Telecommun Syst

DOI 10.1007/s11235-015-9980-1

Residual energy aware mobile data gathering in wireless sensor

networks
Xuan Rao · Hongyu Huang · Jiqiang Tang ·
Huan Zhao

© Springer Science+Business Media New York 2015

Abstract The intrinsic characteristic of wireless sensor position for collecting data depended on residual energy. Fol-
networks is the power limitation of sensor nodes. The most lowing that, we construct a balanced data relay tree in each
difficult challenge is how to save energy of sensor nodes so cluster. Last, we design a shortest path for the MDC. Since
that the lifetime of a sensor network will be prolonged. A the paths of MDC are different in each round, the lifetime of
mobile data collector (MDC) is introduced to achieve this the sensor network can be prolonged. Simulations reveal that
goal. We suppose that all sensor nodes are kept static once our method is better than the existing methods and prolong
deployed, and a single MDC traverses the network to reduce the lifetime of wireless sensor network.
the communication of relaying data among sensors. In gen-
eral, we need to consider two factors when designing a travel- Keywords Residual energy awareness · Spectral
ing path of a MDC, i.e., the data overflow on a sensor node and clustering · Data relay tree · Path scheduling
the timeliness of each data. In this paper, we aim to prolong
lifetime of a sensor network by designing heuristic traveling
paths of the MDC under these two constraints. It is obviously 1 Introduction
that a fixed MDC path leads to a quicker energy consumption
of the nodes near that path. So we propose an iterative scheme Data gathering in Wireless Sensor Networks (WSNs) has
which determines the traveling path of the MDC before each been researched for over ten years. With the development of
round of the data gathering. For each data gathering round, IoT, wireless sensors are playing a more and more crucial
our scheme consists of four steps. First we iteratively parti- role. In general, there are two types of the data generated
tion the network into clusters by spectral clustering, and then by sensor nodes: the periodical data in each duty cycle and
select a cluster head as the polling point which is a special the event-driven data when special event happened or user
request launched. For the former kind of data, they should
be collected and posted to data sink for further processing
X. Rao · H. Huang · J. Tang · H. Zhao
periodically. So there can be lots of data transmission and
Key Laboratory of Dependable Service Computing in Cyber
Physical Society of Ministry of Education, Chongqing University, reception which always consume much energy. As we all
Chongqing, China know, the most significant characteristic of sensor node is
the energy limitation. As sensors are kept static once they
X. Rao · H. Huang (B) · J. Tang · H. Zhao
are deployed, we hope that the sensor network can work as
College of Computer Science,
Chongqing University, Chongqing, China long as possible. The lifetime of a WSN is depended on the
e-mail: [email protected] lifetime of a single sensor node which exhausted its energy
X. Rao earliest. In order to prolong the lifetime of the sensor network,
e-mail: [email protected] we must focus on the energy consumption of a single sensor
J. Tang node.
e-mail: [email protected] When it comes to prolonging the sensor network life-
H. Zhao time, researchers have done noteworthy work. In MAC layer,
e-mail: [email protected] contention-free data collection schemes are proposed. Many

123
 X. Rao et al.

of this kind of data collection scheme are based on TDMA In a word, a good mobile data gathering scheme should
[1], which avoids interference among nodes by assigning a ensure an expected network lifetime and have a bounded data
fixed time slot for each node, such as DEANA and DMAC gathering latency as well [3,4]. In this paper, in order to pro-
[1]. In routing layer, they proposed many data routing proto- vide a scalable data gathering scheme for large-scale static
cols. Due to the different ways of constructing the topol- sensor networks, we also utilize MDC to gather data from
ogy of WSN, these protocols can be divided into three sensors roundly. We suppose that MDC must visit to a cer-
categories [1]: (1) location-based protocols, such as GAF, tain sensor node at least with the interval of the expire time of
GEAR and Span. (2) data-centric protocols, such as SPIN, sensor data. We aim to prolong the lifetime of sensor network
DD and Rumor Routing [1]. (3) Multipath-based proto- under a constrained MDC’s path length. As nodes near the
cols, such as Disjoint Paths and N-to-1 Multipath Discov- MDC path will consume much more than others, so the MDC
ery. These protocols aim to decrease the number of trans- path should be dynamically changed in each data gathering
mitted packets and save the energy consumption. In appli- round. In this case, in each data gathering round, we divide
cation layer, data aggregation is one of the most general the sensor network into small sub-clusters by spectral cluster-
methods. All the intermediate sensor nodes will process the ing iteratively until the path length of the MDC exceeds the
received data and converge them into a single packet. It is system predefined parameter. We judge and weigh two prob-
usually available for reducing the data amount in sensor lems: (1) Which node should be the cluster head; (2)Which
networks. In some other data-sensitive applications, which sub-cluster should be divided further. Then we construct an
mean the applications with high data precision and time- efficient data relay tree in each sub-cluster based on the resid-
liness, however, this kind of data aggregation may lead to ual energy of sensor nodes in that sub-cluster. Last, the MDC
loss of accuracy. Hence constructing a maximum lifetime can start a round of data gathering.
data gathering tree is another way of avoiding unnecessary Our method alleviates the “hot spots” problem, which
energy consumption [2] in this layer. In order to further means the sensor nodes near the data receiver always die
save the energy and prolong the sensor network lifetime, quickly, encountered in data routing scheme by variable trav-
researchers suggest to use a mobile data collector for data eling paths of the MDC, and achieves a balance between the
gathering. uniformity of energy consumption and the long data collec-
A mobile data collector (MDC) is usually a robot or a tion latency in mobile data gathering with single-hop. Most
vehicle with antenna, which can traverse the whole sensor existing papers which involve the MDC do not consider the
network and stop in the neighbourhood of a sensor node and residual energy while it is non-negligible to prolong the sen-
finally return back to the data sink for further data processing. sor network lifetime. We put this factor into a key stage in
We call this a data gathering round. When the MDC comes each step of our algorithm.
close to a certain sensor node, it “polls” the data generated The main contributions consist of the followings:
in the neighbourhood of that sensor node. It can achieve a
contention-free data transmission by an elaborate time slots – We combine the multi-hop sensor network and MDC.
assignment for energy saving. This measure can be a good This is different from [5], focused on single-hop data
help for reducing the data transmission energy consumption transmission on MDC which bound to long delay and
in WSN because many sensor nodes no longer need to send [6], gathered data by clusters lonely which resulted in a
data to the far away sink node. Instead, sensors near the MDC fierce energy consumption;
can send data to MDC directly when the MDC comes close to – Spectral clustering is used in classify the large network
them so that the topology of the sensor network is changed. into small ones. As far as we known, there is no paper tak-
However, the drawbacks of the scheme with MDC consist of ing it into consideration. Although [7] is based on spec-
the followings: (1) Sensor nodes only have limited physical tral clustering, it merely considers the factor of distance
memory so data may overflow if the MDC takes a long time which is not accurate to measure the energy consumption
to visit that sensor node. (2) For some data-sensitive applica- as we do;
tions, data should be processed before they become invalid. If – We construct a data relay tree based on residual energy for
MDC sends the expired data to data sink, it is a waste of time the sake of energy efficient routing which not mentioned
and energy. (3) As the MDC is powered by a battery, when before.
it runs out of energy, the MDC can never return back to data
sink. Based on these observations, a system delay time must The rest of this paper is organized as follows. Section
be constrained due to the introduction of the MDC. It is easy 2 introduces the related work. Section 3 outlines the REA-
to see that the shorter the delay time is, the more energy WSN MDG problem and proves this problem to be NP-hard. An
consumes while longer delay time result in the above three energy consumption model is also introduced here. Section 4
problems. This is because the delay time is strong related to proposes the centrality algorithm and Sect. 5 provides numer-
the path length the MDC travels. ical results. Finally, Sect. 6 concludes this paper.

123
Residual energy aware mobile data gathering in WSNs

2 Related work and the full utilization of SDMA, considered the moving time
of the MDC and the data uploading time of sensors compre-
In this section, we briefly review some recent work on data hensively. For a more deliberate consideration, Zhao et al.
gathering in wireless sensor networks. Based on whether the [16] considered two different cases depending on whether
MDC takes part in the data gathering scheme or not, we can the mobile collector has fixed or variable sojourn time at
divide data gathering into two categories. each anchor point, which is the same as the polling point.
The basic model of the first category, without MDCs, is They formalized the problem as network utility maximiza-
composed of two types of sensor networks which are homo- tion under constraints of guaranteed network lifetime and
geneous network and heterogeneous network respectively data gathering latency. Sometimes the application is sensi-
[2,8–12]. No matter what the type of the network is, the tive for delay, MERP [17] addressed the problem of finding
method of clustering the original network into small net- a set of polling points for MDC to minimize the network
works should be designed. Once the clusters are formulated, cost under the given delay. They constructed the Minimum
cluster heads are responsible for gathering data from clus- Steiner Tree, and designed two different algorithm named
ter members and further they transmit all the data to sink RP-CP and RP-UG.
by single-hop way or multi-hop way. The sensor nodes in In terms of channel capacity, NCM [19] minimized
the former network have the same capability of communica- the network cost by means of scheduling communications
tion and the difference in heterogeneous network is that the between sensors and the MDC and dynamically adjusting
cluster heads in this kind of network are always much more the sojourn time at different anchor points. They defined a
powerful than their members, which resulting better scala- cost function as the data amount a sensor uploads to mobile
bility and less energy consumption. RBCA and GFTSS [2] collector during its sojourn time at a certain anchor point.
address the problem of how fast can information be collected Somasundara et al. [20,21] studied the scheduling of mobile
from a wireless sensor network organized as a tree. Two elements to ensure no data loss due to buffer overflow.
types of converge-cast, raw-data converge-cast and aggre- For a comprehensively consideration, frameworks are also
gated converge-cast, are discussed for drawing the conclu- proposed for mobile data gathering. Considering concur-
sion that the data collection rate is not remain limited by rent data uploading, DaGCM [22] framework is designed
interference but by the topology of the routing tree. In [8], for sharply shorten data gathering latency and significantly
a condition-free schedule is proposed for cluster heads to reduce energy consumption. They solved this problem with
gather data in heterogeneous sensor network. However, clus- the subgradient iteration algorithm. LBC-MU [23] is a three-
ter heads in this paper have high communication capability layer framework to achieve good scalability, long network
that they can sent messages directly to their members. That lifetime and low data collection latency. It consisted of sensor
means the cost of the whole sensor network will be expensive. layer, cluster head layer and mobile collector layer. Results
The mobility of sensor networks has been studied in some showed that this kind of framework behaved a efficient per-
literatures recently [5,6,13–22]. The common ground of formance.
these papers lies in that one or more MDCs are used to tra-
verse the sensor network and stopped at some special places,
which are often called polling points, for data gathering. Sen- 3 Problem formulation
car [13] is the first one to consider the problem of plan-
ning the moving path of a mobile observer and balancing 3.1 Network model and problem formulation
the traffic load to prolong network lifetime, the proposed
schedule showed that MDC can greatly affect network life- For a Given WSN, it can be represented by an undirect graph
time. In terms of relay hops count, SPG-DGA [6] studied the G(V, E), where V is the set of vertices representing sen-
tradeoff between energy saving and data gathering latency by sor nodes, and E is the edge relationship of a pair of nodes.
exploring a balance between relay hops count of local data Each sensor has a unique ID i ∈ {1, . . . , N } for identifi-
aggregation and the moving length of the mobile collector. cation. We assume that the communication range of each
Compared with multi-hop transmission, Ma et al. proposed a sensor node is same denoted by R. We define the neigh-
greedy algorithm [5] which constrained that sensors have to bour N (v) of the sensor node v as the node set that can be
transmit with M-collector in single-hop. It focused on the reached by v in R and suppose each sensor is only able to
problem of minimizing the length of each data gathering communicate with its neighbours. E i j is the edge between
tour. For applications with strict distance/time constraints, node i and node j. If node j is in N (v), then arc E i j is
they consider utilizing multiple M-collectors and minimize added into the graph G(V, E). Let Pi presents the residual
the number of M-collectors. energy of Vi . Due to the battery’s intrinsic characteristics, the
To minimize the total data gathering time, MDG-SDMA initial energy may be unequal. We suppose there is a hard-
[14] explored the tradeoff between the shortest moving tour ware counter in each node to count the residual energy Pi

123
 X. Rao et al.

and euclidean distance d to the data sender after commu-

nication. The pair of sender and receiver can compute the
distance between each other based the signal strength they
received.
Since the MDC has the freedom to move to any location
Fig. 1 Radio energy dissipation model
in the sensor network, an opportunity to plan the optimal
tour is provided. In fact, planning an optimal tour is to deter- traveling tour of MDC in G  corresponds to the TSP cost in
mine a set of path points for MDC to travel. Once these path the same vertices in G. The TSP in G will have a minimum
points are determined, MDC can perform its task by visit- path length if and only if the same path in G  is the minimum
ing each path point with a special sequence. When MDC length for REA-MDG. In addition, we have another objective
moves close enough to one of these path points, it starts to that the energy consumption of each should be balanced so
collect data generated by the neighbours of that path point. that the lifetime of sensor network is prolonged. This suggests
After all the data has been collected, the MDC moves to our problem is harder than the TSP problem. Thus, the REA-
the next one and finally returns back to the data sink. Based MDG problem is NP-hard. 

on the discussion above, we define the optimal tour L as
L =< π0 , π1 , π2 , . . . , πn , π0 >, where π0 is the position of
data sink and πi , i = 1, . . . , n, is the ith path point on L. 3.2 Energy consumption model
We suppose the communication capability of MDC is the
same as the sensor nodes, that is, the MDC can only send We introduce a radio energy dissipation model which is delib-
or receive data with sensor nodes within a distance of R. In erated in [7]. In our work, we use this model to measure and
this case, we define the path points as a subset C of sensor analyse the energy consumption of each sensor node. This
nodes V . Since we define the lifetime of whole sensor net- model is illustrated in Fig. 1, where d is the distance between
work as the duration until first sensor node dies, the factors the transmitter and receiver, L is the length of the data to be
of great concern we should focus on are that energy con- transmitted, E elec denotes the energy per bit dissipated to
sumption and energy residual. In addition, in each data gath- run both the transmitter and the receiver circuits. The spent
ering round, different sensor nodes relay different amount of energy by the radio transmitter is given by:

data resulting more nonuniform residual energy distribution. L ∗ E elec + L ∗ E f s ∗ d 2 d < d0
This makes some bottleneck sensor nodes consume energy E t x (L , d) = (1)
L ∗ E elec + L ∗ E mp ∗ d 4 d ≥ d0
quickly and die. We must dynamically change the network
topology according to sensors residual energy. where E T x (L , d) presents the energy spent to transmit L
In summary, our objectives are (1) to find a subset of sen- bits of data over a distance of d. The parameter E f s and E mp
sors C as the path points; (2) to design the data relay path denote the amplifier energy in a free space model(with d 2
from other sensor nodes to a corresponding node in C dynam- power loss) and in a multi-path fading(with d 4 power loss)
ically; (3) to determine the data gathering tour by visiting channel models respectively. The threshold d0 is defined as:
each path point in C exactly once, such that the tour length 
Efs
of the MDC satisfies the predefined parameter and maximize d0 = (2)
E mp
the lifetime of WSN. We call this the Residual Energy Aware
Mobile Data Gathering, short for REA-MDG. We have the The spent energy by the radio receiver is given by:
following theorem concerning the REA-MDG problem.
Er x (L) = E elec ∗ L (3)
Theorem 1 The REA-MDG problem is NP-hard.
Based on this model, it is easy to find that the transmit dis-
Proof We show the NP-hardness of the REA-MDG problem tance plays a significant role in energy consumption.
by giving a polynomial-time reduction from TSP problem
for any instance of REA-MDG problem. Given a complete
graph G = (V, E) as an instance of TSP. We construct an 4 REAMDC-WSN algorithm
instance of REA-MDG on graph G  = (V  , E  ), which is
topologically identical to G, where V  is consisted of all the In this section, we present a new mobile data gathering proto-
sensor nodes and the sink node, and E  represents the edges col for WSN. We call it Residual Energy Aware Mobile Data
between any two vertices. As we define the path points of Collection scheme in Wireless Sensor Networks, in Wire-
L as the sensor nodes in sensor network, there must exist a less Sensor Network, which is short for REAMDC-WSN.
subset C  ⊆ V  that represents the optimal tour L  . The MDC Our main idea here is, dividing the whole wireless sensor
have to visit each πi in L  so that all the data generated by network into two small clusters and choosing one to be sub-
the neighbour of πi can be collected. Hence, the length of divided. Repeating this iteration until the total running time

123
Residual energy aware mobile data gathering in WSNs

of MDC exceeds a system predefined parameter. After all clusters for MDC to travel. For conciseness, the concrete
clusters are discovered, a residual-energy based data relay procedure of spectral clustering is not mentioned here, which
tree for each clusters is deployed for energy efficient uti- can be found in [24,25].
lization. We divide our algorithm into four correlative sub-
algorithms, which are: (1) How to determine the similarity
4.2 Choose cluster head
matrix for spectral clustering; (2) How to determine the sub-
cluster need to be re-clusterd; (3) How to find a cluster head
In general, the cluster head of a cluster is responsible for
for each cluster; (4) How to design an energy efficient data
collecting all packets generated from its cluster members
relay tree for each cluster. These four sub-algorithms should
and waiting for the arrival of MDC to upload the data. In
be considered jointly.
this case, cluster heads carry heavy burden, as a result they
consume much more energy and die quickly.
4.1 Clustering In our paper, however, with the help of MDC, there are
no traditional cluster heads any more. In fact, the “polling
The objective of the clustering step is to split the considered points”, short for pp, replace them. We will make a simple
graph into two subgraphs in perspective of energy saving. explanation of it. For the sake of brevity, a pp is just a normal
We decided to use spectral clustering for its merits of sim- sensor node. When MDC travels to a certain pp, it will start
ple implementation, efficient solvability and very often out- to harvest data generated by the sensor nodes in the cluster
performs traditional clustering algorithms such as k-means belongs to that pp. The sensor nodes in pp’s one-hop neigh-
algorithm. What’s more, the conception of similarity for two bourhood transmit the packets directly to the MDC rather
elements it used matches with a pair of two sensor nodes so than pp. The main reason for such schedule is the considera-
that we can easily get two subgraphs in an energy efficient tion of energy balance. As the lifetime of the whole wireless
manner. sensor network is limited by the lifetime of single sensor
The next problem we should solve is seeking the suitable node, if we let the nodes in the neighbourhood of that cluster
parameter to describe the similarity among the sensors. The head to average its task, the energy consumption on cluster
reason we choose Gaussian similarity type as the element head must be decreased. Based on this idea, we design a pp,
of spectral cluster similarity matrix is that the energy con- namely cluster head, chosen method to tell the MDC where
sumption model we used. In our energy consumption model, to stop for sojourn.
distance is crucial for energy. Since the distance of two sensor There are two factors for a pp’s chosen. One is the resid-
nodes determines the energy consumption when transmitting ual energy distribution of the pp’s neighbours and the other
packets, obviously we believe, for a pair of sensor nodes, the is the quantity of the neighbours. The former one abides by
less energy they consume, the more similar they are, and they the observation that the most energy consumption nodes in a
should be put together in the same cluster. On the contrary, the cluster is the pp’s neighbour rather than the pps themselves
only reason for two sensor nodes are separated in two clusters (In fact, pp is the least energy consumption node in a clus-
is that they are far away from each other which resulting a ter). The latter factor is mainly for the observation that more
huge energy consumption. With all things above, as we want neighbours will make the energy consumption more bal-
to find a partition of the graph such that the edges between anced. That is because neighbour nodes communicate with
different groups have very low weight (which means that ver- the MDC directly and the direct communication decrease
tices in different clusters) and the edges within a group have the inner-cluster energy consumption. A weight formula is
high weight, distance will be a very good measurement. designed to determine which node is eligible for being a pp.
Let A ∈ R N ×N be the similarity matrix of the graph G. The possibility of a sensor node to be a pp is:
Each value of A is associated to each pair of the graph nodes
(i, j). This value is of gaussian similarity type and the matrix |Nv |
Pi |Nv |
A is given by : ω(v) = (1 − α) ∗ |N1 | +α∗ (5)
v max Nv
 1 Pmax
2 (i, j)
ex p(− d 2σ 2 ) i = j
A = [ai j ] = (4) where v is the one sensor node in a certain cluster, Pi is the
0 other wise
current residual of node v, Pmax is the maximum energy of
In one iteration of spectral clustering, we can get k clusters any node, namely, the initial energy of any node. Nv is the
which may be sub-divided later. Without loss of generality, number of the one-hop neighbour of node v and max Nv is the
we set k = 2. The reason we choose this value is the idea in maximum number of Nv . α is the weight of two factors.Here
dichotomy. That is to say, we start as a whole sensor network we let α equals 0.5, means that they are equivalently impor-
and in each iteration partition one cluster into two clusters, tant. It is reasonable for other α value which only represents
like [7]. After n rounds, we will get n + 1 small enough the different weight of the two factors.

123
 X. Rao et al.

Algorithm 1 Choose Polling Points

Initialization: adjacent matrix An×n , energy vector Vn ;
Iteration:
1: for i = 1 → n do
2: Find the neighbour set Si of node i;
3: Find the residual energy Pi of each node in Si
4: Compute ω(i) through Eq. 5 (a) original data relay (b) corresponding data
5: end for graph relay analysis graph
6: Compute the maximum ω(i), i = 1, . . . , n
7: return The maximum weight value node id v. Fig. 2 a Connection patterns of a sensor network. b Data relay analysis
graph T(S,A) corresponding to the connection patterns of the network

When a data gathering round is finished, the energy dis- construct a data relay analysis graph to calculate this vari-
tribution of sensors in each cluster must be changed. If the able. The data relay analysis graph T (S, A) for calculating
pps in previous round are the same in next round, there must P is constructed as follows: S = s1 , s2 , . . . , sn is the set of
be some sensor nodes become task heavy nodes. In order to all sensor nodes, and A is the set of all links a(i, j), where
avoid this problem, we suppose MDC should compute new i, j ∈ S. For each pair of nodes si , s j ∈ S, if si relays the
pps before each round so that it will visit different polling data generated by s j whether directly or indirectly, then arc
points in each round. Although the computation of new pps a(si , a j ) will be added into set A. Fig. 2 shows a simple
consumes much energy, it is operated on powerful MDC and example.
MDC can be recharged at data sink. So this part of energy For example, In Fig. 2a, s3 take responsibility for relaying
consumption is not considered in our scheme. the data of s4 to s2 , in other words, s4 transmits data to s3
From Eq. 5, we can find the most eligible node as the pp directly and to s2 indirectly. So two arc (s4 , s3 ) and (s4 , s2 )
in each cluster. The algorithm for finding such cluster heads are added in Fig. 2b. Supposing the graph in Fig. 2a is split
is described in algorithm 1. into two clusters in Fig. 2b by the middle dotted line l. The
amount of data packets which do not need to be transmitted
depends on two aspects: one is the number C ∗ of arc that
4.3 Determine subdivision
l goes through; the other is the number N of the isolated
nodes derived from removing arcs that cross with l . So P
Once the polling points are determined, the MDC can move
is defined as:
to these specific sensor nodes to harvest data. If the total time
for MDC to travel all existing pollling points is less than the P = C ∗ + N (7)
system predefined delay value Len, we can further utilise P
Let α = L ,which determines that whether a certain
the mobility of MDC to save sensors energy. To find which
cluster should be sub-divided in a certain iteration. In general,
cluster should be sub-divided in order to dividing big cluster
we always want a subdivision that saves maximum energy
into small clusters, we choose a greedy strategy for getting
while gains minimum path length. So the bigger of the value
around this.
α is, the more likely the cluster should be subdivided.
The basic idea of our proposed greedy algorithm is the
tradeoff of balancing the energy consumption decrement and
4.4 Design data relay tree
MDC path increment. At each stage of the algorithm, two
variables are calculated for measurement. One is P, another
This is the final step of our algorithm. We should design
is L. In this case, suppose that the cluster Ci in iteration t is
a data relay path for all cluster members transmitting their
divided into two clusters Ci  and Ci  in next iteration t +1. As
sensings to MDC stopped at pp. We focus on the idea of bal-
the amount of the cluster heads is increased, the total time
ancing the energy consumption based on the residual energy
for MDC traveling is obviously increased. The MDC path
of each node. In previous literatures,like [6,7], the definition
increment is L, which can be defined as:
of load balance is the data transmission amount is averagely
L = T S P(t + 1) − T S P(t) (6) equal among sensors. This is, however, not appropriate when
considering with the residual energy of sensor nodes. For
where T S P(t) denotes the total path length for MDC to tra- instance, a sensor with lower residual energy has to transmit
verse the all polling points in iteration t. the same quantity of data of higher ones, resulting in the lower
Similarly, the total energy consumption of the sensor one quickly die while the higher one still has enough energy
nodes which belong to original cluster is decreased due to which will be wasted due to the whole sensor network is dead.
subdivision, which result in the sensor nodes located at the In this case, we define load balance as the following: the
boundary of the two sub-clusters do not need to transmit data more residual energy sensor has, the more data it should
any more. P is defined as the saved energy quantity. We transmit. This hypothesis is reasonable. An algorithm is pro-

123
Residual energy aware mobile data gathering in WSNs

posed for solving this load balance problem. As the pps are Algorithm 2 Data relay tree algorithm
determined, we visit each pp in order of breadth-first search Initialization: Residual energy vector Pi
(BFS). The Ui set are recorded as the i th layer, which means Uik ’s siblings Ci (Uik )
the nodes are far away to in the same distance, to sensor Uik ’s parents Ci−1 (Uik )
Iteration:
nodes in BFS, and the corresponding residual energy set 1: for i = n → 1 do
of Ui is Pi . Suppose that there are totally n layers when /************* establish phases **************/
the BFS finished. The Iteration is started from the outer- 2: for k = 1 → si ze(Ui ) do
most nodes with maximum n. let Uik be the k th element in 3: Find U ij−1 in Ci−1 (Uik ) with maximum residual
energy
Ui , k = 1, . . . , si ze(Ui ). N (Uik ) is the one-hop neighbour 4: Add an edge from Uik to Ui−1
j
of Uik . For simplicity of description, we define another sym- 5: Update Pi−1 -= E * Ui .count
k k
bol Ci (Uik ) = Ui N (Ui ), which means the the one-hop 6:
j
Update Ui−1 .count += Uik .count
neighbour in the same layer with Uik . In other words, Ci (Uik ) 7: end for
indicate the possible siblings of Uik and Ci−1 (Uik ) indicate /************ adjust phases *****************/
the possible parents of Uik . 8: for k = 1 → si ze(Ui ) do
9: if Ci−1 (Uik ) == 0 && Ci (Uik ) = 0 then
Uik could be added to the data relay tree in two ways: 
10: Find Uik with maximum residual
energy
j
1. send data to Ui−1 directly, which is a node in Ci−1 (Uik ), 11: Add an edge from Uik to Uik



where j = 1, . . . , si ze(Ci−1 (Uik )), j = k 12: Update Pik -= E * Uik .count
j j
2. send data to Ui in Ci (Uik ) and then Ui relay all the data 13: else if Ci−1 (Uik ) == 1 && Ci (Uik ) == 0 then
p j 14: continue
to Ui−1 in Ci−1 (Ui ), where p = 1, . . . , si ze(Ci−1 (Ui )), 15: else if Ci (Uik ) = 0 then

j = 1, . . . , si ze(Ci (Uik )), j = k 16: Find Uik ∈ Ci (Uik ) with the maximum
residual energy
p 
The two different ways of course have different amount 17: Find Ui−1 ∈ Ci−1 (Uik ) with the maximum
residual energy
of energy consumption. If Uik transmits data via 1), that is to 18:
 
if Pik > (Uik .count + Uik .count) * E
j j
say, Uik sends data to Ui−1 , then updates Pi by subtracting p j 
&& Pi > Pi + (Uik .count + Uik .count)
count ∗E, where count is the total quantity of data packets * E then

to be transmitted and E is the energy consumption per data 19: Add and edge from Uik to Uik
k j
packet computed by the energy consumption model. 2) If the 20: Remove the edge from Ui to u i−1
j
maximum residual energy of any nodes in Ci−1 (Uik ) is less 21: Update Pi−1 += Uik .count ∗ E
p
j 22: Update Pi−1 -= (Uik .count
enough than the node in Ci−1 (Ui ), then Uik should transmit 
j + Uik .count)*E
data to its sibling Ui which the sibling itself relays all the data 23: end if
p
to its parent Ui−1 . By this way, we can achieve the goal of 24: end if
mitigating the burden of less residual energy nodes and aug- 25: end for
menting the more ones. It is worth mentioning that each node 26: end for
return The maximum weight value node id v.
should keep a record of its parent node id and the amount of
data to be transmitted. The algorithm is shown in algorithm 2.
Figure 3 is an example for this step. The solid black nodes
represent the sensor nodes in a cluster. The number in bracket
is the residual energy of each sensor node. S4 can both regard
S3 and S5 as its parent node and sends data to one of them.
The residual energy of S3 is less than the S2 , which is the
parent node of S5 , the sibling node of S4 . In this case, S4 just
sends data to S5 and let S5 relays its data to S2 . Although there
are too many children nodes of S2 , it still the most eligible
sensor nodes to be the indirectly parent node of S4 for its
high residual energy.

4.5 REAMDC-WSN algorithm

(a) original connection (b) constructed efficient
graph data relay tree

Combined with the above four steps, a residual algorithm in Fig. 3 a Connection pattern of a cluster. b The efficient data relay tree
WSN is proposed. The main procedure of the algorithm is corresponding to the connection patterns of the cluster

123
 X. Rao et al.

1400
Algorithm 3 REAMDC-WSN Algorithm
Initialization: graph G(V, E), adjacency matrix A, system predefined
delay Len, T = G, L  = 0 1200

Iteration:
1: while L  < Len do 1000
2: Construct the similarity matrix S for T , divide T into clusters C1

Node Number
and C2 , add C1 ,C2 into C. 800
3: Compute the weight of each cluster head by Algorithm.1
4: Find an approximate shortest tour L  on selected polling points
600
in C.
5: If L  < Len, choose a cluster head c by Algorithm.1, let T = c’s
corresponding similarity matrix. 400
6: end while
7: Compute the data relay tree for all the sub-clusters. 200

0
50 0 50 100 150 200 250 300 350 400 450 500
Sensor Network Lifetime
45
Fig. 5 sensor network lifetime as a function of the MDC tour length
40

35
run out of its energy, we suppose that the number of accumu-
30
lated data gathering rounds is the sensor network lifetime.
25

20 5.1 Tour length and sensor network lifetime

15
The major goal of our REAMDC-WSN algorithm is to pro-
10 long the lifetime with the constraint of a special system delay.
5
Fig. 5 shows the relationship between total lifetime and the
tour length of the MDC. With the tour length increases, the
0 lifetime of sensor network do not increase at a constant rate.
0 5 10 15 20 25 30 35 40 45 50
Instead, at the beginning, it increases fast, and then tends to
Fig. 4 A final clusters distribution of a sensor network be slow and finally keeps steady. This can be interpretted
as follows. When the tour length is constrained as a small
value, like L = 100, which implies clusters are relatively big
summarised in Algorithm 3. Figure 4 shows after REAMDC- and more energy are consumed by relaying data from other
WSN algorithm runs, the clusters are formed and the black sensor nodes. As a result, the lifetime of sensor network
nodes represent the polling points. The high residual energy is short and much energy is wasted. In this case, once the
nodes form the big clusters and low residual energy nodes tour length increased, the cluster size will prominent change
form the small ones. and the lifetime will be improved quickly. Fig. 5 shows that
once we boost L to 400, a significant improvement of sen-
sor network lifetime is shown because the whole sensor net-
5 Evaluation work is clustered into many small clusters, and nearly each
node can reach the cluster head easily. The MDC makes a
This section presents the evaluation of our REAMDC-WSN extremely good contribution to sensor network lifetime. As
algorithm compared with a traditional algorithm LEACH to some extent, more tolerant system delay will not signifi-
[26] and another mobile data gathering algorithm SCRC- cantly impacts the clusters change and lifetime slows down.
WSN proposed by Elbhiri et al. [7]. LEACH is a distrib- If we keep on increase L, no changes will be made due to the
uted data gathering algorithm which choose cluster heads MDC can even visit each sensor node one by one.
in rotation to avoid heavy burden on a single sensor node.
SCRC-WSN is a centralized algorithm that combined spec- 5.2 Average distance among sensor nodes
tral clustering with dynamic cluster head election. In the sim-
ulation, we assume that a bunch of sensor nodes are randomly The Fig. 6 shows the network lifetime and cluster number
deployed in a 50m × 50m sensing field. The Base Station is with different average distance among sensor nodes for a
located at the top-left corner of the sensing field. All the sen- given tour length L. As sensor nodes are randomly distributed
sor nodes have different initial energy. Once a sensor node in sensing field, the average distance of them are different.

123
Residual energy aware mobile data gathering in WSNs

10 100
Cluster Number LEACH
9 Network Lifetime(log) 90 SCRC−WSN
REAMDC−WSN
8 80

Sensor Network Residual Energy

7 70

6 60
Filed

5 50

40
4
30
3
20
2
10
1
0
0 0 200 400 600 800 1000 1200 1400 1600 1800 2000
0 5 10 15 20 25 30 Sensor Network Lifetime
Average Distance
Fig. 8 sensor network lifetime as a function of sensor nodes number
Fig. 6 network lifetime and cluster number as a function of average
distance among sensor nodes The amount of sensor nodes range from 100 to 300. The
REAMDC-WSN represents a significant advance than the
1400 REAMDC−WSN LEACH algorithm and also performs better than SCRC-
SCRC−WSN
LEACH WSN. For N = 300, the total data gathering round of
1200
REAMDC-WSN can achieve 1350 rounds whereas this value
1000
is only 1180 and 650 when LEACH and SCRC-WSN is
Sensor Network Lifetime

used. It follows that even if the nodes number increases, the

800 REAMDC-WSN approach still gives a best results compared
to other algorithms. This can be verified that all four steps of
600 our algorithm are either based on the residual energy or the
energy consumption, a significant lifetime prolong is bound
400
to achieve. The weakness of LEACH is the random selection
200
of cluster heads. This may lead to a unbalanced distribution
of cluster heads and making a large scale of sensor nodes are
0
0 100 200 300
difficult for communication with cluster heads. The problem
Sensor Node Number of SCRC-WSN is that it only considers energy in cluster step
and ignore the most significant data relay step. REAMDC-
Fig. 7 sensor network lifetime as a function of sensor nodes number
WSN takes these factors into consideration and gets a good
performance.
The broken line in figure presents that once the average dis-
tance become larger, the cluster number and network lifetime 5.4 Residual energy in network and sensor network lifetime
are both become smaller. That is because the larger average
distance of sensors implies the sensors are distributed more The Fig. 8 presents the total network remaining energy in
sparse. Since the tour length of the MDC is fixed, the MDC every transmission round with these three algorithm. The
cannot travel far to another cluster to collect data. In this case, network remaining energy decreases rapidly in LEACH and
data can only be transmitted by other sensor nodes which cost SCRC-WSN algorithm than in REAMDC-WSN one. For
extra energy consumption. In addition, as the energy con- LEACH, a slope is approximate 0.094 and 0.072 for SCRC-
sumption we used, the transmission energy consumption is a WSN. That value for REAMDC-WSN, however, is 0.067.
function of the distance of a pair of transmitter and receiver, Much energy is saved during the determining sub-division
the distribution of sensor nodes has a marked impact on the cluster step. As we always select the least cost sub-cluster,
cluster formation and energy consumption. the residual energy must be the most in these three algorithm.

5.3 Sensor node number and sensor network lifetime 6 Conclusion

The Fig. 7 shows that the relationship between sensor In this paper, we have presented the residual energy based
network lifetime and the number of the sensor nodes. approach for utilizing MDC to collect sensor data under tem-

123
 X. Rao et al.

poral constraints. A centralized data gathering algorithm, 10. Younis, O., & Fahmy, S. (2004). Distributed clustering in ad-
REAMDC-WSN, is developed for scenarios where energy hoc sensor networks: A hybird, energy-efficient approachy: IEEE
INFOCOM: Proc.
of sensor node is saved as much as possible so that the life- 11. Zhang, D., Zhang, D., Xiong, H., Hsu, C., & Vasilakos, A. (2014).
time of sensor network is prolonged. In our algorithm, energy BASA: Building mobile ad-hoc social networks on top of android.
is the key concern from beginning to end. We use the spectral IEEE Networks, 28(1), 4–9.
clustering to divide network into clusters and decide whether 12. Manjeshwar, A., and Agrawal, D.P. (2001). Teen: A routing proto-
col for enhanced efficiency in wireless sensor networks, Proc. IEEE
a certain sensor node can be a cluster head or not by two International Parallel and Distributed Processing Symp. (IPDPS),
considerable factors: the number of the neighbour it has and Apr.
the residual energy of its neighbour. We iteratively find the 13. Ma, M., & Yang, Y. (2007). MDC: An energy-efficient data gath-
sub-clusters until most energy save is achieved. Finally, we ering mechanism for larg-scale multihop sensor networks. IEEE
Transaction on Parallel and Distributed System, 18, 1476–1488.
construct efficient data relay tree for each cluster. Our simu- 14. Zhao, M., Ma, M., & Yang, Y. (2011). Efficient data gathering with
lations show that our approach significantly reduces network mobile collectors and space-division multiple access technique in
energy consumption and prolong the network lifetime. No wireless sensor networks. IEEE Transactions on Computers, 60(3),
cluster method, however, can apply to all type of scene. Nei- 400–417.
15. Zhao, M., Ma, M., & Yang, Y. (2011). Data gathering in wire-
ther does Spectral clustering. More scaleable methods should less sensor networks with mulitiple mobile collectors and SDMA
be designed in future. In terms of more strict time constraint technique sensor nerworks. IEEE Transactions On Computers, 60,
scenario, multi MDCs can be considered. 3.
16. Zhao, M., & Yang, Y. (2012). Optimization-based distributed algo-
Acknowledgments This work is supported by the National Nat- rithms for mobile data gathering in wireless sensor networks. IEEE
ural Science Foundation of China (Grant No. 61003247), the 51st Transactions on Computers, 11(10), 1464–1477.
Chinese Postdoc Science Foundation (Grant No. 2012M510932), the 17. Xing, G., Wang, T., Xie, Z., & Jia, W. (2008). Rendezvous planning
Fundamental Research Funds for the Central Universities (Grant No. in wireless sensor networks with mobile elements. IEEE Transac-
106112013CDJZR180004) and the Natural Science Foundation of tions on Mobile Computing, 7(12), 1430–1443.
Chongqing (Grant No. cstc2014jcyjA40030). 18. Di Francesco, M., & Das, S. K. (2011). Data collection in wireless
sensor networks with mobile elements: A survey. ACM Transac-
tions on Sensor Networks, 8(1), 7.
19. Zhao, M., Gong, D., Yang, Y. (2010). A cost minimization algo-
References rithm for mobile data gathering in wireless sensor networks, IEEE
7th International Conference on Mobile Adhoc and Sensor Systems
1. Vaish, G.D. (2012) Routing rotocols in wireless sensor networks: A (MASS), CA.
survey, 2012 Second International Conference on Advanced Com- 20. Somasundara, A. A., Ramamoorthy, A., & Srivastava, M. B. (2007).
puting & Communication Technologies(ACCT), 7–8 Jan, pp. 474– Mobile element scheduling with dynamic deadlines. IEEE Trans-
480. action Mobile Computing, 6(4), 395–410.
2. Incel, O. D., Ghosh, A., Krishnamachari, B., & Chintalapudi, K. 21. Zhang, D., Chen, M., Guizani, M., Xiong, H., & Zhang, D. (2014).
(2012). Fast data collection in tree-based wireless sensor networks. Mobility prediction in telecom cloud using mobile calls. IEEE
IEEE Transctions on Mobile Computing, 11(1), 86–99. Wireless Communications, 21(1), 26–32.
3. Zhang, D., Zhou, J., Guo, M., & Cao, J. (2011). TASA: Tag-free 22. Guo, S., and Yang, Y. (2012). A distributed optimal framework for
activity sensing using RFID tag arrays. IEEE Transactions on Par- mobile data gathering with concurrent data uploading in wireless
allel and Distributed Systems (TPDS), 22, 558–570. sensor networks, Proceedings IEEE INFOCOM.
4. Zhang, Yin, Chen, Min, Mao, S., Hu, L., & Leung, V. (2014). 23. Zhao, M., and Yang, Y. (2011). A framework for mobile data gath-
CAP: Crowd activity prediction based on big data analysis. IEEE ering with load balanced clustering and MIMO uploading, Pro-
Network, 28(4), 52–57. ceedings IEEE INFOCOM.
5. Ma, M., Yang, Y., & M, Zhao. (2013). Tour planning for mobile 24. Von Luxburg, U. (2007). A tutorial on spectral clustering. Statistics
data-gathering mechanisms in wireless sensor networks. IEEE and Computing, 17(4), 395–416.
Transactions on Vehicular Technology, 62, 1472–1483. 25. Ng, A. Y., Jordan, M. I., & Weiss, Y. (2002). On spectral cluster-
6. Zhao, M., & Yang, Y. (2012). Bounded relay hop mobile data gath- ing: Analysis and an algorithm. Advances in Neural Information
ering in wireless sensor networks. IEEE Transactions on Comput- Processing Systems, 2, 849–856.
ers, 61(2), 265–277. 26. Heinzelman, W.R., Chandrakasan, A., and Balakrishnan, H.
7. Elbhiri, B., Fkihi, S. E., Saadane, R., Lasaad, N., Jorio, A., Abouta- (2000). Energy-efficient communication protocol for wireless
jdine, D. (2013). A new spectral classfication for robust clustering microsensro networks, In Proceedings of the 33rd Hawaii Interna-
in wireless sensor networks. IFIP WMNC’. tional Conference on System Sciences(HICSS-33), January (2000).
8. Zhang, Z., Ma, M., & Yuan, Y. (2008). Energy-efficien multihop
polling in clusters of two-layerd heterogeneous sensor networks.
IEEE Transactions On Computer, 57(2), 231–245.
9. Heinzelman, W.R, Chandrakasan, A., Balakrishnan, H. (2000).
Energy-efficient communication protocol for wireless micro-
sensor networks, Proceedings Hawaii International Conference on
System Sciences(HICS).

123
Residual energy aware mobile data gathering in WSNs

Xuan Rao received his B.Engr. Jiqiang Tang received the BEng
degree from the College of Soft- degree from the School of Com-
ware Engineering, Chongqing puter and Information Science,
University, China, in 2012. He Southwest Normal University,
is currently working toward the Chongqing, China, in 2005. He
M.S degree in College of Com- is currently working toward the
puter Science, Chongqing Uni- PhD degree in the College of
versity, Chongqing, China. Since Computer Science, Chongqing
2012, he has been focused on University, Chongqing, China.
mobile data gathering in wireless Since 2005, he has been with
sensor networks. Chongqing University of Tech-
nology, Chongqing, China, as a
software engineer and a teacher.
His research interests include
wireless sensor networks and
mobile computing.

Hongyu Huang received his

B.Engr. degree from the Depart-
ment of Computer Science,
Chongqing Normal University,
China, in 2002. He received his
M.S. degree from the College
of Software, Chongqing Univer- Huan Zhao received his B.Engr.
sity, China, in 2005. Last, he and M.S. degree from the College
received his Ph.D degree from of Computer Science,Chongqing
the department of Computer Sci- University, China, in 2005 and
ence, Shanghai Jiao Tong Uni- 2008 respectively. He is cur-
versity, China, in 2009. Since rently working toward the Ph.D
2009, he has been with the Col- degree in the College of Com-
lege of Computer Science at the puter Science, Chongqing Uni-
Chongqing University, where he versity, Chongqing, China. His
is currently an associate profes- research interests include wire-
sor. Dr. Huang has co-authored over 20 research papers, most of them less sensor networks and energy
are focusing on the research fields of vehicular ad hoc networks and harvesting.
wireless sensor networks. His current research interests include large-
scale distributed systems, cloud computing, and vehicular ad hoc net-
works.

123