Computers and Electrical Engineering

Transcription

1 Computer and Electrical Engineering 39 (213) Content lit available at SciVere ScienceDirect Computer and Electrical Engineering journal homepage: Licened uer activity etimation and track in mobile cognitive radio ad hoc network q Guoqin Ning a,b,, Kauhik R. Chowdhury b, Jiaqi Duan b, Pruayon Nintanavonga b a Department of Information Technology, Central China Normal Univerity, Wuhan 4379, China b Department of Electrical and Computer Engineering, Northeatern Univerity, Boton, MA 2115, USA article info abtract Article hitory: Available online 1 December 212 In cognitive radio (CR) network, a tatic activity model fail to capture the dynamic and time-varying behavior of the licened or primary uer (PU). In thi paper, a ditributed cheme i propoed that allow mobile CR uer to learn about the activity of the PU, and dieminate thi information to the neighboring node that alo function a information repoitorie. In order to guarantee ening preciion and tranmiion efficiency, the propoed method witche between time-intenive fine ening and quick normal ening. Our approach ue the maximum likelihood etimator to learn average buy and idle period in the fine ening tage. Thee identified activity pattern are then ued during normal ening, where the mean quare error (MSE) value of PU on off time i continuouly monitored to enure that the etimation i ufficiently accurate. When PU activity change ignificantly, the MSE i conidered a the indicator to re-tart the fine ening. Simulation reult reveal that our propoed method can efficiently track the dynamic of the PU activity. Ó 212 Elevier Ltd. All right reerved. 1. Introduction Cognitive radio (CR) i enviaged to upport the increaing demand of pectrum for wirele communication, by allowing CR uer (i.e., econdary uer, SU) to operate on the vacant part of the pectrum allocated to licened uer (i.e., primary uer, PU) [1,2]. Here, SU require frequent ening to determine the tate of the radio environment, and adaptively chooe tranmiion parameter for avoiding interfering with PU [3,4]. Many ening cheme require a priori knowledge of the tatitical behavior of PU. The exponential ditribution model i a common example for PU activity [5,6]. In[5], the author model the channel uage through the emi-markov proce, but do not point out how to leverage the detection hitory to improve the etimation of PU activity. In [7], a theoretical framework wa developed to jointly optimize the ening and tranmiion period with the aim of maximizing the pectrum efficiency ubject to interference avoidance contraint. However, the PU activity parameter, uch a the tatitical average idle and buy period, were aumed to be fixed and known a priori to SU. Furthermore, there i a coniderable body of work relying on a contant PU activity tatitical model to optimize the ening in CR network [8 12], without actually obtaining or inferring them. In mot of thee paper, the fixed and known average idle and buy period are vital in the calculation of the detection probability and fale alarm probability. Hence, it i advantageou to obtain the PU activity tatitic during the operation of the network, by collecting periodic meaurement from the environment. Such a primary activity prediction q Review proceed and approved for publication by Editor-in-Chief Dr. Manu Malek. Correponding author at: Department of Information Technology, Central China Normal Univerity, Wuhan 4379, China. Tel.: ; fax: addree: gqning@mail.ccnu.edu.cn (G. Ning), krc@ece.neu.edu (K.R. Chowdhury), jiaqi.duan@gmail.com (J. Duan), pruayonn@coe.neu.edu (P. Nintanavonga) /$ - ee front matter Ó 212 Elevier Ltd. All right reerved.

2 176 G. Ning et al. / Computer and Electrical Engineering 39 (213) approach can be found in [13], where the author ue frame tructure and time lot, and activity prediction wa limited to etimating the remaining time of the idle period, but not the average buy and idle period that together influence the choice of the pectrum. Therefore, in our paper, PU activity parameter are aumed to be unknown to the SU in advance, and our objective i to etimate the average idle and buy period and to track PU activity change. The approach in [14] i mot relevant to our cheme. Here, the author not only utilize the ON/OFF pectrum uage model, but alo preent an etimation technique to proce the detection hitory and learn the traffic pattern of PU. Neverthele, they aume that the SU already have a et of PU uage pattern ample without detailing (i) how to collect thee ample, and (ii) how many ample hould be enough to guarantee the ening preciion. In ummary, in order to enure that the pectrum ening cheme work accurately, deriving an accurate PU activity model i a key concern, which need further attention for the reearch community. Additionally, thi model mut be able to adapt to the change of PU activity over time. We would alo like to point out the relevance of our approach in light of recent ruling by the FCC on pectrum ening in the TV whitepace [15]. The FCC ruling in 211 tate that local ening (without geo-location and pectrum databae acce) i only permitted for certified device. Thi requirement neceitate reviiting the ening model and new approache that are able to function without pre-condition and location-pecific aumption. It i likely that CR device uing model that do not capture the exact pectrum uage activity at the teting ite will be unable to get the neceary permit for future operation. Recent reult in [16] introduced a cheme to model the primary uer activity. However, the author focu on the burty and piky traffic during hort-term activity fluctuation. Each SU i required to monitor the pectrum band and end the oberved ample to the bae tation. A firt-difference clutering and correlation cheme i ued to capture the hort fluctuation tranmiion opportunitie during the PU ON period, which can maximize the CR network performance. However, the idle fluctuation time during ON period i very hort and run a greater rik of interfering with PU tranmiion. Therefore, the hort-term fluctuation i not conidered in thi paper. In addition, we aume a CR mobile ad hoc network, in which there i no bae tation. CR ad hoc network contitute additional challenge in etimating PU activity and modeling on account of node mobility [2], not een in tatic environment, uch a [17]. In addition, many etimation cheme require coniderable memory (in cae a hitory of meaurement need to be tored), which i difficult given the local hardware capability of a ingle node [18]. Mot aforementioned previou work aume a priori knowledge of activity ditribution, and for uch cae, the practical channel uage tatitic are very different when region of overlap exit for the PU. The final deciion baed on uch tatic model in the overlapping region are likely to be wrong. Different from thi, in our propoed approach, we do not care about a ingle PU ditribution. Rather, we bae etimate of the channel availability uing current, real time meaurement. If more than one PU i preent in the region, the ditribution that we learn will be the combination of the two random variable that determine their individual on (or off) time. Without lo of generality and for the implicity, we conider SU cooperatively detect one PU in the CR network [19]. In thi paper, we make three important contribution for etimating the PU activity and tracking the PU activity change. We optimize the number of meaurement needed to accurately etimate the PU activity, which will be ued later to tweak the performance of pectrum ening algorithm. For thi, we define two activity detection duration. The fine ening phae occur in the early tage when the current etimate are totally abent or very coare. The normal ening phae i during the continuou operation of the network, and erve to detect any change in the exiting level of PU activity. Baed on mean quare error (MSE), we devie topping and retarting rule for the fine ening, which allow the SU to witch between thee two phae. We develop a cooperative weighted activity etimation cheme where SU hare their own locally obtained etimate of the PU activity with neighboring SU around the PU activity region. Thu, thee prior meaurement erve a a tarting point for the new SU, which further contribute toward refining the etimate. Our cheme take into account not only the previou activity etimate derived by the predeceor, but alo the number of ample and channel condition that may have contributed to error in calculation of thee earlier etimate. Finally, we alo propoe a ditributed torage mechanim where the ening data i aved in the local region of the PU, thereby reducing memory requirement of each SU. The ret of thi paper i organized a follow: In Section 2, the CR ytem model i preented. Our propoed mobile ening model i decribed in Section 3. Section 4 preent the numerical and imulation reult. Finally, Section 5 conclude the paper. 2. Sytem model In our approach, the PU tranmitter i operating on a licened pectrum band modeled a an ON OFF ource, alternating between ON (buy) and OFF (idle) period. Aume ON and OFF period follow exponential ditribution, with the repective average of a and b [5]. Moreover, a the activity of the PU tranmitter can change over time [2], a and b may alo exhibit long-term variation.

3 G. Ning et al. / Computer and Electrical Engineering 39 (213) Further, we define the PU tranmiion region, called A p, approximated a a circle around the PU tranmitter (the circular white dic around the PU tranmitter in Fig. 1), where the SU mut carefully chooe tranmiion opportunitie. The PU activity etimation become important in thi region A p. The ON and OFF ample period meaurement are gathered by the SU a it move through the activity region (ee path of SU2 and SU3), and on reaching the circular boundary, it dieminate it own etimate to the urrounding node via broadcat meage. We aume that there i a common control channel (CCC) for the information exchange between SU [21]. Conider the haded ring around the PU activity region, denoted by A, which repreent the area in which all SU ave and propagate the broadcat meage of the lat known activity etimate. Thu, a new SU, entering the haded ring A from outide at any point, can obtain the lat few etimate of the PU activity that were broadcat by it predeceor. In turn, thi new SU will update the etimate and pread thi knowledge when it exit the region A p at the other end of it path. In the mobile CR network, only mobile SU who enter the PU tranmiion region A p can ene the licened pectrum band and broadcat the latet detected average a and b. The extent of the broadcat i limited to the other SU within the ring A. The SU undertake two type of ample collection, which are fine ening and normal ening. If there i no information about PU tranmitter at all, or the MSE of detected average a and b i greater than a threhold, the ening cycle mut be hort and repeated often. Thi type of detection i called a fine ening. If SU ha coniderably accurate mean buy period and mean idle period, normal ening will be ued. In either cae, the final etimate will be broadcated at the boundary of the region A and A p. Conider that g SU have dieminated their detection packet including their detected mean etimate in the region A, which are repreented by a i ; b i ; i ¼ 1;...; g, and alo including the number of ON ample N i and the number of OFF ample M i that were ued for thi etimation. Aume that the network ha a memory of g, i.e., a new SU entering the region A can receive at mot g et of the tuple fa i ; b i ; N i ; M i g. The broadcat packet that dieminate the PU activity etimate contain the following field: packet time-tamp, a ening tatu bit ( or 1), final average ON period a, final average OFF period b, g (number of earlier SU who have broadcated the average detection reult and g 6 g ), g et of fa i ; b i ; N i ; M i g, the maximum number of ON period ample (out of total poible g), the actual number of combined ON period ample, all ON ample duration, the maximum number of OFF period ample, the actual number of combined OFF period ample, and all OFF ample duration. Here, packet time-tamp will be ued to avoid broadcat loop, a and b are computed by the g et of fa i ; b i ; N i ; M i g. If the ening tatu bit i et to 1, SU can perform normal ening. Otherwie, fine ening need to be undertaken a the etimate i not accurate. If a SU obtain very few ample, thee ample will be inerted into the packet. Once the combined ample become large, they will be ued to compute the average period, and then, they will be deleted from the packet. Generally, fale alarm and mi detection probabilitie are ued to analye average interference time and lot pectrum opportunity in the pectrum ening [7]. In thi paper, intead, we focu on the etimation of the ON/OFF average period and tracking the PU activity change, which precede the normal operation of the network. 3. Propoed mobile ening model In thi ection, we focu on how to chooe the duration of the fine and normal ening, how to proce and combine ample collected by SU, how to top and retart fine ening, and the haring of information among the SU Sening cycle and type The probability of the period ued by primary uer i give a [4], Fig. 1. Mobile CR ad hoc network.

4 178 G. Ning et al. / Computer and Electrical Engineering 39 (213) P on ¼ a a þ b And the probability of the idle period i P off ¼ b a þ b In addition, the lot pectrum opportunity ratio T L i defined to indicate the expected minimum fraction of the OFF tate undetected by SU, the maximum outage ratio T P i the maximum fraction of interference that primary network can tolerate. Aume that obervation time t in one ening cycle i fixed. When PU tranmitter i idle, tranmiion time in one ening cycle i denoted by t t, then the ening cycle T off in PU idle tate i expreed a T off ¼ t þ t t ð3þ where the maximum tranmiion time t t i bounded by [4] t t 6 l log 1 T P ; l ¼ minða; bþ P ð4þ off In previou work uch a [4,9], only one ening cycle i ued in PU idle and buy tate. However, when PU tranmitter i buy, a different ening cycle T on i defined in thi paper, compoed of obervation time t and quiet time t q, a follow: T on ¼ t þ t q ð5þ where the maximum quiet time t q i bounded by t q 6 l log 1 T L ; l ¼ minða; bþ P ð6þ on If there are everal licened channel, the quiet time t q in a ening cycle can be cheduled to detect other licened channel [7,14]. While our model can be eaily extended in thee cenario, the multiple-channel etimation i out of cope of thi paper. When a SU enter the range A p and receive no information about PU, the SU mut perform fine ening, in which the ening cycle of T off and T on are ame and fixed to a comparatively mall value T min. Thi allow frequent ampling of the environment by the SU in it initial entry into the PU activity region. Later, when the fine ening phae end (i.e., the PU tatitic are accurately known), T off and T on will be computed by Eq. (3) and (5), repectively, which i defined a normal ening. In dynamic environment, however, if the final average a and b deviate in accuracy from a and b beyond a certain threhold a indicated by the following MSE, the ening cycle will be reduced to T on c and T off c, where T on and T off are alo computed by Eq. (3) and (5), and c e (.5,1). Thi force the SU to re-tart the fine ening and quickly gather additional information about the environment Cae I. Etimation for ingle uer When SU i firt move into the region A p, it broadcat a requet meage on the CCC to other neighboring SU to inquire about the PU operational characteritic, namely, the average ON period and OFF period. The neighboring SU who have thi information will end back thi PU activity information, including the ening tatu bit and the lat g etimate, to the requeting SU. The SU will then elect the latet information baed on the packet time-tamp. According to the ening tatu bit, the SU can determine to perform either fine ening or normal ening. During ongoing activity etimation in the area A p,sui record each detected ON period ample and OFF period ample. Aume the number of ON period ample i N i and the number of OFF period ample i M i, which are the repective cardinality value of the et of {t on (1), t on (2),..., t on (N i )} and {t off (1), t off (2),..., t off (M i )} (Fig. 2). Then, the maximum likelihood etimator (MLE) i ued to calculate the average ON period a i and OFF period b i [22,23]. The PU buy time and idle time can be modeled by the exponential ditribution. The probability denity function (PDF) of ON tate i ð1þ ð2þ ON OFF ON OFF ON OFF ON on T t on (1) on T off T off T t off (1) t on (2) t off (2) t on (N i ) t off (M i ) Fig. 2. SU i detect the PU ON/OFF period.

5 G. Ning et al. / Computer and Electrical Engineering 39 (213) pðt on ; aþ ¼ 1 a e ton a ð7þ Then, the likelihood function i Lðt on ; aþ ¼ YN i 1 a e j¼1 tonðjþ a ¼ a N i e X N i 1 a t onðjþ j¼1 ¼ a N i e N i Ton a ð8þ According to MLE algorithm, n d ln LðT on ; aþ ¼ da d N i ln a þ N it on a da Then the optimal mean buy period a i of MLE can be calculated by, o ¼ N i a þ N it on a 2 ¼ ð9þ a i ¼ T on ¼ 1 X N i t on ðjþ N i j¼1 ð1þ Similarly, the optimal mean idle period b i i, b i ¼ T off ¼ 1 X M i t off ðjþ M i j¼1 ð11þ where jn i M i j¼or1. Intuitively, the number of ON and OFF ample are related to the length of PU true a and b, the SU peed and the travel ditance in the area A p.ifsui obtain only a few ample, a i and b i may uffer from coniderable deviation from the true a and b. Therefore, if the et of ample i le than a threhold number N t, SU can broadcat thee ample intead of the average value alone, i.e., fa i ; b i ; N i ; M i g Cae II. Etimation for multiple uer Aume that there are g mobile SU broadcating their individual etimation of PU activity. Generally, the arithmetic mean approach, called a non-weighted average in the paper, can be ued to etimate the final a and b with Eq. (1) and (11) in thi cae. However, ince the number of ON and OFF period ample of variou SU hould be different, we introduce a weighted average approach to compute the final a and b for multiple SU, a follow: a ¼ Xg i¼1 N i a i P g N j¼1 j ð12þ b ¼ Xg i¼1 M i b i P g M j¼1 j ð13þ We aign a weight to the final etimation made by a SU a proportional to the number of ON and OFF ample it ue for the calculation. Moreover, due to the dynamic nature of PU activity and the limited memory, the network only tore the latet g et of fa i ; b i ; N i ; M i g. When PU activity or SU mobility peed i high, then the number of collected ample for the etimation are few. In uch a cae, multiple et of ample from different SU who are paing through the activity region in near-overlapping time can be combined to compoe a ingle et. Thu, if the ith SU get a few ample, then thee ON and OFF ample will be added into the broadcat packet intead of the tatitical value fa i ; b i ; N i ; M i g. After the next SU, i.e., the i + 1th SU, finihe it ening, it own ample are cumulatively conidered for the analyi, along with the ample carried in the previou broadcat packet of the ith SU. If the total number of ample now i larger than the threhold number N t, then the i + 1th SU calculate the average ON and OFF period to generate the average et of fa i ; b i ; N i ; M i g. Henceforth, only thee average are included in the broadcat packet, in place of the actual ample value. Converely, if the total number of ample i le than the threhold number N t, then the new ample of i + 1th SU are alo added into the broadcat packet Chooing tart and top condition for fine ening In thi ection, we eparately derive the tart and top condition that allow the SU to witch between fine ening and normal ening.

6 171 G. Ning et al. / Computer and Electrical Engineering 39 (213) Stop condition for fine ening The fine ening can be topped after the etimated mean of the PU activity converge to the true mean ON/OFF time. Thi implie that with progreive round of broadcat of the etimated value and continued refinement undertaken by ubequent SU, the MSE of the ample will decreae. We need to chooe a MSE threhold E m for topping the fine ening proce (which i time conuming) and conequently, increae the tranmiion efficiency. Aume there are g(6g ) et of fa i ; b i ; N i ; M i g, the MSE r on of the mean ON period can be computed by ffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi 1 X g r on ¼ g ½ a i¼1 i Eða i ÞŠ 2 ð14þ where Eða i Þ i the latet average buy period computed by Eq. (12). Similarly, we can derive the MSE r off of the OFF period, which i given by ffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi 1 X g r off ¼ g ½ b i¼1 i Eðb i ÞŠ 2 ð15þ Only when both r on and r off are le than E m, the fine ening terminate and the ening tatu bit will be et to 1 ignaling the end of the fine ening proce Start condition for fine ening PU activity i dynamic and varie over time, poibly even with the time of the day. Thi require periodically reviing the activity etimation. A a and b vary over time, our approach i able to detect the change and retart the fine ening. A trivial option here i to ue the MSE for detecting the change in PU activity, i.e., MSE will increae if there i a change in PU activity. However, thi technique doe not work efficiently in low-varying environment [24]. Thu, in order to capture the change of PU pectrum uage efficiently and quickly, we introduce a new variable r that exploit the variation of MSE. Here, the current r on (t) and r off (t) will be compared with the previou MSE value of r on (t 1) and r off (t 1). We define ron ðtþ a: ron ðtþ ¼ jr onðtþ r on ðt 1Þj r on ðt 1Þ Similarly, roff ðtþ can be derived a: roff ðtþ ¼ jr off ðtþ r off ðt 1Þj r off ðt 1Þ If ron ðtþ or roff ðtþ i larger than the threhold S TH, the fine ening will be retarted. Once the fine ening top, the etimated a and b will not be changed until the next round of the fine ening phae i triggered. When a SU enter the PU tranmiion range A p, it perform normal ening becaue the ening tatu bit i et to 1. Note that r on (t 1) and r off (t 1) can be calculated through the g et of fa i ; b i ; N i ; M i g. After moving out from the range A p, the oldet et of fa 1 ; b 1 ; N 1 ; M 1 g will be dicarded if g = g (i.e., the limit on the ample hitory i reached), and the SU own etimated value will be tored a the latet et. Thu, the current r on (t) and r off (t) can be calculated followed by ron ðtþ and roff ðtþ from Eq. (16) and (17). If ron ðtþ or roff ðtþ i higher than the threhold S TH, the ening tatu bit will be et to. In addition, only the current ening reult will be broadcated by the SU. Other SU who receive thi broadcat packet will be aware of the variation of PU activity, thereby individually re-tarting their own fine ening. Finally, in cae that SU cannot obtain any information from neighboring SU, it then perform fine ening with ening by Eq. (3) and (5), repectively. The cone- cycle T min. Otherwie, it ue a and b to calculate the ening cycle T on quent ening procedure can be decribed a follow: and T off ð16þ ð17þ (1) If ening tatu bit i, the SU will perform fine ening with T on c and T off c. Once it move out of A p, the new r on and r off will be computed. If r on and r off are le than E m, then ening tatu bit will be et to 1. (2) On the contrary, if ening tatu bit i 1, the SU will perform normal ening. After it exit the range A p, ron and roff will be calculated and compared with threhold S TH.If ron or roff i larger than the threhold S TH, the fine ening will be retarted. The entire mobile cooperative ening procedure i illutrated in the flowchart in Fig Numerical and imulation reult In thi ection, we conduct imulation tudie to evaluate the performance of our propoed PU activity etimation algorithm. In the mobile CR network, there i only one PU operating on a licened pectrum band. The location of the PU tranmitter i at the center coordinate of (2 m, 2 m). The tranmiion radiu of PU tranmitter i 8 m, which cover the region A p while the broadcating area of SU i within the ring A, which pan 12 m further from A p boundary. The peed

7 G. Ning et al. / Computer and Electrical Engineering 39 (213) A SU enter Ha PU info Y Normal detection N Ak info Fine detection N >STH σ Obtain feedback info N MSE<Em N Y Y Y Sening tatu bit= Sening tatu bit=1 N Sening tatu bit=1 Y Broadcat detected info Fig. 3. Sening flowchart for one ingle SU. of mobile SU i uniformly ditributed from 3 m/ to 2 m/, and the moving direction i alo uniformly ditributed in [,2p). The random walk model wa adopted to model the movement of SU. A per the IEEE tandard [25], the obervation time t for one time ening i le than 2 m when uing energy ening method. Conequently, t i et to 2 m, and the minimum fine ening cycle T min i et to 8 m. In order to examine the protocol convergence with regard to the number of ample obtained by individual SU detection, we gradually injected 2 SU into region A p ucceively to detect the PU tranmitter with a =1,b =.5. A a reult, there were 5 SU that obtained more than 1 ample. From Fig. 4, we oberve that when the number of ON period ample i more than 3, the mean ON period of all the 5 SU fluctuate around the true mean ON period by 15%. However, when the number of ample i le than 3, the average ON period a i ha a large deviation from the true a. Therefore, in a practical application, when a SU obtain few ample, thee ample can be broadcat allowing u to et the ample threhold number N t a 3. Moreover, when a SU ha no information about the PU, it can utilize the hortet ening cycle T min for detection. After it obtain more than a certain number of ample, the ening cycle can be increaed to enhance the ening efficiency. The cooperative ening efficiency of multiple SU can be demontrated by letting 5 SU move into the PU tranmiion range A p ucceively. Thi implie that the number of ample collected by individual SU may differ a well. Two previouly mentioned average approache, non-weighted average and weighted average, were employed to evaluate the performance. In the beginning of thi imulation, the firt SU ue fine ening with minimum ening cycle T min. Once the number of ample reache 3, the ening cycle will be et to T on c and T off c, where c =.8. It i clear from Fig. 5 that when more than 1 SU ue fine ening to ene the licened band, the average ON period a i very cloe to the accurate average ON period, which i a =. Additionally, the weighted average approach obviouly outperform the non-weighted average approach throughout the range. It i alo crucial to evaluate the performance of MSE and the relative error in our approach. We define relative error a ja aj=a. A hown in Fig. 6, MSE r on decreae a more SU ene the PU tranmitter and it aymptotically converge to the true mean. Additionally, the relative error alo exhibit the imilar behavior. However, the fluctuation of relative error i much larger than that of MSE, and SU cannot determine the true a when uing relative error. The r on i almot equal to 4% when there are more than 25 SU, which indicate that not all the SU detected information fa i ; b i ; N i ; M i g are required to be tored and broadcated. Thi i another reaon why we introduce g which i the maximum number of fa i ; b i ; N i ; M i g tored.

8 1712 G. Ning et al. / Computer and Electrical Engineering 39 (213) True α=1 Mean value trace for each SU Mean ON period(unit: econd) Number of ON ample Fig. 4. Convergence of the etimated mean ON period of five individual SU (a =1,b =.5 ). Detected average ON period (unit: econd) Accurate average ON period Non-weighted average ON period Weighted average ON period Number of SU Fig. 5. Multiple SU detect PU (a =, b =.5 ). After the fine ening i completed by 2 SU, the final mean ON period and mean OFF period are ued to calculate the ening cycle for the ubequent normal ening. A hown in Fig. 7, it i obviou that baed on the a and b etimated by our propoed ening model, both the maximum outage ratio T P and lot pectrum opportunity ratio T L are below the predefined threhold.4 after a hort period of operation. We next evaluate the detection enitivity. Here g i et to 2. We let 1 SU ucceively detect the PU activity. After the 5th SU finihe it detection, a varie from to.5 and.6. On the contrary, b i contant at.5. A depicted in Fig. 8, MSE of r on i almot contant when the number of SU increae from 2 SU to 5 SU. However, r on tart increaing lowly from the 51th SU becaue of the change of PU activity. Thi i becaue when the 51th SU finihe it detection, the etimated average ON period of thi SU i cloe to the current real average ON time.5 or.6, while other etimated ON period ued converge to. Hence, the impact of thi pecific SU on the etimate a i light, and the final etimated ON period a i till dominated by the previou SU, before the 51th SU. Conequently, if MSE i ued to retart a new fine ening procedure, it will have a coniderable delay and caue interference with PU. It hould be pointed out that the value of r on (5) and r on (51) are very mall, but the increment of r on (51) r on (5) i in the ame order of magnitude and i le than r on (5) and r on (51). On the contrary, the ratio of (r on (51) r on (5))/r on (5) i large. From Fig. 8, we oberve that the value of S ron ð51þ i more than 35%. A a reult, the variation of PU activity can be quickly detected by the 51th SU, which can be ued to retart fine ening within a very hort time. In addition, we can find that when the activity of PU doe not change, r (t) fluctuate within acceptable bound, which i a direct reult of combining multiple ample and reducing the udden impact of an outlier meaurement. Finally, from the 52th SU, fine ening i retarted. Therefore, alternatively we ue r (t)to monitor the PU activity ince it i more enitive to the change in PU activity pattern.

9 G. Ning et al. / Computer and Electrical Engineering 39 (213) Mean quare error Relative error Error of average ON period Number of SU Fig. 6. Multiple SU detect PU (a =, b =.5 )..1 Ratio of interference / lot tranmiion Predefined T P and T L Actual T P Actual T L Simulation time (unit: econd) Fig. 7. Interference ratio and lot communication ratio with T P = T L =.4. In [14], the author utilized moving time window to track the variation of PU activity. The latet ON and OFF period ample will be collected within a time window and the etimation procedure will be executed once every a period of time. In order to further demontrate that our propoed cheme can quickly track the change of PU activity, we compare it with the cheme in [14]. Our propoed cheme i called Scheme 1, the track method in [14] i called Scheme 2. In our propoed cheme, the maximum number of SU that pa through the PU activity region and cooperate to etimate the final a and b i g. Hence, we et the ize of the moving window here i not the individual ample, but the number of collaborating SU at a given time, i.e., g = 2. However, the total number of SU that pa through the PU activity region from the tart to the end of the imulation i 1. Further, we explore the repone time of the ytem when a varie from to.5 halfway through the imulation (i.e., when the 5th SU finihe it detection). From Fig. 9, we can find that our propoed Scheme 1 can quickly track the change of PU activity. Thi i becaue r (t) i ued to retart the fine ening, and the earlier g etimation are dicarded. However, the tracking peed of Scheme 2 i much lower, a the outdated etimation from the 32th to 5th SU are till ued to compute the final average a when the 51th SU finihe it own ening. In addition, from the 7th SU, 2 ON etimation of Scheme 2 converge to.5 ame a thoe of Scheme 1, o final average a of two cheme converge to.5. When PU activity doe not change, i.e., a =, ame method i ued to calculate a in Scheme 1 and Scheme 2, and therefore the curve are overlapping. Hence, we conclude that our method provide enhanced reponivene of the ytem when PU activity i ubject to change. The broadcat overhead i a concern in collaborative CR ening. If all the raw ening data, i.e., ON/OFF period ample, are broadcat between the SU, then the broadcat overhead will be large. Therefore, the packet tructure in Section 2 i ued, which relie only on ending the average value. By uing the tatitical information in place of the actual raw ample, the

10 1714 G. Ning et al. / Computer and Electrical Engineering 39 (213) α=.5.3 α=.6 MSE-σ οn.2.1 MSE-σ οn Number of SU Number of SU σ of ON period α=.5 σ of ON period α= Number of SU Number of SU Fig. 8. MSE and it variation when a change (b =.5 )..55 Detected average ON period (unit: econd).5 5 Accurate average ON period Average ON period of Scheme 1 Average ON period of Scheme Number of SU Fig. 9. Track peed comparion of PU activity (a =?.5, b =.5 ). Average broadcated data(unit: kbit/) Number of SU=5 Number of SU=6 Number of SU= E m for topping fine detection Fig. 1. Broadcating overhead (a =, b =.5 ).

11 G. Ning et al. / Computer and Electrical Engineering 39 (213) overhead i contained. In order to evaluate the overhead incurred by broadcat, we deployed varying number of SU uniformly in the broadcating region A at the beginning of imulation. In addition, we aume that each SU in A will broadcat the latet received detection reult. Once r on and r off reach the threhold E m, the overhead, defined a the total amount of broadcated data divided by the correponding required time, will be computed. A hown in Fig. 1, with the increae in detection preciion, reduction in E m from.12 to.3, the average broadcated data per econd increae. Thi i becaue more time and large number of SU are required to yield the requeted detection preciion. Additionally, the broadcat overhead increae when the total number of SU increae from 5 to 7. Thi implie that more SU enter PU tranmiion region A p, a more detection information need to be dieminated. 5. Concluion In thi paper, we preented and evaluated a novel licened uer activity etimation algorithm to detect the average buy period and idle period of PU in a mobile CR ad hoc network. The MLE and weighted average method were ued to proce the ening reult, and the MSE wa ued to decide top and retart intant of fine ening. Our approach demontrated efficient tracking of the change of PU activity, reulting in convergence to the true activity with an acceptable overhead. The imulation reult indicate that the propoed algorithm can accurately etimate the PU pectrum uage pattern by uing the optimal number of meaurement and through the cooperation of multiple SU. Moreover, our propoed mobile cooperative ening approach can be eaily extended for tatic SU and centralized CR network. Acknowledgment The author would like to thank the anonymou reviewer for their valuable comment and the editor for their time pent in handling the paper. Thi work i upported by Natural Science Foundation of Hubei Province (No. 211CDB164) and Self-determined Reearch Fund of CCNU from the college baic reearch of MOE (No. CCNU9A17). Reference [1] Haykin S. Cognitive radio: brain-empowered wirele communication. IEEE J Select Area Commun 25;23(2):21 2. [2] Akyildiz IF, Lee W, Chowdhury KR. CRAHN: cognitive radio ad hoc network. Ad Hoc Netw 29;7: [3] Cormio C, Chowdhury KR. A urvey on MAC protocol for cognitive radio network. Elevier J Ad Hoc Netw 29;7(7): [4] Song C, Zhang Q. Cooperative pectrum ening with multi-channel coordination in cognitive radio network. In: Proc. IEEE ICC; 21. p [5] Zhao Q, Tong L, Swami A, Chen Y. Decentralized cognitive MAC for opportunitic pectrum acce in ad hoc network: a POMDP framework. IEEE J Select Area Commun 27;5(3): [6] Lee WY, Akyildiz IF. Joint pectrum and power allocation for inter-cell pectrum haring in cognitive radio network. In: Proc IEEE DySPAN; 28. p [7] Lee WY, Akyildiz IF. Optimal pectrum ening framework for cognitive radio network. IEEE Tran Wirele Commun 28;7(1): [8] Chowdhury KR, Felice MD. Search: a routing protocol for mobile cognitive radio ad-hoc network. Comput Commun 29;32(18): [9] Huang S, Liu X, Ding Z. Opportunitic pectrum acce in cognitive radio network. In: Proc IEEE INFOCOM; April 28. p [1] Felice MD, Chowdhury KR, Melei W, Bononi L. To ene or to tranmit: a learning-baed pectrum management cheme for cognitive radio meh network. In: Proc IEEE WIMESH; 21. p [11] Yang L, Cao L, Zheng H. Proactive channel acce in dynamic pectrum network. In: Proc IEEE CrownCom, Orlando, FL, USA; Augut 27. p [12] Zarrin S, Lim TJ. Throughput-ening tradeoff of cognitive radio network baed on quicket ening. In: Proc IEEE ICC; 211. [13] Sung KW, Kim SL, Zander J. Temporal pectrum haring baed on primary uer activity prediction. IEEE Tran Wirele Commun 211;9(12): [14] Kim H, Shin KG. Efficient dicovery of pectrum opportunitie with MAC-layer ening in cognitive radio network. IEEE Tran Mobile Comput 28;7(5): [15] FCC Pre releae; 211. < [16] Canberk B, Akyildiz IF, Oktug S. Primary uer activity modeling uing firt-difference filter clutering and correlation in cognitive radio network. IEEE/ ACM Tran Netw 211;19(1): [17] Miao M, Tang DHK. Impact of channel heterogeneity on pectrum haring in cognitive radio network. In: Proc IEEE ICC; 28. p [18] Park J, Van Der Schaar M. Cognitive MAC protocol uing memory for ditributed pectrum haring under limited pectrum ening. IEEE Tan Commun 211;59(9): [19] Peh ECY, Liang YC, Guan YL, Zeng YH. Cooperative pectrum ening in cognitive radio network with weighted deciion fuion cheme. IEEE Tran Wirele Commun 21;9(12): [2] Purley MB, Royter TC. Low-complexity adaptive tranmiion for cognitive radio in dynamic pectrum acce network. IEEE J Select Area Commun 28;26(1): [21] Chowdhury KR, Akyildiz IF. OFDM baed common control channel deign for cognitive radio ad hoc network. IEEE Tran Mobile Comput 211;1(2): [22] Kay StevenM. Fundamental of tatitical ignal proceing: etimation theory. Prentice Hall; p [23] Myung IJ. Tutorial on maximum likelihood etimation. J Math Pychol 23;47:9 1. [24] Sanna M, Murroni M. Nonconvex optimization of collaborative multiband pectrum ening for cognitive radio with genetic algorithm. Intl. J Digital Multimedia Broadcat 21:12. [25] Cordeiro C, Challapali K, Birru D, Shankar S. IEEE 82.22: the firt worldwide wirele tandard baed on cognitive radio. In: Proc IEEE DySPAN; 25. p Guoqin Ning i an aociate profeor with the Department of Information and Technology, Central China Normal Univerity, Wuhan, China. He received the Ph.D. degree from Huazhong Univerity of Science and Technology in 26. He wa a viiting cholar at Northeatern Univerity, Boton, USA, from 21 to 211. Hi reearch interet are cognitive radio network and radio reource management. Kauhik R. Chowdhury received the M.S. degree in computer cience from Univerity of Cincinnati, in 26, and the Ph.D. degree from the Georgia Intitute of Technology, Atlanta in 29. He i an aitant profeor in the Electrical and Computer Engineering Department at Northeatern Univerity, Boton, USA.

12 1716 G. Ning et al. / Computer and Electrical Engineering 39 (213) Hi reearch interet are cognitive radio network, energy harveting, multimedia communication over enor network. Jiaqi Duan received the B.S. degree and M.S. degree from Northwetern Polytechnical Univerity, Xi an, China, in 26 and 28, repectively. He i puruing Ph.D. degree in School of Electronic and Information, Northwetern Polytechnical Univerity. He wa a viiting cholar at Northeatern Univerity, Boton, USA, from 21 to 211. Hi reearch interet include cognitive radio network and communication ignal proceing. Pruayon Nintanavonga received the M.S. degree in electrical engineering from Boton Univerity, USA, in 26. He i currently working toward the Ph.D. degree in the Electrical and Computer Engineering Department at Northeatern Univerity, Boton, USA. Hi reearch interet include energy harveting enor network, cognitive radio network, radio frequency integrated circuit deign and power management in low-power enor network.