You are on page 1of 10

2015 44th International Conference on Parallel Processing

Joint Wireless Charging and Sensor Activity Management in Wireless


Rechargeable Sensor Networks
Yuan Gao, Cong Wang and Yuanyuan Yang
Department of Electrical and Computer Engineering, Stony Brook University, Stony Brook, NY 11794, USA
Abstract—Recent studies show that the novel wireless charging Although the above work solves the energy replenishment
technology can extend the lifetime of Wireless Sensor Networks problem in WRSNs, they ignore an important factor that how
(WSNs) towards perpetual operations. Recharging Vehicles (RVs) the sensor activity in the network is managed in accordance
can be applied in WSNs to recharge sensors conveniently via
wireless charging devices. Most of existing work focused only on with energy replenishment. Since sensors are usually densely
energy replenishment whereas ignored sensor activity management. deployed, taking advantage of redundancies in sensing coverage
In this paper, we propose a new framework that can jointly schedule could provide further energy saving. For example, in a moni-
sensor activity and recharging to save the traveling energy of RVs. toring task to detect the presence of rare animals, sensors are
First, we propose two schemes to manage sensor activity: balanced densely deployed to guarantee seamless coverage of the field.
clustering and distributed sensor activation schemes. We further
introduce a new metric so that the energy demand in each cluster Since each animal can be monitored by multiple sensors, some
can be managed. Then we formulate the recharging problem into redundant sensors can be switched off to save energy. How to
a Traveling Salesman Problem with Profits, which is NP-hard. For manage sensor activity along with recharge schedules of RVs to
the recharging route schedule, we first study the case of a single save system cost is an interesting and challenging problem.
RV by coordinating sensor activity and energy replenishment, and In this paper, we propose a new framework that consists
then extend it to multiple RVs using two different schemes. The
first scheme focuses on reducing traveling distance of RVs by of comprehensive sensor activity management schemes and
confining their moving scopes and the second one improves the recharge scheduling algorithms. We consider a group of targets
overall system performance by giving RVs a global view over the that move randomly in the field to be monitored continuously
entire network. Finally, we validate the correctness and evaluate by a number of sensor nodes. Based on the sensing coverage
the performance of the sensor activity management schemes along and the location of targets, sensors are first divided into clusters.
with the recharging algorithms by extensive simulations. Our results
indicate that significant reduction on system cost can be achieved. Each cluster contains the nodes that monitor the same target.
The sensor activity management schemes can save traveling energy A balanced clustering algorithm is proposed to organize sensors
of RVs by 16% while maintaining a reliable detection on targets. into clusters of almost equal size around targets. The purpose
Compared with a simple greedy algorithm, the first and the second of balancing cluster size is to prolong the working time of all
recharging schemes can save 41% and 13% traveling distance of clusters so the RVs can recharge sensors less frequently. Once the
RVs, and reduce nonfunctional nodes by 23% and 52%, respectively.
Index Terms—Wireless sensor network; wireless charging; sensor clusters are formed, the sensing load will be distributed evenly to
activity management; clustering; target detection; recharge schedul- each sensor, for example, in a round-robin manner. In this way,
ing. we can reduce energy consumptions on sensors while keeping
targets covered. Moreover, we introduce a parameter (defined in
I. I NTRODUCTION Section III) that can control when and how many sensors are
Recently, Wireless Sensor Networks (WSNs) have found a sending recharge requests, which can improve energy efficiency
wide range of applications in environmental monitoring, security of RVs greatly.
management, disaster relief, combat field surveillance, and space For recharge schedules, we first formulate the problem into
exploration [4]–[6]. With the revolutionary wireless charging a mixed integer programming problem, which is proved to
technology [1], sensor nodes can be recharged in distance be NP-hard. To tackle it, we first provide a simple greedy
by mobile Recharging Vehicles (RVs) equipped with charging algorithm as a baseline. The greedy algorithm dispatches RVs
devices [7]–[13]. Such networks are called Wireless Recharge- to recharge sensors with the maximum recharge profit (defined
able Sensor Networks (WRSNs). Compared with environmental in Section IV) in each step. However, it may incur long traveling
energy harvesting [2], [3], it is more reliable and the network distance for RVs. To overcome this problem, we further propose
performance is no longer affected by the dynamics in ambient a recharging algorithm for a single RV, and then extend it
energy sources. to multiple RVs case by two different schemes. For the case
With this advanced technology, a considerable amount of with a single RV, the algorithm first selects the node with
work has been done recently [7]–[11] to study how to apply the maximum recharge profit as the final destination and then
this technology in WSNs. In [7], joint energy replenishment iteratively inserts other nodes with less but comparable good
and data collection is considered, where a set of representative profits into the recharge route. In this way, the energy efficiency
sensor nodes are selected for recharge and the data from the of RV can be improved significantly. To extend the algorithm for
neighborhood is simultaneously collected by RVs. A framework multiple RVs, we propose two different approaches. In the first
for optimizing data communication in the network is established approach, we use the well-known K-means algorithm to partition
based on the recharging nodes. In [8], the optimization frame- nodes into groups and each RV is responsible for its assigned
work is extended to account for the time-varying recharge time. group of nodes. The second approach, in contrast, schedules the
A weighted sum algorithm for scheduling recharge routes of recharging routes for all RVs based on the global information.
multiple RVs is proposed in [9]. In [10], a recharge scheduling The main contribution of this paper can be summarized as
problem that considers practical constraints from both recharge follows. First, we propose to jointly consider sensor activity
capacity of RVs and node’s battery deadlines is studied. and energy replenishment in WRSNs and provide comprehen-

0190-3918/15 $31.00 © 2015 IEEE 790


789
DOI 10.1109/ICPP.2015.88
sive sensor activity management schemes to save the energy   
of sensors and reduce traveling distance of RVs. Second, we 

formulate the recharge scheduling problem into a mixed integer 
problem and propose recharge route scheduling algorithms to 
cooperate with the sensor activity management schemes. Finally,  

our extensive simulations demonstrate that the proposed algo- 



rithms can improve energy efficiency of RVs significantly and   
reveal possible trade-offs between energy saving and network   
performance. To the best of our knowledge, this is the first 
work that jointly considers wireless charging and sensor activity
management.
The rest of this paper is organized as follows. Section II
describes the network model. Section III introduces the sen-
sor activity management schemes. Section IV studies the joint
recharge scheduling and sensor activity problem and proposes Fig. 1. Network architecture.
the recharge routes scheduling algorithms. Section V validates with energy replenishment, we adjust sensor’s duty cycle by a
the correctness and evaluates the performance of the proposed load balance algorithm and make sensors in a cluster request
framework. Finally, Section VI concludes this paper. recharge at the same time. In this way, an RV can make one
move to recharge several nodes thereby avoiding long distance
II. P RELIMINARIES movement back and forth in the field. The network architecture
In this section, we elaborate our system model, along with is shown in Fig. 1.
some basic assumptions and important definitions. Various prac-
B. Random Sensor Deployment
tical applications of our model can be found in [14], and
some typical examples include detecting leakage of harmful In this paper, a random sensor deployment is adopted to
chemicals and monitoring natural phenomenons (such as seismic cover the sensing field. Although deterministic placement can
or volcanic activities). minimize the number of sensors to provide full coverage [16]–
[18], our choice of random deployment incurs much less labor
A. System Description cost. In the deterministic placement, a node is deployed at a fixed
We consider a network with N sensors uniformly randomly geographical location which requires accurate survey of sensing
distributed over a square field with side length L. To model a and communication coverage. Finding out the exact location
series of events such as natural phenomenons, we assume that to place each sensor node also incurs very high human labor
M targets appear randomly at any location in the field and stay cost. Second, in many applications, deploying nodes manually
for a duration defined as target period before appearing again at in remote or hazardous areas is infeasible. On the other hand,
new locations. A base station is located at the center of the field due to low manufacturing cost of sensor nodes, we can disperse
to collect sensing data from sensors. It also provides the power a large number of low cost sensor nodes from an airplane
source to recharge RV’s battery and performs basic computations or artillery [19]. To estimate the minimal number of sensors
such as calculating node traffic, lifetime, RV’s recharge routes, required to provide full coverage,√we can use results from [20].
etc. We apply m RVs to recharge sensors which consume em per 3 3Sa
N= , (1)
meter while moving at a constant speed vr . RVs replenish their 2π 2 r2
own batteries at the base station when the batteries are nearly where Sa is the area of the sensing field (Sa = L2 ), r is the
depleted. Long range radio modules can be launched on the RVs sensing range of sensors, and N is the minimum number of
to facilitate communications with the base station. sensors required to provide full coverage.
All sensors have the same sensing range of ds . A target
can be monitored by multiple sensors as long as it lies in III. S ENSOR ACTIVITY M ANAGEMENT
their sensing ranges. However, due to physical limits from the In this section, we describe how sensors should manage their
sensing detector, a sensor can only monitor one target at a time activities so that energy saving on the RVs can be achieved.
which consumes es energy per unit time. The sensing data is These activities include how to organize sensors into clusters
generated at a constant rate of λ packets per unit time and regarding dynamic target locations and how to cooperate with
transmitted to the base station in multi-hops. Sensors have the each other to monitor the targets while maximizing the recharge
same communication range of dc and consume et and er energy efficiency from the RVs.
while transmitting and receiving, respectively. All sensors have
capacity Ec and recharge threshold Eth . If a sensor’s energy falls A. Balanced Cluster Formation
below the recharge threshold, it will send out a recharge request The first step is to adaptively organize sensors into clusters
to the base station. The base station also maintains a recharge based on locations of targets. A cluster consists of sensors that
node list to store recharge requests. Then a recharge schedule cover the same target. One of the objectives in cluster formation
is calculated based on the updated recharge node list and the is to balance the number of sensors in each cluster. That is, every
recharge sequences would be disseminated to RVs. The recharge target will be covered by almost the same number of sensors.
time is modeled according to [15]. Sensors monitoring the same If the cluster sizes are unbalanced, smaller clusters will deplete
target are organized into a cluster. To coordinate sensor activity their energy much faster, thus they will send energy requests

790
791
Algorithm 1 Balanced Clustering Algorithm
      Input: Positions of all sensors and targets.


Output: Balanced sensor clusters C.
1: Phase 1:
 2: Initialize A = ∅, and |U| = M , value of zeros, where M is the
number of targets.
3: for i = 1, 2, . . . , M do
 
 
   4: P(i) = ∅.
5: C(i) = ∅.
6: end for
7: for i = 1, 2, . . . , M do
   8: for j = 1, 2, . . . , M do
9: if sensor j can detect target i then
10: Update P(i) = P(i) ∪ j.
Fig. 2. Organization of sensors after cluster formation. 11: end if
earlier and more frequently. This will increase the frequency of 12: end for
13: end for
recharge for smaller clusters and result in higher energy cost on 14: Update A = ∪M i=1 P(i).
RVs. Fig. 2 shows the organization of sensors after a balanced 15: Phase 2:
cluster formation is completed. 16: for i = 1, 2, . . . , |A| do
The clustering algorithm can be described as follows. After 17: [I] = sort(U, ‘ascending  ).
sensors are deployed, they report their locations to the base 18: for j = 1, 2, . . . , |I| do
19: Check if A(i) ∈ P(I(j)).
station as a preprocessing step. Then the base station assigns an 20: if YES then
identification (ID) number to each sensor. These IDs are used to 21: Update C(I(j)) = C(I(j)) ∪ A(i).
identify locations of different sensors. 22: Update U(I(j)) = U(I(j)) + 1.
Once a target is detected by a sensor, it is reported to 23: Break the inner loop, and continue from the outer loop.
24: end if
base station with its location. The base station conducts cluster 25: end for
formation in two phases. In phase one, the base station creates 26: end for
a set A to store sensors that can detect at least one target. The 27: return C.
load of a sensor is defined as the number of targets it can detect.
B. Energy Request Management
A is sorted in an increasing order according to sensors’ load. In
this way, we can give a higher priority to sensors with less load As an integrated part of managing sensor’s activity, when to
as they have fewer choices to join clusters. For each target, it request recharge in each cluster determines the energy efficiency
also creates a set P to store sensors that can detect that target. of the network. In this subsection, we leverage the exact time
A set U is initialized to keep track of the number of sensors to request energy recharge in order to achieve energy saving
in each cluster. In phase two, sensors will be assigned evenly on the RVs. In previous works [7]–[10], [12], [13], sensors
to cover each target. The base station begins the process by will request for energy replenishment once their battery energy
assigning sensors in the sorted set A to the targets one after falls below the threshold. Although such treatment of recharge
another. When the algorithm picks a sensor i ∈ A that is not requests is reasonable, we will show that network performance
assigned, it will check each target set P, and assign sensor i can be further improved by properly managing recharge requests.
to the first set P which contains it. If sensor i is assigned to Since the size of clusters is much smaller than the field size1 ,
target j’s P set, it will monitor target j until the target changes the energy cost on RVs due to intra-cluster movement is usually
location. The checking order of set P is in an ascending size much lower than the energy cost to travel through the entire
order of set U. In this way, we can fill clusters with smaller size field. Thus, if energy demand within a cluster can be adjusted,
first and balance the number of sensors in each cluster. an RV can make one move into the cluster and fulfill all energy
After a sensor is added into a cluster i, the counter U is requests rather than coming back and forth into the cluster at
updated and the algorithm proceeds to consider the next sensor different times. To adjust the energy demand within a cluster,
until all the sensors in A have been assigned. Then the algorithm effective management of when and how many sensors should
outputs a set of clusters C with similar size. After the clusters request energy replenishment is needed.
are generated, sensors will store a list of IDs of other cluster We leverage the time to request recharge by a metric called
members. The pseudocode for the balanced clustering algorithm Energy Request Percentage. It is defined as the maximum
is shown in Algorithm 1. allowable percentage of sensors in a cluster that have battery
The computation complexity of the algorithm can be analyzed energy fallen below the recharge threshold without sending any
as follows. In phase one, it takes O(M N ) time to build set recharge request. Once this percentage is reached, sensors in
P for all targets, where M is the number of targets and N is the cluster will be allowed to send out their energy requests. By
the number of sensors. In phase two, every time the algorithm adjusting this value, we can manage when and how many sensors
proceeds a new node in A, it needs to sort U in an ascending are being added to the recharge node list. A higher energy request
order, and loop the P set for every element in U to check the percentage indicates that sensors will postpone their requests
possibility of assignment. Hence the time complexity for phase until the percentage is reached so that higher energy demand for
two is O(|A| · M log M ). Therefore, the total time complexity the cluster is achieved. Each time an RV is scheduled to recharge
of the clustering algorithm is O(M N + |A| · M log M ), which 1 The diameter of a cluster is at most twice of the sensing range, which is
is bounded by O(M N log M ). about 8-16 m. The sensing field usually extends to several hundreds of meters.

791
792
the cluster, it would be able to recharge more sensors, hence
  
reduce the recharging times for the cluster, and save traveling  

cost on the RVs. !
" 
We now analyze how much energy saving such energy request 
management can bring. Denote the number of sensors in a cluster  

as nc (nc ≥ 1), the distance from the base station to that cluster  

as dist, and the energy request percentage as K(0 ≤ K ≤ 1). 



For the scheme previously used in [7]–[10] in which sensors   

will request energy replenishment as long as their energy drops 



# 
below threshold, the traveling energy for an RV to recharge 
this cluster is 2nc · dist · em in the worst case, where em
is the traveling energy consumption rate of the RV. However,
with our energy request management, the traveling energy is
2nc / max(nc K, 1) · dist · em . Ascan be seen, by adjusting (a) Energy replenishment without sensor activity management.
K values, the traveling energy of an RV can be managed. When
K = 1, the energy cost of an RV is only 1/nc of the former,   


indicating that a considerable amount of energy on RVs can be
saved by energy request management. 

Note that reducing the moving cost on RVs by postponing



energy request time might have an impact on network perfor-  

mance. A larger K will keep more sensors staying in low energy




status and prone to deplete their battery energy. Hence, the metric   
cannot be increased indefinitely. The trade-off between network


performance and energy saving will be examined via simulations. 

C. Sensor Activation Scheme


To take full advantage of the overlaps in sensing coverage,
we adjust the sensing activities of sensors to balance energy
consumption in each cluster. The algorithm activates sensors (b) Energy replenishment with sensor activity management.
to monitor targets in a distributed round-robin fashion. Recall
Fig. 3. Impact of managing activity on recharge routes.
that during cluster formation, each node has stored IDs of other
members in the cluster. The sensor activation scheme begins working sensor, targets are more likely left unattended compared
with the sensor having the lowest ID to monitor the target for a with the case when all sensors are active. Thus, it is more
time slot. Then the node sends out a notification packet to the sensitive to the recharge scheduling. Besides, if K is large, it will
next sensor in the cluster. If there is no acknowledge for the extend the recharge latency after a sensor has actually requested
notification packet, it indicates that the next sensor might have for recharge. When the number of energy requests is large, the
depleted its battery energy. In this case, the node will send out a RVs cannot resolve all the requests at once. In this situation,
new notification packet to the next available node. The process sensors may deplete battery energy and not be able to cover the
continues until the target changes its location. targets. To prevent this problem, clusters with low energy will
Intuitively, compared with activating sensors all the time, be prioritized in the recharge routes so that reliable detection of
which is used in [7]–[10], adopting the round-robin activation targets can be guaranteed.
scheme will reduce the energy consumption rate of sensors in the
cluster by nc times, and thus decrease the recharging times for
RVs. Besides, the round-robin activation manner can balance the IV. J OINT R ECHARGE S CHEDULING AND S ENSOR ACTIVITY
sensing load evenly among all sensors in the cluster, thus achieve M ANAGEMENT
high energy efficiency, which can further reduce the traveling
energy of RVs. In this section, we study the Joint Recharge Scheduling
An example illustrating the advantage of our joint approach and Sensor Activity Management problem (JRSSAM). How to
is given in Fig. 3. As shown in Fig. 3(a), without sensor schedule RVs to recharge directly affects the energy efficiency
activity management, sensors will send energy requests when of the network. As there is only one active sensor in a cluster to
their energy falls below the recharge threshold. A possible case monitor the target, meeting sensor’s battery deadline is particu-
in the example is that sensors 1 to 3 will send requests first, and larly important in guaranteeing coverage. This requires coordi-
followed by sensors 4 to 6 and 7 to 9. In this case, an RV needs nations between sensor activities and energy replenishment. For
to travel 3 rounds to recharge all the nodes. In contrast, with the each recharge, our objective is to maximize the amount of energy
proposed sensor activity management shown in Fig. 3(b), sensors replenished into the network while minimizing the moving
will send a combined recharge request as a cluster and an RV energy cost on RVs. We first formulate the optimization problem
only needs one round to recharge all the sensors. Therefore, a into a mixed integer programming problem and then show that
great amount of moving energy on the RV is saved. it is NP-hard by reducing it from the well-known Traveling
However, the round-robin activation scheme brings a potential Salesman Problem with Profits [22] and propose algorithms to
issue to the network. Since at any moment there is only one cope with it.

792
793
A. Problem Formulation TABLE I
N OTATIONS
Notation Definition
The JRSSAM problem can be formulated into the following
S Set of sensor nodes, |S| = N .
optimization problem. We denote the set of sensors by S =
{s1 , s2 , . . . , sN } and the set of targets by T = {t1 , t2 , . . . , tM }. T Set of targets, |T | = M .
A set of RVs H = {h1 , h2 , . . . , hm } traverse the field for energy H Set of RVs, |H| = m.
replenishment. The energy requests from sensors are stored in a R Recharge node list (nodes need recharge), |R| = n.
recharge node list R = {r1 , r2 , . . . , rn }. di Energy demand of sensor node i, i ∈ R.
The network can be represented by a graph G = (V, E), where Cr Energy capacity of RVs.
V represents the set of sensors and the base station (v0 ) and E Iij Binary variable, equals 1 if target j is covered by sensor
represents the paths among those sensors. Edge (i, j) is the path i; otherwise, it is 0, i ∈ S, j ∈ T .
between sensor i and sensor j, ∀i, j ∈ V. xa Binary variable, equals 1 if edge (i, j) is in the recharg-
ij
Energy demand di of sensor i is defined as the difference ing route of RV a; otherwise, it is 0, i, j ∈ R, a ∈ H.
between its full battery capacity and current energy. Binary yia Binary variable, equals 1 if sensor i is recharged by RV
a; otherwise, it is 0, i ∈ R, a ∈ H.
variable Iij represents whether or not sensor i detects target j.
If target j is detected by sensor i, Iij is 1; otherwise, it is 0. cij Traveling cost of path (i, j), which is proportional to
the length of path (i, j), i, j ∈ S.
For every pair of sensors i and j, binary variable xaij denotes
whether path (i, j) is selected for RV a. If the path is selected, yia ∈ {0, 1}, ∀i ∈ R, ∀a ∈ H (11)
xaij is 1; otherwise, it is 0. yia is a binary variable which equals Iij ∈ {0, 1}, ∀i ∈ S, ∀j ∈ T (12)
1 if sensor i is recharged by RV a; otherwise, it is 0. ui is the
2 ≤ ui ≤ n, ∀i ∈ R (13)
position of sensor i in the recharging path. cij is the traveling
cost of path (i, j), which is proportional to the distance between ui − uj + (n − m)xaij ≤ n − m − 1
sensors i and j. We denote the recharge capacity of each RV as ∀i, j ∈ R, i = j, ∀a ∈ H (14)
Cr . Table I summarizes the definitions of the notations.
In the above formulation, constraint (3) indicates that for every
Our goal is to maximize the energy recharged by RVs into RV, it must start its recharging tour from the base station and end
the network while minimizing their total traveling energy con- at the base station (node v0 ). Constraint (4) ensures that every
sumption, i.e., the total energy demand minus the total traveling node to be recharged by the RVs must have an incoming arc
energy from the RVs (recharge profit). and an outgoing arc. Constraint (5) restricts that every sensor
can only monitor at most one target. Constraint (6) guarantees
 n
m  m 
 n 
n every target will be monitored by at least one sensor. Constraint
Maximize : yia di − cij xaij (2) (7) ensures that the capacity of RV should not be violated.
a=1 i=1 a=1 i=1 j=1 Constraints (8) and (9) state that every sensor node can be
recharged by at most one RV and every RV will recharge at
least one sensor node during a recharging tour. Constraints (10)
Subject to to (12) specify that xaij , yia and Iij are all 0-1 valued. Constraints
(13) and (14) eliminate the subtour in the recharge routes, which
n
 n
 is formulated according to [21].
xa0j = xai0 = 1, ∀a ∈ H (3) A special case of the optimization problem is to allow RVs
j=1 i=1 to have infinite energy capacity so the problem is reduced
n n to the Traveling Salesman Problems (TSP) with Profits. TSP
xaik = xakj = yka , ∀k ∈ R, k = i, j, ∀a ∈ H (4) with Profits defines a group of salesmen who visit cities to
i=1 j=1 collect prizes. The objective is to maximize the total prize minus
M traveling cost. Since a special case of our problem reduces to

Iij ≤ 1, ∀i ∈ S (5) TSPs with Profits, which is known to be NP-hard [22], our
j=1
problem is NP-hard as well. Due to the nature of the problem,
N
seeking optimal solutions could be very difficult. To this end,
 we propose a heuristic algorithm to cope with it.
Iij > 0, ∀j ∈ T (6)
i=1 B. Greedy Algorithm
n n 
 n
di yia + cij xaij ≤ Cr , ∀a ∈ H (7) An intuitive approach to solving the problem is to maximize
i=1 i=1 j=1
the recharge profit at each step. To do that, the algorithm selects
m the node with the maximum recharge profit in each round.
yia ≤ 1, ∀i ∈ R (8) However, this simple greedy approach could be inefficient in
a=1 some cases. For example, in situations that the node with the
n maximum profit happens to be the farthest one from the RV’s
yia > 0, ∀a ∈ H (9) location, The RV must travel a long distance to recharge only
i=1 a single node. A great amount of energy is wasted due to
xaij ∈ {0, 1}, ∀i, j ∈ R, a ∈ H (10) the poor recharge schedule. To provide a baseline for further

793
794
improvement, we present the greedy algorithm in Algorithm 2. moving distance after inserting R(j) between Q(i) and Q(i+1).
The node with the maximum recharge profit is defined as the Positive p(i, j) means an insertion can improve RV’s energy
destination node for the RV. efficiency, while negative p(i, j) should be avoided. From all
the p(i, j) values, the algorithm selects the most positive one,
Algorithm 2 Greedy Recharging Algorithm and inserts the corresponding node from R into the position in
Input: recharge node list R, sensors’ positions, a set of RVs’ current Q that results in maximum profit. After the new node is inserted,
nodes F, |F| = m, and energy demand of sensors D. the RV checks its own energy level. If its battery is low, it returns
1: Initialize every RV’s f lag = 1, indicating every RV has enough
energy. to the base station to replenish its own battery.
2: while R =  ∅ and ∃f lag = 1 do The above procedure is repeated until there are no more nodes
3: for i = 1 to i = m do can be inserted or the recharge capacity of RV is violated. So
4: Set hi ’s f lag = 0. far, a recharging sequence Q is generated for a single RV. After
5: if recharge threshold of hi is not violated then
6: Set hi ’s f lag = 1. the RV finishes its current recharging sequence, the algorithm is
7: Calculate distances from all nodes in R to Fi , and store repeated until all the nodes in R are recharged. The pseudocode
them in set W. of the algorithm is shown in Algorithm 3.
8: Choose the destination node, [dest] = max(D − em W),
here em is the traveling energy consumption rate of RV.
9: Send hi to recharge the dest node. Algorithm 3 Calculating Recharging Sequence Algorithm
10: Update energy information of hi . Input: recharge node list R, sensors’ positions, RV’s current node crt,
11: Update position information of hi . and energy demand of sensors D.
12: Update R = R \ dest. Output: A recharging sequence Q consisting of sensor nodes for an
13: end if RV.
14: end for 1: Calculate distances from all nodes in R to crt, store them in W.
15: end while 2: Choose the dest node with the best recharge profit, [dest] =
max(D − em W), here em is the traveling energy consumption
rate of RV.
3: Initialize recharging sequence as Q = [crt → dest].
Next, we further propose a joint recharge scheduling algorithm 4: Update RV’s information to reserve energy for the dest node.
that takes advantage of sensor activities to improve energy 5: Update R = R \ dest.
efficiencies in the network. We first consider the case with 6: Label insert = 1, meaning one insertion has been performed.
a single RV and then extend it to coordinate multiple RVs 7: while insert = 1 and R = ∅ and RV has enough energy for the
using two different schemes. The first scheme aims to reduce dest node do
8: Label insert = 0.
RV’s traveling distances by confining its moving scope. The 9: for every possible position s from crt to dest do
second scheme focuses on how to improve the overall system 10: for every possible node n in the recharge node list R do
performance by enabling RVs to access global information of 11: Calculate profit difference p(s, n) = D(R(n))−em Δd(s).
the network. 12: end for
13: end for
C. Joint Recharge Scheduling for a Single RV 14: Select the maximum profit difference [s0 , n0 ] = max(p).
15: if p(s0 , n0 ) > 0 then
First, we introduce the algorithm for a single RV. We will 16: Update Q by inserting n0 at position s0 .
extend it to multiple RVs in the next subsection. As discussed 17: Update RV’s energy information.
above, in the greedy algorithm, RVs would spend a large amount 18: Update R = R \ n0 .
of energy on traveling. Thus, how to make use of that part 19: Label insert = 1.
of energy can make a great difference in the network energy 20: end if
21: end while
efficiency. When an RV is heading to the destination node with 22: return Q.
the maximum recharge profit, there could be other sensor nodes
near the traveling path that have lower but comparatively good
recharge profits. If the recharging path can also cover these In addition, there are some other important aspects to consider
nodes, the overall energy efficiency can be improved signifi- while creating the recharging sequence. To guarantee a reliable
cantly. Based on this observation, we propose to insert nodes detection on targets, all energy demands from sensors inside a
iteratively between RV’s current location and the destination cluster are replaced by an aggregated cluster energy demand.
node. Thus, the more energy demand from a cluster, the higher
Specifically, after the destination node dest is decided, it is probability it will be selected to recharge. Further, to prevent
added into the recharging sequence Q = [crt → dest], where RV from moving back and forth to the same cluster, which is
crt is RV’s current position. Each time the algorithm inserts energy inefficient, the RV recharges all the nodes in the cluster
a node between crt and dest. With one more node inserted, that are in the recharge node list R. The recharging tour inside
Q is elongated and one more possible position for the next a cluster is guided by a canonical TSP algorithm, such as the
insertion is available. For example, in the second insertion, nearest neighbor algorithm with time complexity O(n2c ) [24],
we calculate the recharge profit for inserting each node in R where nc is the number of sensors in the cluster.
except dest into the only available location between crt and
dest, and pick the best candidate to perform the insertion. In D. Joint Recharge Scheduling for Multiple RVs
general, for inserting the j-th node in R between Q(i) and
Q(i + 1), the change in recharge profit p(i, j) is calculated as In this subsection, we discuss how the algorithm can be
p(i, j) = D(R(j)) − em Δd(i), where Δd(i) is the increased extended to multiple RVs and present two different schemes.

794
795
1) Partition Nodes into Groups: In the first scheme, we and the destination node. The algorithm needs to iterate through
partition all sensors in the recharge node list into m groups, the remaining n − 1 nodes in the recharge node list and select
with the number of groups equal to the number of RVs. In this the best candidate. Thus, the time complexity of inserting the
way, each RV is only responsible for one sensor group. Within second node is O(n − 1).
each group, the RV can follow Algorithm 3 to perform recharge. Note that with one more node being inserted, one more
The partition restricts the RVs to move within their designated possible position for the next insertion is created, which means
areas and thus prevents them from traveling over long distance. that the number of iterations to test each node is also increased
Henceforth, we call this Partition-Scheme. by one after one insertion. Therefore, the total time complexity
In particular, we adopt the well-known K-means [23] method for building the entire recharging sequence is,
to perform the partition. Suppose there are n sensors in the n + 1 · (n − 1) + 2 · (n − 2) + · · · + (l − 1) · (n − l + 1)
recharge node list, and the number of RVs is m (m < n).
K-means method aims to partition the n sensors into m sets = 2n + 2n + · · · + (l − 1)n − (12 + 22 + · · · (l − 1)2 )
S = {S1 , S2 , . . . , Sm } so that the Within-Cluster Sum of (1 + l)ln (l − 1)l(2l − 1)
= −
Squares (WCSS) is minimized. In other words, its objective is  2 6
to find O(n) if l = O(n),
k  = (16)
argmin ||x − μi ||2 (15) O(n3 ) if l = O(n).
S i=1 x∈Si Then the total time complexity for recharging the complete
where μi is the mean value of sensor positions in Si by taking recharge node list is,

an average of their x, y coordinates and μi is used as the starting O(n · nl ) = O(n2 ) if l = O(n),
position for RV i. As we can see from the above equation, after (17)
O(n3 · nl ) = O(n3 ) if l = O(n).
applying the K-means method, each RV will be assigned with
a group of sensors that are geographically close to each other. Moreover, assume that the maximum number of nodes in clusters
Partition in this manner will further reduce the traveling energy is nc and there are M targets. Then the time complexity of
of RVs. recharging those clusters is O(M n2c ). Hence, the total complex-
2) Combined Scheduling: The second scheme is to schedule ity of the recharging algorithm for a single RV is,

RVs throughout the entire recharge node list. Initially, we find the
O(n2 + M n2c ) if l = O(n),
recharging sequence for the first RV using Algorithm 3. Those T (l, n) = (18)
nodes that are in the recharging sequence of the first RV are O(n3 + M n2c ) if l = O(n).
removed from the recharge node list before finding recharge For the case with multiple RVs, there are two different schemes.
sequence for the second RV. This procedure is repeated for all For the Partition-Scheme, it first divides the sensors into m
other RVs. Whenever an RV finishes recharging all the nodes groups, so that every RV is in charge of one group. Each
in its sequence, the same algorithm is applied again on the group has an average size of n/m, so the time complexity for
remaining nodes until the recharge node list is exhausted. We recharging each group is T (l, n/m). For a total number of m
denote this scheme as Combined-Scheme. groups, the complexity of K-means method is O(nmk) [23],
Note that there is a fundamental difference between the where k is the number of iterations for the K-means algorithm
two schemes. In the Partition-Scheme, RV’s moving scope is to converge. The total time complexity for Partition-Scheme is
geographically confined and each RV is only responsible for the ⎧ n 2

⎪ O(m( m ) + M n2c + nmk)
nodes in a designated area. On the other hand, the Combined- ⎪
⎨ 2
Scheme forms the recharge sequence from a global view so the = O( nm + M n2c + nmk) if l = O(n),
n 3 2 (19)
RVs may potentially consume more energy, but it will bring some ⎪
⎪O(m( m ) +3 M nc + nmk)

⎩ n
performance benefits. The impacts on the network performance = O( m 2
2 + M nc + nmk) if l = O(n).
and cost are compared in the simulations.
For the Combined-Scheme, since each RV’s recharging se-
E. Complexity Analysis quence is scheduled based on the whole recharge node list,
the time complexity of scheduling all RVs is mT (l, n). The
We now analyze the time complexity of the proposed recharge
number of rounds to recharge the whole recharge node list is
scheduling algorithms. Suppose the length of the recharge node
n/(ml). Thus, the total time complexity of Combine-Scheme is
list is n, and the number of RVs is m. For the greedy algorithm,
(n/(ml))mT (l, n) = (n/l)T (l, n), which is
choosing the destination node costs O(n) time, since the algo- ⎧
rithm needs to iterate every node once to select the maximum ⎪
⎪ O(( nl )n2 + M n2c )


recharging profit. Scheduling all RVs costs O(mn) time. To = O(n3 + M n2c ) if l = O(n),
recharge all n nodes in the recharge node list, it takes a total n (20)
⎪O(( l )n3 + M n2c )

of O(mn(n/m)) = O(n2 ) time. ⎪

Next, let us analyze the algorithm complexity for a single RV = O(n3 + M n2c ) if l = O(n).
first. Assume the length of the recharging sequence is l (l ≤ n). From Eqs. (19) and (20), we can see that if the number of
Inserting the first destination node takes O(n) time since it needs iterations until convergence is not too large (say, k < n2 /m),
to calculate the recharge profit for each node iteratively, and the Partition-Scheme often has lower complexity, since the
selects the best among them. For the second insertion, there is Combined-Scheme needs to schedule recharging routes for all
only one available position, i.e., between RV’s current position the RVs globally at each time.

795
796
TABLE II Evaluations on Sensor Activity Management
PARAMETER S ETTINGS
No ERC−Full time
Parameters Values 4

Total Traveling Energy (MJ)


No ERC−With RR
Number of sensors N 500 With ERC−Full time
Number of target M 15 3 With ERC−With RR

Number of RVs m 3
Side length of sensing field L 200m 2

Transmission range dc 12m


1
Sensing range ds 8m
Simulation time 120 days
0
Target period 3 hours Greedy Partition Combined
Threshold for sensors Eth 50% of Ec
Moving consumption rate for RVs em 5.6J/m Fig. 4. Impact of sensor activity management on RV’s moving cost.
Traveling speed of RVs vr 1m/s most, which validates our analysis that sensor activity manage-
ment can reduce energy consumption on RVs.
V. P ERFORMANCE E VALUATION
Compared with activating sensors in full time, the round robin
In this section, we present simulation results on the perfor- scheme also reduces the moving energy cost of RVs as seen
mance of the proposed algorithm based on the system parameters in the difference between “Full time” and “With RR.” This
of real systems. In our simulation, there are a total of N = 500 is because that the round robin activation scheme can balance
sensors uniformly and randomly deployed over a square field sensor’s working load so that the energy consumption of sensors
with the side length L = 200 m. A group of M = 15 targets is decreased and more traveling energy is saved for RVs. The
randomly appear in the field and stay for a period of 3 hours difference between “With ERC” and “No ERC” reveals that the
before appear again at another random location. All sensors ERC can also bring significant improvement on energy efficiency
are equipped with CC2480 communication modules [25] which of RVs, since they can make one move into the cluster and
automatically enters low power mode (< 5μA) in idle period and recharge more sensors. Note that the advantages of the sensor
draws 27mA current while transmitting or receiving a packet. activity management will become more evident if there are more
It has a power supply range from 2.0V to 3.6V . Here we targets, as a larger percentage of sensors will benefit from it.
take 3V , which is equivalent to two AAA Panasonic Ni-MH
batteries [15]. All sensors generate data packets at a constant rate B. Trade-off between Energy Efficiency and Network Perfor-
of λ = 15pkt/min with packet length of 20 bytes. The routing mance
path is calculated using Dijkstra’s shortest path algorithm. To According to our analysis, there is a potential trade-off be-
detect targets, sensors are equipped with Passive Infrared (PIR) tween energy efficiency and network performance. That is, with
motion detector, which draw an average of 10mA current at 3V a higher Energy Request Percentage (ERP), RVs can save more
when they are active, and an average of 170μA current when they traveling energy. On the other hand, higher ERP potentially
are idle [26]. The recharge thresholds for sensors are set to be increases the risks of sensors to deplete battery due to late
50% of their full energy capacity. The RVs consume 5.6J/m recharge, which directly leads to missing targets. Here, we
while moving at a constant speed of 1m/s. The simulation demonstrate this trade-off in Fig. 5. Since our simulation shows
is evaluated over 120 days. Table II summarizes the system that the recharge scheduling algorithms have similar results, we
parameters used in simulations. provide the result of the greedy algorithm as an example.
As we can see, RV’s traveling energy declines with a larger
A. Benefits of Sensor Activity Management
ERP. By contrast, a larger ERP leads to higher target missing
First, we evaluate the benefits brought by the proposed sensor rate because some nodes have depleted battery energy, since
activity management. In particular, we compare the moving a larger ERP value can effectively keep more nodes in a low
energy cost of RVs in four cases. In the first case, all sensors in energy status before calling the next recharge. Therefore, finding
clusters are working full time without Energy Request Control an appropriate ERP value is important in practice to achieve
(ERC). This case is denoted as “No ERC-Full time.” In the a reasonable balance between energy efficiency and network
second case, sensors in a cluster are activated in a Round Robin performance. The results in Fig. 5 indicate that when ERP is
(RR) manner without ERC, denoted as “No ERC-with RR.” higher than 0.6, the target missing rate jumps above zero and
In the third case, all sensors in a cluster are activated at all target detection cannot be guaranteed all the time.
times with ERC. This case is denoted as “With ERC-Full time.”
Finally, in the fourth case, sensors are activated in a round robin C. Evaluations on Recharging Schemes
fashion with ERC, denoted as “With ERC-with RR.” Here, ERC In this subsection, we evaluate and compare the performance
is achieved by setting the Energy Request Percentage (ERP) to of the proposed recharge scheduling algorithms.
a value between 0 and 1, e.g., 0.6. As shown in Fig. 6(a), a larger ERP results in more sensor
Fig. 4 shows the comparisons among the four cases. As we nodes in a cluster to request recharge each time so that each
can see, for all three recharge scheduling algorithms, the scheme movement of the RV can replenish more sensors. In addition, the
“With ERC-with RR” consumes the least traveling energy of RV also has less chance to return to the same cluster for energy
RVs, whereas the scheme “No ERC-Full time” consumes the replenishment in future which leads to less traveling cost overall.

796
797
Evaluations on Trade−off Traveling Energy of RVs vs. ERP Values
3.2 12
3.5
Traveling Energy
Partition−Scheme
Missing Rate 10
Combined−Scheme

Traveling Energy (MJ)


Traveling Energy(MJ)
3 3 Greedy Algorithm

Missing Rate(%)
8

2.8 6 2.5

4
2.6 2
2

2.4 0
1.5
0 0.2 0.4 0.6 0.8 1 0 0.2 0.4 0.6 0.8 1
ERP Value ERP Value
(a) Traveling Energy of RVs.
Fig. 5. Evaluations on trade-off between energy efficiency and coverage ratio.
Average Coverage Ratio vs. ERP Values
It is worth noting that, among all the recharging schemes, the 98
Partition-Scheme saves the most traveling energy on RVs, 41%
96
more than greedy algorithm and 32% more than the Combined-

Coverage Ratio (%)


Scheme. This is because that the K-means algorithm can partition 94
the network into different areas efficiently and the movement of
each RV is confined in its designated area. 100

In Fig. 6(b) and 6(c), with ERP value becomes larger, more 98 Partition−Scheme
sensors stay in low energy status and energy replenishment is
Combined−Scheme
postponed. This may reduce the robustness of the network due to 96
Greedy Algorithm
energy depletion, link error and lost of target coverage. Note that
94
the Combined-Scheme has the least percentage of nonfunctional 0 0.2 0.4 0.6 0.8 1
sensor nodes, 52% less than the greedy scheme and 38% less ERP Value

than Partition-Scheme. This is because that in the Combined- (b) Average Coverage Ratio of Targets.
Scheme, RVs can acquire global knowledge over the entire Avg. Pct. of NS vs. ERP Values
5
recharge node list and have the privilege to insert nodes with
high recharge profit into the recharge sequence.
Nonfunctional Pct. (%)

4
To further evaluate the recharging schemes, we introduce a
new metric called recharging cost, which is defined as the ratio
3
of the total traveling distance of RVs to the average number of
operational sensors over the entire simulation time. Its physical 2
meaning can be interpreted as the average distance RVs have to
travel to maintain perpetual operation of each sensor. The results 1
Partition−Scheme
Combined−Scheme
of recharging cost are shown in Fig. 6(d). It can be seen that the
Greedy Algorithm
Partition-Scheme has much lower recharging cost, which is 32% 0
less than the Combined-Scheme and 41% less than the greedy 0 0.2 0.4 0.6
ERP Value
0.8 1

scheme on average. Note that a higher ERP value also reduces


(c) Average Percentage of Nonfunctional Sensors in
the recharging cost. This is due to that the benefits brought by Recharging Schemes.
energy saving outweigh the drawbacks for nonfunctional nodes. Recharging Cost vs. ERP Values
It also validates our choice to introduce ERP for improving RV’s
energy efficiency. Partition−Scheme
Recharging Cost (m/sensor)

1200 Combined−Scheme
D. Evaluations of Recharge Profit Greedy Algorithm
1000
Finally, we examine the recharge profit (defined by expression
(2) achieved by the recharging schemes. 800
As shown in Fig. 7(a), with a larger ERP value, there will
be fewer nodes added to the recharge node list which leads to
600
less energy recharged into the network. The Combined-Scheme
has the most energy recharged into the network, because it gives
400
RVs a wider range to select nodes for recharging. It is interesting 0 0.2 0.4 0.6 0.8 1
to see that with small ERP values (0 − 0.4), even the greedy ERP Value
scheme can put more energy into the network than the Partition- (d) Recharging Cost of Recharging Schemes.
Scheme. Since in the greedy scheme, the RVs are not constrained
Fig. 6. Performance comparisons between the recharging schemes.
by any particular group of sensors. When more energy requests
are received with a smaller ERP, RVs can select nodes with the
maximum profit. However, when ERP grows larger, the number

797
798
Energy Recharged vs. ERP Values improvements over a simple greedy approach on both energy
112
efficiency and network performance.
Partition−Scheme
Energy Recharged (MJ)
110 Combined−Scheme VII. ACKNOWLEDGMENTS
Greedy Algorithm The work in this paper was supported in part by the grant from
108 US National Science Foundation under grant number ECCS-
1307576.
106 R EFERENCES
[1] A. Kurs, A. Karalis, R. Moffatt, J. D. Joannopoulos, P. Fisher and
104 M. Soljacic, “Wireless Power Transfer via Strongly Coupled Magnetic
Resonances,” Science, vol. 317, pp. 83-86, 2007.
[2] M. Rahimi, H. Shah, G. S. Sukhatme, J. Heideman and D. Estrin, “Studying
102 the Feasibility of Energy Harvesting in a Mobile Sensor Network,” IEEE
0 0.2 0.4 0.6 0.8 1
ERP Value International Conference on Robotics and Automation, Sept. 2003.
[3] D. Niyato, E. Hossain, M. M. Rashid and V. K. Bhargava, “Wireless Sensor
(a) Total Energy Recharged of Three Recharging Schemes. Networks with Energy Harvesting Technologies: A Game-theoretic Ap-
Objective Score vs. ERP Values proach To Optimal Energy Management,” IEEE Wireless Communications,
vol. 14, no. 4, Aug. 2007.
110 [4] M. Ma and Y. Yang, “SenCar: An Energy-efficient Data Gathering Mecha-
Partition−Scheme nism for Large-scale Multihop Sensor Networks,” IEEE Trans. Parallel and
108 Combined−Scheme Distributed Systems, vol. 18, no. 10, pp. 1476-1488, 2007.
Objective Score (MJ)

Greedy Algorithm [5] M. Ma, Y. Yang and M. Zhao, “Tour Planning for Mobile Data-gathering
Mechanisms in Wireless Sensor Networks,” IEEE Trans. Vehicular Technol-
106 ogy, vol. 62, no. 4, pp. 1472-1483, 2013.
[6] C. Ma, Z. Zhang and Y. Yang, “Battery-aware Scheduling in Wireless Mesh
104 Networks,” Mobile Networks and Applications, vol. 13, no. 1-2, pp. 228-241,
2008.
[7] M. Zhao, J. Li and Y. Yang, “A Framework of Joint Mobile Energy Replen-
102 ishment and Data Gathering in Wireless Rechargeable Sensor Networks,”
IEEE Trans. Mobile Computing, vol. 13, no. 12, pp. 2689-2705, Dec. 2014.
100 [8] S. Guo, C. Wang and Y. Yang, “Joint Mobile Data Gathering and Energy
0 0.2 0.4 0.6 0.8 1 Provisioning in Wireless Rechargeable Sensor Networks,” IEEE Tras. Mobile
ERP Value computing, vol. 13, no. 12, pp. 2836-2852, Dec. 2014.
(b) Recharge Profit Achieved by Three Recharging [9] C. Wang, J. Li, F. Ye and Y. Yang, “NETWRAP: An NDN Based Real-
Schemes. Time Wireless Recharging Framework for Wireless Sensor Networks,” IEEE
Trans. Mobile CompSuting, vol. 13, no. 6, pp. 1283-1297, June 2014.
Fig. 7. Evaluation of recharge profit. [10] C. Wang, J. Li, F. Ye and Y. Yang, “Recharging Schedules for Wireless
Sensor Networks with Vehicle Movement Costs and Capacity Constraints,”
of nodes that actively request for recharge is reduced. Thus, IEEE SECON, 2014.
the disadvantage of the greedy scheme is more evident and a [11] Y. Yang and C. Wang, Wireless Rechargeable Sensor Networks, Springer,
2015.
considerable amount of energy is wasted due to traveling. [12] L. He, L. Fu, L. Zheng, Y. Gu, P. Cheng, J. Chen and J. Pan, “ESync: An
Fig. 7(b) shows the recharge profit achieved by the recharging Energy Synchronized Charging Protocol for Rechargeable Wireless Sensor
schemes. We can see that the Combined-Scheme has the highest Networks,” ACM MobiHoc, 2014.
[13] S. He, J. Chen, F. Jiang, Yau, D.K.Y., G. Xing and Y. Sun, “Energy Provi-
profit. When ERP is small, due to less amount of energy being sioning in Wireless Rechargeable Sensor Networks,” IEEE Transactions on
recharged into the network, the Partition-Scheme has a lower Mobile Computing, vol. 12, no. 10, pp. 1931-1942, Oct. 2013.
recharge profit than the greedy scheme. However, when ERP [14] I. F. Akyildiz, W. Su, Y. Sankarasubramaniam and E. Cayirci, “Wireless
Sensor Networks: A Survey,” Computer Networks, vol. 38, no. 4, pp. 393-
goes up, it results in much lower traveling energy and higher 422, Mar. 2002.
recharged energy into the network in the Partition-Scheme. [15] Panasonic Ni-MH Batteries Technical Handbook,
According to Eq. (2), the combined effect of these factors makes “http://www2.renovaar.ee/userfiles/Panasonic Ni-MH Handbook.pdf.”
[16] K. Kar and S. Banerjee, “Node Placement for Connected Coverage in
it outperform the greedy scheme. Sensor Networks,” IEEE WiOpt, 2003.
[17] X. Bai, S. Kumar, D. Xuan, Z. Yun and T. H. Lai, “Deploying Wireless
VI. C ONCLUSIONS Sensors to Achieve Both Coverage and Connectivity,” ACM MobiHoc, 2006.
In this paper, we study joint management of sensor activities [18] Y. Wang and Y. Tseng, “Distributed Deployment Schemes for Mobile
Wireless Sensor Networks to Ensure Multilevel Coverage,” IEEE Trans.
and recharge scheduling in WRSNs. We introduce a load balanc- Parallel and Distributed Computing, Sept. 2008.
ing sensor activation scheme and adjust the energy request time [19] D. Estrin, R. Govindan, J. Heidemann and S. Kumar, “Next Century
in order to provide energy saving on the RVs. We formulate the Challenges: Scalable Coordination in Sensor Networks,” ACM MobiCom,
1999.
recharging scheduling problem as a mixed integer programming [20] R. Williams, “The Geometrical Foundation of Natural Structure: A Source
problem, which is NP-hard. To coordinate with sensor activities, Book of Design,” Dover Pub. Inc. New York, pp. 51-52, 1979.
we present an algorithm for a single RV first to achieve both [21] B. Gavish, “A Note on the Formulation of the m-salesman Traveling
Salesman Problem,” Management Science, 1976.
high energy efficiency and reliable detection on targets, and [22] D. Feillet, P. Dejax and M. Gendreau, “Traveling Salesman Problems with
then extend it to multiple RVs using two different schemes. Profits,” Transportation Science, vol. 39, no. 2, pp. 188-205, May 2005.
The first scheme partitions the sensors into groups for different [23] J. MacQueen, “Some Methods for Classification and Analysis of Multi-
variate Observations,” Proc. of 5th Berkeley Symposium on Math. Statistics
RVs to reduce their traveling distance, whereas the second and Probability, pp. 281-297, 1967.
scheme allows the RVs to access all information about the [24] T.H. Cormen, C. E. Leiserson, R. L. Rivest and C. Stein, “Introduction to
network. Finally, we demonstrate the advantages of our proposed Algorithms,” MIT Press, 2001.
[25] Texas Instruments Technical Documents,
schemes by extensive simulations. The results show that by “http://www.ti.com/lit/ds/swrs074a/swrs074a.pdf.”
managing sensor activities, more energy can be saved on RVs. [26] ON Semiconductor, “http://www.onsemi.com/pub link/Collateral/NCS360
Further, the proposed recharging algorithms achieve significant 00-D.PDF.”

798
799

You might also like