determined, the UAVs
ments for collaboration between UAVs. The deployment loca- in this framework can automatically update the service
tion [29], [30], flight path [31], [32], load balancing [33], and locations according to the distribution of ground devices
resource and task allocation [34], [35] of UAVs are all current and the number of UAVs. Moreover, the UAVs monitor
research hotspots. With the development of machine learn- their own battery power in real time. When some of them
ing and reinforcement learning, many works tend to use AI run out of power, they will ask for new UAVs to replace
technologies to solve related problems [36], [37], which is the current UAVs and ensure the normal operation of
also the development direction of global technology applica- the whole system.
tions. Nevertheless, most of these works ignored the long-term 3) In this framework, the ground devices’ tasks can be
operation requirements of the UAV group. evenly distributed among the UAVs while ensuring that
they can be completed within a specified time. Load
balancing among the flying UAVs can prevent irre-
B. Motivation and Challenges versible damage such as equipment damage caused by
In our opinion, the continuous and long-term service the long-term high-load operation of some UAVs.
of the multi-UAV system is an indispensable requirement. The remainder of this article is organized as follows. We
Many previous studies have focused on energy consumption first introduce the structure of our proposed framework and
optimization, which only prolongs the using time of UAVs and give the relevant system model in Section II. After that, in
does not fundamentally solve the problem of continuous ser- Section III, the definition and formulation of the UAV load-
vice provided by UAVs. The battery power of the UAV itself balancing problem are given, and the schemes for calculating

Authorized licensed use limited to: Bibliothèque ÉTS. Downloaded on November 05,2023 at 07:17:29 UTC from IEEE Xplore. Restrictions apply.
GUO et al.: ACHIEVE LOAD BALANCING IN MULTI-UAV EDGE COMPUTING IoT NETWORKS 18727

Fig. 1. Multi-UAV entry and exit system model. ① The replacement process of a fixed UAV with insufficient battery. ② The entry process of a movable
UAV when the UAV group 2 has insufficient resources. ③ The replacement process of a movable UAV with insufficient battery. ④ The exit process of a
movable UAV when the UAV group 1 has redundant resources. ⑤ The UAV position update process when the number of UAVs changes.

the UAVs’ number, deployment location, and task allocation reaches a critical value, a new UAV will be requested to
are proposed to adapt to the requirements of ground devices. perform replacement operations. We denote pn ∈ {0, 1} as the
Then, Section IV presents extensive simulation results and working condition of UAV n, where pn = 0 means UAV n
analysis to verify the feasibility and superior performance of does not get off the ground, and pn = 1 means UAV n has
our system. Finally, Section V concludes this article. been launched to work as a BS.

II. S YSTEM M ODEL A. Communication Model


As shown in Fig. 1, we consider an IoT scenario with Regarding the communication process between UAVs and
ground devices covered by UAVs, where the UAVs serve as ground devices, the instantaneous communication distance is
aerial BSs to provide services for the ground devices. In this a factor that has to be considered. The Euclidean distance
figure, we show the possible actions of UAV groups in a time between UAV n and ground device m is given as

slot in different situations. In this system, we divide time into  2  2  2
dm,n = xnUAV − xm,t + yUAV − ym,t + zUAV . (1)
T time slots, and each time slot t ∈ T  {1, 2, . . . , T}. For n n
each ground device m ∈ M  {1, 2, . . . , M}, the location We can get the average path loss between the UAVs
avg
and task arrival information of ground device is defined as and ground devices Lm,n = PLos m,n Lm,n + (1 − Pm,n )Lm,n ,
Los Los NLos

Qm,t = {Lm,t , Dm,t }. Lm,t = {xm,t , ym,t , 0} represents the 2-D Los and LNLos are the path loss of LoS and NLoS links,
Lm,n m,n
location information of ground device m at time t, where xm,t respectively [39], which can be expressed by the following
and ym,t define the location information in the x-axis and y-axis formulas:
directions, respectively. Dm,t = {Dm,0 , Dm,1 , . . . , Dm,T } repre- 4π fc dm,n
sents the task arrival information of ground device m, where
Los
Lm,n = 20 lg + Los (2)
c
Dm,t presents the task situation of ground device m in time and
slot t. Moreover, Dm,t = {λm,t , dm,t , cm,t }, where dm,t is the 4π fc dm,n
size of task, cm,t denotes the computing resources required by NLos
Lm,n = 20 lg + NLos (3)
c
this task, and λm,t ∈ {0, 1} defines whether the ground device
m has a task to be processed by UAVs at time slot t. λm,t = 0 where c is the speed of light, fc means the system carrier
means the task does not need UAVs to cope with, and λm,t = 1 frequency, and Los and NLos represent the average addi-
means the task needs UAVs for processing. tional loss, which depend on the environment of the LoS and
For the set of UAVs N  {1, 2, . . . , N}, NLoS links, respectively. Moreover, PLos m,n is the probability of
each UAV n has the following attributes Pn = the LoS link between the UAV n and the ground device m [40],
comp
{pn , pcomm , p
energy UAV UAV
, L , c , p , ptask }, where p
comp which can be modeled as
n n n n n n n
and pcomm are separately the computing power and the 1
n m,n =
PLos      . (4)
communication power, and pn
energy
denotes the UAV’s initial 1 + a · exp −b · arctan zUAV n /dm,n − a
battery power (full battery power). LnUAV = {xnUAV , yUAV n , zUAV
n } a and b in this formula are environment-related parame-
represents the 3-D location information of UAV n, ptask n is ters, which we set as 4.88 and 0.43, respectively. Next, the
the task processing queue on the UAV n, and cUAV n denotes instantaneous communication rate [41] can be simulated as
the current battery power of UAV n, i.e., real-time battery  avg 
power monitoring. It should be noted that when the battery pm · 10−Lm,n /10
rm,n = B log2 1 + . (5)
power required for task processing and the landing of UAV σ2

Authorized licensed use limited to: Bibliothèque ÉTS. Downloaded on November 05,2023 at 07:17:29 UTC from IEEE Xplore. Restrictions apply.
18728 IEEE INTERNET OF THINGS JOURNAL, VOL. 9, NO. 19, 1 OCTOBER 2022

Here, B denotes the channel bandwidth, pm is the communica- Second, the UAV straight flight energy consumption with
tion transmit power of the ground devices, and σ 2 represents fixed height is shown as (11), at the bottom of the page, where
the Gaussian white noise power. c1 = [3/( 2 R2 )], c2 = (ρS/2W), c3 = mu /W, c4 = (ρA/W),
In this way, we can get the communication time when the and c5 = (1/2)d0 ρsA. S is the equivalent plate area of the
ground device m offloads the task to UAV n in time slot t fuselage in m2 , d0 is the fuselage drag ratio, and mu is the
dm,n body mass of UAV in kg.
comm
tm,n,t = . (6) Finally, the UAV vertical climb and descent energy con-
rm,n
sumption can be, respectively, calculated by
Then, we can get the energy consumption of communication  
comm
Em,n,t = tm,n,t
comm
· pcomm . (7) Enup tninitial , tnfinal
n
tnfinal Wv(τ ) − ma(τ )v(τ ) W − ma(τ )
B. Task Computing Model = Q2 + +
tninitial 2 2
According to the above settings of UAVs and ground 
devices, the computing time of the ground device m in time × v(τ )2 + 2(W − ma(τ ))ρA dτ (12)
slot t can be obtained
comp cm,t and
tm,n,t = comp . (8)
pn  
Endown tninitial , tnfinal
The computational energy consumption of UAVs can be
modeled by the following formula [42]: tnfinal Wv(τ ) + ma(τ )v(τ ) W + ma(τ )
 2 = Q2 + +
comp 2 2
Em,n,t = γ cm,t dm,n pcomp
n (9) tninitial

where γ is the related coefficient to the UAVs’ chip × v(τ )2 + 2(W + ma(τ ))ρA dτ. (13)
architecture.

C. Flying Model III. P ROBLEM F ORMULATION AND S OLUTIONS


When considering the flight model, we assume that the In this section, we first formulate the problem of minimizing
UAVs are not affected by external environmental factors and the number of UAVs and the problem of load balancing among
only consider UAVs’ weight problem. We divide the flight sta- UAVs, and then we present our UAV management solutions.
tus of UAV into four types, namely, hovering status, vertical
climb status, vertical descent status, and horizontal flying sta- A. Problem Formulation
tus. In addition, we set the UAVs’ acceleration to a and the
Although more UAVs can provide better services to ground
maximum flight speed to Vmax . During the flight, the UAVs
users, it is not always better. Compared with the communi-
first accelerate to the maximum speed with a constant accel-
cation and computing energy consumption of UAVs, flight
eration a, then they fly at the maximum speed Vmax , next they
energy consumption accounts for a large part of the total
decelerate with the constant acceleration −a, and they finally
energy consumption. Therefore, minimizing the number of
can reach the target positions.
UAVs is an important measure to save UAV resources. The
Then, we will discuss the flight energy consumption of
UAV exit mechanism may play an important role here. In this
UAVs [43]. First, UAV hover energy consumption can be
multi-UAV framework, we should minimize the number of
calculated by
  UAVs under the constraint of maximum allowable processing
Enhover tnhover = (Q0 + Q1 ) · tnhover (10) delay, which can be defined as follows.
√ Definition 1 [Minimizing the Number of UAVs in Multi-UAV
where Q0 = (δ/8)ρsA 3 R3 and Q1 = (1+k)[W 3/2 /( 2 2ρA)]. Network (MNUMN)]: In multi-UAV assisted edge comput-
In these formulas, W is the aircraft weight in Newton, ρ is the ing IoT networks, given the state information of the previous
air density in kg/m3 , R is UAVs’ rotor radius in meter, A is system in each time slot, such as the ground devices’ locations,
the rotor disc area in m2 , A = π R2 , b is the number of UAVs’ the information of ground devices’ tasks (the input data sizes,
blades, c is the blade chord length of UAVs, s is the rotor required CPU cycles, and the maximum permissible process-
solidity, δ is the profile drag coefficient [44], is the blade ing delay). The MNUMN problem is to minimize the number
angular velocity, and finally k is the incremental correction of UAVs while meeting the maximum permissible processing
factor to induced power. delay.

  tnfinal     
Enlf tninitial , tnfinal = Q0 1 + c1 v(τ )2 + Q1 1 + c2 v(τ )2 + c3 a(τ )v(τ )2
tninitial
 
 
· 1 + c2 v(τ )2 + c3 a(τ )v(τ )2 + c24 v(τ )4 − c4 v(τ )2 + c5 v(τ )3 dτ (11)

Authorized licensed use limited to: Bibliothèque ÉTS. Downloaded on November 05,2023 at 07:17:29 UTC from IEEE Xplore. Restrictions apply.
GUO et al.: ACHIEVE LOAD BALANCING IN MULTI-UAV EDGE COMPUTING IoT NETWORKS 18729

On the basis of the problem definition and the system model, C2 : λm,t ∈ {0, 1} ∀ m ∈ M, t ∈ T
the MNUMN problem can be formulated as C4 : tn,i <= tmax ∀ n ∈ N , i ∈ I

N C5 : Eland <= penergy
n ∀ i ∈ N. (15)
P1 : min pn (t)
{N,T} Here, len(∗) represents the queue length of ∗. As mentioned
n=1
s.t. C1 : pn ∈ {0, 1} ∀ n ∈ N before, constraints C1 and C2, respectively, constrain the status
C2 : λm,t ∈ {0, 1} ∀ m ∈ M, t ∈ T of UAVs and ground devices’ tasks. Moreover, C4 indicates
that the processing delay of all tasks on UAVs should meet
M
 t  
N
 comp  the given maximum permissible delay requirement. Constraint
C3 : cm · λm,t < tmax · pn · pn
C5 restricts the power of UAVs to ensure that UAVs can land
m=1 n=1
∀ n ∈ N , m ∈ M, t ∈ T (14) safely after the service is completed. Obviously, problem P2
is a simple convex optimization problem. Considering that the
where the function pn (t) is similar to the pn mentioned in our resources required for task processing are different, we set the
system model, and they both represent the working condition optimization objective to minimize the difference in the num-
of UAV n, except that pn (t) represents the working condition ber of UAVs’ task loads. It should be noted that the number of
of UAV in a specific time slot. pn (t) = 0 means UAV n does tasks and the number of UAVs in each time slot are variable.
not get off the ground at time slot t, and pn (t) = 1 means
UAV n has been launched to work as a BS at time slot t. B. Solutions
Constraints C1 and C2, respectively, constrain the status of
1) UAV Location Optimization: When the number of UAVs
UAVs and ground devices’ tasks. Moreover, C3 ensures that
changes, we need to change the position of UAVs to ensure
the UAV’s computing resources can meet the needs of ground
good communication coverage for the entire ground area. For
users. The result of problem P1 is expressed as a piecewise
these problems, clustering algorithms are often used to deter-
function, in which the computing resource needs of ground
mine where UAVs are deployed. Here, we use fuzzy C-means
users are the independent variable, and the number of UAVs is
(FCMs) to determine the most optimal positions of movable
the dependent variable. The resource requirements within each
UAVs without changing the initial UAV’s (fixed UAV’s) posi-
range correspond to different changes in the number of UAVs.
tion. The initial UAV is deployed in the middle of the area
The solution of problem P1 is to guarantee the service quality
to provide users in the area with the most basic services and
of UAVs. It is necessary to first ensure that the UAVs have
communication coverage. Compared with the hard clustering
sufficient resources to complete the task within the specified
of the K-means algorithm, the soft clustering of the FCM
time, and then it makes sense to consider the problem of task
algorithm determines which cluster the current ground user
offloading problem.
belongs to by assigning membership weight and calculating
In edge computing scenarios, load balancing of edge BSs is
fuzzy relationships.
important, especially for small edge computing servers such
In the FCM algorithm, we use a large-scale path-loss
as UAVs with limited resources. Load balancing among UAVs
component [45] to replace the traditional clustering by the
can prevent individual UAVs from operating at high loads for
Euclidean distance. We set the number of UAVs other than
a long time, avoiding hardware loss and the generation of large
the initial UAV to be c, so the ground devices M will be
amounts of heat. In addition, load balancing among UAVs is
divided into c categories. And there will be c category centers
also conducive to the management of the multi-UAV networks.
C. Cn represents the deployment position of UAV n (or the nth
If each UAV has different characteristics, it is not conducive to
cluster center). First, we set an objective function as follows:
the management of the UAV group. In this article, our goal is
to balance the task load of each UAV while meeting the maxi- 
c 
M
mum latency requirements of ground devices. Specifically, this G= uηn,m τm,n (16)
problem can be defined as follows. n=1 m=1
Definition 2 [Load Balancing in Multi-UAV Networks
(LBMNs)]: Given the information of the ground devices’ and where τm,n = 1/||LnUAV − Lm ||−2
denotes the large-scale path-
UAVs’ locations, the computing resources and maximum per- loss component [46], η is a factor of fuzzy relationship, and
missible processing delay required by ground devices in each un,m represents the fuzzy relationship between device m and
time slot, and the information of UAVs’ situation (computing UAV n. FCM is a process of continuously iterative calculation
capacities, real-time battery power, and task buffer queue), the of fuzzy relationship and cluster centers until they reach the
LBMN problem is to evenly distribute ground devices’ tasks optimum. So, we update them separately according to
to UAVs and balance the task load among UAVs. M η
m=1 un,m Lm
On the basis of the problem definition and the system model, Cn =  M η (17)
the LBMN problem can be formulated as m=1 un,m

N   task  N  task  N 2 and


n=1 len p n − x=1 len p n / p
x=1 n
P2 : min N 1
{N} un,m =   . (18)
n=1 pn c 2
τm,n η−1
s.t. C1 : pn ∈ {0, 1} ∀ n ∈ N k=1 τm,k

Authorized licensed use limited to: Bibliothèque ÉTS. Downloaded on November 05,2023 at 07:17:29 UTC from IEEE Xplore. Restrictions apply.
18730 IEEE INTERNET OF THINGS JOURNAL, VOL. 9, NO. 19, 1 OCTOBER 2022

Algorithm 1 UAV Location Update Algorithm 2) Multi-UAV Autonomously Entry and Exit Solution: In
Input: UAV positions LnUAV , GD positions Lm , Number of this part, we show the specific working process of the entire
iterations TFCM , threshold κ. multi-UAV framework, so as to realize the operation of UAVs’
Output: Category centers C. entry, exit, replacement, position update, task assignment, and
1: Initialize the C according to the current position of UAVs, so on. As shown in Algorithm 2, the whole system runs in time
and calculate un,m by Eq. 18; sequence, and we only set one initial UAV to provide the most
2: G(1) = 1, G(0) = 1, t = 1; basic communication and computing services for the entire
3: while |  G(t) > κ| and t < TFCM do region. It should be noted that before the entire system runs in
4: Update category centers C with Eq. 17; time slots, we use the BP neural network to learn the changes
5: Calculate objective function G(t) using Eq. 16; in regional human flow and form a model to predict changes
6: Update all fuzzy relationships by Eq. 18; in the computing resource requirements of ground users. The
7: G(t) = G(t) − G(t − 1); reason why we use the BP neural network is that it can form a
8: t = t + 1; nonlinear result, and more elements can be added to the input
9: end while layer. In addition to the date Dp , time T p , and the number
of ground devices Nump mentioned in the algorithm, some
other elements can also be added, such as holiday situation,
Algorithm 2 Multi-UAV Network Entry and Exit Algorithm air quality, weather conditions, etc.
(MNEEA) In Algorithm 2, we set a battery power threshold EAlert of
comp
Input: All UAVs’ information Pn , n ∈ N , All ground devices’ UAVs, EAlert = Enhover (tfly ) + Encomm (tfly ) + En (tfly ) + Eland ,
information Qm , m ∈ M, Total time T , Minimum battery where tfly is the time required for the new UAV to launch to the
comp
requirements for UAVs emin , Maximum delay requirement designated position, Enhover (tfly ), Encomm (tfly ), and En (tfly ),
tmax , Previous system situation S = (Dp , T p , Nump ). respectively, represent the flight, communication, and com-
1: Launch of initialized UAV; puting energy consumption of the UAV in this time period,
2: Train the BP neural network model model = BP(S); and Eland is the energy consumption required for the land-
energy
3: for all t ∈ T do ing of UAV. When the remaining power pn of the UAV
4: Determine whether it is necessary to launch new UAVs reaches this threshold, the system starts to request a new UAV
to replace low-power UAVs by power threshold EAlert , for replacement. We ensure the continuity of UAV services
update operational UAVs’ information, and α = 0; through such settings and ensure the safety of UAVs through
5: if the number of ground users has changed then C5 constraint in P2. One more thing needs to be mentioned,
6: β = 0; we set the UAVs to fly vertically first, then they fly straight
7: end if horizontally after reaching the height h, and finally reach the
8: Predict the computing resources required by ground target position. After such settings, the flight energy consump-
users cneed = model(d, t, a); tion of the UAV becomes predictable, which facilitates the
9: Compute the current total resources of the UAVs coffer ; formulation of various decisions of the system. In our experi-
10: while cneed /coffer > tmax do mental simulation, the launch time of one UAV is in the range
11: Add a new UAV to provide services, and α = 1; of [4.898–22.684 s], and the redeployment time of the UAVs
12: end while that have been launched is in the range of [0–17.785 s]. As
13: while [cneed /(coffer − cUAV ) < tmax ] ∩ (β == 0) do the number of UAVs increases, the redeployment positions of
14: Exit the UAV with least remaining battery, and α = the UAVs are closer and closer to the original positions, and
1; the overall redeployment time of the system will be shorter.
15: end while For a long-running system, the flight time is acceptable and
16: if (Suct < 1) ∩ (α == 0) then has to be faced. Moreover, the UAVs that have not reached
17: Add a new UAV to provide services; their destination can still be used as edge BSs, and there will
18: Update α = 1 and β = 1; be no suspension of service.
19: end if In general, the flow of this multi-UAV system in each time
20: if the number of UAVs has changed α == 1 then slot is as follows. First, each UAV checks whether its own
21: Call Algorithm 1 to obtain C; battery power is sufficient to complete the remaining task pro-
22: end if cessing work and UAV landing work. If the battery power is
23: Call Algorithm 3 to assign tasks; insufficient, a new UAV will be launched for replacement.
24: Update the UAV position, battery, task information, task Second, the system predicts the required computing resources
processing success rate Suct ; and available computing resources to determine whether it is
25: end for necessary to join or exit UAVs, so as to save UAV resources
while ensuring the QoSs. Then, the system again determines
whether additional UAVs are needed according to the suc-
cess rate of previous task processing. Such operations can
As shown in Algorithm 1, the FCM algorithm can reach avoid large-scale takeoffs and landings of UAVs and save more
convergence through multiple iterations of Cn and un,m , and resources. Next, if the number of UAVs changes, Algorithm 1
the optimal position of the UAV can be determined. is used to update the best service positions of UAVs. Moreover,

Authorized licensed use limited to: Bibliothèque ÉTS. Downloaded on November 05,2023 at 07:17:29 UTC from IEEE Xplore. Restrictions apply.
GUO et al.: ACHIEVE LOAD BALANCING IN MULTI-UAV EDGE COMPUTING IoT NETWORKS 18731

Algorithm 3 Approximate Load-Balancing Task Allocation C. Analysis of Algorithms


Algorithm
In this part, we analyze the computational complexity of
Input: All UAVs’ information Pn , n ∈ N , All ground devices’ the above algorithms. First, the computational complexity of
information Qm , m ∈ M, Time slot t. Algorithm 1 is O(M · N). In this algorithm, the complexity of
Output: Updated UAV information Pn , n ∈ N . updating category centers in step 4 can be obtained as O(M·N)
1: for all m ∈ M do
by (17), and the complexity of calculating objective function
2: Pre-allocate the tasks to the UAV in the cluster center; in step 5 can be obtained as O(M · N) according to (16).
3: Update operational UAVs’ information; Step 6 calculates all fuzzy relationships between ground users
4: end for
and UAVs, so its complexity is also O(M · N). Based on the
5: Calculate the average number of tasks that each UAV
above analysis, we can get that the complexity of Algorithm
should get av and initialize K; 1 is O(M · N).
6: for all n ∈ N do
Next, for Algorithm 2, its computational complexity is O(T ·
7: n ) > av then
if len(ptask bsize ), where bsize is the number of neurons in the hidden layer
8: UAV n only keeps the first av tasks, and the remain- in a three-layer neural network. In this algorithm, we ignore
ing tasks are stored in K; the influence of Algorithms 1 and 3 on the complexity of
9: else this algorithm. And we can see that the algorithm only loops
10: UAV n only keeps all tasks; between steps 10–12 and 13–15, but the purpose of these two
11: end if loops is to determine the number of UAVs that enter or exit,
12: end for
so we consider its complexity to be O(1). For the BP neural
13: for all n ∈ N do
network in this algorithm, we need to predict the ground user
14: n ) <= av and K! = ∅ do
while len(ptask state of this time slot in step 8, so its time complexity is
15: UAV n chooses the closest task in K; O(bsize ) according to the setting of the neural network model.
16: end while Then, the complexity of Algorithm 2 should be O(T · bsize ).
17: end for
Finally, for Algorithm 3, its computational complexity
18: for all n ∈ N do
should be O(M · log2 (M/N)). Steps 1–4 preassign tasks
19: Sort the task queue in UAV n according to time with a greedy task assignment scheme, and its complexity
sequence; should be O(M). Steps 6–12 select tasks that are suitable
20: end for
for assignment, and its complexity should be O(M). Steps
13–17 allocate the filtered tasks, so the complexity should
be O(len(K)). Finally, steps 18–20 quickly sort the tasks
in each UAV, so the average time complexity should be
O(N·[M/N]·log2 (M/N)) = O(M·log2 (M/N)). To sum up, the
Algorithm 3 is used to allocate the remaining tasks of UAVs
time complexity of Algorithm 3 should be O(M · log2 (M/N)).
and the newly arrived tasks in the time slot, which can real-
ize load balancing among UAVs while guaranteeing service
quality. Finally, according to the computing, communication, IV. P ERFORMANCE E VALUATION
and flight process of each UAV, the battery power, task In this section, we conduct a series of simulation exper-
queue, and real-time location information of each UAV are iments and obtain some numerical results to evaluate the
updated. performance of our proposed solutions. They all prove that
3) Load-Balancing UAV Task Allocation: In this part, we the effectiveness of our schemes for the optimization of UAVs’
will describe our UAV task allocation solution. The determina- quantity control and task load balancing.
tion of task offloading positions has always been an important
research direction in the field of edge computing. A good A. Parameter Settings
offloading position can save more task processing time and Without loss of generality, we assume that the UAV group
energy consumption. Moreover, the research on UAVs’ load- serves a geographic area of size 500 × 500 m2 . For the tasks
balancing problem is also of great significance. Load balancing in ground devices, we assume the size of task is between 0.5
between UAVs can increase the service life of UAVs and and 1 MB, the CPU cycles required by task is between 50
prevent the occurrence of long-time high-load work of individ- and 150 Megacycles, and λm,t is random within 0 and 1. For
ual UAVs. In addition, the high-load operation of individual the UAVs, we assume that all UAVs are isomorphic, and each
UAVs may generate additional heat, resulting in a waste of UAV has the same computing, communication, and flight capa-
resources. bilities. The CPU cycle frequency and communication power
As shown in Algorithm 3, the whole process can be divided of UAV are set to 1 GHz and 1 W separately, and the ini-
into three parts. First, the UAVs screen their preallocated tasks, tial power of UAV is 500 kJ. In terms of communication, the
and the tasks before t time slot are retained in the origi- channel bandwidth B is set to 1 MHz, and the communication
nal UAV to the greatest extent. Then, the UAVs with fewer transmit power pm is set to 30 dBm. In terms of UAV flight,
tasks select the nearest tasks for processing. Finally, each UAV we assume that the UAV has a constant acceleration of 5 m/s2
sorts the tasks by time slot and prioritizes the tasks that arrive and a maximum flight speed of 30 m/s. The aircraft weight
first. in Newton is set to 20 N, the air density ρ = 1.225 kg/m3 ,

Authorized licensed use limited to: Bibliothèque ÉTS. Downloaded on November 05,2023 at 07:17:29 UTC from IEEE Xplore. Restrictions apply.
18732 IEEE INTERNET OF THINGS JOURNAL, VOL. 9, NO. 19, 1 OCTOBER 2022

Fig. 2. Illustration of UAV target locations and ground users’ clustering situation. (a) Clustering scene with 70 users and 2 UAVs. (b) Clustering scene with
70 users and 3 UAVs. (c) Clustering scene with 70 users and 4 UAVs.

TABLE I
PARAMETER S ETTINGS

Fig. 3. Comparison of the predicted number of ground users and the actual
number of ground users. The actual number is updated every 5 min, and the
figure shows changes in the actual number and predicted number at 288 time
points (24 h).

UAVs’ rotor radius R = 0.4 m, the rotor solidity s = 0.05,


profile drag coefficient δ = 0.012, the blade angular velocity Fig. 4. Convergence behaviors of BP neural network under different gradient
descent algorithms.
= 300 radians/s, and the incremental correction factor to
induced power k = 0.1. Table I lists the parameter settings in
our simulation experiments.
device clustering groups, which proves that our Algorithm 1
is effective.
B. Numerical Results Then, in Fig. 3, we compare the predicted and actual number
In Fig. 2, there are 70 ground devices randomly distributed of ground users within 24 h. The data set updates the number
in this area. With the increase of the number of UAVs, we of ground users every 5 min, so the actual number is constantly
show the process of repeated reclustering of ground devices fluctuating around the predicted result curve. From the figure,
and repeated updating of the UAVs’ target positions, which we can see that the BP neural network used in Algorithm 2
demonstrates the results of Algorithm 1 in detail. The posi- can effectively predict the number of ground users in each
tions of black five-pointed star in the figure are the target period. In addition, in Fig. 4, we further show the process
positions of UAVs, and the randomly distributed dots of vari- of minimizing the loss function of the BP neural network
ous colors are the position display of ground devices. Different under different gradient descent algorithms, i.e., stochastic
colors represent different clustering groups. We can find that gradient descent (SGD), root mean-square prop (RMSProp),
the ground devices are classified into several uniform groups and adaptive moment estimation (Adam). Adam combines the
and the UAVs’ target points are all at the centers of the ground advantages of Momentum and RMSProp algorithms, so it can

Authorized licensed use limited to: Bibliothèque ÉTS. Downloaded on November 05,2023 at 07:17:29 UTC from IEEE Xplore. Restrictions apply.
GUO et al.: ACHIEVE LOAD BALANCING IN MULTI-UAV EDGE COMPUTING IoT NETWORKS 18733

Fig. 5. Illustration of the changes in actual required ground resources and the Fig. 6. Comparison of the task processing failure rates under different
number of UAVs under different strategies. The shaded green area indicates strategies. The strategy represented by the red line has four fixed UAVs.
the range of possible variations in ground resource requirements.

not only dynamically adjust the learning rate of each parame-


ter but also has a faster convergence speed. Therefore, we use
the Adam algorithm in our experiment, and it performs well
compared with other algorithms.
Fig. 5 shows how the number of UAVs changes with the
demand for ground resources in our system. Among them, the
green line represents the real-time resource demand change
of ground users, and the green shadow around it represents
the range of possible variations in ground resource require-
ments. Then, the red line in the figure represents the change
Fig. 7. Comparison of the UAVs’ task load standard deviation under different
in the minimum number of UAVs required with the change of strategies. This standard deviation represents the difference between UAVs’
ground resource requirements. The blue line in the figure rep- task loads.
resents the change in the number of UAVs under our scheme.
Through the comparison of the schemes, we can find that the
number of UAVs in our scheme is always 1–2 more than the strategy can basically meet the user resource requirements, but
number of UAVs in the enumeration method, which can meet it cannot fully meet the actual needs of ground users. Finally,
the goal of approximately minimizing the number of UAVs. the solution we proposed is based on the BP neural network
Moreover, our scheme is more stable than the enumeration and refers to the number of real-time ground users to finally
method, and the frequency of changes in the number of UAVs determine the number of UAVs at each moment. We can see
is relatively small, which can save more energy consumption that our solution can ensure that all tasks of users in this area
for UAV redeployment. It turns out that the change in the num- can be processed in time after the system is stable. It should
ber of UAVs in our solutions is indeed related to the change be noted that the success rate in our experiment is not always
in the actual required ground resources, which also proves the 100%, and the success rate may not be 100% when the UAV
feasibility and timeliness of the entry and exit mechanism of group performs entry and exit operations. What we show in
our algorithm. this figure is the task processing success rate after the UAV
Next, as shown in Fig. 6, we show the task processing fail- group is stabilized.
ure rates under different strategies. The blue line represents the After proving the performance of Algorithms 1 and 2, we
task processing failure rate with the fixed number of UAVs. need to prove the correctness of our load-balancing algorithm.
The red line represents the task processing failure rate with Fig. 7 shows the comparison of UAVs’ task load standard devi-
the help of BP neural network prediction. The black line rep- ation under different task allocation schemes. The larger the
resents the task processing failure rate of our solution. First, standard deviation, the more uneven the task distribution in
for a UAV group with a fixed number of UAVs, the change in the UAV group. We compare our scheme with the other two
its task processing failure rate has a great relationship with the schemes. For the greedy algorithm [47], it offloads the task
change in the number of ground users. We can find that the to the nearest UAV with available resources and full cover-
change of the blue line is similar to the change in the num- age. For the enumeration method, it ignores the geographical
ber of users in Fig. 3. Second, for the UAV group assisted by factors of ground devices and UAVs, and simply groups tasks
BP neural network prediction, it can complete the tasks bet- to achieve the most uniform distribution effect. Through the
ter than the UAV group with a fixed number of UAVs. The comparison of the three, we can find that our scheme and
red line often fluctuates around 100%, which is caused by the enumeration method have better task load-balancing effect,
the difference between the predicted and actual ground user but the complexity of the enumeration method is O(M N ) and
number. And it proves that the BP neural network prediction there is a big gap between the O(M · log2 (M/N)) of our

Authorized licensed use limited to: Bibliothèque ÉTS. Downloaded on November 05,2023 at 07:17:29 UTC from IEEE Xplore. Restrictions apply.
18734 IEEE INTERNET OF THINGS JOURNAL, VOL. 9, NO. 19, 1 OCTOBER 2022

Fig. 8. Comparison of the UAVs’ task load status under different solutions. Fig. 9. Comparison of UAVs’ battery power changes under different solu-
(a) Illustration of UAVs’ task load status with our load-balancing algorithm. tions. (a) Illustration of UAVs’ battery power changes with our load-balancing
(b) Illustration of UAVs’ task load status with greedy algorithm. algorithm. (b) Illustration of UAVs’ battery power changes with greedy
algorithm.

energy consumption. Among them, Fig. 9(a) shows the UAV


load-balancing algorithm. Moreover, the enumeration method battery power changes with our load-balancing algorithm. We
ignores the deployment of UAVs and ground devices, and can see that the UAVs’ energy consumption in each time
the communication problem after task assignment is also an period is similar, and the battery level is basically the same
important factor to be considered. It is not worth the extra at each time point. This is because the load-balancing algo-
communication cost for a relatively small load imbalance. rithm we proposed allows the tasks in the UAV group to be
Overall, this figure proves that our Algorithm 3 has a good evenly distributed to each UAV and the task processing energy
effect on load balancing and can achieve near-optimal task consumption of every UAV is basically the same. In addi-
load balancing. tion, Fig. 9(b) shows the UAV battery power changes with
In the previous experiments, we proved the feasibility of our the greedy algorithm. We can see that the battery power gap
algorithms in long-term operation. Below we will intercept the between the UAVs is not large at the beginning, but with the
system operation in a short period of time for analysis. We fix long-term operation, the battery power gap between UAVs is
the number of ground users at 70, and the number of UAVs getting bigger and bigger. The reason for this phenomenon
needed at this time is 5 according to the algorithms above. is the unbalanced task load between UAVs. One UAV with
Next, we will show the UAVs’ internal task load and energy a large task load may consume more power to complete the
consumption situations in a short period of time. tasks compared with other UAVs. What is more, the long-
In Fig. 8, we compare the changes of UAVs’ task load with term high-load operation of individual UAVs may generate
different algorithms. Fig. 8(a) shows the changes of UAVs’ more heat, waste more resources, and even cause damage to
task load with our load-balancing algorithm. In this exper- the UAVs’ hardware facilities. Finally, there is another point
iment, we show the load status of UAVs over a period of that needs to be mentioned is the management of the UAV
time after the number of UAVs stabilizes. We can see that group. The UAV group’s framework we proposed allows each
although the load of UAVs is different at each time, the UAV to maintain a smaller difference and makes them easier
load gap between UAVs is small and the task load between to manage. In contrast, the UAVs with greedy algorithm have
UAVs is basically in a balanced state. In contrast, Fig. 8(b) their own characteristics, which is not conducive to the unified
shows the changes of UAVs’ task load with greedy algorithm, management of the UAV group.
and the task load of each UAV is greatly different. Facts
have proved that the load-balancing algorithm we proposed
is indeed effective. V. C ONCLUSION
As shown in Fig. 9, we show the UAV battery power In this article, we proposed a load-balancing multi-UAV
changes under the two solutions without calculating the flight edge computing framework that supports a dynamic entry and

Authorized licensed use limited to: Bibliothèque ÉTS. Downloaded on November 05,2023 at 07:17:29 UTC from IEEE Xplore. Restrictions apply.
GUO et al.: ACHIEVE LOAD BALANCING IN MULTI-UAV EDGE COMPUTING IoT NETWORKS 18735

exit mechanism. The framework aims to use multiple UAVs to [14] P. K. R. Maddikunta et al., “Unmanned aerial vehicles in smart agri-
provide continuous high-quality edge computing services for culture: Applications, requirements, and challenges,” IEEE Sensors J.,
vol. 21, no. 16, pp. 17608–17619, Aug. 2021.
users within a certain area, and the framework can redeploy the [15] H. O. Cruz, M. Eckert, J. M. Meneses, and J. F. Martínez, “Precise real-
UAVs according to ground devices’ resource requirements and time detection of nonforested areas with UAVs,” IEEE Trans. Geosci.
changes in locations. After simulation experiments, we proved Remote Sens., vol. 55, no. 2, pp. 632–644, Feb. 2017.
[16] H. Kim, J. Ben-Othman, and L. Mokdad, “UDiPP: A framework for dif-
the superiority of the framework in terms of self-adaptation. ferential privacy preserving movements of unmanned aerial vehicles in
The framework can realize the automatic entry, exit, and smart cities,” IEEE Trans. Veh. Technol., vol. 68, no. 4, pp. 3933–3943,
replacement of UAVs, and save UAV resources as much as Apr. 2019.
[17] J. Liu, H. Guo, J. Xiong, N. Kato, J. Zhang, and Y. Zhang, “Smart
possible while ensuring service quality and flying safety of and resilient EV charging in SDN-enhanced vehicular edge computing
UAVs. In addition, the task allocation plan we proposed can networks,” IEEE J. Sel. Areas Commun., vol. 38, no. 1, pp. 217–228,
reasonably allocate tasks according to the urgency of tasks Jan. 2020.
[18] B. Yang, X. Cao, X. Li, Q. Zhang, and L. Qian, “Mobile-edge-
while meeting the delay requirements of ground devices. The computing-based hierarchical machine learning tasks distribution for
load is basically balanced among the UAVs, which prevents IIoT,” IEEE Internet Things J., vol. 7, no. 3, pp. 2169–2180, Mar. 2020.
irreversible damage to hardware equipment and additional [19] J. Xiong, H. Guo, J. Liu, N. Kato, and Y. Zhang, “Collaborative
energy waste caused by the high-load operation of individual computation offloading at UAV-enhanced edge,” in Proc. IEEE Global
Commun. Conf. (GLOBECOM), 2019, pp. 1–6.
UAVs. In the future, we will consider expanding the service [20] L. Zhang and N. Ansari, “Latency-aware IoT service provisioning in
scope, investing more UAVs to provide services for ground UAV-aided mobile-edge computing networks,” IEEE Internet Things J.,
users. Moreover, we will further consider the limited number vol. 7, no. 10, pp. 10573–10580, Oct. 2020.
[21] S. K. Khan et al., “UAV-aided 5G network in suburban, urban, dense
of UAVs and how to allocate the number of UAVs working at urban, and high-rise urban environments,” in Proc. IEEE 19th Int. Symp.
each time slot while ensuring the UAV group’s continuity and Netw. Comput. Appl. (NCA), 2020, pp. 1–4.
task completion rate. [22] F. Qi, X. Zhu, G. Mang, M. Kadoch, and W. Li, “UAV network and IoT
in the sky for future smart cities,” IEEE Netw., vol. 33, no. 2, pp. 96–101,
Mar./Apr. 2019.
[23] N. Zhao et al., “Caching UAV assisted secure transmission in
R EFERENCES hyper-dense networks based on interference alignment,” IEEE Trans.
Commun., vol. 66, no. 5, pp. 2281–2294, May 2018.
[1] H. Wang, M. Daneshmand, and H. Fang, “Artificial intelligence (AI) [24] T. Zhang, Y. Xu, J. Loo, D. Yang, and L. Xiao, “Joint computation and
driven wireless body area networks: Challenges and directions,” in Proc. communication design for UAV-assisted mobile edge computing in IoT,”
IEEE Int. Conf. Ind. Internet (ICII), 2019, pp. 428–429. IEEE Trans. Ind. Informat., vol. 16, no. 8, pp. 5505–5516, Aug. 2020.
[2] B. Mao, F. Tang, Y. Kawamoto, and N. Kato, “Ai models for green [25] Y. Du, K. Yang, K. Wang, G. Zhang, Y. Zhao, and D. Chen, “Joint
communications towards 6G,” IEEE Commun. Surveys Tuts., vol. 24, resources and workflow scheduling in UAV-enabled wirelessly-powered
no. 1, pp. 210–247, 1st Quart., 2022. MEC for IoT systems,” IEEE Trans. Veh. Technol., vol. 68, no. 10,
[3] C.-W. Tsai and S.-J. Liu, “An effective IoT service-to-interface assign- pp. 10187–10200, Oct. 2019.
ment algorithm via search economics,” IEEE Internet Things J., vol. 5, [26] X. Ma, C. Yin, and X. Liu, “Machine learning based joint offloading and
no. 3, pp. 1708–1718, Jun. 2018. trajectory design in UAV based MEC system for IoT devices,” in Proc.
[4] S. Verma, Y. Kawamoto, and N. Kato, “A smart Internet-wide port IEEE 6th Int. Conf. Comput. Commun. (ICCC), 2020, pp. 902–909.
scan approach for improving IoT security under dynamic WLAN [27] H. Guo, J. Li, J. Liu, N. Tian, and N. Kato, “A survey on space-air-
environments,” IEEE Internet Things J., early access, Dec. 3, 2021, ground-sea integrated network security in 6G,” IEEE Commun. Surveys
doi: 10.1109/JIOT.2021.3132389. Tuts., vol. 24, no. 1, pp. 53–87, 1st Quart., 2022.
[5] C. Lin, G. Han, X. Qi, J. Du, T. Xu, and M. Martínez-García, “Energy- [28] W. Chen, J. Liu, H. Guo, and N. Kato, “Toward robust and intelligent
optimal data collection for unmanned aerial vehicle-aided industrial drone swarm: Challenges and future directions,” IEEE Netw., vol. 34,
wireless sensor network-based agricultural monitoring system: A cluster- no. 4, pp. 278–283, Jul./Aug. 2020.
ing compressed sampling approach,” IEEE Trans. Ind. Informat., vol. 17, [29] S. Sun, G. Zhang, H. Mei, K. Wang, and K. Yang, “Optimizing multi-
no. 6, pp. 4411–4420, Jun. 2021. UAV deployment in 3-D space to minimize task completion time in uav-
[6] V. N. Nguyen, R. Jenssen, and D. Roverso, “Intelligent monitoring and enabled mobile edge computing systems,” IEEE Commun. Lett., vol. 25,
inspection of power line components powered by UAVs and deep learn- no. 2, pp. 579–583, Feb. 2021.
ing,” IEEE Power Energy Technol. Syst. J., vol. 6, no. 1, pp. 11–21, [30] N. Babu, C. B. Papadias, and P. Popovski, “Energy-efficient 3-D deploy-
Mar. 2019. ment of aerial access points in a UAV communication system,” IEEE
[7] W. Feng, C. Hu, J. Zhang, and H. Yan, “A forest fire identification Commun. Lett., vol. 24, no. 12, pp. 2883–2887, Dec. 2020.
method for unmanned aerial vehicle monitoring video images,” in Proc. [31] J. Zhang et al., “Stochastic computation offloading and trajectory
IEEE CSAA Guid. Navig. Control Conf. (CGNCC), 2018, pp. 1–6. scheduling for UAV-assisted mobile edge computing,” IEEE Internet
[8] B. Mao, F. Tang, Y. Kawamoto, and N. Kato, “Optimizing computation Things J., vol. 6, no. 2, pp. 3688–3699, Apr. 2019.
offloading in satellite-UAV-served 6G IoT: A deep learning approach,” [32] L. Wang, K. Wang, C. Pan, W. Xu, N. Aslam, and A. Nallanathan,
IEEE Netw., vol. 35, no. 4, pp. 102–108, Jul./Aug. 2021. “Deep reinforcement learning based dynamic trajectory control for UAV-
[9] C. Zhan, H. Hu, X. Sui, Z. Liu, J. Wang, and H. Wang, “Joint resource assisted mobile edge computing,” IEEE Trans. Mobile Comput., early
allocation and 3D aerial trajectory design for video streaming in UAV access, Feb. 16, 2021, doi: 10.1109/TMC.2021.3059691.
communication systems,” IEEE Trans. Circuits Syst. Video Technol., [33] L. Yang, H. Yao, J. Wang, C. Jiang, A. Benslimane, and Y. Liu, “Multi-
vol. 31, no. 8, pp. 3227–3241, Aug. 2021. UAV-enabled load-balance mobile-edge computing for IoT networks,”
[10] Z. Qin et al., “Task selection and scheduling in UAV-enabled IEEE Internet Things J., vol. 7, no. 8, pp. 6898–6908, Aug. 2020.
MEC for reconnaissance with time-varying priorities,” IEEE [34] H. Shimada, Y. Kawamoto, and N. Kato, “Novel computation and
Internet Things J., vol. 8, no. 24, pp. 17290–17307, Dec. 2021, communication resources allocation using relay communications in
doi: 10.1109/JIOT.2021.3078746. UAV-mounted cloudlet systems,” IEEE Trans. Netw. Sci. Eng., vol. 8,
[11] H. Guo, X. Zhou, J. Liu, and Y. Zhang, “Vehicular intelligence in 6G: no. 4, pp. 3140–3151, Oct.–Dec. 2021.
Networking, communications, and computing,” Veh. Commun., vol. 33, [35] R. Duan, J. Wang, C. Jiang, H. Yao, Y. Ren, and Y. Qian, “Resource allo-
Jan. 2022, Art. no. 100399. cation for multi-UAV aided IoT NOMA uplink transmission systems,”
[12] Z. Yu, Y. Gong, S. Gong, and Y. Guo, “Joint task offloading and IEEE Internet Things J., vol. 6, no. 4, pp. 7025–7037, Aug. 2019.
resource allocation in UAV-enabled mobile edge computing,” IEEE [36] W. Shi, J. Li, H. Wu, C. Zhou, N. Cheng, and X. Shen, “Drone-cell
Internet Things J., vol. 7, no. 4, pp. 3147–3159, Apr. 2020. trajectory planning and resource allocation for highly mobile networks:
[13] N. Zhao et al., “UAV-assisted emergency networks in disasters,” IEEE A hierarchical DRL approach,” IEEE Internet Things J., vol. 8, no. 12,
Wireless Commun., vol. 26, no. 1, pp. 45–51, Feb. 2019. pp. 9800–9813, Jun. 2021.

Authorized licensed use limited to: Bibliothèque ÉTS. Downloaded on November 05,2023 at 07:17:29 UTC from IEEE Xplore. Restrictions apply.
18736 IEEE INTERNET OF THINGS JOURNAL, VOL. 9, NO. 19, 1 OCTOBER 2022

[37] S. A. Latif et al., “Ai-empowered, blockchain and SDN integrated secu- Xiaoyi Zhou (Graduate Student Member, IEEE)
rity architecture for IoT network of cyber physical systems,” Comput. received the B.S. degree in cyber engineering from
Commun., vol. 181, pp. 274–283, Jan. 2022. Xidian University, Xi’an, China, in 2019, where he
[38] C.-W. Tsai, S.-W. Wang, K.-C. Hu, and M.-C. Chiang, “A modified is currently pursuing the master’s degree.
multiple-search multi-start framework-enhanced micro genetic algorithm His research interests cover mobile-edge com-
for wlan load balancing,” in Proc. 7th Int. Conf. Ubiquitous Future puting, UAV-aided edge computing, and vehicular
Netw., 2015, pp. 202–207. network.
[39] F. Gao, Y. Zhou, X. Ma, T. Yang, N. Cheng, and N. Lu, “Coverage-
maximization and energy-efficient drone small cell deployment in aerial-
ground collaborative vehicular networks,” in Proc. IEEE 4th Int. Conf.
Comput. Commun. Syst. (ICCCS), 2019, pp. 559–564.
[40] J. Wang, C. Jin, Q. Tang, N. N. Xiong, and G. Srivastava,
“Intelligent ubiquitous network accessibility for wireless-powered MEC
in UAV-assisted B5G,” IEEE Trans. Netw. Sci. Eng., vol. 8, no. 4,
pp. 2801–2813, Oct.–Dec. 2021.
[41] L. Zhang et al., “Task offloading and trajectory control for UAV-
assisted mobile edge computing using deep reinforcement learning,” Yutao Wang (Graduate Student Member, IEEE)
IEEE Access, vol. 9, pp. 53708–53719, 2021. received the B.S. degree in computer science and
[42] H. Guo and J. Liu, “UAV-enhanced intelligent offloading for Internet technology from Southwest University, Chongqing,
of Things at the edge,” IEEE Trans. Ind. Informat., vol. 16, no. 4, China, in 2020. She is currently pursuing the
pp. 2737–2746, Apr. 2020. master’s degree with the School of Cybersecurity,
[43] Z. Yang, W. Xu, and M. Shikh-Bahaei, “Energy efficient UAV commu- Northwestern Polytechnical University, Xi’an,
nication with energy harvesting,” IEEE Trans. Veh. Technol., vol. 69, China.
no. 2, pp. 1913–1927, Feb. 2020. Her research interests cover mobile-edge
[44] Y. Zeng, J. Xu, and R. Zhang, “Energy minimization for wireless com- computing and space–ground cooperative
munication with rotary-wing UAV,” IEEE Trans. Wireless Commun., computing.
vol. 18, no. 4, pp. 2329–2345, Apr. 2019.
[45] X. Chen, “Decentralized computation offloading game for mobile
cloud computing,” IEEE Trans. Parallel Distrib. Syst., vol. 26, no. 4,
pp. 974–983, Apr. 2015.
[46] F. Jiang, K. Wang, L. Dong, C. Pan, W. Xu, and K. Yang, “Deep-
learning-based joint resource scheduling algorithms for hybrid MEC
networks,” IEEE Internet Things J., vol. 7, no. 7, pp. 6252–6265,
Jul. 2020. Jiajia Liu (Senior Member, IEEE) received the B.S.
[47] A. M. Seid, G. O. Boateng, S. Anokye, T. Kwantwi, G. Sun, and G. Liu, degree in computer science from Harbin Institute of
“Collaborative computation offloading and resource allocation in multi- Technology, Harbin, China, in 2004, the M.S. degree
UAV-assisted IoT networks: A deep reinforcement learning approach,” in computer science from Xidian University, Xi’an,
IEEE Internet Things J., vol. 8, no. 15, pp. 12203–12218, Aug. 2021. China, in 2009, and the Ph.D. degree in information
sciences from Tohoku University, Sendai, Japan, in
2012.
He has been a Full Professor (Vice Dean)
with the School of Cybersecurity, Northwestern
Hongzhi Guo (Member, IEEE) received the Polytechnical University, Xi’an. He has published
B.S., M.S., and Ph.D. degrees in computer sci- more than 220 peer-reviewed papers in many high-
ence and technology from Harbin Institute of quality publications, including prestigious IEEE journals and conferences.
Technology, Harbin, China, in 2004, 2006, and His research interests cover a wide range of areas, including intelligent and
2011, respectively. connected vehicles, mobile/edge/cloud computing and storage, IoT security,
He is currently an Associate Professor with wireless and mobile ad hoc networks, and SAGIN.
the School of Cybersecurity, Northwestern Prof. Liu received the IEEE ComSoc Best YP in Academia Award in
Polytechnical University, Xi’an, China. He has 2020, the IEEE VTS Early Career Award in 2019, the IEEE ComSoc Asia–
published more than 40 peer-reviewed papers in Pacific Outstanding Young Researcher Award in 2017, and the IEEE ComSoc
many prestigious IEEE journals and conferences. Asia–Pacific Outstanding Paper Award in 2019. He has been actively join-
His research interests cover edge computing, ing the society activities, such as serving as an Associate Editor for IEEE
SAGSIN, IoT security, AI security, and 5G security. T RANSACTIONS ON W IRELESS C OMMUNICATIONS since May 2018 and an
Dr. Guo was the recipient of the WiMob Best Paper Award in 2019, and Editor for IEEE N ETWORK since July 2015 and IEEE T RANSACTIONS ON
received the IEEE TVT Top Reviewer Award in 2019. He has been actively C OGNITIVE C OMMUNICATIONS AND N ETWORKING since January 2019. He
joining the society activities, such as serving as an Associate Editor for served as an Associate Editor for IEEE T RANSACTIONS ON C OMPUTERS
IEEE T RANSACTIONS ON V EHICULAR T ECHNOLOGY since January 2021 from October 2015 to June 2017 and IEEE T RANSACTIONS ON V EHICULAR
and Frontiers in Communications and Network since January 2021, and an T ECHNOLOGY from January 2016 to December 2020. He is the Chair of IEEE
Editor for International Journal of Multimedia Intelligence and Security IoT-AHSN TC and is a Distinguished Lecturer of the IEEE Communications
since March 2019. Society and Vehicular Technology Society.

Authorized licensed use limited to: Bibliothèque ÉTS. Downloaded on November 05,2023 at 07:17:29 UTC from IEEE Xplore. Restrictions apply.

You might also like