Computers and Operations Research 116 (2020) 104873

Contents lists available at ScienceDirect

Computers and Operations Research

journal homepage: www.elsevier.com/locate/cor

Robust scheduling for target tracking using wireless sensor networks

Florian Delavernhe a,∗, Charly Lersteau b, André Rossi c, Marc Sevaux d
a
Université d’Angers, LERIA, Angers F-49045, France
b
Huazhong University of Science and Technology, State Key Laboratory of Digital Manufacturing Equipment & Technology, Wuhan 430074, China
c
Université Paris-Dauphine, LAMSADE CNRS UMR 7243, Paris F-75016, France
d
Université Bretagne Sud, Lab-STICC CNRS UMR 6285, Lorient F-56321, France

a r t i c l e i n f o a b s t r a c t

Article history: A wireless sensor network (WSN) is a group of sensors deployed in an area, with all of them working
Received 13 February 2019 on a battery and with direct communications inside the network. A fairly common situation, addressed
Revised 16 December 2019
in this work, is to monitor and record data with a WSN about vehicles (planes, terrestrial vehicles, boats,
Accepted 26 December 2019
etc) passing by an area with damaged infrastructures. In such a context, an activation schedule for the
Available online 3 January 2020
sensors ensuring a continuous coverage of all the targets is required. Furthermore, the collected data, in
Keywords: order to be treated, have to be transmitted to a base station in the area, near the sensors. In this work,
Linear programming the future monitoring missions of the network are also taken into account, as well as the energy con-
Sensor network sumption of the current mission. We also consider that the spatial trajectories of the targets are known,
Robust optimization whereas the speed of the targets along their trajectories are estimated, and subject to uncertainty. Hence,
Target tracking the main objective is to seek solutions that can withstand earliness and tardiness from the previsions.
We propose a formulation of the problem with three different objectives and a solution method with
experiments and results. The objectives are treated in a lexicographic order as follows (i) maximize the
robustness schedule to cope with the advances and delaqui leys of the targets, (ii) maximize the mini-
mum of monitoring time we can guarantee in priority areas, (iii) maximize the amount of energy left in
the sensor batteries. We propose new upper bounds on the robustness measure, that are exploited by the
solution approach whose complexity is shown to be pseudo-polynomial. The solution approach is based
on a preprocessing step called discretisation, and the resolution of a series of linear programs.
© 2020 Elsevier Ltd. All rights reserved.

1. Introduction Allen et al. (2006), to monitor volcanoes, the sensors are equipped
with a seismometer, a microphone and a long-range radio. The
Wireless Sensor Networks (WSN) (Akyildiz et al., 2002; Yick sum of their data collected will warn when an eruption is likely
et al., 2008) is a technology becoming more and more prominent to happen. Nonetheless, there are plenty of other different appli-
in industry nowadays, with numerous applications and a bright cations for WSNs (Yick et al., 2008), using their ability to moni-
promising future (Aalsalem et al., 2018; Modieginyane et al., 2018; tor an area or track one or more targets. Many applications have
Rawat et al., 2014; Xu et al., 2018). It involves the deployment of high impact goals, where human lives are at stake. The military
a network in an infrastructure-free area. Each node of the network field is notably using the WSNs (Đurišić et al., 2012). Indeed such
is a small, cheap, easy to configure and reliable sensor, that will networks are easily deployed in enemy territory and are able to
capture data about its environment. Eventually, the purpose of the work while several sensors are discovered and destroyed by the
network is fulfilled, even with a few defective sensors, by gathering enemy. In most of the applications, the networks are deployed in a
and processing together the data collected by the nodes. They may remote area or an area dangerous to access, due to natural disaster
include a wireless communication module, diverse sensing capa- or war conflict. Therefore, with the lack of infrastructure, the bat-
bilities (humidity, temperature, light, movement and many others), teries of the sensors cannot be refilled. Consequently, the lifetime
wheels, different levels of batteries, etc. As an example in Werner- of the network (the period of time during which the network can
fully serve its purpose) is limited. This leads to the implementation
of sensor management optimization methods in order to provide
∗
Corresponding author.
efficient solutions to various problems.
E-mail addresses: fl[email protected] (F. Delavernhe),
[email protected] (C. Lersteau), [email protected] (A. Rossi),
We consider, in this work, the goal of tracking a set of tar-
[email protected] (M. Sevaux). gets by a sensor network (Liu and Liang, 2005). For instance, in

https://doi.org/10.1016/j.cor.2019.104873
0305-0548/© 2020 Elsevier Ltd. All rights reserved.
2 F. Delavernhe, C. Lersteau and A. Rossi et al. / Computers and Operations Research 116 (2020) 104873

the military case, it refers to a WSN randomly deployed in a bat- 2. Related works
tlefield, with one or more targets traversing the battlefield. Once
activated, a sensor is able to monitor the targets inside its sens- The WSN field has plenty of possible applications already ad-
ing range and record data. The sensing range is the maximum dis- dressed in the literature, with different types of wireless networks
tance at which a target can be to be monitored by a sensor. Thus, considered and different optimization problems. We outline here
the basic goal in this problem is to find an activation schedule, several problems divided in two types of WSN: mobile and static
that will alternate between active and inactive state for the sen- ones.
sors. The use of this kind of schedule will highly extend the life- On the one hand, as presented in Mohamed et al. (2017) in a
time of the network compared to a continuous activation of all the mobile WSN, sensors are still able to move after their initial de-
sensors (Benini et al., 20 0 0). Keeping in mind that the schedule ployment as they are motorized. Thus, in these problems, the lit-
has to guarantee a full and continuous monitoring of the targets erature is based mostly on coverage related objectives where the
whenever and wherever the targets are located (at every instant advantage of the mobility of the sensors is used to fill uncovered
a target, if it is under the range of at least one sensor, has to be parts or to assure a minimum level of coverage. As an example
monitored). Though, the network lifetime is limited by the sensing in Elhoseny et al. (2018) the k-coverage problem is addressed. The
ranges and the battery capacities of the sensors. aim is to cover all locations of the field with at least k sensors.
The activity of sensing a given set of targets is called a mission, In an other example (Liu and Liang, 2005) treats the θ -coverage
and to the best of our knowledge, all the contributions in opti- where a full coverage of the zone is impossible and therefore the
mization to wireless sensor networks are focused on a single mis- problem is to cover at least θ % of the area. Moreover, there are dif-
sion for the monitoring of moving target. The problem addressed ferent other objectives. In Patel et al. (2005), the authors addressed
in this paper is to decide when to activate each sensor so as to a cluster-based problem. Cluster-heads are often used in networks
maximize the ability of the resulting schedule to keep being feasi- for performing data fusion on the data collected by the sensors.
ble despite uncertainty affecting the targets speed along their tra- They can be assigned to various potential locations that cover dif-
jectory for the current mission. The secondary objective is to max- ferent sets of sensors. Since the sensors are mobile over the hori-
imize the wireless sensor network ability (i.e. the amount of time zon of time, the quality (data coverage) of a location is fluctuating.
available with the batteries) to monitor a given set of zones of in- Therefore, the network is able to relocate several times the cluster-
terest for future missions. The last objective is to minimize the to- heads to different locations, with a cost for each relocation. Conse-
tal amount of energy required by the current monitoring mission. quently, in their work, the authors proposed a column generation
In the case where targets are terrestrial vehicles moving on heuristic to find an optimal trade-off between the data coverage
roads or tracks, their geographical trajectory can be predicted quite and the relocation costs.
naturally (e.g. buses or trains). For boats or aircrafts, the trajec- On the other hand, in the static WSN, the sensors being de-
tory can be estimated at least for a short amount of time. In this ployed cannot move. A popular application of static WSN is the tar-
work, we assume that such a trajectory prediction is available, but get tracking, where different objectives can be optimized, such as
the considered uncertainty is about the speed at which the targets energy consumption, scalability, fault tolerance and tracking preci-
move along their trajectory. sion.
The main contributions of this paper are the follow- Energy consumption and network’s lifetime are critical since
ing ones. First, a more realistic model than the one in batteries are often assumed to be not refillable. There exists a
Lersteau et al. (2016) is considered for the WSN, where sen- very extensive literature about these two objectives in target track-
sor communication is taken into account in the new model. In ing. For example, there are two prominent protocols used to
most of the applications of WSN, the communication of the data save energy, LEACH (Handy et al., 2002) and HEED (Younis and
collected is mandatory and since the power consumption due Fahmy, 2004). In the case of static targets, Cardei and Du (2005) in-
to communication is much larger than the consumption due to crease the lifetime of the network by organizing the set of sen-
sensing (Anastasi et al., 2009), the new model proposed in this sors into a maximal number of disjoint subsets of sensors that
paper is much more accurate. Furthermore, the previous work cover all the targets. The authors showed that if the sets are
(Lersteau et al., 2016) is also extended to consider multiple targets activated in turn, the lifetime of the network is extended. Fur-
in the problem. As a result, we introduce a new upper bound on thermore, an efficient power management method is presented
the stability radius defined in Section 6, and the previous bounds in Campos-Nañez et al. (2008), using a game-theoretic approach to
introduced in Lersteau et al. (2016) are naturally extended to propose a distributed scheme. The network lifetime is often maxi-
include communication and multiple targets. The stability radius mized with an efficient schedule of the sensors’ activity. For exam-
(Sotskov et al., 1998) is a measure of the ability of the network ple, Castaño et al. (2014) propose a column generation approach
to remain feasible despite uncertainty, it is formally defined in to compute a schedule for maximizing the network lifetime un-
Section 3. Second, this paper proposes an extended and compre- der connectivity and coverage constraints. In Carrabs et al. (2015),
hensive approach to produce robust solutions, refined with the the network lifetime is also maximized, but they consider that
two additional criteria: the results in Lersteau et al. (2016) are each sensor is assigned to a family and each family has a cov-
extended to more than one targets, and hop-communication erage requirement. The lifetime maximisation may also consider
(communication between sensors to transfer the collected data to connectivity constraints (i.e., the data collected is transmitted
a base station) is taken into account. Finally, this approach also to a central processing node) and thus multi-role sensors. In
has energy considerations added to the robust scheduling, with these cases, the power consumption of sensors depends on their
priority areas. roles, i.e., idle, relaying or monitoring, as in Carrabs et al. (2016,
This paper is organized as follows: Section 2 presents the re- 2017); Castaño et al. (2015). Lifetime maximization in the con-
lated work, it is followed by a reminder of the original definition text of target tracking is currently a hot topic: as an example,
of the stability radius in Section 3. Section 4 defines the problem. Alibeiki et al. (2019) propose a genetic-based approach for a di-
Afterwards, we present a discretisation phase, as a preamble to rectional sensor network with adjustable range. The heuristic ap-
the solution method in Section 5 and propose an upper bound for proach finds efficient solution to monitor non-moving target while
the stability radius in Section 6; Section 7 introduces the solution maximizing the lifetime of the network. A sensor is only able to
method; Section 8 presents the experiments and their results and monitor the targets inside its sensing range, that is adjustable, in
finally, Section 9 concludes this paper. its direction of activation. The greater the sensing range, the more
 F. Delavernhe, C. Lersteau and A. Rossi et al. / Computers and Operations Research 116 (2020) 104873 3

energy is consumed, and the sensor has several working directions, events can impact the processing time of a job i with a variation of
but it can use only one of them at a time.  i such that the new processing time is equal to pi ±  i , which will
The criterion of scalability as presented in Kung and also affect the completion time. A schedule has a stability radius of
Vlah (2003); Naderan et al. (2012) is also an important one. With ϱ if and only if it is always the optimal schedule for all the vectors
this objective, a WSN protocol should scale well to large numbers of processing time p ∈ Oϱ (p), with Oϱ (p) the closed ball centered
of sensors or targets, since a dense network or an important on p = { p1 , p2 , . . . , p|Q | } the original processing times and with a
number of targets can significantly increase the communication radius equals to ϱ. It means that the schedule remains feasible if
consumption. Scalable protocols typically use cluster-based or for each job i,  i ≤ ϱ.
distributed approaches instead of centralized ones. The tracking After introducing the problem in the next section, the original
precision criterion optimizes how far the estimated locations or definition of the stability radius is adapted to the target tracking
passing times of the targets can be derived from the real locations problem in Section 5.
or times (Lersteau et al., 2016; Naderan et al., 2012). Literature
covers the probability to lose a target, the recovery process, 4. Problem definition
robustness against delays and advances or trust region for the
location, noise sensitivity, etc. For the fault detection criterion, the The problem of multiple targets tracking by a wireless sensor
algorithms detect flaws in the network or unexpected external network, with hop-communications of data to a base station is
events and reconfigure the network as quickly as possible, as addressed. Hop-communication means that a sensor that cannot
in Jin et al. (2015). communicate directly with the base station can sends its data to
In our study, we extend the work of Lersteau et al. (2016) where the base station through intermediate sensors, that serve as relays.
the authors addressed a target tracking robustness problem in This problem is referred to as Pcn where n is the number of targets
Wireless Sensor Networks. The aim is to find a schedule that cov- and c stands for communication. In this problem, during a given
ers a single target at any time, with the target position supposed time horizon T, a set J of n targets with their spatial trajectories
to be exactly known over a time horizon. However, the targets can already known, will traverse a zone monitored by a set I of m sen-
be subjected to delays or advances and the schedule is protected sors. A function τ j (t) estimates the position of the target j at each
from these perturbations by the stability radius ρ . Indeed, what- instant t.
ever the delays or advances of the targets at any point of their First, the network has to guarantee a constant monitoring of
trajectory, the schedule remains feasible as long as these values each target. If the trajectory of a target is not continuously under
remain less than the stability radius. In the way it is presented the monitoring range of at least one sensor, the problem is con-
in Lersteau et al. (2016), this problem has some common features sidered infeasible. Note that if these targets are neglected and re-
with an assembly line such as in Sotskov et al. (2006), where the moved from our problem, we may obtain a new feasible problem.
aim is to schedule operations on a minimum number of stations Second, the network has to transfer all the data collected to a base
with possible variations in the processing time, under a cycle time station which is connected to a permanent source of energy and is
constraint. The method presented in Lersteau et al. (2016) starts able to forward the data to a remote control center. For that pur-
with a discretisation phase, i.e., the transformation of the geomet- pose, the sensors are equipped with communication modules to
rical problem into data used to model a combinatorial optimization transmit and receive data. Once activated they are able to form a
problem. The covered space is partitioned into faces (this term is path from the monitoring sensors to the base station, with several
defined in Section 5). The target has several time windows as it sensors used as relays if necessary.
moves through different faces, with each transition between two The sensors can only communicate with other sensors or with
faces called a tick. The target trajectory is no longer needed and is the base station if they are under its communication range Rc .
turned into a succession of time windows, with a list of available The sensor can monitor only the targets that are under its sensing
sensors to monitor the target during each such time window. The range Rs , i.e., at an instant t, targets that are more than Rs meters
authors proposed a pseudo-polynomial two-step algorithm. They away from a sensor i cannot be monitored by this sensor. N(i) is
noticed that the increase of the stability radius has an impact on the set of all the sensors in the communication range of the sen-
the time windows, and their solution approach relies on a bisec- sor i, i.e. those who can send and receive information from sensor
tion method to find a feasible set of time windows with the high- i. Each sensor i ∈ I has a limited battery with an energy of Ei joules.
est possible stability radius. Each step of the bisection methods The sensors are multi-role (Castaño et al., 2016) and therefore
solves a transportation problem. Finally, a linear program is solved we consider three kinds of energy consumption:
to maximize the stability radius for the time windows returned
• Monitoring a target requires a power of pS watts (a watt is a
by the bisection method. The present paper also relies on the dis-
joule per second)
cretisation and bisection phases but extends the problem by con-
• Emitting data requires pT watts.
sidering multiple targets and communication constraints. Thus, we
• Receiving data from another sensor requires pR watts.
now need to find a route for the collected data: from the activated
sensors, relayed by several sensors, to a base station. Communi- A monitoring activity by a sensor collects data that is necessar-
cation heavily impacts the sensor batteries and thus the returned ily transmitted to the base station. Thus, if a sensor i is monitoring
solutions. In addition, we have added different objectives, aiming a target for s seconds, i will also transmit s seconds of data, hence
to save energy for future missions, which is again an extension i will consumes ( pS + pT ) × s joules. Likewise, in a relay activity, if
of Lersteau et al. (2016). the sensor i receives s seconds of data, i will transmit s seconds of
data, hence i will consume ( pR + pT ) × s joules. Therefore, moni-
3. Formal definition of the stability radius of a schedule toring a target for s seconds, draws ( pS + pT ) × s joules out of the
battery of the monitoring sensor and ( pR + pT ) × s joules out of
A formal definition of the stability radius of a schedule can be each battery of the sensors used to relay the collected data to the
seen in Sotskov et al. (1998). In the scheduling problem, the stabil- base station. Note that the sensors are always sending the collected
ity radius is an indicator on the greatest variation on the process- information. These activities (sensing, receiving and transmitting)
ing times of the jobs for which the optimal schedule remains opti- can take place in the sensor at the same time. Moreover, if a sen-
mal. In such problems, each job i of the set of jobs Q has a process- sor is monitoring x targets at the same time, it also consumes x
ing time pi and, in a schedule s, a completion time c(s)i . Uncertain times its monitoring power. For example, at an instant t ∈ T, if a
4 F. Delavernhe, C. Lersteau and A. Rossi et al. / Computers and Operations Research 116 (2020) 104873

sensor monitors two targets and receives data from another sensor, 5. Discretisation
it will transmit the data from these three activities. Hence, for this
sensor, the instant power consumption at t is 2 pS + pR + 3 pT watts. The problem input is a set of geographical data. The sensors
i.e., it consumes energy for the monitoring of both targets, plus with their characteristics and their position, are deployed in a
consumes for receiving information, plus consumes three times the zone, along with the priority areas, the targets and their routes.
emitting consumption since it transmits both the data collected However, to determine the schedule of the sensors’ activity and
while sensing and the data received from another sensor. Conse- the routing of the collected data, the problem instance has to be
quently, a non sensing nor transmitting nor receiving sensor is not discretised. Discretisation is based on Lersteau et al. (2016), and
consuming energy. More detailed and complex energy consump- has been extended to the case of multiple targets, with commu-
tion models are presented in Halgamuge et al. (2009); Miller and nications and priority areas. Let us consider the example of Fig. 1,
Vaidya (2005). where the yellow disks represent the sensing range of three sen-
A long-term solution of the problem is expected to preserve sors, the black arrow the route of a target, B the base station and
the network ability to respond to future monitoring missions. the gray disks model the communication range. Two priority areas
Hence in the sequel, we only activate one sensor at a time to are shown as green polygons, they have two different ranks, the
monitor a target since activating multiple sensors will just col- smallest one being the most important one.
lect redundant data and waste energy. However, several sensors A face is defined as the set of all the locations (i.e., spatial points
can be activated at the same instant for the transmission task, or in the zone) monitored by the same subset of sensors (in Fig. 1,
to monitor different targets. Furthermore, we also consider pri- face {1,2,3} is the set of all location points that are under the range
ority areas, also called hot spots in the coverage related litera- of sensors 1, 2 and 3). Each face f is associated a set of candidate
ture (Huang and Tseng, 2005). The aim is to preserve and balance sensors S(f), i.e., any sensor that can monitor f is in S(f). The dis-
the residual capacities of the batteries for future target tracking cretisation phase turns the trajectory of each target as a sequence
missions (Lersteau et al., 2018). The decision makers define mul- of visited faces, as doing so allows to select one sensor in the set
tiple priority areas where the solution should preserve as much of the candidate sensors of a face to monitor the target. The in-
monitoring time as possible. Moreover, because all the areas may stant when a target moves from a face to another one is called a
not be considered equally important, the network managers set a tick. When a target enters the range of a sensor, it defines an en-
rank  ∈ C to each area, where C is the set of ranks and r is the tering tick; it is a leaving tick when it leaves the range of a sensor.
j
maximum rank. The rank of an area is fixed, it does not evolve The kth tick of target j is denoted by tk which also denotes its date
during the monitoring, and the higher the rank is the more im- of appearance, and σk is an integer value which is +1 if the tick
j
portant the area is. The rank expresses a preference for the preser- is entering, and −1 if it is leaving. By convention, the first and last
vation of energy in the considered areas. We treat the problem as ticks are respectively leaving and entering (Lersteau et al., 2016).
a multi-objective one, with a lexicographic order of the objectives. j
In addition, a time window k is the duration between two suc-
The primary objective is to find an activation schedule for the sen- j j
cessive ticks tk and tk+1 .
sors that respects the constraints (continuously monitor the tar-
In the example of Fig. 1, three faces only are considered (the
gets, transfer data, respect battery constraints) and that maximizes
other ones are not visited by the target): {1}, {1,2} and {1,2,3}.
the stability radius. The general notion of stability radius is defined
Hence this defines four ticks as presented in Table 2.
in Section 3 and a definition adapted to our problem in Section 5.
For the priority areas and their given rank, the problem is
Thus, the schedule remains feasible even if the targets are late or
discretized as follows. For a rank, with its priority represented
early to any point of their trajectories, provided that earliness or
by an integer  ∈ {1, . . . , r}, where r is the number of different
lateness stay below the stability radius.
ranks, wherever a potential target appears inside a priority area
The second objective is to maximize the minimal amount of
whose rank is , the network should provide the same monitor-
monitoring time guaranteed in the priority areas in each rank
ing time guaranteed. For each rank  ∈ {1, . . . , r}, F ( ) is the set
 ∈ C, while respecting the priority as expressed by the ranks
of all the faces that have a non-empty intersection with an area
of areas. This means that the guaranteed monitoring time avail-
of rank , and no intersection with an area of higher rank. In-
able anywhere inside a priority area of rank r is maximized, then,
deed, if a face has a non-empty intersection with two areas hav-
among all the solutions that offer this guaranteed coverage, we
ing ranks  and  with  <  , this face is part of F ( ) only
maximize the coverage guarantee everywhere inside the priority
since  is the most important rank. The set of all the sensors
areas of rank r − 1, and so on up to rank 1.
that can monitor at least one face in F ( ) is denoted by T().
Finally, as a third objective, we minimize the total amount of
More formally, T ( ) = ∪ f ∈F ( ) S( f ) for all  ∈ {1, . . . , r}. In order to
energy required by the current mission, while maintaining all the
previous objective to their optimal values.
Table 1 summarizes the notations.

Table 1
Summary of the notations.

I = {1, . . . , m} Set of sensors

J = { 1, . . . , n} Set of targets
Ei Energy of the battery of sensor i
pS Power consumption in watts for sensing
pR Power consumption in watts for receiving data
pT Power consumption in watts for transmitting data
Rc Communication range of a sensor
Rs Sensing range of a sensor
T Time horizon
C = {1, . . . , r } Set of ranks of areas
τ j (t) Position of target j ∈ J at time t ∈ [0, T]
N(i) Set of all the sensors in range of communication of sensor i
Fig. 1. A three-sensor example.
 F. Delavernhe, C. Lersteau and A. Rossi et al. / Computers and Operations Research 116 (2020) 104873 5

Table 2
The four ticks of example of Fig. 1.

Face {1} {1,2} {1,2,3}

Time window length (given) 10 = 5 11 = 10 12 = 2.5

Tick t01 = 0 t11 = 5 t21 = 15 t31 = 17.5
σ 0 = −1
1
σ 1 = +1
1
σ 2 = +1
1
σ31 = +1

The stability radius of a feasible schedule S can be defined us-

ing the following notations (see Sotskov et al., 1998):
• ζ = (ζ jk ∈ R ) j∈J,k∈K j is the set of possible advances and delays,
i.e., ζ jk is the deviation of the actual time spent by target j in
j
time window k compared to the estimated amount of time k .
 c for the example of Fig. 1.
ζ jk < 0 corresponds to an advance and ζ jk > 0 to a delay. Thus
Fig. 2. The communication digraph G
ζ jk ≥ −kj for all j ∈ J and for all k ∈ Kj , because a target cannot
spend a strictly negative amount of time in a time window.
• ( ) is the set of feasible schedules for , with the set of all
λjk such that target j spends λjk units of time in time window
guarantee the same amount of monitoring time wherever a po-
k, for all j ∈ J and for all k ∈ Kj .
tential target might be in an area of rank , we should guaran-
tee the same amount of monitoring time for every face f ∈ F ( ). The stability radius of solution S can be written as
Therefore, the priority areas and ranks are now discretized into
sets of faces F ( ) for each rank . For example, in Fig. 1, if the
ρ (S, ) = max{ε > 0|∀ζ ∈ B(ε ), S ∈ ( + ζ )}

Where B(ε ) = {ζ ∈ ζ | ||ζ ||∞ ≤ ε}, and

j
smallest area has rank 2 and the other one has rank 1, then is the set of all k .
F (1 ) = {{2}, {3}, {1, 2}, {1, 3}, {2, 3}, {1, 2, 3}, } and F (2 ) = {1}. The Hence, the stability radius ρ (S, ) of schedule S is the maxi-
reason why face {1} is not in F (1 ) is that it belongs to F (2 ). mum amount of delay or advance of that the targets can have in
The candidate sensors for the two ranks are T (1 ) = {1, 2, 3}, and each time window, without compromising the ability of S to en-
T (2 ) = {1}, it can be seen that the same sensor may appear in dif- sure a full coverage of all the targets.
ferent T() sets with no inconvenience. Finally, for a given rank ,
we can reduce F ( ) by keeping only the faces with no other faces 6. Upper bound on the stability radius of a schedule
from F ( ) included in them, i.e., we keep only a face f ∈ F ( ) if
∀ f  ∈ F ( ), f  ∈/ f . Indeed, if a set of sensors s is guaranteed an In this section, we present different upper bounds. The fi-
amount of monitoring time t, then each super-set s of s has at nal upper bound used in the sequel is the smallest bound out
least the amount of time t guaranteed. In the example of Fig. 1, we of three upper bounds, two of them are natural extensions
have F (1 ) = {{2}, {3}} and F (2 ) = {{1}}. F (1 ) is reduced to only from Lersteau et al. (2016) and are presented in Section 6.1, the last
two faces since, for example, the face {1, 2, 3} is not considered one is a novel contribution of this work, introduced in Section 6.2.
because it has at least as much covered time guaranteed as the
face {2}. 6.1. Definition of two general upper bounds on the stability radius
The communication between sensors is represented by the
communication digraph G  c = (I ∪ {B}, A ) with B the base station
We aim at producing a schedule of the sensors’ activity that
and A the pairs of sensors that can communicate. For all the col- maximizes the value of the stability radius, in order to offer the
lected data, an optimal flow from the monitoring sensor to the maximum protection against delays and advances. Upper bounds
base station is computed inside G  c . With the example of Fig. 1, we
are useful for achieving this goal, and are part of the solution ap-
obtain the following G  c (Fig. 2):
proach introduced in Section 7.
The new notations introduced in this section appear in Table 3. The two bounds presented in Lersteau et al. (2016) are both
Finally, the stability radius of a feasible schedule of the sensors’ built upon the expansion of the timespan of each time window.
activity is a measure of its ability to maintain a full coverage to the The first one, denoted by UB1 , is based on the time windows and
targets despite their advances and delays. More specifically, we as- their set of common candidate sensors, and the second one, UB2 ,
sume that after the discretisation phase, target j ∈ J crosses Kj ≥ 1 relies on the sensor capacities. These two upper bounds on the sta-
j
time windows, and should spend k units of time in time window bility radius, that have originally been introduced for a single tar-
k, for all k ∈ {1, . . . , K j }. get without communication, are extended to the case of multiple

Table 3
Summary of the notations introduced for discretisation.

tkj Time of tick k of target j


+1 if the tick k of target j is entering,
σkj
−1 otherwise.
T () Potential subsets of candidate sensors that can monitor a target in a face ranked 
r Total number of ranks for areas
F ( ) Set of all faces that have a nonempty intersection with an area of rank 
and an empty intersection with each rank  > 
S(f) Set of sensors covering the face f
6 F. Delavernhe, C. Lersteau and A. Rossi et al. / Computers and Operations Research 116 (2020) 104873

targets and now also consider the energy consumption for com- if two different targets visit the same face, they both could spend
munication. Thus, we obtain two tighter bound: 2ρ units of extra time in that face, with ρ the stability radius. 2ρ
⎧ ⎛  ⎞ ⎫ is the maximal extra time a target can spend in a face while being
⎨ Ei ⎬

1 i∈S j ( k )∩S j ( k ) covered by a schedule with a stability radius ρ . This data transmis-
UB1 = min ⎝ + tkj − tkj+1 ⎠|k < k sion is still possible provided that for each vertex separator Vf ⊂ I
j∈J ⎩2 p +p
S T
⎭  c that partitions I ∪ {B}\Vf into two connected components (the
(k,k )∈K 2j of G
first one containing B, and the second one containing S(f), the sen-
where Sj (k) is the set of candidate sensors for the target j during sors from f), the energy of the sensors of Vf should be sufficient to
j
time window k. tk is the tick delimiting the beginning of the time ensure data reception and transmission, this can be stated as:
window k and tk+1 the leaving tick associated with the time win-
j
⎛ ⎞
dow k.
⎛  ⎞ ⎜ j  ⎟ R 
Ei ⎜2ν ρ + kj ⎟
⎠( p + p ) ≤
T
Ei
UB2 = min⎝
i∈S ( f )
⎠ f ⎝
j∈J k∈K j i∈V f
f ∈F 2 × ( pS + pT ) × n f S j ( k )= S ( f )

where F is the set of faces visited by the targets, S(f) is the set of In order to obtain the tighter possible upper bound on ρ , the
sensors covering the face f and nf the number of times the face f vertex separator Vf may be such that the sum of the energy of
is visited (i.e., the number of times a target enter in the face f). the batteries is minimal. Furthermore, the above inequality can be
tightened by considering that Vf splits the vertices of I ∪ {B}\Vf into
6.2. Upper bound based on the communication digraph
two sets XB and Xf . XB is the vertex set that includes B, and Xf is
the vertex set that includes S(f). Consequently, the vertex separa-
In this section, we introduce a new upper bound on the stabil-
tor Vf separates B from all the faces whose candidate sensors are
ity radius based on the communication consumption and sensing
a subset of Xf . Consequently the sensors of Vf should have enough
consumption. Indeed, the previous bounds have been introduced in
energy to:
a context where communication was not taken into account. The
extensions brought to these bounds in this work are not consid-
ering the data relay tasks, i.e., the reception and transfer of data • Receive and transmit all the data collected from the faces with
from other sensors. These extensions only consider the transfer of sensors in Xf and without any sensors in Vf ,
the data by the sensors that collected it. However, it is more likely • Receive or monitor, and then transmit the data from the faces
that the relay of the data is needed a lot and therefore will im- with sensors in Xf and with some sensors in Vf but not all of
pact significantly the batteries. Since the sensors may need mul- them,
tiples relays to send the data to the base station and considering • Monitor and transmit the data from the faces with all sensors
that the transmission and reception powers are not negligible com- include in Xf and Vf .
pared to sensing, the performance of the computed schedule will
heavily depend on the routes available to transfer the data. We obtain:
In the application context, some subsets of sensors will have to
⎛ ⎞
carry all the communication to reach the base station, typically the
sensors surrounding the base station. For each of these sets of sen- ⎜  ⎟ S
⎜2μ j ρ+ kj ⎟
sors, increasing the lifetime of the network induces more commu- V ⎝ f ⎠p
nication, and then more power consumption. Thus, it is more likely j∈J k∈K j
S j (k )⊆V f
that one of those sets of sensors has first all batteries drained and
⎛ ⎞
depleted, and imposes a limitation on the stability radius value, so
the sum of the battery capacities of all sensors in this set defines ⎜ ⎟
⎜ j  ⎟
an upper bound on the value of the stability radius for any feasi- ⎜2γ ρ + ⎟
⎟ min( p , p ) +
j S R
+ ⎜ V k
ble schedule. Hence, for each face f, we find the set of sensors for
⎝ ⎠
f
j∈J k∈K j
which all the data collected will get relayed, we call it a restrain- S j (k )⊆V f
ing set of f. Moreover, we note that any feasible schedule covers S j (k )∩V f =∅
at least the case with no delay nor advance. In such a case, the ⎛ ⎞
computation of the minimal amount of data that need to be col-
⎜  ⎟
lected in a face is straightforward, i.e., it is equal to the amount ⎜2, ε j , ρ , +, , kj ⎟ pR
of time spent by the targets in that face. Hence, we also compute ⎝ X f \V f ⎠
j∈J k∈K j
the minimum amount of collected data that is then transmitted by S j (k )⊆X f \V f
all the restraining sets. Therefore, we compute an upper bound on ⎛ ⎞
the stability radius based on the battery level of all the restraining
⎜  ⎟ 
sets, and on the minimum amount of collected data. + ⎜2 , ν j , ρ , + , , kj ⎟ pT ≤ Ei
We remind that the communication digraph G  c = ( I ∪ {B}, A ) ⎝ Xf ⎠
j∈J k∈K j i∈V f
models all different paths for the data transmission between the S j (k )⊆X f
sensors and the base station (see Section 5). Let f be a face that is
visited by at least one target for a nonzero duration, then the data
where for all j ∈ J, the following constants μV , γV and εX \V are
j j j
collected
⎛ in f require sensors⎞ for receiving and transmitting data f f f f
defined by:
⎜ j  ⎟
⎜2ν ρ + kj ⎟
⎠ units of time, where ν f is set to one
j
for f ⎝ • μVj is set to one if and only if target j spends a nonzero amount
j∈J k∈K j f
S j ( k )= S ( f ) of time in at least one face whose candidate sensors form a
subset of Vf , otherwise μV = 0,
j
if and only if target j visits the face f, and zero otherwise. Indeed, f
 F. Delavernhe, C. Lersteau and A. Rossi et al. / Computers and Operations Research 116 (2020) 104873 7

• γVj is set to one if and only if target j spends a nonzero amount constraints of the problem have to be updated. It should be noted
f
that there is a discrete set of values for the stability radius that
of time in at least one face whose candidate sensors have a
causes such updates, these values being the distances between en-
non-empty intersection with Vf , but do not form a subset of
tering and leaving ticks belonging to different time windows. We
Vf , otherwise γV = 0,
j
f use the upper bound of the stability radius, presented in Section 6,
• εXj is set to one if and only if target j spends a nonzero to reduce this set of values by removing all the values greater than
f \V f
the upper bound.
amount of time in at least one face whose candidate sensors
The first part in solving P1 consists in finding the maximum
form a subset of Xf \Vf , otherwise εX \V = 0.
j
f f value of the stability radius in this discrete set for which the prob-
As a consequence, UB3 can be defined as:
⎛ ⎛ ⎞⎞
⎜ ⎜ ⎟⎟
⎜ ⎜     ⎟
j T ⎟⎟
⎜ Ei − ⎜ k p +
j S
k min( p , p ) +
j S R
k p +
j R
k p ⎟ ⎟
⎜ i∈V ⎜
⎜ f j∈J ⎝ k∈K j k∈K j k∈K j k∈K j ⎠⎟⎟
⎜ S j (k )⊆V f S j (k )⊆V f S j (k )⊆X f \V f S j (k )⊆X f ⎟
⎜ S j (k )∩V f =∅ ⎟
⎜
ρ ≤ UB3 = min∗ ⎜  j  ⎟
⎟
f ∈F
⎜ 2 μV f p + γV f min( p , p ) + εX f \V f p + νX f p
S j S R j R j T
⎟
⎜ ⎟
⎜ j∈J ⎟
⎜ ⎟
⎜ ⎟
⎝ ⎠

7. Solution method lem is feasible. This is done with a bisection method that checks
the existence of a feasible schedule for a given value in this set,
Since the problem has three objectives handled in a lexico- until the largest one is found. The first value tested is always the
graphical order, we separate it in three successive problems: P1 upper bound of the stability radius. Indeed, if a feasible schedule
for maximizing the stability radius, P2 for maximizing the time with such a stability radius is found, then it is optimal for P1. If
guaranteed in priority areas, and P3 for minimizing the energy this value is not feasible, a stability radius of 0 is tested. If no
consumption. They are solved sequentially, starting from P1 to P3 solution is found in that case, then there is no feasible schedule
where the objective of a problem becomes a constraint in the next for P1 and the algorithm stops. This typically happens when a tar-
one. Therefore, the solution method has three successive steps. get gets out of the range of any sensor, or when the network is
First, a schedule with a maximum stability radius is sought (P1 is so sparse that the collected data cannot be sent to the base sta-
presented in details in Section 7.1). Since P1 cannot be addressed tion. If a solution is found with a nonnegative stability radius, then
by simply solving a linear program because the set of faces to be the bisection method is used to find the largest stability radius
covered depends on the stability radius, we solve this problem us-
ing first a bisection method. Each step creates and solves a new
decision problem for a fixed value of the stability radius  with
a linear program (LP) called LP . Once the maximum value of , Algorithm 1: Solving P1 , P2 then P3 .
called opt , is found in a discrete set of potential values with a bi- // P1 - Maximization of the robustness
section method, a final linear program called LPρ is solved. It finds while ( ← getNextValueDichotomy()) = null do
the optimal value of the stability radius ρ = opt + δ, with δ the MakeLP ()
objective function of the optimal solution of LPρ . SolveLP ()
Next, P2 is addressed. For each rank , from the highest one if IsFeasibleLP () then
to the lowest one, we find a solution maximizing the time during opt ← 
which a target in a priority area of this rank is guaranteed to be end if
monitored. To reach this goal, for each face of a rank, a linear pro- end while
gram denoted by LP is solved. MakeLPρ (LPopt )
The third step is to solve P3: the energy consumption to mon- SolveLPρ ()
itor all the targets is minimized. This is done by solving a linear // P2 - Maximization of the guarantees in the priority
program denoted by LPE . areas
Algorithm 1 summarizes the overall approach. for all  ∈ C do
if  = 0 then
7.1. P1: Maximization of the stability radius MakeLP (LPρ , )
else
The maximization of the stability radius is achieved by enhanc- MakeLP (LP−1 , )
ing the bisection method presented in Lersteau et al. (2016). In end if
this subsection, we give an overview of this method while pro- SolveLP ()
viding more details on the proposed adaptation to our problem. end for
The main observation is that increasing the stability radius leads to // P3 - Minimization of the energy consumption
postpone the entering ticks and advance the leaving ticks. When- MakeLPE (LP|C| )
ever two ticks from different time windows interchange their or- SolveLPE ()
der of appearance, a time window disappears, and a new one is return GetSolutionLPE ()
created, with a reduced set of candidate sensors. As a result, the
8 F. Delavernhe, C. Lersteau and A. Rossi et al. / Computers and Operations Research 116 (2020) 104873

Model 1. LP , the linear program solved at each iteration of the bisection method.

ing P1 is then to maximize δ with 0 ≤ δ < (189 − 97 ). Solving LPρ

yields the optimal value of ρ , which is equal to 97 + δ in this ex-
ample.

7.2. P2: Maximizing the coverage guarantee in the priority areas

Model 2. LPρ solved to maximize the stability radius. Now that a feasible solution with the best stability radius is
found, we search for a schedule that maximizes the coverage guar-
antee in the priority areas. Let us remind that a rank corresponds
in the discrete set. This method differs from the one introduced to a set of faces, and a face can only appear in one rank.
in Lersteau et al. (2016) in the search of a feasible schedule. In- In this phase, for each rank from the highest one to the low-
deed, due to communication requirements, we have to solve a lin- est one, we create a new problem, where we maximize T , i.e., the
ear program LP instead of a transportation problem. This linear monitoring time guaranteed in all areas whose rank is  after the
program has no objective function. It is shown in Model 1, and is current mission. For each rank, a linear program denoted by LP
addressed with a solver. is addressed. It is recalled that a rank  fixes the optimal values
The decision variables are xjik , the monitoring time of the target from the previous problems solved (the stability radius and the T
j during its time window k by the sensor i, and fii the amount for all  > ), and is focused on the maximization of T . Therefore
of data transferred from sensor i to sensor i . We introduce H ij as the linear programs solved are built incrementally, by adding and
the set of all time windows of target j for which i is a candidate modifying constraints and variables from the previous linear pro-
sensor, i.e., the set of all k ∈ Kj such that i belongs to Sj (k). The grams solved (LP+1 if  < r, LPρ otherwise). The model solved is
linear program LP (with no objective function) is the following: presented in Model 3.
The first constraints (1) represent the limitation of the battery From the previous solved linear program, we fix first its optimal
for each sensor, (2) model the transfer of all data to the base sta- value in a new constraint to keep the optimal values of the previ-
tion (it is a flow conservation equation). Finally (3) enforce the cov- ous phases. We then change the objective function to maximize T ,
erage constraints, i.e., each target is continuously covered by a sen- the coverage guaranteed in the areas of rank . Next, we add vir-
sor at any time. tual targets in all the faces included in F ( ) and monitor them.
At the end of the bisection method, we obtain the maximum These targets all have the same time guarantee, T , and model the
value opt for which the problem is feasible. Afterwards, while covering requirement induced by the faces of rank . Furthermore,
considering the time windows in that case, we maximize the sta- a virtual routing of the data (new set of flow constraints) is added
bility radius increase δ , by solving a linear program similar to the to route the potential data recorded from the virtual targets. Thus,
one used for determining opt in LP . The new linear program is when monitoring those faces will be required, the set of candidate
called LPρ , presented in Model 2, it is identical to the one used in sensors will have enough residual energy after the current mission
the bisection method, with in addition an objective function (6), a to track targets in those faces for T units of time, and there will
constraint (7) and an updated constraint (3’). be a path of sensors with enough energy left to transmit the data
Since δ is the stability radius increase from opt , it impacts the to the base station.
The new decision variables are the following:
duration of the time windows in (3’). σk is the value of the k-th
j

tick of the target j (−1 if leaving, 1 if entering). The optimal stabil- • xif the time during which sensor i ∈ S(f) monitors a fictitious
ity radius numerical value is then ρ = opt + δ . target in face f having rank .
The example of Fig. 3 illustrates the method. • fii1 the amount of data generated by a fictitious target, trans-
In this example, with different values of ρ there are different mitted from i to i .
time windows, with different candidate sensors. The stability ra-
The LP model for a rank  is the following:
dius ρ may be in the interval [0,97), or in [97,189), or in [189,
Constraint (1’) is the constraints (1) modified to take into ac-
Upper bound), or be equal to Upper bound. During the bisection
count the virtual flow and fictive targets that are added. It can be
method, we find opt = 97, which means that there is a feasible
seen that by comparison to LPρ , sensors can now be used to per-
schedule with a stability radius of 97. Therefore all schedules with
form sensing and data transmission after the current mission, to
a stability radius which is strictly less than 97 are not considered
insure the network ability to monitor the priority areas.
anymore. This results also implies that there is no solution either
Constraints (2), (3’), (4), (5) are the same as in P1
with a value larger than or equal to 189. The second step in solv-
Constraint (8) sets δ to δ opt , the optimal objective value of LPρ .
Constraints (9) ensure the tracking of the fictive targets for a
rank  for a duration of T units of time.
Constraints (10) ensure the flow balance for the data collected
generated by fictive targets, at each sensor i.
Fig. 3. Illustration of the solution method for P1. Constraints (12) set the value for the previous ranks  .
 F. Delavernhe, C. Lersteau and A. Rossi et al. / Computers and Operations Research 116 (2020) 104873 9

Model 3. LP Model solved for a rank .

7.3. P3: Minimizing energy consumption dichotomy is done on a set of discrete values, that in the worst
case is equals to all possible intersections between the ticks. And,
Problem P3 is to minimize f3 , the total energy consumed to if the targets have polygonal trajectories, the number of ticks can-
achieve the current mission. In order to solve it, the linear pro- not exceed 2qm, with q the number of segments in the trajectories
gram LPE is built from the last linear program solved in P2, by fix- and m still the number of sensors. In P2, we solve as many linear
ing its objective value in a new constraint, changing the objective programs as the number of ranks and finally in P3, there is only
function and resolving it. one linear program to solve.
While solving P3, the schedule and the routing can be changed,
but the stability radius value and the coverage guarantees found in 8. Numerical experiments
P1 and P2 are maintained to their optimal respective values.
The energy consumed by a sensor is the left-hand side of con- 8.1. Description of the protocol
straints (1), this quantity can be written as:
 In this last part, we present our experiments, results and anal-
   
x jik pS + pR fi i + pT fii ysis. We study the behavior of the solution method and the im-
i∈I j∈J k∈H i
j
i ∈N ( i ) i ∈N ( i ) pact of different parameters like the number of sensors, targets
 and ranks. Moreover, we evaluate the efficiency of the upper bound
  
+p S
xi f  + p R
fi1 i +p T
fii1 introduced in this work compared to the other ones extended
 from Lersteau et al. (2016). To this end, we design four experiments
f∈ F (   )| i∈S ( f ) i ∈N ( i ) i ∈N ( i )
≤ ≤r on different sets of instances, each of them investigates the impact
The virtual flow and fictive targets are not actual consumptions of the problem from the following:
since they are not used during the current mission, but are spared • Impact of the sensing and communication powers. We vary
for possible future missions, so they are subtracted from the pre- them in order to compare UB3 to the two other upper bounds.
vious quantity, leading to: In this experiment, we study the efficiency of this new upper
⎛ ⎞
bound on the stability radius.
   
⎝ x jik pS + pR fi i + pT fii ⎠
• Impact of the sensor density. In the experiment, only the num-
ber of sensors is varied.
i∈I j∈J k∈H i i ∈N ( i ) i ∈N ( i )
j • Impact of the number of targets. More targets to monitor in-
Finally, since the total recording time of the target is a constant duces more data, send or receive.
(the stability radius is fixed), we just have to minimize the trans- • Impact of the number of ranks and areas.
ferred data in the current mission, so f3 , the objective function of • Impact of the communication.
P3 is:
 Although each experiment has its proper set of instances,
f3 = Minimize f i i they are all generated using the instance generator presented in
i∈I,i ∈N (i ) Section 8.2.

7.4. Complexity analysis

8.2. Dataset

There are two important parts in the proposed algorithms, the

All of our instances are generated with the same algorithm. The
discretisation phase and the solution method (i.e., the solution of
generator’s inputs are:
P1, P2 and P3).
First, the discretisation can be achieved using a pseudo- • The number of sensors m
polynomial algorithm. A precise analyze of the complexity of the • The surface of the rectangular zone L1 × L2
discretization can be found in Lersteau et al. (2016), that shows • The number of targets n to monitor
that the number of faces is bounded by a polynomial on the num- • The minimum Emin and maximum Emax level of energy initially
ber of sensors. Thus, the modifications brought in this work (com- available in the batteries of the sensors
munication, priority areas and multiple targets), that still rely on • The powers associated with the different tasks (pS , pT and pR )
the numbers of sensors and faces, does not change the complexity. • The communication range Rc and sensing range Rs
Secondly, P1 is a dichotomy that requires a logarithmic number • The number of priority areas q and ranks r
of iterations, with at each iteration, a linear program solved. The • The maximum radius of the priority areas Ra
10 F. Delavernhe, C. Lersteau and A. Rossi et al. / Computers and Operations Research 116 (2020) 104873

Table 4
Default values of the parameters in the instances generator.

Parameter Value Parameter Value

Number of sensors m 400 Energy of the sensors [Emin , Emax ] [350,400]

Number of targets n 2 pS 2.8
Size of the area L1 × L2 300 × 300 pR 1
Sensing range Rs 35 pT 1
Communication range Rc 70 Number of areas q 5
Maximum radius of the priority areas Ra 50 Number of ranks r 2

First of all, the sensors are randomly deployed in the zone. Each Table 5
Domination of UB3 over UB1 and UB2 .
sensor has a random level of battery picked between the two val-
ues given as parameters. Second, the journeys of the targets are pS pR = pT # { min(UB1 , UB2 ) > UB3 }
drawn also randomly in the zone. Their paths are simple routes 3 0.5 2 / 50
made of three segments. Third, the priority areas are randomly de- 3 3 20 / 50
ployed and their rank are also randomly chosen, with at least one 3 5 26 / 50
priority area per rank. Each priority area is a disc whose radius is
selected randomly between fifty percent and one hundred percent
of the maximum value given as parameters. Finally, the base sta- low communication costs, by contrast with UB3 which is mostly
tion is randomly deployed in the zone. based on these costs. However, for an average consumption, with
The default parameters of our instances are presented in pR = pT = pS = 3, UB3 is almost as good as the two other upper
Table 4. bounds combined. In such a case, in 20 instances out of 50, UB3 is
strictly the best upper bound on the stability radius, and is there-
8.3. Results and analysis fore really useful in the bisection method. Finally, we observe that
when power consumption due to communication is much more
We present in this section the results and analysis of the differ- significant than sensing power consumption, which is the most re-
ent experiments. The software is coded in C++ and all the experi- alistic situation (Anastasi et al., 2009), our new upper bound dom-
ments were run on a computer with Ubuntu 16.04 and Intel Core inates the other two ones in 26 cases out of 50, and is therefore
i7-6700HQ CPU @ 2.60 GHz × 8 cores and 16 GBytes of RAM. We significantly more efficient. Naturally, this dominance gets stronger
use the version 12.7 of IBM CLPEX for solving the linear programs. when the power consumption due to communication increases.
The CPU times reported are in seconds. Though, the two other bounds should still be considered since
they are useful in almost half of the instances (note that for the
highest communication costs, UB1 was the best bound in 12 in-
8.3.1. Impact of the sensing and communication powers
stances, same for UB2 ).
In this first experiment, we study the quality of the new up-
To conclude, the new introduced bound, UB3 , has good perfor-
per bound. The objective is to compare it to the previous upper
mances compared to the two other ones and it makes a signifi-
bounds and see if the method benefits from the addition of UB3 .
cant contribution to the approach, in reducing the number of lin-
Indeed, the two previous bounds are still valid in our robustness
ear programs to solve in the bisection method. Though, it is less
problem and have been extended to consider multiple targets and
useful when low communication costs are considered.
communication costs. Though, they are both only considering in a
face f, the consumption induced by the coverage of the targets in-
8.3.2. Impact of sensor density
side this face (i.e., sensing the targets inside f and transmitting the
In this second part, we study the impact of sensor density on
data collected). They are not considering the consumption induced
the method and the values found. For that purpose, we gener-
by forwarding the data collected in other faces. Hence, the addi-
ate a set of fifty feasible instances using the default parameters
tion of our new upper bound based on such principle seems to be
(Table 4), except for the number of sensors which is fixed to 200
a good opportunity to help the solution process. Indeed, the value
sensors. Afterwards, we add a few sensors to each instance in or-
of the upper bound is important since it may reduce drastically
der to study the impact of sensor density in the network. Each sen-
the number of iterations in the bisection method in P1 , thus the
sor added is generated in the same way as the initial sensors, i.e.,
number of linear programs to be solved. Clearly, it is expected to
they have random positions in the L1 × L2 area and a random level
perform better when the communication costs are significant com-
of initial battery in [Emin , Emax ]. The method is run on the instances
pared to sensing. In this experiment, we are studying the efficiency
when the number of sensors m is in {200, 250, 300, 350, 400, 700}.
of the new bound compared to the two extended previous bounds
In Table 6, we report the average number of time windows and
when working on WSN with several significant ratios between the
the number of faces included in a priority area. Table 7 reports
sensing power pS and the communication powers pR and pT .
the computational effort required by each problem of the solution
We test three different sets of powers. In all of them, pR = pT
and pS = 3. First, we set pR = pT = 0.5. In the second set, we have
pR = pT = pS = 3. And in the third set, pR = pT = 5. We generate Table 6
a set of fifty instances, and we use the three different powers Average number of time windows and faces in priority
areas with different numbers of sensors .
on each of them. All other parameters are fixed as presented in
Table 4, except for the number of sensors and targets fixed to 300 #Sensors # windows # faces in priority areas
and 4 respectively. 200 103.48 85.94
We report the number of times where UB3 dominates the two 250 128.94 133.58
other bounds in Table 5. 300 154.58 188.26
Table 5 shows that with a low consumption for communica- 350 179.94 254.40
400 204.94 324.80
tion, UB3 is dominated. Indeed, it reaches the best value in only 700 357.38 973.34
2 instances out of 50. Clearly, UB1 and UB2 are less impacted by
 F. Delavernhe, C. Lersteau and A. Rossi et al. / Computers and Operations Research 116 (2020) 104873 11

Table 7
Average computational effort with different numbers of sensors.

#Sensors CPU (Discretisation) CPU (UB) CPU (P1) LP solved in P1 CPU (P2) CPU (P3) Overall CPU time

200 0.02 0.04 0.28 3.44 1.22 0.49 2.05

250 0.02 0.06 0.16 1.22 2.38 1.14 3.75
300 0.04 0.13 0.41 2.30 4.01 1.86 6.45
350 0.06 0.20 0.35 1.22 6.39 2.62 9.62
400 0.10 0.33 0.53 1.48 8.94 3.63 13.54
700 0.86 2.47 1.03 1.00 59.65 19.52 83.53

Table 8 As expected, adding sensors to a network extends the robust-

Evolution of the objectives with the number of sensors.
ness. It adds more opportunities for the network to monitor the
#Sensors Stability radius T2 T1 f3 targets and to transfer the collected data. Thus, it is less restrained
200 79.26 40.24 5.84 10816.16 by the battery and by the constraint of having only one sensor
250 102.98 43.74 4.87 10667.43 monitoring a target at a time. This observation is also explaining
300 113.56 45.58 3.33 10509.07 the decrease of the last objective. With more sensors, the net-
350 119.75 47.58 1.56 10342.56 work has more opportunities for shorter and more direct routes
400 124.12 45.30 1.35 10312.58
to the base station. It induces less transfer of data between the
700 143.20 32.01 1.23 10179.60
sensors and therefore less energy consumption. The objectives on
the times guaranteed in the priority areas are affected differently.
Indeed, with more sensors, there is more energy in the network
method with the average number of linear programs solved in P1. and it seems that it gives more opportunity to increase these ob-
Finally, in Table 8 we report the average objectives values. The CPU jectives. However, for each rank  in C, T is constrained by the
times are in seconds. time guaranteed in each face f in F ( ), hence T is constrained by
Table 6 shows that when sensor density increases, the num- the less covered face. Adding sensors in the network is not bring-
ber of time windows increases also. Consequently, each step of the ing necessarily more sensors in each face already in F ( ). Thus,
solving method takes more time to solve, with more data to pro- it can still be the same face (i.e., the exact same set of candidate
cess, more constraints and more variables in the models. Further- sensors) that restrain the value of T after adding sensors to the
more, there is also a fast increase of the number of faces inside network. However it still adds more faces to cover in the priority
the priority areas. Therefore, the models for P2 and P3 are becom- areas (Table 6). Consequently, the network is having less coverage
ing even more complex with more constraints and more variables. time guaranteed in the priority areas.
Each row of Table 7 presents the average results over fifty in-
stances. The first column is the number of sensors. The second
one is the CPU time (in seconds) spent by the discretisation step, 8.3.3. Impact of the number of targets
CPU (UB) is the average time (in seconds) for computing the upper In this third experiment, we analyze the impact of the number
bound (6). Column 4, 6 and 7 represent the CPU times for solv- of targets. As in the second experiment, all parameters are fixed
ing P1, P2 and P3 respectively. Column 5 represents the average (see Table 4) except the one we study. The number of targets varies
number of linear programs when solving P1. The last column is in the set {1, . . . , 7}. An initial set of feasible instances is generated
the average overall CPU time (s) required to solve an instance. and solved with 7 targets. Afterwards, we just remove the targets,
As expected, the density of the network impacts a lot the solu- one by one, to the instances to test on different numbers of targets.
tion time. Indeed, the results show that, when the number of sen- We report in our results the average times for all the instances.
sors increases, the computational effort required for every step and Results are summarized in Table 9. The values of the objectives are
thus the overall CPU time are increasing quickly. It is explained by reported in Table 10. The CPU time reported are in seconds.
Table 6, that shows that the number of time windows is increasing. In this experiment, we observe that P1, P2 and P3 have increas-
Consequently, the CPU times required by the discretisation and the ing CPU times resulting in an overall growth of the computational
computation of the upper bound are increasing quickly, since they effort. However the results can be analyzed more deeply. First, for
are mostly dependent on the number of time windows. The solu- the discretisation, only the part of the computation of the routes
tion time of P1 also depends on the number of linear programs to of the targets is impacted, and not the discretization of the com-
be solved. This is why the CPU time for P1 is not always increas- munication or the priority areas. Consequently the time used by
ing when the density increases, since there are often less linear this phase is increasing only slowly. Secondly, the most increasing
programs solved in P1 (because the upper is more often a feasible times are for P1 and the computation of the upper bounds as they
solution) though they are more complex to solve. The solution pro- are totally dependent on the number of targets: more targets imply
cesses of P2 and P3 are heavily impacted by the number of faces more time windows and a stability radius more complex to com-
in the priority areas, hence their running times are also increasing pute. Finally, the time needed for P2 and P3 is increasing slowly
significantly when the sensor density is higher. since only the part of the model inherited from P1 is impacted.
The Table 8 shows that the stability radius is increasing with To conclude on the computational effort, as expected, each part
the number of sensors. The second objective seems to increase of the solution method needs more time when targets are added.
at first but with more than 350 it decreases. The following ob- The values of the objectives in Table 10 show that with more
jective, T1 , is impacted negatively by the increase of the number targets, the stability radius is obviously decreasing. Indeed, with
of sensors, with the covering time left in the priority areas rank more targets the network is more limited, because it consumes
1 overall decreasing. Finally, surprisingly, the last objectives (i.e., more to monitor the targets. Furthermore, if we increase the stabil-
minimization of the communication consumption) is getting better ity radius, we increase it for every target. Hence, we consume more
with more sensors even if a higher stability radius induces more energy when increasing the stability radius for problems with an
data collected and transferred. increasing number of targets.
12 F. Delavernhe, C. Lersteau and A. Rossi et al. / Computers and Operations Research 116 (2020) 104873

Table 9
Results with different numbers of targets.

#Targets CPU (Discretisation) CPU (UB) CPU (P1) CPU (P2) CPU (P3) Overall CPU time

1 0.10 0.15 0.55 9.09 3.02 12.90

2 0.10 0.28 0.49 9.33 3.22 13.41
3 0.10 0.39 0.50 9.30 3.30 13.57
4 0.10 0.49 0.75 9.75 3.38 14.48
5 0.10 0.61 1.19 10.38 3.60 15.88
6 0.10 0.74 1.84 11.25 3.66 17.59
7 0.10 0.89 5.30 12.13 4.18 22.61

Table 10 priority. For low priority, there is often no time guaranteed. We

Evolution of the objectives with the number of targets.
advise to only use up to three or four ranks.
#Targets Stability radius T2 T1 f3 Secondly, we notice that the energy consumed is globally not
1 207.60 55.84 5.93 5629.14 varying. Indeed, it seems that the overall time guaranteed in all the
2 127.61 51.30 4.63 9568.11 priority areas is not considerably rising because it is probably lim-
3 105.98 48.30 3.25 13606.10 ited by the same region of sensors whatever the number of ranks.
4 98.79 43.40 1.95 18139.19 Therefore, the state of the network and the energy left after solv-
5 93.77 35.90 1.05 22820.62
ing P2 does not differ much with more or less ranks. The process
6 87.98 26.07 0.67 28297.36
7 72.55 20.31 0.36 33337.56 of minimizing the energy consumption obtains similar values.

8.3.5. Impact of the communication in the objectives

The same observation can be made with the other objectives. In this experiment, we study the impact of the communica-
The time guaranteed in the priority areas is decreasing and the to- tion. i.e., how the mandatory data transfer to the base station has
tal energy consumption is also increasing. It shows that the net- impacted the solution quality and the computational effort. Tak-
work has still a lot of energy after solving P1 for a few targets. In ing communication into account makes the problem more com-
such cases, the stability radius is bounded by a specific region of plex to solve and is restraining the objectives. In order to study
the network or by a constraint on the number of sensors activated such assumptions, we compare the objectives of the solution ob-
at the same time. Therefore the total amount of energy consumed tained by the method with and without considering the communi-
(minimized in P3) is limited, and more energy is available to guar- cation on a set of instances. Several different communication costs
antee a higher time in the priority areas. However, with more tar- are tested to represent different impacts in the network. Fifty in-
gets, the network consumes more for the stability radius, so the stances are generated with the default parameters presented be-
third objective is increasing and moreover there is less energy for fore (Section 8.2). Our method is run multiple times on each of
the priority areas. Therefore, with more targets, even with the sta- these instances, each time with a different communication power
bility radius decreasing, we consume more energy and can guaran- consumption. The values used for pR and pT are {1, 2, 3}, with
tee less time for the priority areas. pT = pR . The value of pS is still equals to the default value. After-
wards, for each instance, we run our method without considering
the communications costs (i.e., pT = pR = 0). We report the average
8.3.4. Impact of the number of ranks and areas value of the objectives in Table 13, for each cost tested.
In this set of experiments, we study the impact of the priority Afterwards, a second set of feasible instances is generated,
areas by varying the number of areas generated with the number with the default parameters except for the number of targets
of ranks. As for the experiment on targets, we first generate and fixed to 1 and no priority area. Hence, on these instances, we
solve a set of fifty instances with only two areas of the same rank. can compare our method and the method developed in previous
Afterwards, we add areas and ranks on the same instances. There- work (Lersteau et al., 2016). It is recalled that communication is ig-
fore, we obtain different versions. For each instance, we have ver- nored in this article. Several communication costs are tested again
sions with 1 to 7 ranks for respectively 2, 4 to 14 areas with at for our method, with pR = pT and pR in {0, 1, 2, 3}. The CPU times
least one area per rank. Average times are presented in Table 11 (still in second) needed for the computation of UB and P1 are re-
and objective values for P2 and P3 (the stability radius is not vary- ported in Table 14.
ing) in Table 12. The CPU times are in seconds. As expected, all the values of the objectives are worsening
This experiment shows that the numbers of ranks and areas in- when increasing the communications costs. Although, for the sta-
crease the computational effort for the discretisation and for the bility radius, the difference is really small between no communi-
problem P2. Moreover, as expected, the computation of the upper cation cost ( pR = pT = 0) and low costs. However, there is a bigger
bound on the stability radius and problem P1 are not impacted by difference when comparing no cost to the most important costs
the areas. We can observe that, with the increase of the number ( pR = pT = 3), with a loss of approximately 10% of the stability
of ranks, the overall CPU time is increasing, up to five times more radius. The coverage times left in the priority areas, for the two
from one to seven ranks. ranks, are also quickly decreasing with higher costs. When com-
Table 12 shows that the time guaranteed in an area is heav- munication costs are added, the priority areas of rank 1 are losing
ily depending on the priority assigned to it. Indeed, as an exam- almost all their coverage time, and the areas of rank 2 an impor-
ple the areas of rank 1 have a lot of time guaranteed when they tant part of it. Finally, the overall energy consumed for communi-
are the only priority areas. However, for the same areas, the time cation (i.e., f3 ) is obviously continuously increasing with the costs.
is dropping when we add higher ranks. Moreover, with a lot of These results confirm that, as expected, the communication
ranks, there is a really small time guaranteed for the less impor- is restraining the objectives. The value of the stability radius is
tant ranks. Therefore, we conclude that the time guaranteed in an impacted, since more energy are necessary to cover the targets.
area is heavily depending of its rank. For the most critical areas, Though, it does not vary a lot with low communication costs. In-
the time guaranteed is huge but will heavily decrease along the deed, the stability radius is also constrained by the constraints
 F. Delavernhe, C. Lersteau and A. Rossi et al. / Computers and Operations Research 116 (2020) 104873 13

Table 11
Results with different numbers of ranks.

#Ranks #Areas CPU (Discretisation) CPU (UB) CPU (P1) CPU (P2) CPU (P3) Overall CPU time

1 2 0.06 0.33 0.55 4.59 3.38 8.90

2 4 0.08 0.36 0.57 10.77 4.17 15.96
3 6 0.12 0.33 0.56 14.82 4.09 19.92
4 8 0.16 0.33 0.56 19.12 4.12 24.29
5 10 0.20 0.32 0.55 23.94 3.95 28.96
6 12 0.25 0.33 0.57 29.39 4.04 34.57
7 14 0.29 0.32 0.57 34.51 4.11 39.80

Table 12
Evolution of the objectives with the number of ranks.

#Ranks T1 T2 T3 T4 T5 T6 T7 f3

1 46.93 10058.05
2 7.58 50.56 10158.19
3 2.97 12.61 64.61 10383.87
4 1.44 3.94 24.39 68.00 10425.95
5 0.09 0.91 6.48 25.07 65.57 10197.06
6 0.00 0.00 1.31 6.04 22.87 66.62 10407.70
7 0.00 0.00 0.16 1.22 6.77 26.86 70.71 10480.30

Table 13 These results were expected, since the problem solved is much
Impact of communication on the objective values.
more complex, and even with null communication costs, the
pR = pT Stability radius T2 T1 f3 communication is still computed. The linear program solved in
0 124.60 90.52 40.68 0 the bisection method in the present work require much more
1 124.27 52.94 4.58 9559.63 computational effort to solve than the transportation problem
2 122.38 27.79 0.84 19975.08 in Lersteau et al. (2016). The upper bound may also be less re-
3 111.92 14.63 0.10 30292.13 straining thus increasing the running time of P1. Finally, the new
bound added, is obviously increasing the solution time of UB.
Table 14
9. Conclusion
Impact of communication on the solution times.

Problem pR = pT CPU (UB) CPU (P1) In this paper, we extended the original problem, treated
Our 0 0.17 0.13 in Lersteau et al. (2016). Their objective was to find an activation
Problem 1 0.17 0.44 schedule to track a target with a WSN. We developed it to make
2 0.17 1.05
the problem more generic and it now handles cases with multiple
3 0.17 1.19
Lersteau et al. (2016) 0 0.004 0.08
targets to be tracked at the same time by the same WSN. Further-
more, we added communications between sensors, with the task
to transfer all the data gathered to a base station. The communica-
tion is a great generalization since the impact on a WSN is impor-
forcing to only have one sensor activated per target for sensing tant, and in some applications may be even more consuming that
at any time. Thus, in these cases, adding communication costs is the sensing. These two extensions change the solution process. In-
not necessarily reducing the stability radius. That is why the val- deed, the discretisation is modified to deal with the communica-
ues found for the stability radius are similar in our instances with tion and the targets. Likewise, the optimization of the stability ra-
no communication costs and low costs. Though, the stability radius dius (problem P1) is more general. We solved a new model by lin-
is decreasing a lot more with high costs. The others objectives are ear programming instead of a transportation problem. We adapted
worsening a lot more when the communication costs are added. the previous upper bounds to multiple targets and communication
The energy consumption is increased a lot, and not only the sen- and introduced a new bound, based on energy consumption that is
sors in range of a target are consuming energy. The energy left in due to communication. The relevance of this new bound has also
the network for the priority areas is a lot smaller, thus restraining been checked. Afterwards, we added two objectives in our prob-
these objectives. lem, optimized after P1. The new objectives are the maximization
To conclude on the objective values, the addition of the com- of the time guaranteed inside the priority areas (P2) and the maxi-
munication is certainly impacting the objectives. However, the loss mization of the overall energy left (P3). We added two steps to the
of robustness is small, especially for low communication costs. solution method for these objectives, based on new models formu-
As expected, the CPU times in Table 14 are worst with our prob- lated as linear programs.
lem, since the communication are considered. Both the computa- We designed different experiments to test and analyze the so-
tion of the upper bound and P1 are longer to compute. For the lution method. The results show that increasing the numbers of
upper bound, it does not vary with the different tested costs. How- sensors, targets, ranks or areas is contributing to the rise of the
ever, the difference between the time needed for solving P1, with overall running time. The number of sensors, i.e., the density, is
communication and without it, is becoming more and more im- the most significant parameter in terms of CPU time. Finally, we
portant with higher costs. It also shows that solving the problem also showed that the new upper bound based on communication,
without considering the communication is faster than solving the is efficient and is used in the solution process.
same problem while considering the communication but without To go further, several extensions could be considered. First,
any costs, which, again, is not surprising. what if one targets has a delay or an advance outside the stabil-
14 F. Delavernhe, C. Lersteau and A. Rossi et al. / Computers and Operations Research 116 (2020) 104873

ity radius? How to dynamically change the solution to avoid losing Halgamuge, M.N., Zukerman, M., Ramamohanarao, K., Vu, H.L., 2009. An estimation
targets? And how to achieve this at low computational cost so as of sensor energy consumption. Prog. Electromagnet. Res. 12, 259–295.
Handy, M., Haase, M., Timmermann, D., 2002. Low energy adaptive clustering hier-
to meet real time constraints? In order to test various realistic sit- archy with deterministic cluster-head selection. In: Mobile and Wireless Com-
uations, we intend to use CupCarbon simulator (CupCarbon,Mehdi munications Network, 2002. 4th International Workshop on. IEEE, pp. 368–372.
et al., 2014). Huang, C.-F., Tseng, Y.-C., 2005. The coverage problem in a wireless sensor network.
Mob. Netw. Appl. 10 (4), 519–528.
Jin, Y., Ding, Y., Hao, K., Jin, Y., 2015. An endocrine-based intelligent distributed co-
CRediT authorship contribution statement operative algorithm for target tracking in wireless sensor networks. Soft Com-
put. 19 (5), 1427–1441.
Kung, H.-T., Vlah, D., 2003. Efficient location tracking using sensor networks. In:
Florian Delavernhe: Conceptualization, Methodology, Software,
Wireless Communications and Networking, 2003. WCNC 2003. 2003 IEEE, 3.
Validation, Investigation, Writing - original draft, Writing - review IEEE, pp. 1954–1961.
& editing. Charly Lersteau: Methodology, Resources. André Rossi: Lersteau, C., Rossi, A., Sevaux, M., 2016. Robust scheduling of wireless sensor net-
works for target tracking under uncertainty. Eur. J. Oper. Res. 252 (2), 407–417.
Investigation, Writing - original draft, Writing - review & editing,
Lersteau, C., Rossi, A., Sevaux, M., 2018. Minimum energy target tracking with cov-
Supervision, Project administration. Marc Sevaux: Writing - origi- erage guarantee in wireless sensor networks. Eur. J. Oper. Res. 265 (3), 882–894.
nal draft, Writing - review & editing, Supervision. Liu, Y., Liang, W., 2005. Approximate coverage in wireless sensor networks. In: Lo-
cal Computer Networks, 2005. 30th Anniversary. The IEEE Conference on. IEEE,
References pp. 68–75.
Mehdi, K., Lounis, M., Bounceur, A., Kechadi, T., 2014. Cupcarbon: a multi-agent and
discrete event wireless sensor network design and simulation tool. In: 7th In-
Aalsalem, M.Y., Khan, W.Z., Gharibi, W., Khan, M.K., Arshad, Q., 2018. Wireless sensor ternational ICST Conference on Simulation Tools and Techniques, Lisbon, Portu-
networks in oil and gas industry: recent advances, taxonomy, requirements, and gal, 17–19 March 2014. Institute for Computer Science, Social Informatics and
open challenges. J. Netw. Comput. Appl. 113, 87–97. Telecommunications Engineering (ICST), pp. 126–131.
Akyildiz, I.F., Su, W., Sankarasubramaniam, Y., Cayirci, E., 2002. A survey on sensor Miller, M.J., Vaidya, N.H., 2005. A mac protocol to reduce sensor network energy
networks. IEEE Commun. Mag. 40 (8), 102–114. consumption using a wakeup radio. IEEE Trans. Mob. Comput. 4 (3), 228–242.
Alibeiki, A., Motameni, H., Mohamadi, H., 2019. A new genetic-based approach for Modieginyane, K.M., Letswamotse, B.B., Malekian, R., Abu-Mahfouz, A.M., 2018. Soft-
maximizing network lifetime in directional sensor networks with adjustable ware defined wireless sensor networks application opportunities for efficient
sensing ranges. Pervasive Mob. Comput. 52, 1–12. network management: a survey. Comput. Electr. Eng. 66, 274–287.
Anastasi, G., Conti, M., Di Francesco, M., Passarella, A., 2009. Energy conservation in Mohamed, S.M., Hamza, H.S., Saroit, I.A., 2017. Coverage in mobile wireless sensor
wireless sensor networks: asurvey. Ad Hoc Netw. 7 (3), 537–568. networks (M-WSN): a survey. Comput. Commun. 110, 133–150.
Benini, L., Castelli, G., Macii, A., Macii, E., Poncino, M., Scarsi, R., 20 0 0. A discrete– Naderan, M., Dehghan, M., Pedram, H., Hakami, V., 2012. Survey of mobile object
time battery model for high-level power estimation. In: Proceedings of the Con- tracking protocols in wireless sensor networks: a network–centric perspective.
ference on Design, Automation and Test in Europe. ACM, pp. 35–41. Int. J. Ad Hoc Ubiquitous Comput. 11 (1), 34–63.
Campos-Nañez, E., Garcia, A., Li, C., 2008. A game-theoretic approach to efficient Patel, D.J., Batta, R., Nagi, R., 2005. Clustering sensors in wireless ad hoc networks
power management in sensor networks. Oper. Res. 56 (3), 552–561. operating in a threat environment. Oper. Res. 53 (3), 432–442.
Cardei, M., Du, D.-Z., 2005. Improving wireless sensor network lifetime through Rawat, P., Singh, K.D., Chaouchi, H., Bonnin, J.M., 2014. Wireless sensor networks: a
power aware organization. Wirel. Netw. 11 (3), 333–340. survey on recent developments and potential synergies. J. Supercomput. 68 (1),
Carrabs, F., Cerulli, R., D’Ambrosio, C., Gentili, M., Raiconi, A., 2015. Maximizing life- 1–48.
time in wireless sensor networks with multiple sensor families. Comput. Oper. Sotskov, Y.N., Dolgui, A., Portmann, M.-C., 2006. Stability analysis of an optimal
Res. 60, 121–137. balance for an assembly line with fixed cycle time. Eur. J. Oper. Res. 168 (3),
Carrabs, F., Cerulli, R., D’Ambrosio, C., Raiconi, A., 2016. Extending lifetime through 783–797.
partial coverage and roles allocation in connectivity-constrained sensor net- Sotskov, Y.N., Tanaev, V.S., Werner, F., 1998. Stability radius of an optimal schedule:
works. IFAC-PapersOnLine 49 (12), 973–978. A survey and recent developments. In: Industrial applications of combinatorial
Carrabs, F., Cerulli, R., D’Ambrosio, C., Raiconi, A., 2017. An exact algorithm to ex- optimization. Springer, pp. 72–108.
tend lifetime through roles allocation in sensor networks with connectivity con- Đurišić, M.P., Tafa, Z., Dimić, G., Milutinović, V., 2012. A survey of military appli-
straints. Optim. Lett. 11 (7), 1341–1356. cations of wireless sensor networks. In: Embedded Computing (MECO), 2012
Castaño, F., Bourreau, É., Rossi, A., Sevaux, M., Velasco, N., 2016. Partial target cover- Mediterranean Conference on. IEEE, pp. 196–199.
age to extend the lifetime in wireless multi-role sensor networks. Networks 68 Werner-Allen, G., Lorincz, K., Ruiz, M., Marcillo, O., Johnson, J., Lees, J., Welsh, M.,
(1), 34–53. doi:10.1002/net.21682. 2006. Deploying a wireless sensor network on an active volcano. IEEE Internet
Castaño, F., Bourreau, É., Velasco, N., Rossi, A., Sevaux, M., 2015. Exact approaches Comput. 10 (2), 18–25.
for lifetime maximization in connectivity constrained wireless multi-role sensor Xu, L.D., Xu, E.L., Li, L., 2018. Industry 4.0: state of the art and future trends. Int. J.
networks. Eur. J. Oper. Res. 241 (1), 28–38. Prod. Res. 56 (8), 2941–2962.
Castaño, F., Rossi, A., Sevaux, M., Velasco, N., 2014. A column generation approach Yick, J., Mukherjee, B., Ghosal, D., 2008. Wireless sensor network survey. Comput.
to extend lifetime in wireless sensor networks with coverage and connectivity Netw. 52 (12), 2292–2330.
constraints. Comput. Oper. Res. 52, 220–230. Younis, O., Fahmy, S., 2004. Heed: a hybrid, energy-efficient, distributed clustering
CupCarbon, Cupcarbon. http://www.cupcarbon.com/. Accessed: 14-05-2018. approach for ad hoc sensor networks. IEEE Trans. Mob. Comput. 3 (4), 366–379.
Elhoseny, M., Tharwat, A., Yuan, X., Hassanien, A.E., 2018. Optimizing k-coverage of
mobile wsns. Expert Syst. Appl. 92, 142–153.