This article has been accepted for publication in a future issue of this journal, but has not been

fully edited. Content may change prior to final publication. Citation information: DOI 10.1109/TII.2016.2627478, IEEE
Transactions on Industrial Informatics
1

Decentralized control algorithm for fast monitoring

and efficient energy consumption in energy
harvesting wireless sensor networks
Sabato Manfredi∗† , Member, IEEE, Edmondo Di Tucci∗
∗ Department of Electrical Engineering and Information Technology, University of Naples Federico II, Naples, Italy
† Control and Power Group, Electrical and Electronic Engineering Department, Imperial College, London

Email: sabato.manfredi, [email protected]

Abstract—The paper proposes a fully decentralized algorithm the algebraic connectivity of the network graph affects the
to control nodes transmission radii for simultaneously coping convergence rate of the consensus algorithm. Therefore in a
with energy-type requirement (e.g. network power consumption) wireless networked communication system the convergence
and application-type requirement (e.g. fast monitoring and esti-
mation of variables of interest). The stability of the closed loop rate can be improved by increasing the node’s communication
network under the proposed control law is proved. Simulation radii (i.e. the network connectivity). This however causes the
validations show the effectiveness of the proposed approach in increase of the energy consumption and thus the reduction of
nominal scenario as well as in the presence of uncertain power the EH-WSN lifetime ([11]). Although energy consumption
requirements and harvesting system supply, outperforming a well and algorithm responsiveness are in tradeoff, the most part
known related approach existing in the literature.
of the existing literature focuses on network autonomy maxi-
Index Terms—Wireless Sensor Networks, Consensus Algo- mization or on convergence rate minimization separately.
rithms, Optimization of Industrial Wireless Networks, Energy
Specifically, in ([12]) it is proposed an online adaptive
Efficiency, Energy Harvesting.
energy efficient algorithm where each node autonomously
adjusts its transmission radius on the base of residual power.
I. I NTRODUCTION Differently, in ([13]) the authors introduce a topology main-

T HE recent development of low cost smart sensors and

micro-controllers supporting network connectivity ([1])
is facilitating the implementation of distributed monitoring and
tenance algorithm that dynamically controls the transmission
radii to extend the lifetime of the network while preserving
the coverage of the monitored area. In these works the
control algorithms in many applications (i.e. military, agricul- convergence rate of the distributed algorithm has not been
ture, environment monitoring, home automation, automotive, taken into account. Differently, other works achieve network
industrial applications) by architectures based on WSN ([2]). connectivity regulation by controlling the transmission radius
Indeed wireless data communication networks present reduced in order to satisfy consensus algorithm requirements (e.g.
costs, better power management, ease in maintenance and ef- convergence rate). In [14] authors present the estimation and
fortless deployment in remote and hard-to-reach areas. Energy control of graph connectivity in a decentralized manner. In
Harvesting Wireless Sensor Network (EH-WSN) architecture [15], [16] authors implement a decentralized control strategy
is composed of sensor nodes supplied by harvesting systems that adaptively steers the transmission radii in order to en-
([3]) that manage the energy provided by the battery and/or sure network connectivity maintenance. In the above works,
renewable source (e.g. RF energy, vibrations, thermal energy) the network connectivity is estimated and controlled, while
in order to extend the network autonomy ([5], [7]). Different network lifetime requirement is neglected.
approaches have been proposed to specifically improve the In this paper we formulate a decentralized control law
energy performance of EH-WSN by power control (see e.g. of node transmission radius in order to simultaneously a)
[8] and references therein). EH-WSNs support applications minimize the network energy consumption and b) maximize
in different emerging monitoring and control fields such as the network connectivity. The former aim a) allows to extend
building thermal monitoring, bridge structural health monitor- the lifetime of the overall EH-WSN system, while the latter
ing, earthquake forecast, surveillance, wireless robot systems aim b) increases the responsiveness of the monitoring/control
([2], [9]). A well known distributed protocol aimed to estimate consensus algorithm.
the average of monitored signals (e.g. spatial environmental Only few works investigate the optimal choice of transmis-
temperature) is the consensus algorithm suitable for linear sion radii by exploiting the existing tradeoff between network
and nonlinear estimations ([4], [6], [10]). An important design lifetime and graph connectivity. Specifically, in Sardellitti et
requirement is the responsiveness of the above estimation al. ([17]) it is proposed an offline centralized method that
algorithm for fast monitoring purpose. It is well known that evaluates the node transmission range in order to optimize both
energy and connectivity performance. The proposed solution
Copyright 2009
c IEEE. Personal use of this material is permitted. How-
ever, permission to use this material for any other purposes must be obtained considers all radii being equal and the optimal solution is
from the IEEE by sending a request to [email protected] evaluated independently on the actual available node energy.

1551-3203 (c) 2016 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.
This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication. Citation information: DOI 10.1109/TII.2016.2627478, IEEE
Transactions on Industrial Informatics
2

A similar framework is considered in [18] where radii may The main advantages in a WSN decentralized scheme are
be different and are steered in a centralized way in order to high reliability and network responsiveness and this will be
optimize both energy and consensus algorithm performance. the architecture considered in the paper.
It is worth of note that an interesting and effective online de- Specifically each network node implements a decentralized
centralized algorithm version, namely Decentralized Heuristic cooperative estimator (e.g. consensus algorithm) in order to
Algorithm (DHA in the following) is presented in [19]. Herein, distributively extract a global information (or take a global
we propose a fully Distributed Energy-based Radii Control decision) on the base of local measurements and radio-
algorithm (DERC in the following) to simultaneously optimize acquired data from its neighbours. We also assume that nodes
both energy and consensus performance thus encompassing are supplied by batteries and can eventually be supported by
all offline centralized approaches presented in the literature. renewable energy sources.
Hence the proposed approach is similar to the framework
proposed in [19] optimizing the same objective function to A. EH-WSN model
cope with aims a) and b) with the following notable differ- A EH-WSN composed of n sensors may be modelled
ences. Firstly, the method allows each node distributively to by a graph G with adjacency matrix A ∈ Rn×n , whose
implement a descend gradient algorithm in order to find the elements aij are assumed to be equal to 1 if a source node
optimal value of the transmission radii that minimizes the i is in communication with node j, and 0 otherwise. Let
objective function. Each node tunes its transmission radius pi = [pix , piy ] the coordinates of node i in a Cartesian
on the base of both the above descend algorithm result reference system, rij = ||pi − pj ||2 the euclidean distance
and the node available energy. This strongly improves the between nodes i and j, ρi the i-th transmission radius, aij
network lifetime because the actual node available energy is can be formulated as follows: aij = 1 if ρi ≥ rij and 0
explicitly taken into account in the radii controller. Secondly otherwise.
the above feature allows to get high performance on system A direct graph G is connected if there exists at least one
autonomy mainly in the presence of the energy harvesting, path for any source-destination couple of nodes. The graph
that is a feature not considered by the other approaches in is balanced if it is connected and the number of input links
literature. Additionally a closed-loop network stability analysis equals the number of output links at each node. The network
is analytically carried out. Finally, our formulation is suitable graph is symmetric if aij = aji for all (i, j). The
also for asymmetric direct WSN network graph. P Laplacian
L = [lij ] is the matrix with element lij = aik if i =
The paper is organized as follows. In Section II the problem k6=i
statement is presented. In Section III the proposed distributed j; lij = −aij if i 6= j.
control law, the closed loop network stability analysis and For connected graph the Laplacian matrix L has one eigen-
the implementation features are detailed. In Section IV an value equal to zero and the others n − 1 non null, namely
experimental simulation validation of the proposed approach 0 = |λ1 | ≤ |λ2 | ≤ ... ≤ |λn |. The second smallest no-zero
under different scenarios is carried out. Conclusions are given eigenvalue λ2 is referred as the algebraic connectivity ([14],
in Section V. [16]) because is related to the connectivity property of the
Notation: In this manuscript x = [xi ] indicates a vector of graph.
scalars xi , the capital X = [xij ] denotes the matrix of elements In WSN, the algebraic connectivity of the graph can be
xij and (·)T is the transpose of its argument. 1 indicates a tuned by controlling the communication radii of nodes. The
vector of all ones. Let X be a real square matrix, the symbol relation between the transmission power and the covered
vi denotes the i-th eigenvector of X and λi the corresponding distance derives from the well known path loss formula ([20]).
eigenvalue. xre and xim denote respectively the real and the Let PTi (ρi ) the amount of power consumed by the trans-
imaginary part of the complex variable x = xre + ixim , where mitter node i and PT (ρi ) the signal power at distance ρi , it
 β
symbol i is the imaginary unit. |x| denotes the absolute value results PTi (ρi ) = PT (ρi ) · rρ0i where β > 0 is the channel
of x, N and R the sets of natural and real numbers. Moreover, power loss coefficient and r0 is a normalization factor. We
the following notations are defined: ||x||1 = x1 +x2 +···+x n
n
; denote with P Rj the lower threshold level of power at the
2 2 2
x +x +···+x
p
2 2 2 2
||x ||1 = 1 2
n
n
; ||x||2 = x1 + x2 + · · · + xn ; x
e receiver node j below of which it is impossible to distinguish
denotes an estimation of the variable x. between noise and signal. P Rj is assumed to be known or
estimated as it depends on the specific adopted transmission
II. P ROBLEM S TATEMENT
coding and the signal-to-noise ratio at the receiver. In order to
A WSN centralized system scheme is based on the wireless establish a communication between nodes i and j at distance
infrastructure to retrieve information from the environment and rij , two conditions have to hold: 1) the transmission radius of
to communicate the control decision to actuators, while the i has to be at least equal to the distance between nodes i and
elaboration of a specific decision is taken by a central control j, namely ρi ≥ rij and 2) the signal received by node j has
system (fusion center). Although this scheme has been exten- to be detectable, namely PT (rij ) ≥ P Rj . Herein, we consider
sively used in the past, it suffers from low reliability and high node i connected to node j at distance rij (i.e. aij = 1) if
latency. Differently, in a WSN decentralized system scheme,  ρ β
each node has to cooperate with the neighbouring nodes in i
PTi (ρi ) ≥ P Rj · (1)
order to evaluate in a decentralized way the elaboration carried r0
out by the fusion center in the case of WSN centralized system. with ρi ≥ rij .

1551-3203 (c) 2016 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.
This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication. Citation information: DOI 10.1109/TII.2016.2627478, IEEE
Transactions on Industrial Informatics
3

It is well known ([21]) that with the modern technology, the temperature monitoring) where each node needs to estimate in
average power consumption due to the transmission of data is a distributed way the average value of the measures of interest
always one or two orders of magnitude greater than wireless (e.g. average temperature) carried out by the network nodes
reception and computation contributes. Therefore in this paper, ei (t) is the estimation of x = ||x||1 ). In this case each
(i.e. x
it is assumed that the power consumed by node during time is node implements a dynamic average consensus algorithm. As
only due to the power transmission PTi (ρi ), while any other stated above, a necessary assumption for which the agreement
contribution is neglected. Anyway, in the simulation section, value x is the estimation of the average of the measured
we will validate the performance of the proposed algorithm values x is that the graph is balanced or symmetric [22]. The
also in the presence of uncertain power requirements. Modern latter assumptions are actually restrictive from the practical
wireless transmitter allows to vary the transmission radius, point of view, mainly in the considered framework where the
therefore herein we consider ρi = ρi (t). Moreover in the network graph is time varying due to the variation of the
following we omit the explicit dependence on time t when node radii. Therefore in the following we make a weaker
this does not make any ambiguity. assumption of connected digraph (i.e. the graph can be direct,
The dynamic of battery Penergy ξi (t)
may be modelled by
asymmetric and unbalanced). Additionally we are interested
ξ˙i (t) = −PTi (ρi (t)) + Ns δ t − ts where Ns δ t − ts in the estimation of the average value of input signals (i.e.
s∈Ni x = ||x||1 ), in the implementation of the proposed distributed
represents the Dirac function with amplitude Ns and centered
control algorithm we leverage on a decentralized balancing
at the time instant ts ∈ R and Ni ⊆ N modelling the effect of

procedure proposed in [16]. Hence we make the following
charging and discharging phenomena. Specifically, Ns δ t−ts
assumption:
is the net equivalent energy charged or spent by the node in
a time interval (ts−1 , ts ] respectively due to harvesting (i.e. Assumption 1. The network graph is connected.
Ns > 0) and power requirement event (i.e. Ns < 0).
In the following we model the dynamic of ξi (t) in
the C. Performance indexes and objective function of the EH-
presence of a recharge or discharge event Ns δ t − ts by WSN
ξ˙i (t) = −PTi (ρi (t)) with initial state condition at time instant Recently the analysis of the energy consumption perfor-
ts equivalent to ξi (ts ) = ξi (ts−1 ) + Ns . mance has received particular attention in the design of EH-
WSN. Indeed, the bad management of supply resource can
B. Consensus algorithms affect both the network lifetime and reliability. Any battery
The term consensus protocol denotes an algorithm distribu- is characterized by a threshold ξT H so that the battery is
tively implemented among the agents that allows each agent’s able to supply the node if ξi > ξT H . The node lifetime Tli
state estimation x ei to evolve to a global common value x can then be defined as the time instant at which the battery
(named the consensus value). The dynamic consensus law is: exhausts its energy, namely Tli : ξi (Tli ) ≤ ξT H . The
X network lifetime TL is defined as the shortest node lifetime
e˙ i (t) = −α
x xi (t) − x
aij (e ej (t)) + ẋi (t) (2) TL = mini=1,...,n Tli [11]. As a necessary condition for
j consensus algorithm convergence is the connectivity of the
where α is a gain factor, x ei is an estimation of a consensus network graph, a further performance of interest is the time at
value x̄ = g(x), with g is a function of a vector of measures which the network becomes disconnected. Due to the impor-
of interest x = [xi ]. Algorithm (2) can be recast in compact tance of this event, herein with the term network disconnection
form as x e˙ (t) = −αLe x(t) + ẋ(t) where L is the Laplacian time performance TD we refer to the time at which the
matrix. This algorithm refers to 1-hop dynamic consensus network becomes disconnected. Let PT (ρ) = [PTi (ρi )] be the
algorithm, according to the idea that the node updates the nodes transmission power, a reduction in the average power
local state variable just on the base of information of its 1-hop consumption ||PT (ρ)||1 yields to an increase of the network
neighbours. Notice that the stability analysis of the algorithm lifetime TL . As a consequence, TL is affected by ρ as it results
(2) may be recast in terms of stability analysis of the following TL ∝ ||PT 1(ρ)||1 .
agreement error system around the zero equilibrium point: Let TC the convergence time of the consensus algorithm,
it is defined as the time spent by each node estimation x
ei to
e˙ (t) = −αLe
e(t) (3)

 xei (TC )−x 
converge to a value x, TC : maxi   ≤ 0.05. It is
e
x
where e e = x e − x̄1 is the the agreement error (refer to [4], well known that TC decreases for the increasing of algebraic
[22] for further details). The connectivity of the network graph connectivity ([4], [22]). It is also clear that the increment of
is a necessary condition to assess the convergence of the the transmission radii ρ can yield to the increasing of the
estimation xi to an agreement value x (i.e. eei (t) to 0 for all i). algebraic connectivity of the network. Consequently, ρ can be
Additionally, the time needed by the estimation x ei to converge also used to affect TC because it results that TC ∝ λ2re1(ρ) .
to x (i.e. the convergence time) depends on the λ2re . Therefore, In conclusion ρ may be designed to tradeoff TC and TL .
an increase of λ2re reduces the time spent by node i-th to get an An unified objective function that takes into account both TC
estimation of x. Finally, for balanced or symmetric connected and TL can be defined as J(ρ) = ||Pλ2re T (ρ)||1 TC
(ρ) ∝ TL .
graph λ2 = λ2re and the agreement value x is the average The aim is to find the vector ρ that minimizes J in
value of the input signals xi ([22]), namely x = g(x) = ||x||1 . order to achieve the minimum consensus convergence time
Herein we are interested in control engineering scenario (e.g. at the maximum network lifetime. It is well known ([17],

1551-3203 (c) 2016 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.
This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication. Citation information: DOI 10.1109/TII.2016.2627478, IEEE
Transactions on Industrial Informatics
4

0.025 where ρ = [ρi (ρ)] ∀i. Due to the type of dependence of the
objective function J on ||PT (ρ)||1 and λ2re (ρ) detailed in
Section II-C, it results that J(ρ) is a convex function of ρ
0.02
assuming maximum value for lowest and highest ρ values.
Indeed small values cause the network to be disconnected and
0.015 high values increase network consumption. The steepest de-
J (ρ)

scent method is suitable to solve this kind of convex problem.

Indeed any initial value of ρi can dynamically approach to the
0.01
optimal ρ∗ according to the following equation:



0.005
ρ̇ = −γ∇J ρ(ρ) = −γ∇J ρ (5)

In (5), ∇ is the gradient operator, γ is a gain factor and

0 ρ = [ρi (ρ)]. ∇ operates on the function J, while the elements
0 10 20 30 40 50 of the vector ρ are given by equation (4).
ρ̄
Fig. 1. Objective function J(ρ) for ρ = [ρi ] with ρi = ρ ∀i. The function
has a minimum for ρ = 25. A. Closed-loop network stability analysis
From the energy node model and consensus algorithm
equation (3) and taking into account the control law (4), the
[19]) that the objective function for the considered scenario closed loop network model is:
has a minimum value. For instance, assuming a vector of
uniform transmission radii ρ = [ρi ] with ρi = ρ ∀i, the ξ˙i = −PTi (ρi ) (6a)
objective function J(ρ) is shown in Fig. 1 with a minimum e˙ = −αLee (6b)
at ρ = ρ∗ ' 25.
e
Indeed, starting from the minimum value of ρ assuring ρ̇i (ρ) = Ėi ρ + Ei ρ̇ = Ėi ρ − Ei γ∇J(ρ) (6c)
network connectivity, an increase of ρ strongly increases the e = [e
with e ei ] for i = 1, . . . , n.
network connectivity and this implies a decreasing of J(ρ) as h iT
shown in Fig. 1 for ρ ∈ [7, 25]. Further increasing of ρ implies Defined the vector of state variable z = ξ T e eT ρT with
an increasing of J(ρ) due to the high increase of ||PT (ρ)||1 ξ = [ξi ], ρ = [ρi ], the closed loop network equilibrium point
h T iT
(Fig. 1 for ρ ∈ [25, 50]). is ẑ = ξ̂ 0T ρ̂T where 0 is a zero vector, ξ̂ = [ξˆi ], ξˆi =
 µ
Assumption 2. The function J(ρ) is a differentiable function −ξ ∗
ξT H , ρ̂ = [ρ̂i ], ρ̂i = Êi ρ∗ , Êi = Ei (ξT H ) = ξξTMH−ξ ∗ + 1 .
with a minimum ρ∗ . i
We state the following result:

III. A FULLY DISTRIBUTED CONTROL LAW FORMULATION Theorem 1. If Assumptions 1 and 2 hold then the equilibrium
point ẑ of the closed-loop network (6) is asymptotically stable
The condition of uniform radii ρi = ρ ∀i represents a
and the state z asymptotically converges to ẑ.
suboptimal solution from the energy consumption point of
view. Indeed, nodes with lower energy availability spend the Proof. Let’s consider the candidate Lyapunov function:
same power of nodes with higher battery energy, thus reducing 1 X ˆ 2 + 1e

the overall network lifetime. V (z) = (ρi − Ei ρ∗ )2 + (ξi − ξ) eT e
e
2 i 2
The lifetime of the network can be significantly extended
if nodes with high and low energy availability respectively Notice that V (z) > 0 ∀z 6= ẑ and V (ẑ) = 0. Additionally it
increase and decrease their transmission radii. Let ρ the results:
reference radius for all nodes, the idea is to tune the radius ρi X X
at the node i-th proportionally to ξi . Namely, we consider at V̇ (z) = (ρi − Ei ρ∗ )ρ̇i − (ρi − Ei ρ∗ )ρ∗ Ėi
the node i the following control law: i i
T ˙
X X
µ + ˙
(ξi − ξT H )ξi + e
e ee= (ρ − ρ∗ )Ei ρ̇i
ξi (t) − ξ ∗

ρi (ρ) = + 1 ρ = Ei ρ (4) i
X i
ξMi − ξ ∗
e˙
X
− (ρ − ρ∗ )ρ∗ Ei Ėi + (ξi − ξT H )ξ˙i + e
eT e
where ξMi is the node maximum energy, ξ ∗ and µ are i
∗ Xi
controller parameters to be tuned such that ξξiM(t)−ξ−ξ ∗ + 1 >0 = −γ(ρ − ρ )∇J(ρ) ∗
Ei2 (7a)
i
for all t > 0. The latter condition may be fulfilled for instance i
if ξ ∗ < mini (ξMi + ξT H )/2. X
+ (ρ − ρ∗ )2 Ei Ėi (7b)
In this way, nodes with higher residual energy concur to in-
i
crease network connectivity and so improving the convergence X
time TC : this balances the reduction of TC due to those nodes − (ξi − ξT H )PTi (ρi ) (7c)
i
with lower residual energy. The reference radius ρ in (4) is T
the solution of the optimization problem ρ∗ = arg min ρ J(ρ) −e
e Le
e (7d)

1551-3203 (c) 2016 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.
This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication. Citation information: DOI 10.1109/TII.2016.2627478, IEEE
Transactions on Industrial Informatics
5

In the following we will analyse the terms (7a)-(7d) com- the authors in ([16]) is to compute a diagonal balancing
posing PV̇ (z), ∀z 6= ẑ. The term (7a) is negative definite being gain matrix W = diag{[w1 , . . . , wn ]} such that the matrix
γ > 0, i Ei2 > 0 and taking into account Assumption 2. The LB = WL is a balanced Laplacian version of L with both
term (7b) is negative definite as well because Ei Ėi < 0 for rows and columns zero sum. They formulate a distributed
all i and (ρ − ρ∗ )2 > 0. The term (7c) is negative definite algorithm to evaluate at each node i-th the gain wi by using
being ξi > ξT H for all i and PTi (ρi ) > 0. To show the auxiliary variables ψij according to the following algorithm:
definite negativity of the term (7d) we need to recall an X
extension of algebraic connectivity to the case of directed ψij (k + 1) = ψij (k) − η aih (ψij (k) − ψhj (k)) , ∀j 6= i
h
graph G with non symmetric and unbalanced Laplacian matrix
1 X
L. Specifically, in [23] it was assessed a refinement of the ψii (k) = lji ψij (k) (9)
the algebraic connectivity ã(G) which preserves non-negativity lii
j6=i
T
for directed graph, namely ã(G) = min||v||2 =1,v⊥1 v||v|| HLv
2
with η is a design parameter affecting the algorithm respon-
2
with v generic space vector, H = diag(h), and h = [hi ] siveness. At steady state the gain at the node i-th can be
being the unique positive vector such that hT L = 0 and evaluated as wi = ψii ∀i. We used the above procedure
maxi hi = 1. Therein it was proved that if the network allowing each node to compute gain wi and use it in the
graph is connected then ã(G) > 0. Therefore from the average consensus algorithm (2) to correctly estimate the
definition ofPnLaplacian matrix L and ã(G), we may infer average value. Specifically the following discrete version is
eT Le
e e = 12 i,j lij ||eej − eei ||22 ≥ e
eT HLe e||22 . From
e ≥ ã(G)||e implemented:
Assumption 1 the graph is connected then ã(G) > 0 ([23]) n
and it follows −e eT Le e < 0. Therefore V̇ (z) < 0 ∀z 6= ẑ and x ei (k) − αwi
ei (k + 1) = x
X
xi (k) − x
aij (e ej (k))
the result follows. j=1
+ (xi (k + 1) − xi (k)) (10)
B. A cooperative control law implementation
In this section, we present a distributed implementation Therefore algorithms (9) and (10) allow each node to take
of the proposed control law (4). Specifically, each node can local actions to work on a balanced version LB of Laplacian
locally regulate the transmission radius ρi according to (4) matrix L still guaranteeing the convergence of node estimation
provided it evaluates ρ̄. In this respect the following discrete ei (t) to the average value x = ||x||1 despite the assumption
x
version of the gradient-based equation (5) may be used: of connected direct graph. Notice that Assumption 1 may be
fulfilled by assigning at each node a minimum radius ρimin
−γ
ρ(k+1) − ρ(k) = · that guarantees network connectivity. For instance a procedure
ρ(k) − ρ(k−1) given in [25] may be implemented such that each node can
! (8)
||PT (ρ(k))||1 ||PT (ρ(k−1))||1 distributively set its minimum radius to assess the existence
·
−

λ2re ρ(k) λ2re ρ(k−1) of a minimum spanning tree (i.e. network connectivity).
B.2 Distribute control algorithm
where for the sake of notation ρ(k) represents the value of ρ at Now we are ready to describe the steps to implement in a
the time step k, ρ(k) = [ρi (ρ(k))] is the vector of transmission fully distributed way the control law (4) by the Algorithm 1.
radii ρi at time step k and it has been assumed unitary the In the following k represents the time step at which the radius
duration of time interval between two steps k and k + 1. control (4) is applied. Specifically, after an initialization phase,
Equation (8) can be implemented in a distributed way if a one-shot distributed balancing algorithm (9) is implemented
each node
can locally estimate the terms ||PT (ρ(k))||1 and (Algorithm 1 - step 4). This operation is preliminary to
λ2re ρ(k) . A distributed estimation of the two above terms guarantee the effectiveness of the average algorithm (10) used
can be assessed by using average consensus algorithms-type in the subsequent steps. In order to estimate ρ̄, equation (8)
(2). In this respect, in following we will firstly present how to
is
used and therefore estimation ||P e T (ρ(k))||1 and λe2re ρ(k) of
implement average consensus algorithms (2) under a weaker

respectively ||PT (ρ(k))||1 and λ2re ρ(k) have to be evaluated.
assumption of directed connect network (in place of the Between two control application time steps k
and k+1 each
stronger one of balanced or symmetric graph). Secondly we node estimates of ||P e T (ρ(k))||1 and λe2re ρ(k) by using the
will recall an average

consensus-based algorithm to distribute average consensus algorithm (10). Specifically we may derive
estimate λ2re ρ(k) . Finally we will illustrate step by step a the following estimation:
fully distributed implementation of the proposed control lever-
aging on the presented average consensus-based algorithms. ||P
e T (ρ(k))||1 = ||ξ(k)−ξ(k −1)||1 = ||e ξ(k)||1 −||e
ξ(k −1)||1
B.1 Average consensus in connected directed networks (11)
As stated in Subsection II-B, the distributed consensus algo- where ||eξ(·)||1 is evaluated in a distributed way by using the
rithm (2) allows each node to evaluate an estimation x ei (t) consensus algorithm (10) applied for x ei = ξei and xi = ξi .
of the average value x = ||x||1 provided the network graph In this way, each node evaluates a local estimation ξei of
is balanced or symmetric. Herein, we would like to still ||ξ||1 (Algorithm 1 - step 5), and therefore the estimation
guarantee convergence of algorithm (2) to the average value ||P
e T (ρ(k))||1 by using (11) (Algorithm 1 - step 6). Then
x = ||x||1 under the weaker Assumption 1 of direct connected the distributed estimation of λ e2re (Algorithm 1 - step 7) is
digraph by using the procedure in ([16]). The main idea of implemented using the well known Power Iteration method

1551-3203 (c) 2016 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.
This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication. Citation information: DOI 10.1109/TII.2016.2627478, IEEE
Transactions on Industrial Informatics
6

Algorithm 1 Distributed Energy-based Radii Control algo-

rithm (DERC) at node i-th 55
1: {Initialization steps}
50
2: repeat
3: k =k+1 45
4: wi ← balancing gains (eq.(9))
5: ||e
ξ(k)||1 ← average consensus on ξ(k) (eq.(10)) 40
6: ||Pe T (ρ(k))||1 ← ||e ξ(k)||1 − ||e
ξ(k −

1)||1 (eq.(11))

ρ


7: λ2re ρ(k) ← Distributed λ2re ρ(k) estimation
e 35
8: ρei (k + 1) ← Steepest descent on J (eq.(8))
ρ(k + 1) = ||e 30
9: ρ||1 ← average µconsensus on ρ e (eq.(10))
ξi (k)−ξ ∗
10: ρi (k + 1) ← ξM −ξ∗ + 1 · ρ(k + 1) (eq.(4)) 25
i
11: until ξi (k) ≤ ξT H
20
0 1 2 3 4 5 6 7
k x 10
4

for a symmetric network ([14]) extended to the more general

case of directed graph ([16]). Fig. 2. DERC dynamic evolution of radii ρi . The radii converge close to
ρ∗ = 25 value.
As the value of candidate optimal ray ρ evaluated by (8)
(Algorithm 1 step 8) may result in a different value ρei at
each node i-th (due to numerical or noisy effect), an average ρ∗ = 25 ∀i. In this case each node transmits with a power
consensus algorithm is then implemented to assess at each consumption PTi (ρ∗ ) that is constant during time such that
node an unique estimated value ρ = ||e ρi ||1 (Algorithm 1 step the rate of discharge is constant.In Table I are reported the
9). Finally the control law (4) to regulate the radius at node performance in this scenario under different strategies when
i-th is implemented by step 10 of the Algorithm 1. the optimal network configuration is assessed. In particular for
the above static radii design it results: TC = 15, TL = 40539
IV. S IMULATION VALIDATION and TD = 45562.
In this section we present a campaign of simulations under Then we consider the network when the distributed control
the M ATLAB environment to validate the proposed DERC algorithm DHA proposed in [19] is implemented. This algo-
algorithm in different scenarios. We consider a WSN to rithm tries to optimize the same our objective function J(ρ).
monitor a temperature of a field. The network is composed of The algorithm starts from a full connected network topology
50 nodes randomly distributed onto a 100 × 100 m2 . In path and at each control application step detects the node at highest
loss formula we set r0 = 0.0199 and β = 2.25 according degree and removes a link that mostly reduces the objective
to the indication given in [20]. The consensus algorithms function. The aim of this operation is to simultaneously reduce
are designed to assure convergence and stability properties both the objective function and λn . To estimate the objective
following the indications given in the original papers ([14], function and carry out the above operation, a power iteration
[16], [22]). In order to show the effectiveness of our approach based algorithm is used to compute λ2 and vn assuming a
in practical scenarios, we assume that the network performs symmetric graph topology. Further details about the DHA
a distributed temperature monitoring application and uses algorithm may be found in the original paper [19].
consensus algorithm to evaluate the average value of the In Table I are summarized the performance obtained for
environment temperature. The temperature over the area is DHA algorithm. Notice that the time TO to get the optimal
assumed to be a random variable uniformly distributed in the solution is expected to be long because the algorithm removes
range [17, 27] ◦ C and is sampled with a rate of 0.01 Hz. one link per control step. This decreases the convergence rate
Herein we will show the performance of the proposed towards the optimal solution that in turn decrease the lifetime
DERC algorithm compared than ones obtained under a static TL (2.7 · 104 ).
network radii design and DHA algorithm proposed in [19]. Now we consider the same above network scenario when the
The objective function value and the following performance proposed DERC Algorithm 1 is applied initializing ρ(0) = 50.
indexes are evaluated: network lifetime TL , the network dis- The dynamic evolution of ρi (k) is reported in Fig. 2
connection time TD , consensus convergence time TC and the showing that each node controls in a decentralized way its
time taken by the algorithm to get the optimal solution TO . transmission radius ρi (ρ) on the base of node residual energy,
so that it converges close to the optimal value ρ∗ = 25.
Therefore the final value of radii ρi (ρ) may be different
A. Nominal scenario for each node according to its residual battery energy: the
In this scenario we set the initial energies vector ξ(0) to be effect is to balance the available energy of network nodes.
randomly distributed in the range [8, 10] · 103 Joule. This strongly improves the performance in terms of lifetime
Firstly we consider a static network radii design where the and disconnection time with respect to the other algorithms
node radii ρi are assumed to be the same and tuned in order as shown in Table I (i.e. TL = 69294, TD = 69371 and
to get the minimum of J(ρ) in Fig. 1, namely ρi = ρ = TC = 15). Additionally the time TO to get the optimal value

1551-3203 (c) 2016 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.
This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication. Citation information: DOI 10.1109/TII.2016.2627478, IEEE
Transactions on Industrial Informatics
7

Static design DHA DERC DERC-SINR

k̄ by computing the solution ρ∗ when ξ(k) = ξ(k̄). In case
TL (×103 ) 40 27 69 43
TD (×103 ) 46 32 69 43
of uncertain available energy variations due to the harvesting
TC 15 19 15 15 extra power consumption processes, the optimal frontier value
TO (×103 ) – 250 33 39.7 strongly varies during time. Notice that the proposed algorithm
¯
d(A) 44.8 12.48 43.4 42.1 is effective to get the objective function value close to the
λ2re 29.2 9.33 28.3 28.1
λnre 50 22.71 50 50
optimal frontier by online and decentralized operations.
||PT (ρ)||1 0.12 0.35 0.12 0.12 The performance of the proposed control algorithm can
J(ρ) (×10−3 ) 4.1 37.5 4.4 4.69 be improved by using a multi-hop version of the consensus
TABLE I
P ERFORMANCE INDEXES UNDER DIFFERENT RADII DESIGN AND CONTROL
algorithm (2) ad (10) (see [22], [26]).
STRATEGIES .

is lower than DHA algorithm. As expected J(ρ) is higher than V. C ONCLUSION

the optimal value obtained under the static radii design. In this paper, we have formulated a decentralized control
On the other side DHA algorithm is effective in reducing law of node transmission radii in order to simultaneously
λn = λnre and average node degree d(A)¯ than the proposed minimize the network energy power consumption and max-
DERC algorithm. imize the network connectivity. A key feature of the proposed
As further scenario, we consider the presence of the inter- control law is to be cooperative and fully decentralized such
ference phenomena among nodes due to their transmission that each node adapts its transmission radius just on the base
power. Specifically we consider the standard interference of information received from its neighbours and the available
SINR (SIgnal to Noise Ratio) model ([24]) such that the energy. A stability analysis of the closed loop network under
signal sent by node i-th and received at node j-th is affected the proposed radii control is formulated. An extensive simu-
by the interference of the other nodes transmitting to node lation validation has shown the effectiveness of the proposed
j. Therefore the path loss formula
  is modified according to approach both in terms of network lifetime and monitoring
P (ρ ) β
P Ti i
N0 + k6=j,i PP
Tk (rkj )
≥ P Rj · rρ0i where N0 is a baseband algorithm responsiveness when compared with a well known
decentralized strategy control law presented in literature.
noisy and k6=j,i PTk (rkj ) is the interference due to the
transmission of nodes different than i. In this way the higher
is the nodes transmission power, the more is the interference R EFERENCES
and therefore the fault of a link between i and j. Therefore, [1] Industrial Wireless Technology for the 21st Century, Report, Technology
the equation (1) can be increased to improve the network Foresight, 2004.
connectivity but taking into account the drawback due to the [2] J. Neuzil, O. Kreibich, R. Smid, A Distributed Fault Detection System
Based on IWSN for Machine Condition Monitoring, IEEE Transactions
effect of node interference. From the results of Table I it is on Industrial Informatics, 10,2, 2014.
appear that the effect of nodes interference is mitigated by [3] X. Jiang, J. Polastre, and D. Culler, Perpetual environmentally powered
the proposed algorithm. Indeed the controller tries to make sensor networks, Proceedings of the 4th International Symposium on
uniform the nodes transmission radii, thus avoiding exces- Information processing in sensor networks, 2005
[4] R. Saber, R. Murray, Consensus Problems in Networks of Agents with
sively high value of transmission power that could amplify Switching Topology and Time-Delays, IEEE Transactions on Automatic
the interference phenomena. Control, 49,9, 2004.
[5] A. J. Dickson, S. Burton, M. Shepertycky, Y. F. Liu and Q. Li, Digitally
Controlled Energy Harvesting Power Management System, in IEEE
B. Robustness in presence of uncertain energy variations Journal of Emerging and Selected Topics in Power Electronics, 4,1, 2016.
[6] S Manfredi, A consensus based rate control scheme for ATM networks,
Now we evaluate the robustness of the considered algo- International Journal of Control, Automation, And Systems, vol. 10, p.
817-823, ISSN: 1598-6446, 2012
rithms when random charge and discharge of the batteries
[7] S. Manfredi, M. Pagano, R. Raimo, Ultracapacitor-based Distributed
are considered. Specifically, we assume that 5% of nodes are Energy Resources to support time-varying smart-grid power flows, Proc.
charged by the harvesting systems (i.e. Ns > 0) while 5% International Symposium IEEE Speedam 2012. vol. I, p. 1148-1153,
of nodes are characterized by an extra power consumption IEEE, ISBN: 9781467313001, Sorrento (Italy), 20-22 June 2012
[8] A. Aprem, C. R. Murthy and N. B. Mehta, Transmit Power Control
requirements (i.e. Ns < 0). In Fig. 3(a) it is shown the Policies for Energy Harvesting Sensors With Retransmissions, IEEE
network nodes energy under the static radii design. Notice Journal of Selected Topics in Signal Processing, 7 5, pp. 895-906, 2013.
that the node with extra power consumption heavily affects [9] S. Manfredi, An Algorithm for Fast Rendezvous Seeking of Wireless
Networked Robotic Systems, Ad Hoc Networks, vol. 11, 7, p. 1942-1950,
the lifetime (TL = 34281). In Fig. 3(b) and Fig. 3(c) are ISSN: 1570-8705, 2013
respectively depicted the dynamic evolution of nodes energy [10] S Manfredi, D Angeli, Frozen state conditions for exponential consensus
for the DHA and DERC algorithm. Note how the DERC of time-varying cooperative nonlinear networks, Automatica 64, 182-189,
2016
algorithm is effective in balancing nodes available energy [11] A. Aziz, P. Fitzpatrick, M. Ivanovich A Survey on Distributed Topol-
because nodes with extra power consumption are controlled to ogy Control Techniques for Extending the Lifetime of Battery Powered
have lower transmission radius than nodes with extra available Wireless Sensor Networks, IEEE Communication Surveys and Tutorials,
15,1, 2013
energy. This allows to get a larger lifetime (TL = 74955) than [12] F. Ge and Y. Wang, An adaptive energy efficient topology for wireless
the other strategies. sensor networks, Second International Conference on Future Generation
In Fig. 4 the resulting objective function J(ρ(ρ)) (continu- Communication and Networking, FGCN 2008.
[13] L. Shiu, F. Lin, C. Lee and C. Yang A Distributed Reliable and
ous line) is compared with the optimal frontier (dashed line). Energy-Efficient Topology Control Algorithm in Wireless Sensor Network,
Each point of the frontier is offline evaluated at each time step International Conference on Information Science and Applications, 2012

1551-3203 (c) 2016 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.
This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication. Citation information: DOI 10.1109/TII.2016.2627478, IEEE
Transactions on Industrial Informatics
8

10000
10000 10000

8000
8000 8000

6000
6000 6000

ξ
ξ

ξ
4000
4000 4000

2000
2000 2000

0
0 0 1 2 3 4 0
0 0.5 1 1.5 2 2.5 3 3.5 0 2 4 6 8
k x 10
4 k ×104 k 4
x 10

(a) (b) (c)

Fig. 3. Dynamic evolution of the network nodes energy ξi in presence of uncertain energy variations: (a) static radii design, (b) DHA algorithm and (c)
DERC algorithm.

−3
x 10 SINR: The scheduling complexity of arbitrary topologies, Proceedings of
12
the 7th ACM International Symposium on Mobile Ad Hoc Networking
and Computing, MobiHoc06, 2006.
11
[25] F. J. Ovalle-Martnez, I. Stojmenovi, F. Garca-Nocetti, J. Solano-Gonzlez,
Finding minimum transmission radii for preserving connectivity and
10
constructing minimal spanning trees in ad hoc and sensor networks,
Journal of Parallel and Distributed Computing, 65,2, 2005
9
[26] S. Manfredi, E. Di Tucci, Distributed control for energy efficient and fast
J (ρ)

consensus in wireless sensor networks, International Journal of Control,

8 DOI: 10.1080/00207179.2016.1201218 in press, 2016
7

5
Sabato Manfredi is Assistant Professor of Au-
4 tomatic Control and Adjunct Professor of System
0 2 4 6 8 Dynamics and Industrial Automation at the De-
k 4
x 10 partment of Electrical Engineering and Informa-
tion Technology, University of Naples Federico II,
Fig. 4. Dynamic evolution of the objective function J(ρ) (continuous line) Italy. He is also a member of the Control and
and the optimal frontier (dashed line) for the DERC algorithm. The DERC Power Group, Electrical and Electronic Engineering
algorithm still track the dynamic optimal frontier in presence of energy Department, Imperial College - London, UK. His
available variations. research interests are primarily in automatic control,
with a special emphasis on the analysis and dis-
tributed/decentralized control of linear and nonlinear
time-varying multi agent systems, complex networks and Cyber-Physical
[14] P. Yang, R. Freeman, G. Gordon, K. Lynch, S. Srinivasa and R. Suk-
Systems, communication and sensor/robot networks. He has published more
thankar, Decentralized estimation and control of graph connectivity in
than 70 scientific publications including 14 single-author papers and is author
mobile sensor networks, American Control Conference, 2008.
of the book: Multilayer Control of Networked Cyber-Physical Systems. Appli-
[15] F. Knorn, R. Stanojevic, M. Corless, R. Shorten, A Framework for cation to Monitoring, Autonomous and Robot Systems (Advances in Industrial
Decentralised Feedback Connectivity Control with Application to Sensor Control Series, Springer 2016). He collaborates with many international
Networks, International Journal of Control, 82,11, 2009 universities and companies, holds European and Italian patents, is a founding
[16] L. Sabattini, C. Secchi, N. Chopra, Decentralized control for main- member of an academic spin-off, and is involved in a range of academic and
tenance of strong connectivity for directed graphs, 21st Mediterranean industrial projects.
Conference on Control & Automation (MED), 2013
[17] S. Sardellitti, S. Barbarossa and A. Swami, Optimal Topology Control
and Power Allocation for Minimum Energy Consumption in Consensus
Networks, IEEE Transactions On Signal Processing, 60,1, 2012.
[18] C. Asensio-Marco, B. Beferull-Lozano, Energy Efficient Consensus
Over Complex Networks, IEEE Journal of Selected Topics in Signal
Processing, 9,2, 2015 Edmondo Di Tucci receives in 2010 the Master
[19] C. Asensio-Marco, B. Beferull-Lozano, Optimal topology design for degree in Informatic Engineering (automatic control
energy efficient consensus in broadcast wireless sensor networks, IEEE curriculum) at University of Naples Federico II and
International Conference on Communications (ICC), 2014 in 2012 a postgraduate in Robotics and Intelligent
[20] J. Song and Y. Tan, Energy consumption analysis of ZigBee-based energy Systems. He is familiar in programming microcon-
harvesting wireless sensor networks, IEEE ICC Systems, 2012 trollers and FPGA prototyping systems and with
[21] P. Moravek, D. Komosny , Energy Demands of 802.15.4/ZigBee Com- the design of electronic circuits. Since 2003 he
munication with IRIS Sensor Motes, Transactions on signal processing, works at www.smartsoft.it as Developer and System
978-1-4577-1411-5/11, 2011. Integrator in the field of telecommunications He also
[22] S. Manfredi, Design of a multi-hop dynamic consensus algorithm over works on various research project at University of
wireless sensor networks, Control Engineering Practice, 21,4, 2013 Naples Federico II with many publications in the
field of network control and synchronization. During last 4 years he developed
[23] Chai Wah Wu, On RayleighRitz ratios of a generalized Laplacian matrix
many iOS and Android applications.
of directed graphs., Linear Algebra and its Applications, 402, 2005.
[24] T. Moscibroda, R.Wattenhofer, A. Zollinger, Topology control meets

1551-3203 (c) 2016 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.