energy harvesting-aided spectrum sensing and data transmission...

0018-9545 (c) 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.

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/TVT.2016.2551721, IEEETransactions on Vehicular Technology

Energy Harvesting-Aided Spectrum Sensing and Data Transmission in HeterogeneousCognitive Radio Sensor Network

Deyu Zhang, Zhigang Chen, Member, IEEE, Ju Ren, Student Member, IEEE, Ning Zhang, Member, IEEE,Mohamad Khattar Awad, Member, IEEE, Haibo Zhou, Member, IEEE, Xuemin (Sherman) Shen, Fellow, IEEE

Correspondent Author: Zhigang ChenSchool of SoftwareCentral South UniversityChangshaEmail: [email protected]



1

Energy Harvesting-Aided Spectrum Sensing andData Transmission in Heterogeneous Cognitive

Radio Sensor NetworkDeyu Zhang, Zhigang Chen, Member, IEEE, Ju Ren, Student Member, IEEE, Ning Zhang, Member, IEEE, Mohamad

Khattar Awad, Member, IEEE, Haibo Zhou, Member, IEEE, Xuemin (Sherman) Shen, Fellow, IEEE

Abstract—The incorporation of Cognitive Radio (CR) andEnergy Harvesting (EH) capabilities in wireless sensor networksenables spectrum and energy efficient heterogeneous cognitiveradio sensor networks (HCRSNs). The new networking paradigmof HCRSNs consists of EH-enabled spectrum sensors and battery-powered data sensors. Spectrum sensors can cooperatively scanthe licensed spectrum for available channels, while data sensorsmonitor an area of interest and transmit sensed data to thesink over those channels. In this work, we propose a resourceallocation solution for the HCRSN to achieve the sustainabilityof spectrum sensors and conserve energy of data sensors. Theproposed solution is achieved by two algorithms that operatein tandem, a spectrum sensor scheduling algorithm and adata sensor resource allocation algorithm. The spectrum sensorscheduling algorithm allocates channels to spectrum sensorssuch that the average detected available time for the channelsis maximized, while the EH dynamics are considered and PUtransmissions are protected. The data sensor resource allocationalgorithm allocates the transmission time, power and channelssuch that the energy consumption of the data sensors is mini-mized. Extensive simulation results demonstrate that the energyconsumption of the data sensors can be significantly reducedwhile maintaining the sustainability of the spectrum sensors.

Index Terms—Wireless sensor network, energy harvesting,cognitive radio, energy efficiency, multiple channels

I. INTRODUCTION

Wireless sensor networks (WSNs) have become a prevalentsolution to a wide range of applications including environ-mental monitoring, patient monitoring and smart homes [1].Typically, WSN uses the unlicensed Industrial, Scientific, andMedical (ISM) band for data transmission. However, with theexponential growth in the number of wireless devices operat-ing in this band, WSNs suffer from severe interference [2].Cognitive Radio (CR) has emerged as a promising technologyto allow secondary unlicensed users to opportunistically access

Copyright (c) 2015 IEEE. Personal use of this material is permitted.However, permission to use this material for any other purposes must beobtained from the IEEE by sending a request to [email protected].

D. Zhang and J. Ren are with the School of Information Science andEngineering, Central South University, Changsha, China, 410083. D. Zhangand J. Ren are also visiting scholars at the University of Waterloo now.(e-mail:{zdy876, ren ju}@csu.edu.cn)

Z. Chen is with the School of Software, Central South University, Chang-sha, China, 410083, and Z. Chen is the corresponding author (e-mail:[email protected]).

M. K. Awad is with the Computer Engineering Department at KuwaitUniversity, Kuwait City, Kuwait (e-mail: [email protected]).

N. Zhang, H. Zhou and X. (S.) Shen are with the Department of Electricaland Computer Engineering, University of Waterloo, Canada, N2L 3G1 (e-mail: {n35zhang, h53zhou, xshen}@uwaterloo.ca).

the underutilized spectrum that is licensed to the primaryusers (PUs) [3]. Therefore, CR can reduce the interference andimprove spectrum utilization. The integration of CR functionsinto WSNs leads to Cognitive Radio Sensor Network (CRSN).

In CRSN, spectrum sensors frequently scan the spectrum toobtain higher-resolution estimates of the spectrum availabilityand guarantee PU protection against interference [4]. However,this frequent scanning increases the energy consumption ofan energy-constrained network, which traditionally operatespowered by batteries. Consequently, energy conservation be-comes a critical design issue for CRSNs [5], [6], [7]. Energyharvesting (EH) is considered as one of the effective approach-es for improving the energy efficiency of WSNs. EH-enabledsensors can harvest energy from either radio signals or ambientenergy sources which enable them to operate continuouslywithout battery replacement [8]. In the literature, extensiveresearch efforts have been devoted to improving the energyefficiency of CRSNs. Energy-efficient cooperative spectrumsensing is investigated in [9], [10]. Shah et al. limit the numberof sensors that perform spectrum sensing to minimize theenergy consumption by exploiting the spatial correlation ofthe sensors [9]. Deng et al. investigate the network lifetimeextension of dedicated sensor networks for spectrum sensing[10]. Despite the importance of these efforts, limits remainfor improving the energy efficiency of battery-powered datasensors with low data sensing and limited data transmissionrates, for two reasons. First, unlike data sensors, spectrumsensors perform spectrum scanning at a much higher rate thandata sensing which depletes the battery energy much fasterthan the data sensors. Second, harvested energy is sporadic andunstable, whereas battery-stored energy is static and stable,which makes schemes that are developed for battery-poweredsensors inapplicable for EH-enabled sensors.

In addition to inefficient spectrum and energy utilization,inaccurate spectrum sensing is another limitation of traditionalsensor networks. The spectrum-scanning results of a singlespectrum sensor are prone to detection error due to thespatially large-scale effect of shadowing and small-scale effectof multipath fading [11]. Alternatively, cooperative spectrumsensing can be performed to enhance the accuracy of spec-trum sensing [12]. In cooperative spectrum sensing, multiplespectrum sensors sense the same channel and coordinatetheir decisions on the availability of a given channel. Hence,the incorporation of energy harvesting and cognitive radiotechniques in addition to cooperative spectrum sensing brings



2

major improvements to traditional WSNs. Energy-efficientcooperative spectrum sensing has been the focus of severalresearch activities. In [13], Cheng et al. schedule a group ofspectrum sensors between the active and inactive states toimprove the performance of spectrum sensing. Consideringthe impact of frequent state switching on sensors’ stability, theschedule minimizes the sensors’ switching frequency amongthe states. In [14], Zhang et al. design a distributed cooperativespectrum sensing scheme, wherein the Secondary Users (SUs)only exchange their measurements with the one-hop neigh-bours. In [15], Khan et al. propose a selection scheme to findthe sensors with the best detection performance for cooperativespectrum sensing, without requiring a priori knowledge of theprimary-user-signal-to-noise ratio (SNR). In [16], Liu et al.propose an ant colony-based algorithm for a dedicated sensornetwork, whereby spectrum sensing is performed to supportthe operation of a secondary network. Throughput of thesecondary network is optimized by scheduling the spectrum-sensing activities according to the residual energy of eachsensor. Additionally, to achieve energy efficient cooperativespectrum sensing, parameters are optimized, such as the detec-tion threshold [17], sensing duration [18], and switch cost [19].EH-aided cooperative spectrum sensing should take the EHdynamics of the sensors into consideration. For EH-aidedspectrum sensing, the objective is to explore as many of theavailable channels as possible while maintaining the sustain-ability of the spectrum sensors, considering the diverse energy-harvesting capabilities of the spectrum sensors. However, theaforementioned energy-efficient cooperative spectrum-sensingschemes which only focus on the minimization of the energyconsumption of spectrum sensing, cannot be directly appliedto EH-aided cooperative spectrum sensing.

In this paper, we propose a resource allocation solution toaddress the gaps that are identified above in the existing works,namely, spectrum under-utilization, energy inefficiency andspectrum-sensing inaccuracy. Specifically, for a heterogeneouscognitive radio sensor network (HCRSN) that is composedof EH-enabled spectrum sensors and battery-powered datasensors, we develop a solution that can jointly guarantee thesustainability of spectrum sensors, the energy efficiency ofthe data sensors and the accuracy of spectrum sensing. TheHCRSN operates over two phases, i.e., a spectrum-sensingphase followed by a data transmission phase. In the spectrum-sensing phase, EH-enabled spectrum sensors cooperativelysense the spectrum to detect underutilized channels that arelicensed to the primary network. Spectrum-sensing schedulingis optimized to maximize the detected channel’s available timeconsidering the dynamics of EH. In the data transmissionphase, the available channels are utilized by data sensorsfor sensed data transmission. We combine the resource man-agement and allocation of each phase in a unified solution.Despite the physical independence of spectrum sensors anddata sensors, a unified solution is necessary to optimize theoverall energy efficiency and performance of HCRSNs. Theimbalance of energy replenishment and consumption at eitherthe spectrum or data sensors results in nodes failure anddeteriorates the network performance; thus, energy shouldbe managed under one unified setup. Furthermore, the per-

formance of spectrum sensor scheduling in the first phasehighly impacts the energy efficiency of data sensors in thesecond phase. A longer channel available time detected in thefirst phase increases the channel access time and decreasesthe probability of collision in the second phase. This causalimpact of spectrum sensor scheduling performance in the firstphase on the performance of data sensor resource allocationin the second phase necessitates a unified solution. On theother hand, it is practically infeasible to allocate channels,transmission time, and transmission power before the spectrumsensors identifies the available channels. Therefore, spectrumsensor scheduling and data sensor resource allocation have tobe addressed over two coupled problems operating in tandembut under one unified setup. Summarily, the contributions ofthis paper are twofold:

1) We formulate the EH-aided spectrum-sensing problemas a nonlinear integer programming problem and pro-pose a Cross-Entropy-based algorithm to maximize theaverage available time of the channel under the protec-tion for PUs.

2) We propose a joint time and power allocation algorithmto minimize the energy consumption of the data sensors,based on the analysis of the channel fading and theexponential ON-OFF model of the PUs’ behavior.

It is imperative to mention here that the literature schemes,which consider only energy minimization of spectrum sen-sors [9], [10], [15], [20] rather than channel available timemaximization are inapplicable in HCRSN. Furthermore, unlikeexisting solutions that separately consider channels allocation[21] [22] and power control [23] [24], the proposed solutionjointly allocates time, frequency and power to data sensors;hence, improve the energy efficiency of data sensors.

The remainder of this paper is organized as follows. Thenetwork architecture and cognitive radio model are detailedin Section II. A mathematical formulation and the proposedsolutions of the spectrum sensor scheduling problem and datasensors resource allocation problem are detailed in Section III.Performance evaluation results that demonstrate the efficiencyof the proposed algorithms are presented in Section IV.Conclusions are drawn in Section V.

II. SYSTEM MODEL

A. Network Architecture

We consider a HCRSN that consists of three types of nodes:N battery-powered data sensors, M EH-enabled spectrumsensors and a sink node, as shown in Fig. 1. The HCRSNcoexists with a network of PUs that have access to the licensedspectrum. The licensed spectrum is divided into K orthogonalchannels that have equal bandwidth W . Spectrum sensors aredeployed to sense and identify available channels that are notutilized by PUs, whereas data sensors collect data from anarea of interest. The data is then transmitted over the availablechannels to the sink.

The considered HCRSN operates as follows: First, the sinkschedules spectrum sensors to detect the PUs’ presence overchannels using energy detection [3]. A PU is determined tobe active, i.e., channel unavailable, if at least one scheduled



3

Primary User

Primary Base Station or Access Point

Battery-poweredData Sensor

Sink

EH-enabledSpectrum Sensor

Fig. 1. An illustration of heterogeneous cognitive radio sensor network.

spectrum sensor reports it to be present on the channel [10].The energy consumption of a spectrum sensor used to detectone channel is determined by es = τs · Ps, where Ps isthe power consumption of the spectrum sensing. We assumethat the EH rate is known a priori and is stable over T[25]. To guarantee the sustainability of the m-th spectrumsensor, its energy consumption should not exceed the amountof harvested energy in one period, πm · T , where πm denotesthe average EH rate of the m-th spectrum sensor. Second, thesink assigns the available channels to the data sensors for datatransmission.

DataSensing Data Transmission

SpectrumSensing Idle

τs T − τsT

DataSensing

tn,k

τs′

Idle

τs T − τsT

SpectrumSensor

DataSensor

time

Fig. 2. Timing diagram and frame structure of the HCRSN.

Fig. 2 shows the timing diagram and frame structure ofthe considered network. The HCRSN operates periodicallyover time slots of duration T . Each time slot is divided intotwo phases: the spectrum sensing phase and data transmissionphase. In the spectrum sensing phase, the spectrum sensors co-operatively identify the presence of PUs, while the data sensorscollect information from the area of interest. The duration ofthe spectrum sensing phase is τs, which is further divided intomini-slots of duration τs′ over which a single spectrum sensorsenses one channel. After the spectrum sensing phase, the sinkcollects the results from all the scheduled spectrum sensorsand estimates the availability of the channels. Then, the sinkoptimizes the data transmission scheduling of data sensors toconserve their energy. The data sensors transmit data according

to the schedule in the subsequent data transmission phase withduration T − τs divided over the time slots of duration tn,k inwhich the n-th data sensor transmits data to the sink over thek-th channel.

With respect to the notation, the following holds: a bold-facesmall-case symbol always refers to a vector; and a non-italicbold-face large-case symbol always symbolizes a matrix.

B. Cognitive Radio Model

All of the channels experience slow and flat Rayleigh fadingwith similar fading characteristics. The PU behavior overeach channel is modeled as a stationary exponential ON-OFFrandom process, in which the ON/Active and OFF/Inactivestates represent the presence and absence of a PU over achannel, respectively . We use λk to denote the transitionrate from the state Active to the state Inactive on the k-thchannel and µk to denote the transition rate in the reversedirection. The estimation of λk and µk is out of the scopeof this work; however, they can be obtained by the channelparameter estimation schemes, similar to the ones proposed in[26] and [27]. The channel usage changes from one PU to theother and, hence, affects the transition rates.

Spectrum sensors perform binary hypothesis testing to de-tect the presence of PU signals over channels. Hypothesis0 (H0) proposes that the PU is Inactive and the channel isavailable, while Hypothesis 1 (H1) proposes that the PU isActive and the channel is unavailable. The spectrum sensorreceives a sampled version of the PU signal. The number ofsamples is given by U = τsfs, where fs is the samplingfrequency. An energy detector is applied to measure the energythat is associated with the received signal. The output ofthe energy detector, i.e., the test statistic, is compared tothe detection threshold ε, to make a decision on the stateof the PU, Active or Inactive. The test statistic evaluates toYm,k = 1

U

∑Uu=1 |ym,k(u)|2, where ym,k(u) is the u-th sample

of the received signal at the m-th spectrum sensor on the k-thchannel. We assume that the PU signal is a complex-valuedPSK signal and the noise is circularly symmetric complexGaussian with zero mean and σ2 variance [28].

The performance of the energy detector is evaluated by thethe following performance metrics under hypothesis testing[29]:• The false alarm probability pf (m, k): The probability that

the m-th spectrum sensor detects a PU to be present onthe k-th channel when it is not present in fact, i.e., H0

is true. The false alarm probability is given by [28],

pf (m, k) = Pr(Ym,k > ε|H0) = Q(( ε

σ2− 1)√

U),

(1)where Q(·) is the complementary distribution functionof the standard Gaussian. Without loss of generality, weset the detection threshold to be the same for all ofthe spectrum sensors; hence, the false alarm probabilitybecomes fixed for all of the sensors and is denoted bypf .

• The detection probability pd(m, k): The probability thatthe m-th spectrum sensor detects the presence of a PU on



4

the k-th channel while H1 is true. This probability wasfound to be [28]

pd(m, k) = Pr(Ym,k > ε|H1) = Q

(Q−1(pf )−

√Uγm,k√

2γm,k + 1

),

(2)

where γm,k denotes the received signal-to-noise ratio (SNR)from the PU on the k-th channel. To reduce the communicationoverhead and delay, each spectrum sensor sends the final 1-bitdecision (e.g., 0 or 1 represents the Active or Inactive state,respectively) to the sink. The sink makes the final decisionon the presence of a PU following the Logic-OR rule [28],[10]. Under this rule, the PU is considered to be presentif at least one of the scheduled sensors reports that it ispresent. Therefore, the final false alarm probability F kf andfinal detection probability F kd can be written as

F kf = 1−Πm∈Mk(1− pf ), and (3)

F kd = 1−Πm∈Mk(1− pd(m, k)), (4)

where Mk represents the set of spectrum sensors that isscheduled to detect the k-th channel.

III. PROBLEM STATEMENT AND PROPOSED SOLUTIONS

In an HCRSN with the above-described architecture, cogni-tive radio models and EH dynamics, the problems of schedul-ing the spectrum sensors and allocating the resources for thedata sensors become challenging. In the first problem, thespectrum sensor scheduling (SSS) problem, the sink schedulesthe spectrum sensors to sense the presence of the PUs overthe channels in such a way that the channel availability ismaximized while respecting the EH dynamics and PUs’ prior-ities in accessing the channels. Solving this problem makes theavailable channels known to the sink which allocates them tothe battery-powered data sensors along with the transmissiontime and power allocation, with the objective of minimizingthe data sensors’ energy consumption. This resource allocationproblem is referred to as the data sensor resource allocation(DSRA) problem.

Fig. 3 shows the two problems in tandem, the spectrum-sensing and data-sensing phases, and the data flows amongthem. In the following two subsections, we present prob-lem formulations and solutions for both problems. The firstproblem is formulated as a nonlinear integer programmingproblem, while the second problem is formulated as a biconvexoptimization problem.

A. Spectrum Sensors Scheduling

In this subsection, we investigate the SSS problem which isposed as a nonlinear integer programming problem. Througha Cross-Entropy-based solution, the channel availability ismaximized while guaranteeing EH-enabled spectrum sensorssustainability and PUs protection.

1) Problem Formulation: Three factors impact the averagedetected available time of the channel: the actual averageavailable time, the final false alarm probability complement(1 − F kf ) and the final detection probability complement(1 − F kd ). The actual average available time of the k-th

channel is the product of the mean sojourn time and thestationary probability of the k-th channel. Let LkActive = 1

λk

and LkInactive = 1µk

denote the mean sojourn time of the Activestate and the Inactive state on the k-th channel, respectively.Moreover, the stationary probabilities of the Active and Inac-tive states are given by

P kActive =µk

λk + µk, P kInactive =

λkλk + µk

. (5)

Therefore, the k-th channel average actual available time isgiven by,

αk = LkInactive · PkInactive. (6)

Let J be an M × K matrix with binary elements [J]m,k. Abinary element of 1 indicates the assignment of the m-th spec-trum sensor to detect the k-th channel and 0 otherwise. Giventhat the PU on the k-th channel is inactive, the probability thatthe k-th channel is available is equivalent to the complementof the final false alarm probability, which can be written as

1− F kf = Πm∈Mk(1− pf ) = (1− pf )

∑Mm=1[J]m,k (7)

The data sensor transmission interferes with the PU transmis-sion if the spectrum sensors do not detect the PU presencewhile it is present. The chance of this event is captured bythe mis-detection probability 1− F kd . To protect the PU fromsuch interference, we consider detection decisions with a mis-detection probability of less than MDthr. A binary variableIkd is introduced to indicate whether the protection requirementis satisfied or not and is given by,

Ikd =

{1, if 1− F kd < MDthr,0, otherwise.

(8)

If the mis-detection probability of the k-th channel exceedsMDthr, the detection is considered to be unreliable, and thek-th channel is not accessed by data sensors. Substituting Eqn.(2), (4) into (8) yields,

Ikd =

1, if Πm∈Mk

(1−Q

(Q−1(pf )−

√Uγm,k√

2γm,k+1

))< MDthr,

0, otherwise.(9)

The objective function of SSS that maximizes the averagedetected available time of a channel while protecting the PUcan be written as follows:

K∑k=1

αk(1− pf )∑M

m=1[J]m,kIkd . (10)

The SSS is subject to two constraints; the first constraint isrelated to the EH dynamics, whereas the second constraint isrelated to the frame structure (see Fig. 2). In a given frameT , to maintain the sustainability of the spectrum sensors,the energy consumption of each sensor should not exceedits harvested energy. This arrangement can be mathematicallywritten as

(

K∑k=1

[J]m,k)es ≤ πmT. (11)

Moreover, the time that is used for sensing the k-th channel



5

SSSDSRA

SpectrumSensing

DataSensing

Channel Gain, δn,k

Data Amount, Dn

Max TransmissionPower, pmax

max AccessTime, αk

Data Transmission Time, T − τs

Time, Power, andchannel Allocation

DataTransmission

AvailableChannels

PU Received SNR

Spectrum Sensingduration, τs

EnergyHarvestingRate, πm

Mis-detection Probability

Threshold, MDthr

Spectrum Sensor toChannel Allocation

Matrix, J

Fig. 3. A block diagram of the proposed system. The dashed line separates the optimization plane from the sensing plane.

is bounded by the duration of the spectrum-sensing phase τsin one period, namely,

(

M∑m=1

[J]m,k)τs′ ≤ τs. (12)

Then, the spectrum sensor scheduling problem becomes acombinatorial problem of optimizing the sensor-to-channelassignment matrix J and can be written as follows:

(SSS) maxJ

K∑k=1

αk(1− pf )∑M

m=1[J]m,kIkd

s.t.

(∑Kk=1[J]m,k)es ≤ πmT, ∀m,

(∑Mm=1[J]m,k)τs′ ≤ τs,∀m,

[J]m,k = {0, 1} ∀m, k.

The term αk has a constant value over a given channel. Asmore channels are assigned to a given spectrum sensor, i.e.,as∑Mm=1[J]m,k increases, the value of (1 − pf )

∑Mm=1[J]m,k

decreases, and Ikd tends to take a unit value. Therefore,there exists a trade-off between (1 − pf )

∑Mm=1[J]m,k and Ikd .

However, the assignment [J]m,k exists in the exponential partof (1 − pf )

∑Mm=1 Jm,k and affects Ikd through Mk. These

structures make the SSS an integer programming problem.Intuitively, the objective function in Eqn. (10) can be optimizedby performing an exhaustive search over the space that is char-acterized by the constraints of SSS. However, this arrangementleads to a search space of size 2MK which is computationallyprohibitive especially for the resource-limited sensor network.In the following subsection, we apply the Cross-Entropy-basedalgorithm (C-E algorithm) [30] to address (SSS). Althoughthe performance bound of the C-E algorithm remains an opentheoretical issue [31], it has been shown effective in solvinga similar combinatorial optimization problem [3].

2) Cross Entropy-based Algorithm: The basic idea of theC-E algorithm lies in the transformation of a deterministicproblem into the related stochastic optimization problem suchthat rare-event simulation techniques can be applied. Morespecifically, an associated stochastic problem (ASP) is definedfor the deterministic problem, and then, the ASP problemis solved using an adaptive scheme. The adaptive schemegenerates random solutions that converges stochastically to theoptimal or near-optimal solution of the original deterministicproblem.

Before introducing the C-E algorithm, we transform the con-strained problem into an unconstrained problem by applyinga penalty method. Let ω = −

∑Kk=1 α

k be the penalty forviolating any of the constraints, and then, the SSS problemtransforms to

O = ω · I(∑Mm=1[J]m,k·es>πmT ) + ω · I(∑K

k=1[J]m,k·τs′>τs)

+K∑k=1

αk(1− pf )∑M

m=1[J]m,kIkd .

(13)For a positive constant penalty of ω, the unconstrained ob-jective function evaluates to a negative value for all of theinfeasible solutions that violate constraints (11) and (12). Theindicator function, I(·), takes the value of 1 for true evaluationsof (·) and zero otherwise.

Recall that the sink schedules the spectrum sensors to detectthe presence of PU on certain licensed channels. Therefore,the row vectors of J are drawn from a set, C, of channelassignment vectors that hold a sequence of binary numbers,C = {1, · · · , c, · · · ,C}, and the cardinality of the set isC = |C| = 2K . Mathematically, [J]m,1:K ∈ C. AlthoughC grows exponentially with K, we focus on a single hopnetwork in which the number of potential channels is limited,e.g., 4− 6; hence, the cardinality C is also limited. Next, weallocate channel assignment vectors to the spectrum sensors



6

rather than individual channels as in J. Define a channelassignment vector to the spectrum sensors binary assignmentmatrix, Vz = {vzm,c | 1 ≤ m ≤ M, c ∈ C}, of sizeM×C, where a value of 1 for vzm,c indicates that the channelassignment vector c is allocated to the m-th spectrum sensor.In one of the steps of the C-E algorithm, random samples ofthis matrix are generated, and the superscript z is introducedto denote the sample number.

The Vz samples are generated following a probability massfunction (p.m.f) that is denoted by matrix Qi, which is definedas Qi := {qim,c | 1 ≤ m ≤ M, c ∈ C}, where qim,c denotesthe probability that m is scheduled to sense the channels invector c. The C-E algorithm operates iteratively, and in everystep, the p.m.f matrix is updated. The superscript (·)i denotesthe iteration number. Each iteration of the C-E algorithmconsists of the following steps:

1) Initialization: Set the iteration counter to i = 1 andthe maximum iteration number to imax. Set the initialstochastic policy of all of the spectrum sensors to be auniform distribution on the channel assignment vectorset C, such that m chooses vector c with probabilityq1m,c = 1/C, ∀m, c.

2) Generation of Sample Solutions: Generate Z samplesof the matrix Vz based on the p.m.f matrix Qi. Notethat each spectrum sensor is randomly assigned onechannel assignment vector that holds several channels,i.e.,

∑C1 v

zm,c = 1, ∀z ∀m.

3) Performance Evaluation: Substitute the Z samples ofVz into Eqn. (13) to obtain an objective function valueOz for each sample; the superscript (·)z has beenintroduced to denote the sample number. Sort the Zvalues of Oz in descending order. Set ρ to be a fractionof the sorted objective values to retain, and then, takethe largest dρZe values of the sorted set and ignore allof the others. Moreover, set η to be the smallest valuein the sorted and retained set.

4) p.m.f. Update: Update the p.m.f. based on the retainedobjective function values. The value of qi+1

m,c is deter-mined by

qi+1m,c =

∑Zz=1 v

zm,cIOz≥η

dρZe, (14)

In this step, the channel vector assignment probabilityqim,c is updated by increasing the probability of assign-ments that are generating large objective function valuesover the various randomly generated samples.

5) Stopping Criterion: The algorithm stops iterating if themaximum number of iterations imax is reached or thefollowing inequality stands

||Qi+1 −Qi||Fr ≤ ε, (15)

where || · ||Fr denotes the Frobenius norm1. Otherwise,increment the iteration counter i and go back to Step

1The Frobenius norm is defined as the square root of the sum of the absolutesquares of the elements of the matrix. For example, if

A =

[a11 a12a21 a22

],

2. Eqn. (15) represents the convergence condition ofp.m.f Qi. It was shown in [32] that the sequence ofp.m.f converges with probability 1 to a unit mass that islocated at one of the samples.Note that fine tuning the values ε and imax impacts theconvergence speed of the algorithm and the quality ofthe obtained solution. A large value of ε results in fasterconvergence but a shorter average available time of thechannel. Additionally, a larger value for imax leads toa slower convergence speed but also leads to a longeraverage available time for the channel.

6) Solution Selection: When the algorithm terminates, se-lect the solution Vz that generates the largest objectivevalue Oz . Set the values of J based on the assignmentssolution in Vz . In other words, the channel-vector-to-the-spectrum-sensors assignment in Vz is mapped to thechannels-to-spectrum-sensors assignment in J which isa solution to the original problem SSS.

The sink schedules spectrum sensors to detect the licensedchannels according to the solution obtained in Step-6. Afterthe spectrum-sensing phase, spectrum sensors report theirdecisions on the channel availability to the sink. The sinkestimates the availability of each channel based on the Logic-OR rule and utilizes the available channels to collect datafrom the data sensors. In the following, we investigate thedata sensor resource allocation (DSRA) problem.

B. Transmission Time and Power Allocation in the DataTransmission Phase

In the data transmission phase, data sensors report thecollected data to the sink. Because a data sensor is battery-powered, minimizing its energy consumption becomes criticalto prolong its lifetime. To accomplish this goal, we firstformulate the problem of the data sensors’ transmission timeand power allocation as a biconvex optimization problem, andthen, we propose a joint time and power allocation (JTPA)algorithm to obtain a solution.

1) Problem Formulation: Available channels detected bythe spectrum sensors are allocated to the B cognitive radiotransceivers that are mounted on the sink. If the number ofavailable channels is less than B, then all of the availablechannels are allocated. Alternatively, the available channels aresorted with respect to their sojourn time, and the channels withthe largest sojourn time values are allocated to transceivers.Let K be the number of allocated channels, and note thatK ≤ B. Because all of the channels have the same bandwidthand average power gain, a long average sojourn time impliesa large capacity.

Recall that αk is the k-th channel’s available time. However,scheduling the data sensors to transmit for the entire αk

increases the chance of collision between the data sensor andthe returning PU. Let αk be the maximum access time of thek-th channel, where αk < αk. It is important to design αk

then||A||Fr =

√|a11|2 + |a12|2 + |a21|2 + |a22|2.



7

such that a low collision probability pkcoll(αk) is maintained

on the k-th channel. Given that the PU behavior on eachchannel is a stationary exponential ON-OFF random process,the probability of collision pkcoll(α

k) is given by [27]

pkcoll(αk) = P kInactive · (1− e

−µkαk

). (16)

where P kInactive is the probability that PU is not present on thek-th channel at the beginning of the data transmission phase,and (1 − e−µkα

k

) captures the probability that PU returns in[0, αk]. The detailed derivation of Eqn. (16) is provided inAppendix A.

To maintain a target collision probability pkcoll, the channelaccess time should not exceed,

αk ≤− ln(1− pkcoll/P kInactive)

µk. (17)

Furthermore, αk is bounded by the duration of the datatransmission phase T − τs. Thus,

αk = min

(− ln(1− pkcoll/P kInactive)

µk, T − τs

). (18)

Let T and P with elements tn,k and pn,k denote thetransmission time and power allocation matrices of size N×K.Let tn,k and pn,k denote the transmission time and power ofthe n-th data sensor over the k-th channel, respectively. Thetotal energy consumption of the data sensors is determined by

N∑n=1

K∑k=1

tn,kpn,k. (19)

The transmission time of all of the data sensors over the k-thchannel is limited by the channel access time αk,

N∑n=1

tn,k ≤ αk,∀k. (20)

Furthermore, the transmission time of the n-th data sensoris bounded by the duration of the data transmission phase,namely,

K∑k=1

tn,k ≤ T − τs,∀n. (21)

The data amount that is required from the n-th data sensoris denoted by Dn. During the data transmission phase, the n-th data sensor transmits sensed data over the k-th channel tothe sink at a transmission power of pn,k and duration of tn,k.The data transmission rate is given by

Rn,k = W log2 (1 + δn,kpn,k) , (22)

where δn,k represents the n-th sensor channel gain over thek-th channel at the sink. The allocated rate should be suffi-ciently large to support the generated data. This relationshipis captured by

K∑k=1

tn,kRn,k ≥ Dn. (23)

The transmission time tn,k and power pn,k are nonnegative.

Additionally, pn,k is constrained by the maximum transmissionpower pmax. Thus, we have

tn,k ≥ 0,∀k, ∀n and (24)0 ≤ pn,k ≤ pmax. (25)

We allocate the transmission time T and power P tominimize the energy consumption of all of the data senors,which can be formulated as:

(DSRA) minT,P

N∑n=1

K∑k=1

tn,kpn,k

s.t.

∑Nn=1 tn,k ≤ αk,∀k,∑Kk=1 tn,k ≤ T − τs,∀n,∑Kk=1 tn,kW log2(1 + δn,kpn,k) ≥ Dn,∀n,

tn,k ≥ 0,∀k, n,0 ≤ pn,k ≤ pmax,∀k, n.

The amount of data to transmit is determined by the productof the transmission time tn,k and logarithm of the power pn,k.These structures lead to the non-convexity of the problemDSRA with potentially multiple local optima and generallyimplies difficulty in determining the global optimal solution[33]. However, by showing that DSRA is biconvex, we gainaccess to algorithms that efficiently solve biconvex problems[34]; see Appendix B.

2) Joint Time and Power Allocation (JTPA) Algorithm:Because DSRA is biconvex, the variable space is divided intotwo disjoint subspaces. Therefore, the problem is divided intotwo convex subproblems that can be solved efficiently: timeallocation (DSRA-1) and power allocation (DSRA-2). Thetime allocation problem is given by

(DSRA-1) minT

N∑n=1

K∑k=1

tn,kpn,k

s.t. (20)(21)(23)(24),

while the power allocation problem is given by,

(DSRA-2) minP

N∑n=1

K∑k=1

tn,kpn,k

s.t. (23)(25),

In the following, we adopt the Alternate Convex Search in [34]to solve the DSRA problem. In every step of the proposedalgorithm, one of the variables is fixed, and the other isoptimized, and vice versa in the subsequent step. The proposedalgorithm solves the two problems iteratively and convergesto a partially optimal solution2. The detailed procedure of theproposed algorithm is given as follows:

2The definition of a partial optimal solution is given as follows:

Definition 1. Let f : S → R be a given function and let (x∗, y∗) ∈ S.Therefore, (x∗, y∗) is called a partial optimum of f on S, if

f(x∗, y∗) ≤ f(x, y∗) ∀x ∈ Sy∗ ,

f(x∗, y∗) ≤ f(x∗, y) ∀y ∈ Sx∗ .

Sy∗ and Sx∗ denote the y∗- and x∗-sections of S [34].



8

Algorithm 1 Proposed Algorithm JTPAInput: Network parameters, stopping criterion ε and maxi-

mum number of iterations imax.Output: The optimal (T∗,P∗).

1: Choose an arbitrary starting point (T0,P0) and set theiteration index as i = 0, and the initial solution as z0 = 0;

2: repeat3: Fix Pi and determine the optimal Ti+1 by solving

DSRA-1 via the Simplex method [35];4: Fix Ti+1, determine the optimal Pi+1 and objective

function value zi by solving DSRA-2 via the InteriorPoint method [36];

5: i = i+ 1;6: until zi+1 − zi−1 < ε or i ≥ imax

7: return (Ti+1,Pi+1)

The convergence of the proposed algorithm to the globaloptimum is not guaranteed since DSRA is biconvex and couldhave several local optima. However, because the objectivefunction is differentiable and biconvex over a biconvex set,convergence to a stationary point that is partially optimal isguaranteed [34]. Data sensors transmit their data to the sinkusing the transmission time and power that is determined bythe proposed JTPA algorithm.

IV. PERFORMANCE EVALUATION

We evaluate the performance of the C-E algorithm inthe spectrum sensing phase and the JTPA algorithm in thedata transmission phase through performing simulations. Thesimulation results are obtained through Matlab on a computerwith intel core(TM) i7-4510u [email protected] 2.6GHz, 8 GBRAM.

A. Simulation Setup

We simulate an HCRSN that consists of M = 10 spectrumsensors and N = 30 data sensors. The sensors are randomlyplaced in a circular area with a radius of 20 meters. The sink islocated at the center of this circular area. The HCRSN coexistswith a primary network that is deployed over an area that hasa radius of 200 meters. The PUs’ transmission power is 1 mW,and the noise power is −80 dB. The PU’s channel gain at thesensor is simulated based on 1/d3.5, where d is the distancebetween the PU and spectrum sensor. The target false alarmprobability for all of the spectrum sensors pf is set to 0.1. PUstransmit QPSK modulated signals, with each over a 6 MHzbandwidth W . The default number of licensed channels isseven unless specified otherwise. Over the seven channels, sev-en PUs operate over one channel exclusively. Their transitionrates λk, k = 1, · · · , 7, are 0.6, 0.8, 1, 1.2, 1.4, 1.6, 1.8, respec-tively. Additionally, the transition rates µk, k = 1, · · · , 7, are0.4, 0.8, 0.6, 1.6, 1.2, 1.4, 1.8, respectively. The network oper-ates periodically over slots of length T = 100 ms [37] (see Fig.2). The maximum transmission power is set to pmax = 100mW [38]. The remaining parameters are set according to TableI unless specified otherwise. In the following two subsections,we evaluate the performances of the proposed algorithms.

TABLE IPARAMETER SETTINGS

Parameter SettingsUpper bound of the transmission power pmax 100 mWBandwidth of the licensed channel W [3] 6 MHzSampling rate of the EH spectrum sensor U [28] 6000False alarm probability pf 0.1Energy consumption per spectrum sensing 0.11 mJTime consumption per spectrum sensing τs′ 1 msDuration of the spectrum-sensing phase τs 5 msDuration of each slot T [37] 100 msUpper bound of the collision probability pkcoll(α

k) 0.1Mis-detection probability threshold MDthr 0.9Fraction of samples retained in C-E ρ 0.6Stopping threshold of C-E ε 10−3

B. Performance Evaluation of the C-E Algorithm

The following simulation results provide insights into theperformance of the C-E algorithm over the spectrum sensingphase. Metrics of interest include the convergence speed andquality of the obtained solution. Furthermore, we study theimpact of the stopping criterion parameters ρ and ε on thosemetrics. The performance of the proposed algorithm is alsocompared to the performance of a candidate greedy algorithm.

In Fig. 4, we show the optimality of the C-E algorithmin a scenario that has 3 spectrum sensors and 2-4 licensedchannels. We reduce the number of spectrum sensors in sucha way that an exhaustive search can be efficiently performed.The EH rate and sensing time are set to be sufficiently largethat any assignment would be feasible. The C-E algorithm’soptimal solution, i.e., the Detected Average Available Time ofChannels (DAATC), is compared to that obtained by randomassignment and exhaustive search. The random assignmentrandomly assigns licensed channels to the spectrum sensors,while the exhaustive search traverses all of the possible assign-ments. As shown in Fig. 4, the expected detected channel’savailable time obtained by the C-E algorithm is close to thatof the exhaustive search and is able to achieve 87% − 94%of it. The proposed algorithm’s computed solution is 2 to 3times larger than that of the random assignment.

For a network of 10 spectrum sensors with 7 channels, thestability of the C-E algorithm is shown in Figs. 5 and 6. Fig. 5shows that the convergence of the C-E algorithm with respectto the EH rate ranges from 3 mW to 7 mW3. It can be seen thatthe value of the objective function fluctuates during the startupphase and then converges to the maximum DAATC after30 iterations. Moreover, the value of the objective functionincreases by one-third for the case in which EH rate = 7 mW,while it doubles for the case in which the EH rate = 3 mW.This finding demonstrates the responsiveness of the stochasticpolicy updating strategy defined by Eqn. (14). Moreover, itcan be clearly seen that the DAATC increases with the EHrate.

Fig. 6 shows the convergence results for the C-E algorithmwith respect to the spectrum-sensing duration τs range of 2

3In [25], the real experimental data obtained from the Baseline Measure-ment System (BMS) of the Solar Radiation Research Laboratory (SRRL)shows that, the EH rate ranges from 0 mW to 100 mW for most of the day.



9

Random AssignmentC-E

Exhaustive Search

Number of spectrumsensors = 3

Number of Channels

DA

AT

C(s

ec)

2 3 40

0.5

1

1.5

2

2.5

3

Fig. 4. The comparison of C-E algorithm’s performance and the performanceof random assignment and exhaustive search in terms of the DAATC.

EH rate = 3 mWEH rate = 5 mWEH rate = 7 mW

Number of Iterations

DA

AT

C(s

ec)

0 5 10 15 20 25

1.5

2

2.5

3

3.5

4

Fig. 5. Convergence of the C-E algorithm for three different EH rates, πm

ms to 6 ms and EH rate of 7 mW. As we can see from thefigure, the value of DAATC fluctuates at the startup phase.This is because the samples of channel assignment vectorsare generated according to the uniform distribution at theinitialization step of the C-E algorithm. As the C-E algorithmexecutes, the probability to generate samples that bring higherDAATC increases. At last, the algorithm converges to astable solution that leads to highest DAATC in 30 iterations.Furthermore, the DAATC increases with the length of thespectrum sensing phase τs, because more channels can bedetected by the spectrum sensors with larger τs.

The C-E algorithm stops iterating if the inequality in Eqn.(15) holds, or the maximum number of iterations is reached.Figs. 7 and 8 show the impact of fine tuning the algorithmparameters, ε and ρ, on the convergence speed and qualityof the obtained solution. It can be seen from Fig. 7 that alarge number of iterations is required to satisfy the stoppingcriterion, and a larger DAATC can be obtained for a small ε.

τs = 6 msτs = 4 msτs = 2 ms

Number of Iterations

EH rate = 7 (mW)

DA

AT

C(s

ec)

0 5 10 15 20 25 302

2.5

3

3.5

4

4.5

Fig. 6. Convergence of the C-E algorithm for three different spectrum-sensingdurations, τs.

Furthermore, the algorithm converges in less than 100 iterationeven for the small value of ε = 10−6. Fig. 8 shows the impactof the fraction of samples that is retained (i.e., ρ) in each stepon the algorithm performance. The C-E algorithm convergesfaster with small ρ. Moreover, the DAATC peaks at one valueof ρ and then starts falling. For the parameters that consideredin this study, ρ peaks at 0.6. The fraction ρ should be optimizedto obtain a larger DAATC.

C-E

C-E

Stopping Criterion Threshold ε

Num

ber

ofIt

erat

ions

for

Con

verg

ence

DA

AT

C(s

ec)

10−6 10−5 10−4 10−33.0

3.5

4.0

4.5

40

60

80

Fig. 7. The effect of ε on the performance of the C-E algorithm.

Figs. 9 and 10 show the comparison between the perfor-mance of the C-E algorithm and that of the greedy algorithm.The greedy algorithm corresponds to the algorithm proposedin [39]; it picks the spectrum sensors sequentially and assignsthem the channels that bring the largest DAATC. It can be seenfrom Fig. 9 that the C-E algorithm outperforms the greedyalgorithm in terms of the obtained DAATC over a range ofEH rates. A similar result can be seen in Fig. 10, where thenumber of spectrum sensors varies for a fixed EH rate of 7mW.



10

C-E

C-E

The fraction of samples retained in each iteration ρ

Num

ber

ofIt

erat

ions

for

Con

verg

ence

DA

AT

C(s

ec)

0.2 0.4 0.6 0.83.0

3.5

4.0

4.510

20

30

Fig. 8. The impact of the fraction of retained samples ρ on the performanceof the proposed C-E algorithm.

GreedyC-E

EH rate in mW , πm

DA

AT

C(s

ec)

2 4 6 80

1

2

3

4

5

Fig. 9. A comparison of the C-E algorithm and the Greedy algorithmperformance for a range of EH rates.

C. Performance Evaluation of the JTPA Algorithm

In this subsection, we evaluate the performance of the JTPAalgorithm. For a network of three data sensors with threechannels, we first verify the optimality of JTPA by comparingits performance to that of random scheme and optimal schemein Fig. 11. The random scheme randomly assigns channelsto the data sensors, while the optimal scheme searches overthe complete space. Once the channels are assigned, a Matlaboptimization toolbox is used to allocate the time and power. Asshown in Fig. 11, JTPA consumes 5% to 14% more energythan the optimal scheme. However, JTPA conserves 18% to31% more energy than the random assignment scheme.

The convergence of JTPA is evaluated in a network of tenspectrum sensors and thirty data sensors with five channels.The transition rates λ1:5 and µ1:5 are set to 0.6, 0.8, 1, 1.2, 1.4and 0.4, 0.8, 0.6, 1.6, 1.2, respectively. The spectrum sensingduration τs is set to 5 ms. Fig. 12 shows the convergenceperformance of JTPA with respect to the data amount (Dn)

GreedyC-E

Number of Spectrum Sensors

DA

AT

C(s

ec)

EH rate = 7 mW

2 4 6 80

1

2

3

4

Fig. 10. A comparison of the C-E algorithm and the Greedy algorithmperformance for a number of spectrum sensors.

Optimal schemeJTPARandom scheme

Number of channels = 3Number of data sensors = 3

Data amount of each data sensor (Kb), Dn

Ene

r gy

cons

umpt

ion

ofal

lda

tase

nsor

s(J

)

1 3 50

0.5

1

1.5

2

2.5

3

3.5×10−5

Fig. 11. A comparison of JTPA with the random scheme and optimal scheme

that is transmitted from each data sensor to the sink. It can beobserved that the JTPA algorithm converges after 10 iterationsand the energy consumption decreases 97% during the first6 iterations which implies the efficiency of the proposedalgorithm.

In Figs. 13 and Fig. 14, we compare the energy consumptionof data transmission under the JTPA algorithm and the pmaxscheme. In the pmax scheme, the data sensors transmit at themaximum available power pmax, and the transmission time isdetermined by solving the linear programming problem JTPA-1. The pmax scheme is comparable to the channel allocationscheme proposed in [21], in which data sensors transmit dataat fixed transmission power. Fig. 13 shows the comparison ofthe energy consumption performance with respect to variousrequired amount of data, while pmax is set to 5 mW. Becausethe JTPA algorithm jointly allocates the transmission time andpower over the available channels, JTPA consumes less energythan pmax scheme for different data amount. Fig. 14 shows the



11

Dn = 3 KbDn = 2 KbDn = 1 Kb

pmax = 100 (mW)

Iterations

Ene

rgy

cons

umpt

ion

ofal

lda

tase

nsor

s(J

)

0 5 10 150

10−2

10−1

1

×10−4

Fig. 12. The convergence of JTPA for a three data amounts, Dn = 1,2,3Kb.

JTPA

pmax Scheme

pmax = 5 (mW)

Data amount, Dn (Kb)

Ene

rgy

cons

umpt

ion

ofal

lda

tase

nsor

s(J

)

3 4 5 6 7

4

8

12

16

20×10−6

Fig. 13. A comparison of JTPA and the pmax scheme for a range of dataamounts.

comparison of the JTPA algorithm against the pmax schemefor various values of pmax and data amount Dn = 3 Kb∀n. The energy consumption of the JTPA algorithm decreaseswith an increase in pmax because data sensors can adjust thetransmission power in a larger space for a larger pmax. Similarto the results shown in Fig. 13, JTPA consumes less energythan that of the pmax scheme due to the joint allocation oftransmission time and power.

V. CONCLUSIONS

In this paper, a novel resource allocation solution for het-erogeneous cognitive radio sensor networks (HCRSNs) hasbeen proposed. The proposed solution assigns channels tospectrum sensors in such a way that the detected availabletime of the channels is maximized. Furthermore, it efficientlyallocates the available channels to the data sensors alongwith the transmission time and power in order to prolongtheir lifetime. Extensive simulation results have demonstrated

JTPApmax Scheme

Data Amount, Dn = 3 (Kb)

Maximum transmission power, pmax (mW)

Ene

rgy

cons

umpt

ion

ofal

lda

tase

nsor

s(J

)

1 2 3 4 5

2

4

6

8

×10−6

Fig. 14. A comparison of JTPA and the pmax Scheme for various pmax

values.

the optimality and efficiency of the proposed algorithms.The solution presented in this work enables using primarynetworks channels efficiently while adapting in real time to theavailability of harvested energy, and optimizes the allocation ofthe battery-powered data sensors’ scarce resources. This yieldssignificantly higher spectral and energy-efficient HCRSNs.

For the future work, we plan to investigate the channelallocation and routing protocol design in EH-aided multi-hop HCRSNs, considering the time-varying EH rate and theadaptive detection threshold of sensors.

ACKNOWLEDGMENT

This work was supported by the Fundamental ResearchFunds for the Central Universities of the Central South Univer-sity (No. 2013zzts043). The project was supported partially bythe Kuwait Foundation for the Advancement of Sciences underproject code: P314-35EO-01. This work was also supportedby National Natural Science Foundation of China (61379057,61272149) and NSERC, Canada.

APPENDIX ADERIVATION OF THE COLLISION PROBABILITY pkcoll(α

k)

Let T kInactive be the sojourn time of a OFF/Inactive periodwith the probability density function (p.d.f) fTk

Inactive(α).

Given the exponentially distributed ON/OFF period, the p.d.fof the Inactive period is equal to [27]

fTk

Inactive(α) = µke

−µkα

The probability that the OFF/Inactive period is less than αk,i.e., the PU on channel k returns in [0, αk], can be derived tobe

Pr(T kInactive < αk) =

∫ αk

0

fTk

Inactive(α) dα

=1− e−µkαk

.



12

Since channel k is available with probability P kInactive, theprobability of collision pkcoll(α

k) is given by:

pkcoll(αk) = P kInactive · (1− e

−µkαk

).

APPENDIX BBI-CONVEXITY OF THE DSRA PROBLEM

In the following we show that the DSRA problem is bi-convex.

Theorem 1. If we fix one set of variables in T or P, thenDSRA is convex with respect to the other set of variables.Thus, DSRA is biconvex.

Proof: We first determine a feasible P, and then, DSRAbecomes a problem of determining T to satisfy

(DSRA-1) minT

N∑n=1

K∑k=1

tn,kpn,k

s.t. (20)(21)(23)(24).

which is linear and convex due to the linear objective functionand linear feasible set. DSRA-1 can be solved using the Sim-plex method [35]. Additionally, by fixing T, DSRA becomesa problem of determining P to satisfy

(DSRA-2) minP

N∑n=1

K∑k=1

tn,kpn,k

s.t. (23)(25),

DSRA-2 can be solved by the interior point method. BothDSRA-1 and DSRA-2 are convex and can be solved efficient-ly. Therefore, the objective function

∑Nn=1

∑Kk=1 tn,kpn,k is

biconvex on the feasible set which makes DSRA a biconvexproblem.

REFERENCES

[1] L. Borges, F. Velez, and A. Lebres, “Survey on the characterization andclassification of wireless sensor network applications,” IEEE Commun.Surveys Tuts, vol. 16, no. 4, pp. 1860–1890, 2014.

[2] O. Akan, O. Karli, and O. Ergul, “Cognitive radio sensor networks,”IEEE Network, vol. 23, no. 4, pp. 34–40, 2009.

[3] N. Zhang, H. Liang, N. Cheng, Y. Tang, J. Mark, and X. Shen, “Dynamicspectrum access in multi-channel cognitive radio networks,” IEEE J.Select. Areas Commun., vol. 32, no. 11, pp. 2053–2064, 2014.

[4] C.-C. Chang and H.-Y. Lin, “Spectrum reconstruction for on-chipspectrum sensor array using a novel blind nonuniformity correctionmethod,” IEEE Sensors J., vol. 12, no. 8, pp. 2586–2592, 2012.

[5] J. Ren, Y. Zhang, K. Zhang, A. Liu, J. Chen, and X. Shen, “Lifetimeand energy hole evolution analysis in data-gathering wireless sensornetworks,” IEEE Trans. Ind. Informat., to appear, DOI: 10.1109/TI-I.2015.2411231.

[6] J. Ren, Y. Zhang, N. Zhang, D. Zhang, and X. S. Shen, “Dy-namic channel access to improve energy efficiency in cognitive ra-dio sensor networks,” IEEE Trans. Wireless Commun., to appear,DOI: 10.1109/TWC.2016.2517618.

[7] Z. Su, Q. Xu, and Q. Qi, “Big data in mobile social networks: a qoe-oriented framework,” IEEE Network, vol. 30, no. 1, pp. 52–57, 2016.

[8] Y. Zhang, S. He, and J. Chen, “Data gathering optimization by dynamicsensing and routing in rechargeable sensor networks,” IEEE/ACM Trans.Netw., to appear, DOI: 10.1109/TNET.2015.2425146.

[9] G. Shah and O. Akan, “Cognitive adaptive medium access control incognitive radio sensor networks,” IEEE Trans. Veh. Technol., vol. 64,no. 2, pp. 757–767, 2015.

[10] R. Deng, J. Chen, C. Yuen, P. Cheng, and Y. Sun, “Energy-efficientcooperative spectrum sensing by optimal scheduling in sensor-aidedcognitive radio networks,” IEEE Trans. Veh. Technol., vol. 61, no. 2,pp. 716–725, 2012.

[11] I. F. Akyildiz, B. F. Lo, and R. Balakrishnan, “Cooperativespectrum sensing in cognitive radio networks: A survey,” PhysicalCommunication, vol. 4, no. 1, pp. 40 – 62, 2011. [Online]. Available:http://www.sciencedirect.com/science/article/pii/S187449071000039X

[12] N. Zhang, N. Cheng, N. Lu, H. Zhou, J. W. Mark, and X. S. Shen,“Risk-aware cooperative spectrum access for multi-channel cognitiveradio networks,” IEEE J. Select. Areas Commun., vol. 32, no. 3, pp.516–527, March 2014.

[13] P. Cheng, R. Deng, and J. Chen, “Energy-efficient cooperative spectrumsensing in sensor-aided cognitive radio networks,” Wireless Commun.,vol. 19, no. 6, pp. 100–105, 2012.

[14] W. Zhang, Y. Guo, H. Liu, Y. Chen, Z. Wang, and J. Mitola, “Distributedconsensus-based weight design for cooperative spectrum sensing,” IEEETrans. Parallel Distrib. Syst., vol. 26, no. 1, pp. 54–64, 2015.

[15] Z. Khan, J. Lehtomaki, K. Umebayashi, and J. Vartiainen, “On theselection of the best detection performance sensors for cognitive radionetworks,” IEEE Signal Processing Lett., vol. 17, no. 4, pp. 359–362,2010.

[16] X. Liu, B. Evans, and K. Moessner, “Energy-efficient sensor schedulingalgorithm in cognitive radio networks employing heterogeneous sen-sors,” IEEE Trans. Veh. Technol., vol. 64, no. 3, pp. 1243–1249, 2013.

[17] A. Ebrahimzadeh, M. Najimi, S. Andargoli, and A. Fallahi, “Sensorselection and optimal energy detection threshold for efficient cooperativespectrum sensing,” IEEE Trans. Veh. Technol., vol. 64, no. 4, pp. 1565–1577, 2015.

[18] S. Eryigit, S. Bayhan, and T. Tugcu, “Energy-efficient multichannelcooperative sensing scheduling with heterogeneous channel conditionsfor cognitive radio networks,” IEEE Trans. Veh. Technol., vol. 62, no. 6,pp. 2690–2699, 2013.

[19] S. Bayhan and F. Alagoz, “Scheduling in centralized cognitive radionetworks for energy efficiency,” IEEE Trans. Veh. Technol., vol. 62,no. 2, pp. 582–595, 2013.

[20] G. Shah, F. Alagoz, E. Fadel, and O. Akan, “A spectrum-aware clusteringfor efficient multimedia routing in cognitive radio sensor networks,”IEEE Trans. Veh. Technol., vol. 63, no. 7, pp. 3369–3380, 2014.

[21] S.-S. Byun, I. Balasingham, and X. Liang, “Dynamic spectrum allocationin wireless cognitive sensor networks: Improving fairness and energyefficiency,” in Proc. IEEE VTC, 2008, pp. 1–5.

[22] Z. Hu, Y. Sun, and Y. Ji, “A dynamic spectrum access strategy based onreal-time usability in cognitive radio sensor networks,” in Proc. IEEEMSN, 2011, pp. 318–322.

[23] J. Ayala Solares, Z. Rezki, and M. Alouini, “Optimal power allocationof a single transmitter-multiple receivers channel in a cognitive sensornetwork,” in Proc. IEEE ICWCUCA, 2012, pp. 1–6.

[24] M. Naeem, K. Illanko, A. Karmokar, A. Anpalagan, and M. Jaseemud-din, “Energy-efficient cognitive radio sensor networks: Parametric andconvex transformations,” Sensors, vol. 13, no. 8, pp. 11 032–11 050,2013.

[25] Y. Zhang, S. He, J. Chen, Y. Sun, and X. Shen, “Distributed samplingrate control for rechargeable sensor nodes with limited battery capacity,”IEEE Trans. Wireless Commun., vol. 12, no. 6, pp. 3096–3106, 2013.

[26] P. Tehrani, L. Tong, and Q. Zhao, “Asymptotically efficient multichannelestimation for opportunistic spectrum access,” Signal Processing, IEEETransactions on, vol. 60, no. 10, pp. 5347–5360, 2012.

[27] H. Kim and K. Shin, “Efficient discovery of spectrum opportunities withmac-layer sensing in cognitive radio networks,” IEEE Trans. MobileComput., vol. 7, no. 5, pp. 533–545, 2008.

[28] Y.-C. Liang, Y. Zeng, E. Peh, and A. T. Hoang, “Sensing-throughputtradeoff for cognitive radio networks,” IEEE Trans. Wireless Commun.,vol. 7, no. 4, pp. 1326–1337, 2008.

[29] S. Atapattu, C. Tellambura, and H. Jiang, Energy Detection for SpectrumSensing in Cognitive Radio, ser. SpringerBriefs in Computer Science.Springer, 2014.

[30] R. Rubinstein, “The cross-entropy method for combinatorial and contin-uous optimization,” Methodol. Comput. Appl., vol. 1, no. 2, pp. 127–190,1999.

[31] P.-T. De Boer, D. P. Kroese, S. Mannor, and R. Y. Rubinstein, “A tutorialon the cross-entropy method,” Annals of operations research, vol. 134,no. 1, pp. 19–67, 2005.

[32] A. Costa, O. D. Jones, and D. Kroese, “Convergence properties of thecross-entropy method for discrete optimization,” Operations ResearchLetters, vol. 35, no. 5, pp. 573–580, 2007.



13

[33] C. Floudas and V. Visweswaran, Comput. Chem. Eng., vol. 14, no. 12,pp. 1397 – 1417, 1990.

[34] J. Gorski, F. Pfeuffer, and K. Klamroth, “Biconvex sets and optimizationwith biconvex functions: a survey and extensions,” Math. Method. Oper.Res., vol. 66, no. 3, pp. 373–407, 2007.

[35] G. B. Dantzig and M. N. Thapa, “Linear programming 2: Theory andextensions,” vol. 2, 2003.

[36] S. Boyd and L. Vandenberghe, “Convex optimization,” Cambridgeuniversity press, 2004.

[37] Y. Pei, Y.-C. Liang, K. Teh, and K. H. Li, “Energy-efficient design ofsequential channel sensing in cognitive radio networks: Optimal sensingstrategy, power allocation, and sensing order,” IEEE J. Select. AreasCommun., vol. 29, no. 8, pp. 1648–1659, 2011.

[38] T. Shu, M. Krunz, and S. Vrudhula, “Joint optimization of transmitpower-time and bit energy efficiency in cdma wireless sensor networks,”IEEE Trans. Wireless Commun., vol. 5, no. 11, pp. 3109–3118, 2006.

[39] H. Yu, W. Tang, and S. Li, “Optimization of cooperative spectrumsensing in multiple-channel cognitive radio networks,” in Proc. IEEEGLOBECOM, 2011, pp. 1–5.

Deyu Zhang recevied the B.Sc. degree in Com-munication Engineering from PLA Information En-gineering University in 2005; and M. Sc degreefrom Central South University in 2012, China, allin Communication Engineering. He is pursuing hisPh.D degree in Central South University in computerscience. He has been a visiting scholar with theDepartment of Electrical and Computer Engineering,University of Waterloo, ON, Canada from 2014to 2016. His research interests include stochasticresource allocation in wireless sensors network and

Cloud Radio Access Networks.

Zhigang Chen (M’04) received B.Sc., M.Sc. andPh.D degrees from Central South University, China,in 1984, 1987 and 1998, all in computer science,respectively. He is a professor and Ph.D. Supervisorwith CSU. His research interests are in networkcomputing and distributed processing.

Ju Ren (S’13)received his B.Sc. and M.Sc. degrees,both in computer science, from Central South Uni-versity, China, in 2009 and 2012, respectively. Heis currently a Ph.D. candidate at the Departmen-t of Computer Science, Central South University,China. From Aug. 2013 to Sept. 2015, he was avisiting Ph.D. student in the Department of Electricaland Computer Engineering, University of Waterloo,Canada. His research interests include wireless sen-sor networks, mobile sensing/computing, and cloudcomputing.

Ning Zhang (S’12, M’16) earned the Ph.D degreefrom University of Waterloo in 2015. He received hisB.Sc. degree from Beijing Jiaotong University andthe M.Sc. degree from Beijing University of Postsand Telecommunications, Beijing, China, in 2007and 2010, respectively. He is now a postdoc researchfellow at BBCR lab in University of Waterloo.His current research interests include next genera-tion wireless networks, software defined networking,green communication, and physical layer security.

Mohamad Khattar Awad (S’02, M’09), earnedthe B.A.Sc. in electrical and computer engineering(communications option) from the University ofWindsor, Ontario, Canada, in 2004 and the M.A.Sc.and Ph.D. in electrical and computer engineeringfrom the University of Waterloo, Ontario, Canada,in 2006 and 2009, respectively.

From 2004 to 2009 he was a research assistantin the Broadband Communications Research Group(BBCR), University of Waterloo. In 2009 to 2012,he was an Assistant Professor of Electrical and

Computer Engineering at the American University of Kuwait. Since 2012,he has been with Kuwait University as an Assistant Professor of ComputerEngineering.

Dr. Awad’s research interest includes wireless and wired communications,software-defined networks resource allocation, wireless networks resourceallocation, and acoustic vector-sensor signal processing. He received theOntario Research & Development Challenge Fund Bell Scholarship in 2008and 2009, the University of Waterloo Graduate Scholarship in 2009, and afellowship award from the Dartmouth College, Hanover, NH in 2011. In 2015,he received the Kuwait University Teaching Excellence Award.

Haibo Zhou (M’14) received the Ph.D. degree inInformation and Communication Engineering fromShanghai Jiaotong University, Shanghai, China, in2014. He is currently a Post-Doctoral Fellow withthe Broadband Communications Research (BBCR)Group, University of Waterloo. His current researchinterests include resource management and protocoldesign in cognitive radio networks and vehicularnetworks.



14

Xuemin(Sherman) Shen (IEEE M’97-SM’02-F’09)received the B.Sc.(1982) degree from Dalian Mar-itime University (China) and the M.Sc. (1987) andPh.D. degrees (1990) from Rutgers University, NewJersey (USA), all in electrical engineering. He isa Professor and University Research Chair, De-partment of Electrical and Computer Engineering,University of Waterloo, Canada. He is also the Asso-ciate Chair for Graduate Studies. Dr. Shens researchfocuses on resource management in interconnectedwireless/wired networks, wireless network security,

social networks, smart grid, and vehicular ad hoc and sensor networks. Heis an elected member of IEEE ComSoc Board of Governor, and the Chair ofDistinguished Lecturers Selection Committee. Dr. Shen served as the Techni-cal Program Committee Chair/Co-Chair for IEEE Globecom16, Infocom14,IEEE VTC10 Fall, and Globecom07, the Symposia Chair for IEEE ICC10,the Tutorial Chair for IEEE VTC’11 Spring and IEEE ICC08, the General Co-Chair for ACM Mobihoc15, Chinacom07 and QShine06, the Chair for IEEECommunications Society Technical Committee on Wireless Communications,and P2P Communications and Networking. He also serves/served as theEditor-in-Chief for IEEE Network, Peer-to-Peer Networking and Application,and IET Communications; a Founding Area Editor for IEEE Transactions onWireless Communications; an Associate Editor for IEEE Transactions on Ve-hicular Technology, Computer Networks, and ACM/Wireless Networks, etc.;and the Guest Editor for IEEE JSAC, IEEE Wireless Communications, IEEECommunications Magazine, and ACM Mobile Networks and Applications,etc. Dr. Shen received the Excellent Graduate Supervision Award in 2006,and the Outstanding Performance Award in 2004, 2007, 2010, and 2014from the University of Waterloo, the Premiers Research Excellence Award(PREA) in 2003 from the Province of Ontario, Canada, and the DistinguishedPerformance Award in 2002 and 2007 from the Faculty of Engineering,University of Waterloo. Dr. Shen is a registered Professional Engineer ofOntario, Canada, an IEEE Fellow, an Engineering Institute of Canada Fellow,a Canadian Academy of Engineering Fellow, a Royal Society of CanadaFellow, and a Distinguished Lecturer of IEEE Vehicular Technology Societyand Communications Society.

energy harvesting-aided spectrum sensing and data transmission...

Documents