IEEE INTERNET OF THINGS JOURNAL, VOL. 3, NO.

3, JUNE 2016 257

Cloud-Assisted Data Fusion and Sensor

Selection for Internet of Things
Farshid Hassani Bijarbooneh, Wei Du, Edith C.-H. Ngai, Xiaoming Fu, Senior Member, IEEE,
and Jiangchuan Liu, Senior Member, IEEE

Abstract—The Internet of Things (IoT) is connecting people and devices, such as wireless sensor networks (WSNs) [4]–[6].
smart devices on a scale that was once unimaginable. One major Machine-to-machine automation with wireless sensors is being
challenge for the IoT is to handle vast amount of sensing data widely deployed, but usually in islands of disparate systems.
generated from the smart devices that are resource-limited and
subject to missing data due to link or node failures. By exploring The evolution of IoT attempts to connect these existing sys-
cloud computing with the IoT, we present a cloud-based solution tems to the cloud, which enables advanced data fusion, storage,
that takes into account the link quality and spatio-temporal corre- and coordination capability for achieving higher data quality
lation of data to minimize energy consumption by selecting sensors and energy efficiency. The upcoming challenge of IoT lies in
for sampling and relaying data. We propose a multiphase adaptive handling volumes of data generated from enormous amount of
sensing algorithm with belief propagation (BP) protocol (ASBP),
which can provide high data quality and reduce energy consump- devices, which is known as the big data problem.
tion by turning on only a small number of nodes in the network. The wireless sensors in many IoT applications are bat-
We formulate the sensor selection problem and solve it using both tery powered, resulting in extreme energy constraints on their
constraint programming (CP) and greedy search. We then use our operations, such as sampling, data processing and radio com-
message passing algorithm (BP) for performing inference to recon- munications. To conserve energy and achieve longer network
struct the missing sensing data. ASBP is evaluated based on the
data collected from real sensors. The results show that while main- lifetime, the costs of sensor sampling, processing, and radio
taining a satisfactory level of data quality and prediction accuracy, communications have to be minimized. It is often the case that
ASBP can provide load balancing among sensors successfully and sensor readings in the same spatial regions are highly corre-
preserves 80% more energy compared with the case where all lated. Depending on the application, the sensor readings are
sensor nodes are actively involved. temporally correlated as well. By leveraging the computation
Index Terms—Belief propagation, constraint optimization, capability of the cloud, data fusion can be performed to increase
Internet of Things (IoT), quantization, wireless sensor networks. the data quality by exploring the spatial and temporal cor-
relation of data. The wireless sensors can be coordinated by
I. I NTRODUCTION the cloud to be ON and OFF according to the change in the
environment. In this paper, we explore a seamless solution by
T HE INTERNET has enabled an explosive growth of infor-
mation sharing. With the advent of embedded and sensing
technology, the number of smart devices including sensors,
integrating cloud and IoT to provide comprehensive data fusion
and coordination of sensors to improve data quality and reduce
mobile phones, RF identifications (RFIDs), and smart grids energy consumption.
has grown rapidly in recent years. Ericsson and Cisco pre- Belief propagation (BP) [7]–[9] is a technique for solving
dicted that 50 billion small embedded sensors and actuators inference problems. In the IoT context, the belief of a sensor
will be connected to the Internet by 2020 [1] forming a new node is the data measurement of an event in the environment,
Internet paradigm called Internet of Things (IoT). IoT can sup- and BP provides an iterative algorithm (also called the sum-
port a wide range of applications in different domains, such as product algorithm) to infer the measurements of the sensor
health care, smart cities, pollution monitoring, transportation nodes, especially in cases where the data are missing, because
and logistics, factory process optimization, home safety and of packet losses or because there are no data available at some
security [2], [3]. selectively disabled sensor nodes (mainly to conserve energy
In the past decade, many studies have contributed to and reduce radio inference). In BP, each sensor node deter-
the hardware, software, and protocol design of the smart mines its belief by incorporating its local measurement with the
beliefs of its neighbor sensor nodes (spatial cooperation), and
Manuscript received August 27, 2015; revised October 18, 2015; accepted
November 02, 20015. Date of publication November 19, 2015; date of current its beliefs obtained in the past (temporal cooperation). In such
version May 10, 2016. inference problems, the assumption that the data are spatio-
F. H. Bijarbooneh and E. C.-H. Ngai are with the Department of temporally correlated significantly improves the accuracy of
Information Technology, Uppsala University, 751 05 Uppsala, Sweden (e-mail: data inference using BP in WSNs.
[email protected]; [email protected]).
W. Du is with CITI Lab, INSA-Lyon, 69621 Villeurbanne, France (e-mail: In monitoring applications for the IoT, the data are collected
[email protected]). and put in an environment matrix (EM) [10], where the data
X. Fu is with the Institute of Computer Science, University of Göttingen, readings for each sensor node are stored in one row of the
37073 Göttingen, Germany (e-mail: [email protected]).
matrix and each column index represents a timestamp for the
J. Liu is with the School of Computing Science, Simon Fraser University,
Burnaby, BC V5A 1S6, Canada (e-mail: [email protected]). interval at which the data were sampled. Hence, an EM is a
Digital Object Identifier 10.1109/JIOT.2015.2502182 matrix of size N × T where N is the number of sensor nodes
2327-4662 © 2015 IEEE. Personal use is permitted, but republication/redistribution requires IEEE permission.
See http://www.ieee.org/publications_standards/publications/rights/index.html for more information.
258 IEEE INTERNET OF THINGS JOURNAL, VOL. 3, NO. 3, JUNE 2016

and T the number of time intervals, and the time dimension consumption, data utility, and accuracy (error prediction
T is expanding as more data are collected. BP performs the of the missing data).
inference iteratively from the stream of data that are stored in This paper is organized as follows. In Section II, we dis-
EM based on the current and past data. Therefore, unlike the cuss the related work. In Section III, we give the system
compressed sensing (CS) [11] approach, BP does not require overview. In Section IV, we describe the formulation of our
a complete EM for the whole duration of the time interval to optimization problem on sensor selection, and we solve it using
perform inference. two approaches (CP and heuristic-based greedy algorithm). In
In this paper, we explore cloud-assisted adaptive sensing and Section VI, we conduct simulations to evaluate our solutions
data fusion to reduce energy consumption and improve data based on a real deployment of a WSN. Finally, we summarize
quality for the IoT. We propose an adaptive sensing BP protocol and conclude this paper in Section VII.
(ASBP), where the data are collected in several rounds (a round
is a fixed time interval where the network repeats the same
behavior) by active sensors (sensors that are collecting data in
each round). We formulate and solve an optimization problem II. R ELATED W ORK
that selects the active sensors in each round, by maximizing the The information industry benefits greatly from the techno-
data utility while maintaining energy load balancing. We define logical advancements brought by the IoT [13], [14]. The IoT
data utility as the sum of the qualities of the path links from creates a bridge between many available and recent technolo-
the selected active sensor nodes to the base station, subtracted gies, such as WSNs, cloud computing, and information sensing
by the sum of the correlations of the selected active sensors. [14]–[16]. In monitoring and data acquisition IoT-based sys-
If the selected active sensor nodes are located on a path with tems, it is necessary to collect data effectively and efficiently
greater link quality, then the value of the data utility increases. [14], [15], [17], [18]. The IoT provides a platform for WSNs
Likewise, if the selected active sensor nodes result in a lower to connect to Internet and benefit from the power of cloud
data correlation, then the data utility is increased. In each round computing and data fusion. Therefore, it is necessary to study
of ASBP, the minimum number of selected active sensor nodes data collection schemes that can seamlessly integrate with the
(which is a parameter of our sensor selection optimization prob- cloud and IoT systems. Data collection has been widely studied
lem) is adaptively tuned based on the performance of the BP for stationary WSNs. Gnawali et al. [19] present the state-
inference (data prediction accuracy) throughout the previous of-the-art routing protocol for a sensor network where the
round. In addition to BP, we also use data quantization to further nodes are forwarding data directly to a sink. They consider
compress the data and reduce the transmission costs. stationary WSNs that have static routes from the wireless sen-
In our active sensor selection formulation, we consider non- sors to the sink. Madden et al. [20] introduced a distributed
linear multihop routing protocol constraints. To model the query processing paradigm called acquisitional query process-
sensor selection problem effectively, we use both constraint ing (ACQP) for sensor network data collection. The goal was to
programming (CP) [12] and heuristic-based greedy algorithm. ensure a flexible tasking of motes via a relational query inter-
CP is a powerful framework to model and solve combinatorial face, while providing lifetime constraints, data prioritisation,
problems. A CP model consists of variables, variable domains, event batching, and rate adaptation.
and constraints, as well as objective function (if required), in Prediction-based energy-efficient approaches aim at predict-
which the constraints express the relation between the variables. ing the data to minimize the number of transmissions. Chou
The core concept in CP is constraint propagation. Constraint et al. [21] proposed a distributed compression based on source
propagation performs reasoning on a subset of variables, vari- coding, which highly relies on the correlation of the data, and
able domains, and constraints to infer more restrictive variable it compresses the sensor readings with respect to the sensor
domains, such that the restricted domains still contain all solu- past readings, and the reading measured by the other sensor
tions to the problem. CP combines constraint propagation with nodes. They used adaptive prediction to track the correla-
search procedure to find a local or global optimum (using tion of the data, which is used to estimate the number of
branch-and-bound search space exploration) to an optimization bits needed in source coding for data compression. Recent
problem. work in WSN addressed the use of compressive sensing [11].
The contributions of this paper are as follows. The authors use compressive sensing to exploit the tempo-
1) We present a novel data collection scheme (ASBP) that ral stability, spatial correlation, and the low-rank structure of
utilizes highly correlated spatio-temporal data in the net- the EM. They propose an environmental space–time-improved
work and uses BP to reconstruct the missing data due to compressive sensing (ESTI-CS) algorithm to improve the miss-
packet losses and the sensor selection strategy. ing data estimation. Although compressive sensing achieved
2) We formulate the active sensor selection optimization good accuracy on the estimation of the missing data, it does
problem, and propose two approaches, namely CP and only consider implicit spatio-temporal correlation in the data.
a heuristic-based greedy algorithm to solve the problem. Furthermore, compressive sensing approaches rely on the con-
The CP approach solves the problem to optimality. struction of a data matrix and thus require the synchronization
3) We conduct extensive simulation with a real deploy- of the sensors on the data collection. However, in our work,
ment of a sensor network and the collected data to we present a BP approach for the prediction of missing data,
evaluate the impact of our proposed solution (for both where the spatio-temporal correlation is explicitly enforced and
CP and heuristic-based algorithm) on the overall energy the inference is performed online and iteratively as the data are
BIJARBOONEH et al.: CLOUD-ASSISTED DATA FUSION AND SENSOR SELECTION FOR IoT 259

Fig. 2. Map of the Intel Berkeley Research Lab, with the hexagon-shaped
nodes indicating the locations and the ids of the sensor nodes, which are
deployed to monitor temperature, humidity, and light intensity. The value of the
Fig. 1. Network architecture, where the nodes in an IoT application forward aggregated link quality is represented with the thickness of the link between the
the data to the cloud. The servers perform node coordination to improve data sensor nodes.
quality and save energy, while the data centers stores the collected data as the
data fusion and the data loss prediction is performed.
B. Protocol Design
In our setup, the sensor nodes collect and report the data peri-
received at the base station. In addition to the above, to the
odically (typically every 30 s). Our protocol operates in several
best of our knowledge, there has been no work addressing a
rounds (a round is a time interval where the network repeats
CP approach for energy-efficient sensor selection with dynamic
the same behavior), and each round includes two phases. The
routing, while considering the link quality and correlation of the
first phase is used to collect the minimum required information,
data.
which is used in the second phase to improve energy-efficiency,
energy load balancing, and the data quality. The two phases in
each round are as follows.
III. S YSTEM OVERVIEW Phase 1: Phase one begins as all sensor nodes become active,
and starts collecting and forwarding a fixed number of quan-
A. Network Model tized data to the base station (typically 20 sensor readings).
In our IoT application, stationary sensor nodes collect envi- Throughout this phase, the routing protocol estimates the link
ronmental data, such as temperature, humidity, light intensity, quality for the shortest routes between the sensor nodes and
and noise level. Fig. 1 shows the network architecture of our the base station. The base station then computes the corre-
data collection in IoT applications. We support heterogenous lation coefficient matrix from the sensor data, and also uses
networks, where data can be collected from various devices. the routing tables to compute all the shortest paths from the
The network supports multihop routing and the gateways col- sensor nodes to the base station. These data (link quality, cor-
lect the data and forward the data to the cloud, where the data relation, and shortest routes) are then used as an input to solve
fusion is performed to further analyze the data, predict miss- our sensor selection optimization problem (further explained in
ing data, and store the data in the data centers. The computation Section IV) and select a subset of sensor nodes to be active
power of the servers in the cloud is used to improve data quality during the second phase. The active sensor nodes are the only
and save energy of the sensor nodes using our ASBP protocol sensor nodes in the network that are participating in the data
(to be discussed further in Section III-B). The sensor nodes peri- collection and relaying the data to the base station. The sensor
odically sample data, which is forwarded to the cloud using a selection problem is solved using either CP or a heuristic-based
multihop routing protocol (the ACQP system in TinyDB [20], greedy algorithm to select a set of active sensor nodes, such that
or the collection tree protocol [19]). In this work, we use the it maximizes the spatio-temporal correlation with the inactive
real data collected at the Intel Berkeley Research Lab [22]. sensor nodes, while considering link quality and the dynamic
Fig. 2 shows the map of the Intel Berkeley Research Lab, and routing.
the location of the deployed sensor nodes, which are marked Phase 2: The base station broadcast, a message that informs a
with hexagon shapes, and the sensor id. The link thickness subset of the sensor nodes to become inactive (sleep mode with
between the sensor nodes represents the value of the link quality no radio activity) for a given period of time (typically 2 h). In
aggregated throughout the experiment. this phase, the base station performs the BP algorithm [8], [9] to
The data are collected at the cloud using the gateways associ- infer incrementally the missing data due to the inactive sensor
ated with different applications of IoT. The gateway only relays nodes and packet losses (further explained in Section V). BP
the data to the servers in the cloud, and it is at least aware of captures the high spatio-temporal correlation in the data using
the routing tables of the sensor nodes. In this paper, we refer to a graphical model, which is taken into account in modeling our
the gateway and the base station as the same entity; however, sensor selection optimization problem. As the second phase is
the actual computations (the CP solver and greedy algorithm in completed, the base station continues to use BP during the first
Sections IV-A and IV-B) are performed on the cloud, and all phase of the next round. This allows us to compare the infer-
coordinations are relayed by the gateway. ence results during the first phase with the ground truth, and to
260 IEEE INTERNET OF THINGS JOURNAL, VOL. 3, NO. 3, JUNE 2016

compute the error in prediction. This error is then used by our

protocol to give feedback (on the minimum number of selected
sensor nodes) to the sensor selection optimization problem of
the next round. This allows a dynamic control over the accu-
racy of the data prediction in phase two. Throughout this paper,
we say ASBP to refer to the protocol design above.

IV. P ROBLEM F ORMULATION

We present our CP model for the sensor selection prob-
lem, followed by our heuristic-based greedy algorithm. The CP
Fig. 3. Sensor node 1 sending data to the base station B.
model finds a global optimum solution to the problem, whereas
our heuristic-based algorithm finds a good quality local opti- of the missing data construction
mum solution. The CP model selects the routes for packet relays 
dynamically. Dynamic routing is essential for networks with minimize x[s] · x[s ] · C[s, s ]. (2)
multiple shortest paths to the base station, large varieties in 
s,s ∈S,
the link quality, and energy-efficiency concerns. However, a s=s

heuristic-based greedy algorithm with good quality solutions We define the data utility u[s] to be the weighted linear sum of
is well suited for networks where the sensor selection problem the two objective terms in (1) and (2) for sensor s
cannot be solved in a centralized way, and data accuracy is of
less concern. ∀s ∈ S, u[s] = ω1 · x[s] · q[s]

− ω2 · x[s] · x[s ] · C[s, s ] (3)
s ∈S,
A. CP Model With Dynamic Routing s =s

As we mentioned in Section III-B, throughout phase one of where ω1 and ω2 are non-negative weight coefficients used to
ASBP, the data are collected in the EM, which is used to com- normalize and allow preference adjustment between the path
pute the correlation coefficient matrix. We also estimate the link quality and the aggregated correlation of the data for sensor s
quality from the packet reception rate during phase one. We versus all the other sensors in the network.
then have the following constants in our sensor selection model. The combined objective considering the data utility u[s] and
1) Let S be the set of WSN sensor nodes, with |S| = N . the residual energy E[s] of the sensor nodes becomes
2) Let L[s1 , s2 ] be the link quality between neighbor sensor 
nodes s1 and s2 , indicating the probability of receiving a maximize E[s]α · u[s] (4)
packet sent from s1 to s2 , with s1 , s2 ∈ S. If s1 is not the s∈S

direct neighbor of s2 , then L[s1 , s2 ] = 0. where α is a parameter to adjust the weight of the energy
3) Let B[s] be the link quality between the base station and coefficient on the data utility (typically α is set to 0.5).
a direct neighbor sensor node s, and otherwise B[s] = 0. The path quality constraint enforces that the path quality q[s]
4) Let C[s1 , s2 ] be the absolute value of the correlation of from a selected sensor to the base station must exceed a given
the data between sensor nodes s1 and s2 , with C[s1 , s2 ] ∈ threshold τ
[0, 1]. ∀s ∈ S, q[s] ≥ x[s] · τ (5)
5) Let P [s] be the set of all shortest paths from the sensor
s to the base station, where a path p ∈ P [s] of length n where the threshold τ is adjusted according to the link quality
is denoted by p : (s1 , s2 ), (s2 , s3 ), · · · , (sn−1 , sn ) with to provide a consistent packet delivery on a path to the base
s1 = s and sn is directly linked to the base station. station (typically τ ∈ [0.3, 0.7]).
6) Let E[s] be the residual energy of the sensor s at the end The routing constraint enforces that a path with higher qual-
of the first phase in ASBP protocol. ity is preferred in selecting the active sensors and all sensors on
Let x[s] be a Boolean variable with value 1 if the sensor such a path must be active. For example, Fig. 3 shows two paths
node s is selected for the data collection, and 0 otherwise. Let p1 and p2 from sensor node 1 to the base station B. We assume
q[s] represent the maximum achievable path quality among all that the link quality between sensor nodes on a path to the base
possible shortest paths from sensor s to the base station, in a station is an independent random variable. Therefore, path qual-
solution to the sensor selection problem. ity is the joint probability of the link quality probabilities along
We require the maximization of the path quality a path to the base station
 qp1 [1] = L[1, 5] · L[5, 6] · L[6, 4] · B[4]
maximize x[s] · q[s]. (1)
s∈S
= 0.78 · 0.88 · 0.65 · 0.91
= 0.406
A second objective is to minimize the correlation of the data
qp2 [1] = L[1, 5] · L[5, 6] · L[6, 4] · B[4]
between the selected sensors. This objective implies that data
from the inactive sensors are more likely to have a high corre- = 0.86 · 0.75 · 0.66 · 0.91
lation with the enabled sensors, hence improving the accuracy = 0.387
BIJARBOONEH et al.: CLOUD-ASSISTED DATA FUSION AND SENSOR SELECTION FOR IoT 261

⎛ ⎞
where qpi [s] denotes the path quality of the path pi originated at 
sensor s. The path p1 has a higher path quality (qp1 [1] > qp2 [1]). ∀s ∈ S q[s] = max ⎝B[np ] · (x[s ] · L[s , s ])⎠
p∈P [s]
Hence, when maximizing the path quality (1), the path p1 is (s ,s )∈p
preferred for routing the data, and to enforce that all sensors on ∀s ∈ S, q[s] ≥ x[s] · τ
the path must be selected, the path quality q1 is constructed as 
follows: x[s] ≥ μ.
s∈S
q[1] = max (x[5] · x[6] · x[4] · qp1 [1], x[2] · x[3] · x[4] · qp2 [1]) .
(6) This CP model is directly expressed and solved in our chosen
CP solver without further transformation to the formulation. For
Assuming that sensor 1 is selected (x[1] = 1), the path qual- our CP implementation of this model, we derive implied con-
ity constraint (5) requires that q[1] ≥ τ > 0, and according to straints from the routing constraints (7), to reduce the search
(6), all sensors on either path p1 or p2 must be active (x[5] = effort needed to solve the problem. We observe that some sensor
x[6] = x[4] = 1 or x[2] = x[3] = x[4] = 1). Note that the rout- nodes are often shared along the shortest paths from the origin
ing constraint (6) must be enforced only if the origin sensor 1 sensor node s to the base station. For example, in Fig. 3, sensor
is selected (x[1] = 1), and otherwise the value of the path link node 4 is shared by both paths p1 and p2 . If sensor node 1 is
should not be included in the objective function (4). Therefore, selected, it implies that sensor 4 must be also selected regard-
the nonlinear term x[s] · q[s] is used in the construction of data less of which path is used in forwarding the data to the base
utility (3). In general, the routing constraint becomes ∀s ∈ S station. We incorporate these implied constraints in our model
⎛ ⎞ to help improve the performance of the solver

q[s] = max ⎝B[np ] · (x[s ] · L[s , s ])⎠ (7) 
p∈P [s] ∀s ∈ S, (x[s] = 1) =⇒ ∧s ∈P∩ [s] x[s ] = 1 (9)
(s ,s )∈p

where np is the last sensor on the path p, and s , s are two adja- where P∩ [s] is the intersection set of all sensor nodes on the
cent sensors on the path p. For example, in Fig. 3 for the paths paths from s to the base station. The implication (9) states that
p1 and p2 , we have np1 = np2 = 4. The n-ary constraint max if the sensor node s is selected (x[s] = 1), then the conjunction
is essential in our CP implementation of the routing constraints of all shared sensor nodes on the paths from s to the base station
(7). must be 1, enforcing that all the shared sensor nodes are part of
The active sensor constraint enforces that the minimum the solution (x[s ] = 1, s ∈ P∩ [s]).
number of active sensors is at least μ Our custom search procedure branches on the x[s] decision
 variables. It selects a sensor with the largest mid-value in the
x[s] ≥ μ (8) domain of the data utility u[s]. The mid-value is often a better
s∈S choice when the domain range is large, which is the case at the
where μ provides a tradeoff between energy efficiency and data beginning of the search. The search procedure breaks ties by
quality (BP inference error). selecting the closest sensor to the base station with hop-count
In summary, our CP model for the sensor selection problem as the metric. We then set the value of x[s] to 1 on the left
is defined as follows. branch and 0 on the right branch.
Inputs:
1) L: link quality estimations; B. Heuristic-Based Greedy Algorithm
2) B: base station link quality estimations;
3) C: correlation coefficient matrix; Instead of using CP to solve the sensor selection problem
4) P : shortest routes to the base station; optimally as described above, we also designed a heuristic-
5) E: residual energy. based algorithm built upon a simple greedy search strategy. The
Outputs: intuition behind is that we should remove a sensor if: 1) the data
1) x: selected sensors with x[s] = 1 iff sensor s is selected from the sensor are strongly correlated with the others, mean-
for data collection and x[s] = 0 otherwise; ing that we can predict fairly accurately the reading from that
2) u[s]: data utility of sensor node s; sensor; 2) the sensor is already overused, meaning that the sen-
3) q[s]: path quality achieved in the routing of data from sor has a low energy; and 3) the sensor has a poor connection to
sensor node s to the base station. the base station, meaning that the data transmission from that
Objective: sensor has a high risk to fail. Thus, we do a greedy selection
 by taking all three aspects into consideration and remove sen-
maximize E[s]α · u[s]. sors one by one until we are left with the required number of
s∈S sensors. While simple, the heuristic algorithm may only find a
Such that local optimum to the sensor selection problem, which might be
far from the global optimum.
∀s ∈ S, u[s] = ω1 · x[s] · q[s] Our heuristic algorithm returns a set idSelected of sensor

− ω2 · x[s] · x[s ] · C[s, s ] nodes to be selected during the phase two of each round in
s ∈S,
ASBP protocol. The algorithm takes the constant set of sensor
s =s nodes S, link quality L, base station link quality B, correlation
262 IEEE INTERNET OF THINGS JOURNAL, VOL. 3, NO. 3, JUNE 2016

Algorithm 1. The heuristic-based greedy algorithm with base station link quality, and correlation for those sensor
dynamic routing nodes in idNonReachable are set to 0 from the correspond-
ing data using the function SetZero (lines 5–7). The function
input: S, L, B, C, E, μ, τ SetZero(A, Ids) takes an n × n matrix A, and a set of indices
output: idSelected Ids, and for each index i in Ids sets the value of every possible
1 idSelected ← S pair of (i, j) 1 ≤ j ≤ n in A to zero (A(i, j) = 0 ∧ A(j, i) =
2 q ← BestShortestPath(L, B) 0, 1 ≤ j ≤ n), and if A is a one dimensional array, then it only
3 idN onReachable ← {s ∈ S|q[s] < τ } sets A(i) = 0. In other words, SetZero reflects the unreacha-
4 idSelected ← idSelected − idN onReachable bility of the sensor nodes in idNonReachable into the network
5 SetZero(L, idN onReachable) data structures (link quality and correlation).
6 SetZero(B, idN onReachable) The main loop of the algorithm (line 9) iteratively selects
7 SetZero(C, idN onReachable) a sensor node that contributes the least value to the objective
8 idF easible ← idSelected (4) (equivalent to a sensor node with the lowest data utility
9 while idF easible = ∅ ∧ |idSelected| > μ do weighted by the initial energy), and performs a lookahead move
10 L ← L (lines 10–20) to detect if removing this sensor node violates any
11 B ← B of the constraints. The set idFeasible of feasible sensor nodes is
12 {idM in} ← min arg min (E[s]α · u[s]) initialized with the set of selected sensor nodes idSelected (line
s∈idF easible 8). The set idFeasible is used to keep the track of the sensor
13 SetZero(L , {idM in}) nodes that are potentially removable from the set of selected
14 SetZero(B  , {idM in}) sensor nodes idSelected . A lookahead move is performed, by
15 q ← BestShortestPath(L , B  ) first creating copies L and B  from L and B, respectively (lines
16 idN onReachable ← idSelected ∩ {s ∈ S|q[s] < τ } 10–11). We then select a least contributing sensor {idMin} that
17 idP otential ← idSelected − idN onReachable minimizes the value of the objective function (lines 12). To per-
18 if |idP otential| < μ then form the lookahead move, the link quality data for the sensor
19 idF easible ← idF easible − {idM in} {idMin} is set to zero (lines 13–14), and then the path quality
20 continue q is updated (line 15) to discover the nonreachable sensor nodes
–
21 idSelected ← idP otential idNonReachable (line 16).
22 idF easible ← idP otential If removing the nonreachable sensor nodes in the set
23 SetZero(L, idN onReachable) idNonReachable from the set of selected sensor nodes
24 SetZero(B, idN onReachable) idSelected (line 17) that causes the violation of the active
25 SetZero(C, idN onReachable) sensor constraint (line 18), then the sensor node {idMin} is
– removed from the set idFeasible of feasible sensor nodes (line
26 return idSelected 19), and we skip to the next iteration (line 20). If the looka-
head move does not violate the active sensor constraint, then
we replace the set of selected sensors idSelected and the set of
C, and initial energy E as an input, in addition to the parameters feasible sensor nodes idFeasible with the potential sensor set
μ and τ representing the minimum threshold on the number of idPotential , and we set the values of link quality and correla-
selected sensors and the link quality, respectively. Our heuristic- tion for the nonreachable sensor nodes idNonReachable to zero
based algorithm is listed in Algorithm 1. In our algorithm, the using the function SetZero (lines 21–25). The algorithm ends
identifier of a variable is written with italic font, and the iden- if there are no more feasible sensor nodes (idFeasible = ∅) or
tifier of a function is written with typewriter font. Here, the the active sensor constraint is violated.
variables are imperative programming variables as opposed to
the CP decision variables of Section IV-A.
V. BAYESIAN I NFERENCE AND DATA Q UANTIZATION
The heuristic algorithm creates a set of selected sensors
idSelected (line 1), and initialize it with all the possible sen- This section describes how to use BP to infer the missing data
sor ids. The function BestShortestPath (line 2) takes the because of the inactive sensor nodes and the data transmission
link quality matrix L and base station link quality array B losses of the active sensor nodes throughout the second phase
as an input, and returns an array q of path quality values for of our ASBP protocol.
the shortest path from each sensor node to the base station.
The implementation of BestShortestPath is trivial, as it uses
Dijkstra’s algorithm [23] to compute the shortest paths, while A. Introduction to BP
respecting the path quality constraint (5). BP is a classic algorithm for performing inference on graph-
The heuristic algorithm maintains a set idNonReachable of ical models [8], [9]. In general, it assumes that some observa-
sensor nodes that are not able to reach the base station due tions are made and the task is to infer the underlying events
to the violation of the path quality constraints (5) (line 3). behind these observations. Denote yi the observation at node i
Before entering the main loop of the algorithm, any sensor and xi the underlying event, i = 1, . . . , N . For the application
nodes in the set idNonReachable are removed from the set of of IoT, yi is the reading of sensor i about some phenomenon
selected sensor nodes (line 4), and the values of link quality, that is being monitored, such as the temperature, and xi is
BIJARBOONEH et al.: CLOUD-ASSISTED DATA FUSION AND SENSOR SELECTION FOR IoT 263

on tree graphs, loopy BP has been shown to work well in most

cases for graphs with loops [24]. In addition, there are two gen-
eral BP variations which are sum-product and max-product BP,
respectively [25]. The latter is adopted in this paper because of
its efficiency.

Fig. 4. Example of a graphical model. B. BP for Inference on IoT

In using BP for inferring the missing data in IoT, we need
to construct a graph to model the correlations between sensor
readings. There are two types of correlations in sensor network.
1) Spatial correlation: Data from different sensors may be
correlated with each other. Note that we do not assume
that strong correlations always exist between data from
nearby sensors. Instead, we compute the correlation coef-
ficients between each pair of sensor nodes from the
Fig. 5. Graphical depiction of message passing from nodes p and q to the node observed data. We claim spatial correlations only when
i in BP The updated message mij (xj ) is then sent to the node j. we see large correlation coefficients, regardless of the
spatial distance between two sensors.
the true reading of the phenomenon. Clearly, there are some 2) Temporal correlation: Data from the same sensor may be
statistical dependencies between yi and xi , encoded in a so- correlated over time. Here, we simply assume that the sen-
called evidence function φi (xi , yi ). Very often, we consider sor reading at time t is strongly correlated with that at time
the observation yi to be fixed and write φi (xi ) as a short-hand t − 1.
of φi (xi , yi ). Furthermore, there are also statistical dependen- Thus, we built our graph as illustrated in Fig. 6 where
cies between the several underlying events xi , encoded in a xti denotes the true reading of sensor i at time t. The link
so-called potential function φij (xi , xj ). In IoT, the potential between xti and xt−1
i represents the temporal correlations, with
function captures spatial correlations between the readings at a temporal potential function defined as
nearby sensors.
Given the above notation, the inference of the xi can be for- (xti − xt−1 )2
φti (xti , xt−1 ) = exp − i
.
mulated as the maximization of the following belief function:
i
σi2
 
b({xi }Ni=1 ) = φij (xi , xj ) φi (xi ). Similarly, the link between xti and xtj represents the spatial
ij i correlations, with a spatial potential function associated and
defined as
A graphical depiction of this model is shown in Fig. 4. The rect-
angles are the observation nodes yi and the circles represent (xti − xtj )2
the underlying events xi . The potential functions are associ- φsij (xti , xt−1
i ) = exp − 2 .
σij
ated with the links between xi and the evidence functions are
associated with the links between yi and xi . Note that the noisy sensor reading yit is omitted from the graph
BP performs inference by passing messages between nodes for the purpose of simplification, and the evidence function
in the graph. The message from i to j is defined as associated with the link between xti and yit is defined as
 
mij (xj ) = φi (xi )φij (xi , xj ) mki (xi ) (xti − yit )2
φei (xti , yit ) = exp − .
xi k∈N (i),k=j σi2
where N (i) denotes the neighbors of node i. The message yit can be missing for two reasons: either sensor i is in the
essentially integrates all messages from the neighbors of i, sleep mode or the packet failed to reach the base station. When
except j, as well as the local evidence seen at i. Intuitively, it is missing, we turn the evidence function into a constant,
such a message models how likely it is at node i that node j i.e., φei (xti , yit ) = 1, for all possible values of xti . Such a con-
will be in the state of xj when node i is in the state xi . Thus, stant evidence function essentially treats everything as equally
BP performs message passing between nodes until reached con- as possible. Intuitively, BP handles missing sensor readings by
vergence, and the inference is done by maximizing the belief at reasoning from the past data and the sensor nodes with cor-
each node, which is to gather all incoming messages and the related data. Note that σi and σij are parameters that can be
local belief, i.e., learned from some training data [26].
 In comparison with approaches such as the CS-based
bi (xi ) = φi (xi ) mji (xi ).
approach in [11], BP based on the graph in Fig. 6 is advan-
j∈N (i)
tageous for several reasons.
The message passing process in BP is illustrated in Fig. 5. 1) BP captures the spatial and temporal correlations between
BP is well established in both theory and practice. For exam- sensors explicitly via a graphical model which is updated
ple, while it is known that BP is only guaranteed to converge over time. For example, we can disconnect the sensor
264 IEEE INTERNET OF THINGS JOURNAL, VOL. 3, NO. 3, JUNE 2016

Fig. 6. Graphical model built for WSNs.

nodes when the correlation coefficients drop below some

Fig. 7. 2-bit uniform quantization on the data x1 , . . . , xn that partitions the
threshold. interval {rmin , rmax } into four equal bins using τ = {τ1 , τ2 , τ3 }. Each bin is
2) BP allows the incremental inference that infers the miss- represented by the centroid point of the bin, which is stored in a codebook. A
ing data at time t from the available data at just time t and metric value is then mapped into a bin index, encoded into 2 bits.
t − 1. In contrast, the CS-based approach in [11] takes
as an input data matrix with missing entries, and thus can The length of each partition τi − τi−1 is either uniform with
−vmin
only perform inference in a batch mode for a time interval. τi − τi−1 = vmaxR or nonuniform. In general, the thresh-
We will demonstrate these advantages and the inference olds τ are chosen according to the requirements of the applica-
accuracy in Section VI. tion, adaptively adjusted, or learned from a set of training data.
For example, consider an indoor temperature monitoring, where
the temperature varies at most between 0◦ and 50◦ . Given 0.2◦
C. Data Quantization temperature accuracy requirement of the application, the min-
Quantization is a classic technique in signal processing that imum number of quantization level is (vmax − vmin )/0.2 =
has been widely used for data compression [27]. Quantization 50/0.2 = 250, which implies that at least 8-bit quantization res-
of network data saves storage as it encodes the data into olution (28 = 256 bins) is necessary to satisfy the requirement
fewer bits. It requires fewer number of transmissions and of the application.
smaller packet size. In many applications, a quantized mea- As we mentioned in this section, the data quantization is
sure is informative enough to represent aspects of the network. lossy with the error defined as
For example, many heating, ventilation, and air conditioning ε(x) = x − code(bi ), for Q(x) = bi
(HVAC) sensors only react if temperature or humidity falls
within certain thresholds. In summary, quantized measures are where code maps the code bi to the metric value of the data
less fine-grained and lossy; however, there are many advantages x (typically, the centroid point of the bin). The error is upper
in using a quantized measure. bounded by the bin length, given by
1) A quantized measure is informative enough for describing
the correlation between the data. ε(x) < τi − τi−1 .
2) A quantized measure can be encoded into a few bits, The quantization error is inversely proportional to R, whereby a
saving storage and transmission costs. smaller R leads to a larger ε(x). When R is as large as vmax −
3) A quantized measure is coarse and thus cheaper to obtain. vmin , the quantization becomes equivalent to the rounding of
It is also stable and highly adjustable to match the needs the real value, which is almost lossless.
of the network application.
Let the metric to be quantized take on values in the range
[rmin , rmax ], and values outside this interval are mapped either VI. E XPERIMENTS
to rmin or rmax . The quantization is done by partitioning the
A. Experimental Setup
interval into R bins using R − 1 thresholds, denoted by τ =
{τ1 , . . . , τR−1 }. Each bin is represented by a value within the We experiment with the real data collected from 54 sen-
range of the bin, e.g., the centroid point of the bin’s range. Let sor nodes deployed in the Intel Berkeley Research Laboratory
the value bi represent the ith bin. A look-up table is used to map [22]. The data are collected by a base station, and includes
the metric value to bi according to the bin threshold temperature, humidity, light intensity, and voltage values once
every 30 s, throughout a time span of 36 days. The data set
Q(x) = bi , if τi−1 < x ≤ τi , i = 1, . . . , R. (10) also includes aggregated connectivity data, representing the link
quality between any two sensor nodes, and between sensor
where τ0 = rmin and τR = rmax . The bin index values nodes and the base station. In our simulations of the ASBP pro-
{b1 , . . . , bR } are stored in a codebook, and a metric tocol, we selected a time interval of 10 h, consisting of 5 rounds
value can then be represented by a bin index that is of 2 h, such that at least 30% of data are transmitted successfully
encoded into few bits. For example, Fig. 7 shows six data to the base station.
x1 , . . . , xn quantized into four bins with 2-bit binary indices We apply a uniform quantization on the temperature data
b1 = (00)2 = 0, . . . , b4 = (11)2 = 3 according to (10). between 10th and 90th percentile into 256 bins, where each bin
BIJARBOONEH et al.: CLOUD-ASSISTED DATA FUSION AND SENSOR SELECTION FOR IoT 265

to 70% affects the routing of the data in the multihop data col-
lection. It increases the size of the data collection to five hops,
which requires the sensor nodes closer to the base station to
relay also the data for the nodes further away. Hence, the path
quality q[s] is decreased, and the total data utility is reduced.
In our results, the CP sensor selection achieves the optimum
data utility, and the greedy heuristic-based algorithm manages
to find a satisfactory local optimum. The results show that the
general traditional random approach does perform very poorly
compared to the global optimum. The results for the random
Fig. 8. Correlation coefficient matrix represented with jet color map, with sensor selection are computed by taking the mean of the data
absolute values varying in range [0, 1]. utility and energy consumption for ten random sensor selec-
tions. In all cases, the solution of the sensor selection problem
for CP and the heuristic-based algorithm were found in less
is represented with an 8-bit value in the codebook. The values
than 1 min. We observe that the data utility increases up to 25
outside the interval are mapped to the minimum and maximum
selected nodes and then decreases. This is because of the trade-
of the interval accordingly.
off between the path quality and the correlation. As the number
Fig. 8 demonstrates the absolute values for the correlation
of selected sensors increases, the sum of the data correlation
coefficient matrix using the jet color map. We expect that
between a selected sensor node and all the other sensor nodes
since the data are highly correlated, the uniform quantized
becomes a larger factor in the data utility term (3) compare to
data are also highly correlated. Our experiments show that 8-
the path quality term; hence, the data utility decreases. We con-
bit quantization resolution introduces at most 15% error in the
clude that an efficient sensor selection strategy should select 25
data correlation. However, it does not affect the BP prediction
sensor nodes to maintain a balance between the path quality and
results, as the data are already quantized when received at the
the data correlation.
base station.
The heuristic-based strategy in Fig. 9(b) fails to find a solu-
In our energy consumption evaluations, we consider 14-mA
tion for more than 30 selected sensor nodes, because our
transmission cost, as reported for the Mica2Dot mote [28], used
requirement for reaching the base station is limited to at least
in the Intel Berkeley Research Laboratory deployment.
70% link quality, and without backtracking, the greedy algo-
In our simulations, each round is 2 h, where phase one of
rithm fails at maintaining a route to the base station for all
a round ends if at least 20 data readings are collected at the
selected sensor nodes.
base station from all the sensor nodes. The weights ω1 and ω2
The total energy consumption (in terms of the number of
in data utility (3) are chosen to normalize the path quality and
transmission for data collection and node coordination) for the
correlation. We expect that at least μ sensor nodes are selected;
data transmissions with both settings 30% and 70% on the
hence, the path quality is scaled by the minimum number μ (8)
minimum base station link quality is shown in Fig. 9(c). The
of sensor nodes (ω1 = μ), because the sum of the correlation is
minimum base station link quality is denoted in the legend of
at least μ we set ω2 = 1. The threshold τ of (5) is set to 0.7.
the plot. We observe that at the same threshold on the base sta-
The base station then solves the sensor selection optimization
tion link quality, the energy consumption is almost independent
problem and initiates the second phase of the ASBP protocol.
of the sensor selection strategy. However, the energy consump-
tion is almost doubled as the base station link quality threshold
is increased to 70%, which is due to the additional multihop
B. Results and Analysis
relay of the data required to reach the base station.
We evaluate the performance of our ASBP in terms of data Fig. 10(a) shows the BP results with the CP model, heuristic-
utility, energy efficiency, and data prediction accuracy. We based algorithm, and random sensor selection strategies, upon
compare the data prediction error of the results of our CP varying the minimum number μ of selected sensor nodes. We
model, heuristic-based algorithm, and a random sensor selec- first compute the mean square error (MSE) of the predicted data
tion. On the inference accuracy, we compare with the CS-based versus the ground truth for each sensor node in the temporal
approach in [11] which we consider as the state-of-the-art. Our domain. The result is an array of 54 MSE values on the sensor
simulation of the ASBP protocol is implemented in C++, and node predicted data. We then plot the mean of the MSE error
the CP model is implemented using the CP solver Gecode [29] in Fig. 10(a). The results for the random sensor selection are
(revision 4.2.1), and runs under Mac OS X 10.9.2 64 bit on an computed by taking the average of ten runs. The standard devi-
Intel Core i5 2.6 GHz with 3 MB L2 cache and 8 GB RAM. ation (SD) of CP and the heuristic-based algorithm is at most
Fig. 9(a) and (b) compares the total data utility and energy 12%. The CP model with μ = 10 has an average error of about
consumption achieved in one round by the ASBP protocol using 5%, which indicates that in the temporal domain, in average the
CP, our heuristic-based algorithm, and random sensor selection, prediction of the BP deviates 5% from the ground truth. At the
with a minimum of 30% and 70% for the base station link qual- same data point, the SD is about 12%, and increasing the num-
ity, respectively. For each result, we vary the parameter μ in (8) ber of selected sensor nodes μ always drops the value of SD.
to control the total number of selected sensor nodes for data col- As we expected, the best sensor selection (by CP) achieves the
lection. The increase in the minimum base station link quality minimum error, whereas the random sensor selection does not
266 IEEE INTERNET OF THINGS JOURNAL, VOL. 3, NO. 3, JUNE 2016

Fig. 9. Data utility and energy consumption for data transmission obtained by simulating the ASBP protocol in one round and solving the sensor selection problem
with the CP model, our heuristic-based algorithm, and random sensor selection. Minimum thresholds of: (a) 30% and (b) 70% were used for the base station link
quality, upon varying the minimum number μ of selected sensor nodes. (c) Energy consumption.

Fig. 10. Prediction MSE of our BP-based approach and the CS-based approach in [11] using the CP model, the heuristic-based algorithm, and the random sensor
selection strategies, upon varying the minimum number μ of active nodes. (a) Prediction error of BP with CP, Heuristic, and random node selection. (b) Prediction
error of BP versus CS using our heuristic-based node selection algorithm. (c) Prediction error of BP versus CS using random node selection.

consider the correlation of the data, and as a result has a higher to the explicit modeling of the data correlations in time and in
prediction error. space in the potential functions [9].
The results compared with the energy consumption in Fig. 10(b) and (c) shows the comparisons between our BP-
Fig. 9(c) show that we can save up to 80% energy by select- based approach and the CS-based approach in [11] using the
ing only 10 sensor nodes to be active for the data collection heuristic-based and random node selection, respectively. It can
in each round, while maintaining at most the satisfactory aver- be seen that on the heuristic-based node selection, BP is strictly
age error of 5% with an SD of 12% in the prediction accuracy. better than CS. For example, BP achieves 16% lower predic-
In our approach, depending on the application and the required tion error compared to CS when μ = 10. On the random node
accuracy, we can adjust the selected number of sensor nodes selection, the two perform similarly. Note that the results on
as a tradeoff between the energy consumption and data quality random node selection are the average of ten runs. Such results
(accuracy of the BP). reveal the advantage of BP that the spatio-temorpal correlations
On the inference accuracy, we compared our BP-based are explicitly encoded in the graph structure and in the potential
approach with the CS-based approach in [11]. In particular, [11] functions, which leads to the better accuracy in Fig. 10(b). On
modeled the estimation of the lost data as a problem of matrix the other hand, in Fig. 10(c), BP builds the graph and learns
completion, where an EM matrix is constructed by recording the potential functions on randomly selected nodes without
the data reading of a particular sensor at a particular time. The considering the correlations, whereas CS assumes the random
EM matrix is incomplete because some data are lost during sampling of the data which hold here. Even in such scenarios,
transmission and some sensors are inactive, i.e., not selected, BP still achieves a similar performance as CS.
during some time periods. By applying the matrix completion
techniques developed in CS, the missing data in the EM matrix
VII. C ONCLUSION
can also be estimated. While interesting, a drawback of the
matrix completion formulation in [11] is that in order to con- By exploring cloud computing with the IoT, we present a
struct the EM matrix, data must be collected in different sensors cloud-based solution that takes into account the link quality
regularly and in a synchronized way, so that the data in the time and spatio-temporal correlation of data to minimize energy
dimension are consistent. In contrast, our BP-based approach consumption by selecting sensors for sampling and relaying
makes no such assumption and allows the sensors to collect data data. We have presented a novel cloud-based ASBP protocol
at irregular frequencies or even randomly. This is possible due with energy-efficient data collection for the IoT applications.
BIJARBOONEH et al.: CLOUD-ASSISTED DATA FUSION AND SENSOR SELECTION FOR IoT 267

ASBP solves an optimisation problem to select an optimal set of [17] A. Ulusoy, O. Gurbuz, and A. Onat, “Wireless model-based predic-
active sensor nodes that maximizes the data utility and achieves tive networked control system over cooperative wireless network,” IEEE
Trans. Ind. Informat., vol. 7, no. 1, pp. 41–51, Feb. 2011.
energy load balancing. In our protocol, BP iteratively infers [18] M. Jongerden, A. Mereacre, H. Bohnenkamp, B. Haverkort, and
the values of the missing data from the stream of active sen- J. Katoen, “Computing optimal schedules of battery usage in embedded
sor readings. We have also compared our BP prediction results systems,” IEEE Trans. Ind. Informat., vol. 6, no. 3, pp. 276–286, Aug.
2010.
with the widely used compressive sensing technique [11], and [19] O. Gnawali, R. Fonseca, K. Jamieson, D. Moss, and P. Levis, “Collection
show that our BP algorithm significantly outperforms com- tree protocol,” in Proc. 7th ACM Conf. Embedded Netw. Sensor Syst.,
pressive sensing. We formulate and solve the active sensor 2009, pp. 1–14.
[20] S. R. Madden, M. J. Franklin, J. M. Hellerstein, and W. Hong, “Tinydb:
selection optimization problem using CP, and compare it with An acquisitional query processing system for sensor networks,” ACM
our heuristic-based greedy algorithm. Trans. Database Syst., vol. 30, no. 1, pp. 122–173, Mar. 2005.
We have evaluated the performance of our ASBP proto- [21] J. Chou, D. Petrovic, and K. Ramachandran, “A distributed and adaptive
signal processing approach to reducing energy consumption in sensor net-
col by extensive simulations using real data collected at the works,” in Proc. 22nd Annu. Joint Conf. IEEE Comput. Commun. IEEE
Intel Berkeley Research Lab sensor deployment and their link Soc. (INFOCOM’03), 2003, vol. 2, pp. 1054–1062.
quality estimates. The simulation results show that our ASBP [22] S. Madden. (2014). “Intel Lab data, 2004,” [Online]. Available from
http://www.select.cs.cmu.edu/data/labapp3/index.html
protocol can greatly improve energy-efficiency up to 80%, with [23] T. Cormen, C. Leiserson, R. Rivest, and C. Stein, Introduction to
the optimal CP active sensor selection, while maintaining in Algorithms, 3rd ed. Cambridge, MA, USA: MIT Press, 2009.
average 5% error in the BP data inference. [24] A. T. Ihler, J. W. Fischer III, and A. S. Willsky, “Loopy belief propaga-
tion: Convergence and effects of message errors,” J. Mach. Learn. Res.,
As future work, we plan to extend our ASBP protocol to a vol. 6, pp. 905–936, Dec. 2005.
fully distributed implementation for real deployment, and com- [25] Y. Weiss and W. T. Freeman, “On the optimality of solutions of the max-
pare versus our current optimal results. We are also interested product belief-propagation algorithm in arbitrary graphs,” IEEE Trans.
Inf. Theory, vol. 47, no. 2, pp. 736–744, Sep. 2006.
to integrate adaptive sampling rate into our current results, as [26] J. Su, H. Zhang, C. X. Ling, and S. Matwin, “Discriminative parameter
well as investigating multisink scenarios. learning for Bayesian networks,” in Proc. 25th Int. Conf. Mach. Learn.
(ICML’08), 2008, pp. 1016–1023.
[27] A. Gersho and R. M. Gray, Vector Quantization and Signal Compression.
R EFERENCES Norwell, MA, USA: Kluwer, 1991.
[28] G. Anastasi, A. Falchi, A. Passarella, M. Conti, and E. Gregori,
[1] Cisco. (2011). “The Internet of Things,” [Online]. Available: “Performance measurements of motes sensor networks,” in Proc. 7th
http://share.cisco.com/internet-of-things.html ACM Int. Symp. Model. Anal. Simul. Wireless Mobile Syst. (MSWiM’04),
[2] L. Atzori, A. Iera, and G. Morabito, “The Internet of Things: A survey,” 2004, pp. 174–181.
Comput. Netw., vol. 54, no. 15, pp. 2787–2805, Oct. 2010. [29] Gecode Team. (2006). “Gecode: A generic constraint development envi-
[3] O. Vermesan et al., “Internet of Things strategic research roadmap,” in ronment,” [Online]. Available: http://www.gecode.org/
Internet of Things-Global Technological and Societal Trends. Delft, The
Netherlands: River Pub., 2011, pp. 9–52.
[4] J. Amaro, F. J. T. E. Ferreira, R. Cortesao, N. Vinagre, and R. Bras, “Low Farshid Hassani Bijarbooneh received the
cost wireless sensor network for in-field operation monitoring of induc- Bachelor’s degree in applied mathematics from
tion motors,” in Proc. IEEE Int. Conf. Ind. Technol. (ICIT), Mar. 2010, the Iran University of Science and Technology,
pp. 1044–1049. Tehran, Iran, in 1999, and the Master’s degree in
[5] S. Madden, M. J. Franklin, J. M. Hellerstein, and W. Hong, “Tag: A computer science and Ph.D. degree in constraint
tiny aggregation service for ad-hoc sensor networks,” SIGOPS Oper. Syst. programming for wireless sensor networks from
Rev., vol. 36, no. SI, pp. 131–146, Dec. 2002. Uppsala University, Uppsala, Sweden, in 2009 and
[6] S. Madden, R. Szewczyk, M. Franklin, and D. Culler, “Supporting aggre- 2015, respectively.
gate queries over ad-hoc wireless sensor networks,” in Proc. 4th IEEE He was with the Mobility and Astra Research
Workshop Mobile Comput. Syst. Appl., 2002, pp. 49–58. Groups, Uppsala University. He is currently a
[7] J. Pearl, Probabilistic Reasoning in Intelligent Systems: Networks of Postdoctoral Researcher with the SyMLab Research
Plausible Inference. San Mateo, CA, USA: Morgan Kaufmann, 1988. Group, Hong Kong University of Science and Technology, Clear Water Bay,
[8] J. S. Yedidia, W. T. Freeman, and Y. Weiss, “Understanding belief propa- Hong Kong. He has visited and collaborated with researchers from the
gation and its generalizations,” in Exploring Artificial Intelligence in the Insight Centre for Data Analytics, Cork University, Cork, Ireland; INRIA
New Millennium, G. Lakemeyer and B. Nebel, Eds. San Mateo, CA, USA: Paris-Rocquencourt, Paris, France; Computer Networks (NET) Research
Morgan Kaufmann, 2003, pp. 239–269. Group, Göttingen University, Göttingen, Germany; and the Research Unit in
[9] F. V. Jensen, Introduction to Bayesian Networks, 1st ed. Berlin, Germany: Networking (RUN), University of Liège, Liège, Belgium. His research interests
Springer-Verlag, 1996. include optimization and constraint programming, cloud computing, Internet of
[10] L. Kong, D. Jiang, and M.-Y. Wu, “Optimizing the spatio-temporal Things, and sensor networks.
distribution of cyber-physical systems for environment abstraction,” in
Proc. IEEE 30th Int. Conf. Distrib. Comput. Syst. (ICDCS), Jun. 2010,
pp. 179–188.
[11] L. Kong, M. Xia, X.-Y. Liu, M.-Y. Wu, and X. Liu, “Data loss and
reconstruction in sensor networks,” in Proc. IEEE INFOCOM, 2013,
pp. 1654–1662. Wei Du received the B.S. degree in computer sci-
[12] F. Rossi, P. van Beek, and T. Walsh, Eds., Handbook of Constraint ence from Tianjin University, Tianjin, China, in 1997,
Programming. Amsterdam, The Netherlands: Elsevier, 2006. and the Ph.D. degree in computer science from the
[13] L. D. Xu, “Enterprise systems: State-of-the-art and future trends,” IEEE Chinese Academy of Sciences, Beijing, China, in
Trans. Ind. Informat., vol. 7, no. 4, pp. 630–640, Nov. 2011. 2002.
[14] J. Zheng, D. Simplot-Ryl, C. Bisdikian, and H. Mouftah, “The Internet of He is a Postdoctoral Researcher with the CITI
Things,” IEEE Commun. Mag., vol. 49, no. 11, pp. 30–31, Nov. 2011. Laboratory, INSALyon, Villeurbanne, France. Since
[15] L. Palopoli, R. Passerone, and T. Rizano, “Scalable offline optimization of 2012, he has been a Postdoctoral Researcher with
industrial wireless sensor networks,” IEEE Trans. Ind. Informat., vol. 7, INRIA, Rocquencourt, France; Hamburg University,
no. 2, pp. 328–339, May 2011. Hamburg, Germany; the University of Liège, Liège,
[16] J. Haupt, W. Bajwa, M. Rabbat, and R. Nowak, “Compressed sensing for Belgium; the University of Innsbruck, Innsbruck,
networked data,” IEEE Signal Process. Mag., vol. 25, no. 2, pp. 92–101, Austria; and the University of Göttingen, Göttingen, Germany. His research
Mar. 2008. interests include applications of machine learning on computer networking.
268 IEEE INTERNET OF THINGS JOURNAL, VOL. 3, NO. 3, JUNE 2016

Edith C.-H. Ngai received the Ph.D. degree from Jiangchuan Liu (S’01–M’03–SM’08) received the
the Chinese University of Hong Kong, Shatin, Hong B.Eng. degree in computer science (cum laude) from
Kong, in 2007. Tsinghua University, Beijing, China, in 1999, and
She is currently an Associate Professor with the Ph.D. degree in computer science from the Hong
the Department of Information Technology, Uppsala Kong University of Science and Technology, Clear
University, Uppsala, Sweden. She was a Postdoctoral Water Bay, Hong Kong, in 2003.
Researcher with Imperial College London, London, He is a University Professor with the School
U.K., from 2007 to 2008. Since 2015, she has been of Computing Science, Simon Fraser University,
a Visiting Professor with Ericsson Research Sweden. Burnaby, BC, Canada, and an NSERC E. W. R.
Her research interests include wireless sensor and Steacie Memorial Fellow. He is an EMC-Endowed
mobile networks, Internet of Things, network security Visiting Chair Professor with Tsinghua University
and privacy, smart city, and e-health applications. (2013–2016). From 2003 to 2004, he was an Assistant Professor with the
Dr. Ngai is a member of the ACM. She has served as a TPC Member Chinese University of Hong Kong. His research interests include multimedia
in leading networking conferences, including IEEE ICDCS, IEEE Infocom, systems and networks, cloud computing, social networking, online gaming, big
IEEE ICC, IEEE Globecom, IEEE/ACM IWQoS, IEEE CloudCom, etc. data computing, wireless sensor networks, and peer-to-peer networks.
She was a TPC Co-Chair of the Swedish National Computer Networking Dr. Liu has served on the Editorial Boards of the IEEE T RANSACTIONS
Workshop (SNCNW’12) and QShine’14. She is a Program Chair of ACM ON B IG DATA , the IEEE T RANSACTIONS ON M ULTIMEDIA , IEEE
womENcourage 2015, a TPC Co-Chair of IEEE SmartCity 2015 and IEEE C OMMUNICATIONS S URVEYS AND T UTORIALS, IEEE ACCESS, the IEEE
ISSNIP 2015. She has served as a Guest Editor for a special issue of I NTERNET OF T HINGS J OURNAL, Computer Communications, and Wiley’s
the IEEE I NTERNET OF T HINGS J OURNAL, the IEEE T RANSACTIONS ON Wireless Communications and Mobile Computing. He is the Steering
I NDUSTRIAL I NFORMATICS, Springer Mobile Networks and Applications Committee Chair of the IEEE/ACM IWQoS from 2015 to 2017. He was the
(MONET), and the EURASIP Journal on Wireless Communications and corecipient of the inaugural Test of Time Paper Award of the IEEE INFOCOM
Networking. She is a VINNMER Fellow (2009) awarded by VINNOVA, (2015), ACM TOMCCAP Nicolas D. Georganas Best Paper Award (2013),
Sweden. Her coauthored papers have received the Best Paper Runner-up ACM Multimedia Best Paper Award (2012), the IEEE Globecom Best Paper
Awards of IEEE IWQoS 2010 and ACM/IEEE IPSN 2013. Award (2011), and the IEEE Communications Society Best Paper Award on
Multimedia Communications (2009).
Xiaoming Fu (M’02–SM’09) received the Ph.D.
degree in computer science from Tsinghua
University, Beijing, China, in 2000.
He was a Research Staff with the Technical
University Berlin, Berlin, Germany, until joining
the University of Göttingen, Göttingen, Germany, in
2002, where he has been a Professor of computer sci-
ence and the Head of the Computer Networks Group
since 2007. His research interests include network
architectures, protocols, and applications.
Dr. Fu is a Distinguished Lecturer of the IEEE
Communications Society. He is currently an Editorial Board Member of
the IEEE C OMMUNICATIONS M AGAZINE, the IEEE T RANSACTIONS ON
N ETWORK AND S ERVICE M ANAGEMENT, Elsevier Computer Networks, and
Computer Communications, and has served on the Organization or Program
Committees of leading conferences such as INFOCOM, ICNP, ICDCS,
MOBICOM, MOBIHOC, CoNEXT, ANCS, ICN, and COSN. He has served
as a Secretary (2008–2010) and a Vice Chair (2010–2012) of the IEEE
Communications Society Technical Committee on Computer Communications
(TCCC), and Chair (2011–2013) of the Internet Technical Committee (ITC) of
the IEEE Communications Society and the Internet Society. He has been
involved in EU FP6 ENABLE, VIDIOS, Daidalos-II, and MING-T projects
and is the Coordinator of the FP7 GreenICN, MobileCloud, and CleanSky
projects. He was the recipient of the ACM ICN 2014 Best Paper Award, the
IEEE LANMAN 2013 Best Paper Award, and the 2005 University of Göttingen
Foundation Award for Exceptional Publications by Young Scholars.