Physical layer metrics for vertical handover toward OFDM-based networks

Transcription

1 RESEARCH Open Access Physica ayer metrics for vertica handover toward OFDM-based networks Mohamed Rabie Ouarbi *, Francois-Xavier Socheeau, Sebastien Houcke and Abdedjai Aïssa-E-Bey Abstract The emerging trend to provide users with ubiquitous seamess wireess access eads to the deveopment of mutimode terminas abe to smarty switch between heterogeneous wireess networks. This switching process known as vertica handover requires the termina to first measure various network metrics reevant to decide whether to trigger a vertica handover (VHO) or not. This paper focuses on current and next-generation networks that rey on an OFDM physica ayer with either a CSMA/CA or an OFDMA mutipe-access technique. Synthesis of severa signa feature estimators is presented in a unified way in order to propose a set of compementary metrics (SNR, channe occupancy rate, coision rate) reevant as inputs of vertica handover decision agorithms. A the proposed estimators are non-data aided and ony rey on a physica ayer processing so that they do not require mutimode terminas to be first connected to the handover candidate networks. Resuts based on a detaied performance study are presented to demonstrate the efficiency of the proposed agorithms. In addition, some experimenta resuts have been performed on a RF patform to vaidate one of the proposed approaches on rea signas. 1 Introduction Nowadays, we are facing a wide depoyment of wireess networks such as 3G (LTE), WiMAX, Wifi, etc. These networks use different radio access technoogies and communication protocos and beong to different administrative domains; their coexistence makes the radio environment heterogeneous. In such environment, one possibe approach to overcome the spectrum scarcity is to deveop mutimode terminas abe to smarty switch from one wireess interface to another whie maintaining IP or voice connectivity and required quaity of service (QoS). This switching processisknownasvertica handover or vertica handoff. This new concept wi not ony provide the user with a great fexibiity for network access and connectivity but aso generate the chaenging probem of mobiity support among different networks. Users wi expect to continue their connections without any disruption when they move from one network to another. The vertica handover process can be divided into three main steps [1,2], namey system discovery, handoff decision, and handoff execution. During the system * Correspondence: mohamed.ouarbi@teecom-bretagne.eu Institut Téécom, Téécom Bretagne, UMR CNRS 3192 Lab-STICC Université Europenne de Bretagne, Brest, France discovery step, the mobie terminas equipped with mutipe interfaces have to determine which networks can be used and the services avaiabe in each network. These wireess networks may aso advertise the supported data rates for different services. During the handoff decision step, the mobie device determines which network it shoud connect to. The decision may depend on various parameters or handoff metrics incuding the avaiabe bandwidth, deay, jitter, access cost, transmit power, current battery status of the mobie device, and even the user s preferences. Finay, during the handoff execution step, the connections need to be re-routed from the existing network to the new network in a seamess manner [3]. Cognitive radio appears as a highy promising soution to this combined probems. Cognitive radio systems can sense their RF environment and react, either proactivey or reactivey, to externa stimui [4-7]. By the term react, it is impied that the systems have the abiity to reconfigure the agorithms and its communication parameters to better adapt to environment conditions. Thus, in principe, the operation of a cognitive radio system incudes two stages: sense and decide [8]. This paper focuses on the sensing task. Indeed, we dea with the passive estimation of metrics that hep to 211 Ouarbi et a; icensee Springer. This is an Open Access artice distributed under the terms of the Creative Commons Attribution License ( which permits unrestricted use, distribution, and reproduction in any medium, provided the origina work is propery cited.

2 Page 2 of 25 trigger a vertica handover toward OFDM -based systems such as WiFi, WiMAX, or 3G(LTE). It shoud be noted that the decision step and the handoff execution are not treated in this paper. These tasks may need interaction with the higher ayers to guarantee a seamess and proactive vertica handover, which is beyond the scope of this paper. In the context of vertica handover, ony the passive estimation is reevant since the termina seeks to know a priori whether a network satisfies its QoS needs without wasting time and power to get connected to this network. The main contribution of this work reies on the fact that a the proposed metrics are estimated from the physica ayer signa and require no connection to the system, no signa demoduation, and no frame decoding. To the best of our knowedge, various VHO decision agorithms based on a MAC-ayer sensing have been proposed [1,2,9-12], but none have been investigated on the PHY ayer. Three reevant and compementary metrics are presented. First, we propose a method to estimate the downink signa-to-noise ratio (SNR). The SNR is an indicator commony used to evauate the quaity of a communication ink. The proposed method expoits the correation as we as the cycostationarity induced by the OFDM cycic prefix (CP) to estimate the noise as we as the signa power of OFDM signas transmitted through unknown muti-path fading channe. In addition to the downink signa quaity, some knowedge on the traffic activity can be very informative since it is a good indicator of the network oad. Measures of traffic activity strongy depend on the medium access technique of the sensed network. Today, OFDM wireess networks rey either on CSMA/CA (carrier sense mutipe-access/ coision avoidance), see Wifi networks for instance, or on OFDMA (orthogona frequency division mutipe access), see WiMAX and 3G(LTE). Concerning the CSMA/CA protoco, we propose to estimate the channe occupancy rate (combined upink and downink) and the upink coision rate, which are two reevant metrics of network oad. These metrics can be estimated at the signa eve providing that the termina is equipped of severa receiving antennas. For the OFDMA access techniques, the network traffic is estimated through the downink time-frequency activity rate of the channe. Since OFDMA networks use either synchronous time division dupexing or frequency division dupexing, no coision occurs so that the coision rate metric is irreevant a. The rest of the paper is organized as foows: First, we dea with metrics dedicated to CSMA/CA-based networks. In Section 2.1, we present a SNR estimator dedicated to OFDM-based physica ayers. Section 2.2 describes the proposed agorithms to estimate the channe occupancy rate of a CSMA/CA-based network. A first agorithm is presented in Section Then, due to some imitations of the atter, in Section 2.2.5, we propose a second agorithm based on a Parzen estimator, which shown its robustness thanks to simuations. As a compementary metric, in the congested networks, we propose to estimate the channe occupancy rate. The agorithm is derived in Section 2.3, for channes with different engths on the antennas. Section 3 deas with OFDMA-based systems. In Section 3.1, we show how the proposed SNR estimator can aso be appied for OFDMA-based systems, and in Section 3.2, we describe the proposed agorithm for the estimation of the time-frequency activity rate of OFDMA signas. A proposed architecture of the receiver, based on software-defined radio is described in Section 4. A the proposed agorithms are evauated thanks to computer simuationsinsection5.inaddition, some experimenta resuts for the channe occupancy rate are aso presented in this Section These resuts are presented for the first time; many scenarios have been driven to show how the channe occupancy rate is informative about the QoS avaiabe in a sensed networks. Furthermore, thanks to these experimentations, we are now abe to say that for the case of congested networks, the channe occupancy rate itsef is not sufficient enough to decide whether to trigger the handover or not and that the coision rate is a necessary compementary metric. Finay, we outine some concusions in Section 6. 2 Metrics for CSMA/CA based networks CSMA/CA is a protoco for carrier transmission in some wireess networks. Unike CSMA/CD (carrier sense mutipe-access/coision detect), which deas with transmissions after a coision has occurred, CSMA/CA acts to prevent coisions before they happen. In CSMA/CA, as soon as a node receives a packet to be sent, it checks whether the channe is ide (no other node is transmitting at the time). If the channe is sensed ide, then the node is permitted to begin the transmission process. If the channe is sensed as busy, the node defers its transmission for a random period of time caed backoff. If the channe is ide when the backoff counter reaches zero, the node transmits the packet. If the channe is occupied when the backoff counter reaches zero, the backoff factor is set again, and the process is repeated. In this section, we dea with CSMA/CA networks whose physica ayer is based on the OFDM moduation scheme. First, we present an agorithm for SNR estimation, then we propose a method for estimating the channe occupancy rate and finay a coision rate estimator is detaied.

3 Page 3 of OFDM signas SNR estimation SNR is an important metric that indicates the ink quaity. We propose a bind estimation approach, based on the correation and the cycostationarity induced by the OFDM CP. Assuming that an OFDM symbo consists of N sc subcarriers, the discrete-time baseband equivaent transmitted signa is given by Ms 1 E s Nsc 1 n x(m) = a k,n e 2iπ (m D k(nsc+d)) N sc g(m k(nsc + D)). (1) N sc k= n= where M s denotes the number of OFDM symbos in the observation window, E s is the average avaiabe power, and a k, n are the transmitted data symbos at the nth subcarrier of the kth OFDM bock. These data symbos are assumed to be independent identicay distributed (i.i.d), D is the cycic prefix (CP) ength, and m g (m) is the puse shaping fiter. Let {h()} =,..., L-1 be a baseband equivaent discretetime Rayeigh fading channe impuse response of ength L with L<D. The received sampes of the OFDM signa are then expressed as L 1 y(m) = h()x(m )+w(m), (2) = where w(m) is an additive white Gaussian noise such that w(m) CN (, σ 2 w). The signa-to-noise ratio (SNR) is expressed as SNR = S σw 2, (3) L 1 S = E s E[ a k,n 2 ] σh() 2. (4) = where E[.] stands for the expectation operator. To get the SNR, first we have to estimate the noise power σ 2 w, and then, the power of the received signa S Noise power estimation To estimate the noise variance, we propose to take advantage of OFDM signas structure. More precisey, redundancy was induced by the CP; in fact, the CP eads to x(k(n sc + D) + m) = x(k(n sc + D) + N sc + m), k Z, and m Î {,..., D-1}. Assuming a perfect synchronization and a time-invariant channe over an OFDM symbo duration, we can get D-Lnoise variance estimates defined as M ˆσ w,u 2 = 1 s 1 D 1 y(k(n sc + D)+m) 2M s (D u) k= m=u (5) y(k(n sc + D) + N sc + m) 2, L u D 1. The estimator with the smaest variance is found for u = L. The difficuty is then to estimate L. In[13],we proposed an estimator of L inspired from maximum ikeihood estimation. This estimator has the major advantage of being independent of any threshod eve and shows good performance compared to the threshod-based technique proposed in [14]. Here presented method has a computationa compexity (C.C) of O(M s.d 2 ) Signa power estimation We here propose to use the cycostationary statistics induced by the CP [15] to estimate the signa power. A signa power estimate can be given by Nc 1 Ŝ = sin(πqα ) 2N c +1 y (N ) α sin(πqα D) eiπqα(d 1), (6) q= N c ˆR qα where α =1/(N sc + D) and ˆR qα y (N sc )= Ms (N sc +D) 1 m= y(m)y (m + N sc )e 2iπmqα. M s (N sc + D) N c represents the number of considered cyce frequencies to estimate the signa power. The choice of N c is a trade-off between the estimator bias and variance. In [13], we show that we must choose qa within the coherence bandwidth of the channe B c. As the channe impuse response is unknown at reception, B c is approximated as ˆB c =1/(ρˆL) where r is a coefficient expressing the desired correation rate within B c. Consequenty, we choose N c = min, N ) ( Nsc + D sc.asshown ρ ˆL 2D in [13], r s choice has ony a very itte infuence on the estimator performance. The signa power C.C is estimated to be O(N c M s (N sc + D)). OFDM synchronization can be performed in a nondata-aided context by the mean of agorithms such as [16] and [17] for instance. The compexity of these agorithms is O(M s.(n sc + D).D) for [16] and O(M s.(n sc + D).D 2 ) for [17]. Miss-synchronization ony impacts the noise variance estimator and has the foowing effects. If the symbo synchronization is not we performed, signa sampes may be incuded in the noise variance estimator, eading to an overestimation of the noise variance. If the carrier frequency offset is not we mitigated, the phase of y(k(n sc + D) + m) and y(k(n sc + D) + N sc + m) wi be different so that the redundancy induced by the CP wi not be we expoited, eading once again to an overestimation of the noise variance. To put it in a nutshe, both events wi ead to an underestimation of the signa-to-noise ratio, which is not so dramatic for the vertica handover process. Indeed, underestimating the SNR and not

4 Page 4 of 25 connecting to the access point are much better than overestimating it, and then we find that the QoS does not satisfy our needs and wasting time again finding other potentia candidates. We point out that the method presented in [14], as our method, aso requires a perfect time-frequency synchronization. 2.2 Channe occupancy rate estimation In [12,18], it has been highighted that the usage of the channe bandwidth in a CSMA/CA system such as WiFi can be approximated as the ratio between the time in which the channe status is busy according to the NAV (network aocation vector) settings and the considered time interva. Indeed, prior to transmitting a frame, a station computes the amount of time necessary to send the frame based on the frame s engthand data rate. This vaue is paced in the duration fied in theheaderoftheframe.byreadingthisfie,wehave access to the traffic oad. The higher the traffic, the arger the NAV busy occupation, and vice versa. Then, once we read a NAV vaue during a certain time window, the avaiabe bandwidth and access deay can be estimated given a certain packet ength [19]. The main drawback with this method is that it requires to be connected to the access point in order to have access to the NAV duration from the header. This may increase the decision time if many standards or access points (AP) are detected. In this section, we propose a method that requires no connection to the AP and no NAV duration reading. This method [2] is based on a physica ayer sensing: Considering that the medium is free when ony noise is observed and occupied when signa pus noise sampes are observed (data frame), we use a ikeihood function that can distinguish the signa pus noise sampes from the one corresponding to noise ony. Once we get the number of signa pus noise sampes, a simpe ratio processing provides the network occupancy rate Mode structure In this section, we assume that CSMA/CA-based access points are detected. Between two consecutive frames we have different inter frame spacing (IFS) intervas, which guarantee different types of priority. At the receiver side, the observed signa is a succession of frames of noise sampes corresponding to the IFS intervas or ide periods and of data frames (Figure 1). For carity reason, we assume in this section that we have ony one data frame in the observation duration (N s sampes), and Section expains the proposed agorithm to ocate it. Consider that our receiver is doted of N antennas b, and et y i =[y i (1),..., y i (N s )] be a set of N s observations on the ith antenna such that y i (m) =w i (m) 1 m m 1 1 y i (m) = Li 1 = h i()x(m m 1 )+w i (m) m 1 m m 2 (7) y i (m) =w i (m) m 2 +1 m N s where the x(m) is an OFDM source signa expressed as in (1), h i () is the channe response from source signa to the ith antenna, and L i is the order of the channe h i. The process w i (m) is a compex additive white Gaussian noise with zero mean and variance σw 2. The variance σ w 2 is assumed to be known or at east estimated by a subspace-based agorithm [21], where mutipe antennas at reception are required Frame ocaization As presented in the previous section, the vector y i can be divided into three parts: noise, signa pus noise, and noise. Starting from the set of observation y i,wewoud ike to find which sampes correspond to noise and which ones correspond to signa pus noise. This probem is a cassica signa detection probem. Signa detection theory is a we-known probem in signa processing. This probem deas with the detectabiity of signas from noise. Many works have been done in this fied, and a arge iterature exists ([22-24],...). A maximum a posteriori testing, a Bayes criterion, a Neyman Pearson, or an energy detector [25] can be used. Here, we use another approach, since the sampes are supposed to be independent in the noise areas and correated in the signa pus noise area due to the channe effect and their OFDM structure. We propose to use a ikeihood function that provides an information about the independence of the processed sampe, and we are seeing ater that this approach is cose to a constant fase aarm rate detector, when its main advantage reies Figure 1 Physica versus MAC ayer.

5 Page 5 of 25 on the fact that it does not need to set a threshod vaue to the detector. Let now Y i (u) denotes the foowing set of observations: Y i (u) =[y i (u),..., y i (N s )] 1 u < N s (8) And et us define f Y the joint probabiity density function of Y i (u). If Y i (u) is composed of ony noise sampes N s f Y (Y i (u)) = f w (y i (m)), (9) m=u where f w is the probabiity density function of a compex norma aw centered and variance σw 2, given by f w (x) = 1 πσw 2 e x 2 /σw 2, (1) The og-ikeihood that the vector Y i (u) isformedof (N s -u) noise-independent sampes is expressed as [ Ns ] L i (u) =og f w (y i (m)) (11) m=u Computing the mean of the N og-ikeihood functions expressed on each sensor, we get a criterion J (u) to provide an information about the nature of the processed sampes J (u) = 1 N N L i (u) i=1 = (N s u)og(πσ 2 w ) 1 Nσ 2 w (12) N N s y i (m) 2 i=1 m=u As u varies in the interva [1, m 1 ), the number of noise sampes composing Y i (u) decreases and so does J (u) unti it reaches a minimum bound at m 1 (see Figure 2). However, for u varying from m 1 to m 2, the number of signa pus noise sampes decreases; therefore, the ratio of noise sampes to signa pus noise sampes increases and by the way J (u) increases. It reaches its maximum vaue if and ony if Y i (u) contains ony noise sampes, i. e., when u = m 2. Finay, for m 2 <u<n s, J (u) decreases again for the same reason that the one expained for 1 <u<m 1. We concude that the edges of the detected frame can be estimated as { [ ] ˆm1 = arg min J (u) u [ ] (13) ˆm 2 =argmax J (u) u Estimation of the channe occupancy rate When we have ony one data frame in the observed window, the occupancy rate can easiy be estimated.8 yi(u) Sampe index 15 J (u) 1 5 ˆm 2 5 ˆm Sampe index Figure 2 Exampe with one frame and corresponding criterion behavior.

6 Page 6 of 25 thanks to the previous criterion by ˆm 2 ˆm 1.However, the assumption to have ony one frame in the observation window is too restrictive. In practice, we may get a signa as shown in Figure 3 or with more frames. Based on the behavior of J (u), we can ceary see (Figure 3b) that the sope of J (u) is positive when u corresponds to the index of a signa pus noise sampe and negative when u corresponds to the index of a noise sampe. Therefore, we can take advantage of the gradient of J (u) to distinguish the nature of the observed sampes. Introducing the function F(u) such that (u) = 1 [ ] sign{ (J (u))} +1. (14) 2 Here,wedenoteby the gradient of J (u) processed using the centra difference method, such that the derivative for any point of index u {1, N s } is processed as N s and zero when it is ony noise, and the channe occupancy rate is estimated by Ĉ or = 1 N s N s u=1 (u). (15) Criterion vaidation imits In this section, we propose to investigate the imits of the proposed criterion J (u). The aim is to find the dynamic where J (u) we behaves, i.e., where its sope is positive for signa pus noise sampes and negative for noise sampes. For 1 u m 1 : J (u) decreases ony if E[J (u)] <, and therefore if u E[J (u)] = (N s u)og(πσw 2 ) 1 σw 2 [(m 1 u)σw 2 +(m2 m1)(σ w 2 + S)+(Ns m2)σ w 2 ] (J (u)) = 1 (J (u +1) J (u 1)). 2 For the first point, we use the forward finite difference such that (J (1)) = J (2) J (1). Finay, at the right end eement, a backward difference is used the derivative costs: we get σ 2 w < 1 πe E[J (u)] u =og(πσw 2 )+1, and (16) (J (N s )) = J (N s ) J (N s 1). sign{.} denotes the sign operator. According to this, F (u) equas 1 when signa pus noise sampes are present For m 1 u m 2 : J (u) is an increasing function E[J (u)] ony if >, then if u.6 y(u) x 1 4 (a) J (u) Figure 3 (a) Absoute vaue of a wifi signa, (b) corresponding behavior of the criterion J (u). (b)

7 Page 7 of 25 E[J (u)] = (N s u)og(πσw 2 ) 1 σw 2 [(m 2 u)(σw 2 + S)+(N s m 2 )σw 2 ] the partia derivative is E[J (u)] u =og(πσw 2 )+ 1 σw 2 (σw 2 + S), (17) and J (u) increases ony if σ 2 w > 1 πe (1+γ ) (18) where γ = S σ 2 w is the signa-to-noise ratio. For m 2 u N s : we get the same resut as in (16). As a concusion for an optima behavior of J (u), the noise variance must satisfy 1 πe (1+γ ) <σ2 w < 1 πe. (19) This inequaity represents the imits of the proposed criterion. It means that the performance of the proposed method depends on the noise variance vaue and aso on the signa-to-noise ratio. Therefore, if the noise variance does not satisfy Equation (19), we can think to adjust it appying a certain gain on the received signa. Indeed, by mutipying the whoe vector of observation y by a gain η, the noise variance is no onger σw 2 but ησ w 2, where h must be chosen such that it satisfies 1 πe 1+γ <ησ2 w < 1 πe. (2) The right part of the inequaity is easy to satisfy, but unfortunatey the eft part requires the knowedge of the signa-to-noise ratio, which is not avaiabe in our case. Another approach is to introduce a new criterion that overcomes this drawback; this criterion is the distance between J (u), a Parzen estimator-based criterion introduced in the next section Parzen estimator-based criterion The proposed soution consists in processing a new criterion that aims to minimize the distance between the true probabiity density function of the noise and a Parzen-estimated probabiity density function of the observed sampes [26,27]. The main advantage of this new criterion is that it does not rey on Equation (19). We see in Section 5.1 that its performance remains constant for any vaue of σw 2. Starting from the set of observations = {R{y i (m)}, I{y i (m)}}, i {1,..., N}, m {1,..., N s },(21) where R{.} and I{.} denotes the rea and imaginary part of the sampe. We get 2NN s sampes avaiabe for estimating the Parzen window density distribution. Given a sampe y i (m) =p i (m)+j.q i (m), its Parzen window distribution is given by ˆf(y i (m)) = ˆf (p i (m)).ˆf (q i (m)), (22) where ˆf(z) = 2NN 1 s 1 K 2NN s F k= ( z zk F ). (23) Such that K is the Parzen window kerne and F is a smoothing parameter caed the bandwidth. This kerne has to be a suitabe p.d.f function. We use Gaussian kernes with standard deviation one. The new processed criterion is [ J K (u) = 1 N Ns ] og ˆf(y i (m)). (24) N i=1 m=u Once we get J K (u), wemeasurethedistancebetween J (u) and J K (u) to obtain a new criterion K(u) = J (u) J K (u). (25) Substituting J (u) by K(u) in Equation 14, the function F(u) is processed to be then used to find the channe occupancy rate Equation (15) Fuctuations probem The difficuty is to estimate the channe occupancy rate accuratey for ow signa-to-noise ratio. In fact, there are fuctuations that can misead the decision for a given sampe (Figure 4). To fix this probem, we propose to use a smoothing technique. The choice of the ength of the smoothing window W is very important. We choose W equa to the ength of a SIFS (for Short IFS), which is the smaest interframe interva. Thus, theoreticay, we can not get a set of successive noise sampes of a ength ess than a SIFS. Then, if we met a set of noise-ony sampes of ength ess than an SIFS, it means that the agorithm took the wrong decision and F(u) wi be forced to 1 for those sampes Reation with the CFAR method We can demonstrate that there is a direct reation between our method and the CFAR (Constant Fase AarmRate[28])method.Themaindifferenceofthe proposed technique is that it does not rey on a fase aarm probabiity P fa. Indeed, the proposed approach ony depends on the noise variance vaue.

8 Page 8 of 25 2 J (u) (a) 1.8 Φ(u) Figure 4 (a) J (u), (b) corresponding F(u) (b) First of a, et us consider the case of the Gaussian noise. The CFAR approach reies on a threshod associated with a fase aarm P fa. Considering the foowing hypothesis test { H : y i (m) =w i (m) H 1 : y i (m) = L i 1 = h (26) i()x j (m )+w i (m) and a given threshod, the probabiity of fase aarm can be expressed as P fa =Pr{ y i (m) 2 λ H }. Since the noise is supposed Gaussian, its absoute ( ) σw vaue foows a Rayeigh distribution R and 2 P fa =2 λ = exp y i (m) σ 2 w ( λ2 σ 2 w ( ) exp y i(m) 2 σw 2 dy i (m), ). (27) Therefore, an observed sampe is considered as signa pus noise sampe if and ony if y i (m) 2 > σ 2 w og(p fa). In our case, considering that (L i (m)) = L i (m +1) L i (m), wehavethefoowing expression (L i (m)) = og(πσw 2 )+ 1 σw 2 y i (m) 2. As said previousy, the symbos are considered as signa pus noise if and ony if the gradient is positive. It foows that y i (m) 2 > σ 2 w og(πσ2 w ). We obtain the same criteria with the CFAR if we choose a P fa = πσw 2, providing that Equation (19) is satisfied. The main advantage of the proposed approach reies on the fact that the choice of the P fa is automatic and achieves good performance when Equation (19) is satisfied. As there is a recursive reation between two consecutive sampes of J (u), such that ( ) J (u 1) = J (u) og(πσw 2 ) 1 N Nσw 2 y i (u) 2. (28) To reduce the computationa cost, we propose to compute the criterion in the backward sense, i.e., from its ast eement and then deducing the other eements recursivey. In this case, the CC is reduced to O(NN s ). The whoe agorithm is described in Agorithm 1. Agorithm 1 Channe Occupancy Rate Estimation Observe N s sampes on the desired channe; J (N s )= 1 N i=1 y i(n s ) 2 ; Nσ 2 w for u = N s - 1: -1: 1 ( doj (u) =J (u +1) og(πσw 2)+ 1 ) N Nσw 2 i=1 y i(u) 2 ) end for Compute the functions F(u) vaues using (14); i=1

9 Page 9 of 25 Smooth F(u) thanks to the described procedure in 2.2.6; Deduce the C or thanks to (15). As the number of users increases, the oad increases and the coision probabiity too. To maintain a good QoS and to avoid the coisions, the backoff intervas are increased in an exponentia manner. This eads to injecting a arge amount of white spaces in the communication exchange For congested networks, i.e., whereathenodeshaveaframereadytobesentin their buffers, we remark that the channe occupancy rate decreases. In order to avoid a VHO in that particuar case, it is reevant to have access to another reevant metric in such situation, which is the coision rate. 2.3 Frame coision detection The contention-based access mechanism in WiFi impies that a the stations have to isten to the channe before competing for the access in order to avoid coision between the frames. Unfortunatey, as the number of competing stations increases, the coision probabiity increases and the throughput decreases affecting the QoS. Then, the coision rate is a good metric for both horizonta handover where many access points are avaiabe and aso vertica handover if we wish to hand off from any standard to an OFDM access point. A proposed method [29,3] for coision detection in a WiFi system suggests that the AP of a basic service set (BSS) measures RF energy duration on the channe and broadcasts this resut. Then, stations can detect coisions by checking the duration against their previous transmission schedues, if they are different it means that a coision occurs. This method assumes that the mobie is abe to measure this time duration and requires to be connected and synchronized with the access point. Within this framework, we propose a method for coision detection that requires no connection to the AP. Once the data frames are detected thanks to the agorithm presented in Section 2.2.2, we use an information theoretic criterion to get the rank of the autocorreation matrix of the observed frame. Unfortunatey, to estimate the number of sources, the channe ength is necessary. To skip this step, we propose to expoit the OFDM structure of the signas: since the channe ength is aways ess than the cycic prefix, using a smoothing window for the autocorreation matrix of a ength equa to the cycic prefix, we can get the number of sources and decide whether a coision occurred or not (number of sources greater than 1). In this case, the number of antennas must be greater than the number of source, so we need at east 3 antennas to detect the coision. The signa mode is said to be MIMO for mutipe input mutipe output. We consider that M sources are emitting and that the receiver is doted of N antennas. The observed signa on the ith antenna is expressed as y i (m) = M j=1 L ij 1 = h ij ()x j (n )+w i (m), (29) where the x j (m) forj = 1,..., M areofdmsourcesignas expressed as in (1), h ij () is the channe impuse response from source signa j to the ith antenna, and L ij is the order of the channe h ij. Consider that we detected a data frame of ength N f, and et L j =max i (L ij ) be the ongest impuse response of the channe, zero-padding h ij () if necessary. First, defining the foowing vectors y(m) =[y 1 (m), y 2 (m),..., y N (m)] T, (3) h j (m) =[h 1j (m), h 2j (m),..., h Nj (m)] T, (31) w(m) =[w 1 (m), w 2 (m),..., w N (m)] T, (32) we can express the signa mode as y(m) = L M j 1 h j ()x j (m )+w(m), (33) j=1 = Considering an observation window of d sampes and defining y d (m) = [ y T (m),..., y T (m d +1) ] T, (34) x d (m) = [ x 1 (m),..., x 1 (m d L +1),..., x M (m),..., x M (m L d +1) ] T, (35) w d (m) = [ w T (m),..., w T (m d +1) ] T, (36) we get y d (m) =Hx d (m)+w d (m), (37) where H is Nd (L + Md) (L def = M 1 L j) Syvester matrix defined as H = [H 1, H 2,..., H M ], (38) H j = h j () h j (L j ) h j () h j (L j ). (39) Note that the dimension of H j is Nd (L j + d).

10 Page 1 of 25 Defining the statistica covariance matrices of the signas and noise as [ R y = E y d (m)y d (m) H], (4) [ R x = E x d (m)x d (m) H], (41) [ R w = E w d (m)w d (m) H], (42) we have the foowing reation R y = HR x H H + σ 2 w I Nd, (43) where I Nd is the identity matrix of order Nd and (.) H is the transpose conjugate operator. Assuming that the channes have no common zeros, and for a arge enough observation window of a size d, we estabish that the rank of R x is r = min{(md + L), dn}. (44) Using an information theoretic criterion, ike AIC or MDL [31], it is possibe to get an estimate of r, such that AIC(k) = 2og Nd i=k+1 1 Nd k λ 1/(Nd k) i Nd i=k+1 λ i (Nd k)nf Nd (Nd k)nf λ 1/(Nd k) i i=k+1 MDL(k) = og 1 Nd i=k+1 Nd k λ i +2k(2Nd k), (45) + k 2 (2Nd k)ogn f, (46) where the i for i = 1,..., Nd are the sorted eigenvaues of R y, N f represents the ength of the detected frame. The rank of the autocorreation matrix R y ˆr is determined as the vaue of k Î {,..., Nd - 1} for which either the AIC or the MDL is minimized. ˆr AIC = arg min[aic] k ˆr MDL = arg min k [MDL] (47) Therefore, according to Equation (44), the number of sources M is estimated as the nearest integer to r L d. Unfortunatey, the channe ength L is unknown, and we shoud have it to estimate M. To avoid this step, we propose to expoit the properties of the OFDM signas. We know that the ength of the cycic prefix is aways chosen to be greater than L ij. So, if the smoothing factor d is defined as equa to the cycic prefix, we are sure that L ij <d. We can generaize that to estimate a number of sources greater than one. In fact, if r = Md + L then L = r-md.sincel = M j=1 max (L ij ), we are sure that L< i Md and by the way r-md<md.thus,r/m <2d, and therefore M > r 2d.Weconcudethat ˆM is the nearest integer greater than r 2d. If this vaue equas 1, it means that there is indeed one source, otherwise more than one source is present and a coision occurs. The agorithm is described in Agorithm 2. For each frame, we have to compute the eigenvaue decomposition (EVD) andthenperformaicormdl.asthec.cofthese two agorithms is negigibe compared to the EVD, the computationa cost is proportiona to an EVD. Agorithm 2 Coision detection agorithm nb_coision = ; Run agorithm described in Section 2.2.2; for each detected data frame do Process the autocorreation matrix R y ; Compute r thanks to (45) or (46); if cei(r/2d) >1then nb_coision = nb_coision+ 1; end if end for nb coision coision rate = the number of detected frames 3 Metrics for OFDMA-based networks Orthogona frequency division mutipe access (OFDMA) is a muti-access technique based on orthogona frequency division mutipexing (OFDM) digita moduation scheme. Mutipe access is achieved in OFDMA by assigning subsets of subcarriers to individua users in a given time sot. This technique aows to support differentiated quaity of service (QoS), i.e., to contro the data rate and error probabiity individuay for each user. First, we propose to appy the agorithm presented in Section 2.1 to get an estimate of the downink SNR in an OFDMA-based network. Then, we propose an aternative approach to estimate the time frequency activity rate, which is a simiar metric of the channe occupancy rate for CSMA/CA-based systems. Concerning the coision rate, as said previousy, since OFDMA-based systems are fu dupex, no coision occurs and it has no meaning as a metric. 3.1 SNR estimation for OFDMA based systems Assuming that an OFDMA symbo consists of up to N sc active subcarriers, we can modify Equation (1) to get the expression of an OFDMA signa x(m) = E s N sc k Z Nsc 1 n= n ε k,n a k,n e 2iπ (m D k(nsc+d)) N sc g(m k(nsc + D)).

11 Page 11 of 25 In this case, ε k, n is a set of i.i.d random variabe vaued in {, 1}, expressing the absence or presence of signa activity in the (k, n) time frequency sot. The received signa is expressed as in Equation (2), and the SNR is expressed as SNR = S σw 2, (48) with L 1 S = E s E[ ε k,n a k,n 2 ] σh() 2. (49) = The whoe agorithm presented in Section 2.1 stays vaid for OFDMA signas. 3.2 Time-frequency activity rate estimation for OFDMA system In OFDMA-based systems, when the number of active subcarriers is sma, the data traffic shoud aso be. Therefore, providing a satisfying downink signa strength, it is better for a muti-mode termina to connect on such a base station rather than on one where the data traffic is high (high number of active subcarrier). In this section, we focus on the passive estimation of the aocation rate of OFDMA physica channes timefrequency sots. The aocation rate is defined as the number of active sots (aocated symbos) divided by the tota number of sots per frame. In some networks such as WiMAX, the physica channes aocation rate is reguary broadcasted by the base station so that it can be known by any termina. However, this requires a muti-mode termina that istens to the surrounding networks to intercept every frame preambe. If the muti-mode termina has to decode every intercepted preambe to get this information, the vertica handover can be a very time- and power-consuming process. An aternative approach deveoped in this section is to get the OFDMA physica channes aocation rate by bindy estimating the time-frequency activity rate of OFDMA physica signas. Such approach focuses on the signa properties and therefore does not require any message decoding (assuming this message is made avaiabe by the base station, which may not be the case in a OFDMA networks). To the best of our knowedge, there is no agorithm pubished to date that addresses the bind estimation of the time-frequency activity rate ofofdmasignas.weproposeamethod[32]witha ow computationa cost to estimate the time frequency activity rate of a WiMAX networks. This method is based on the estimation of the first- and second-order moments of the received signa. The received signa is expressed as in Equation (2). We assume that the receiver is synchronized with the transmitter in time and in frequency. This synchronization can be reaized thanks to the frame preambe or thanks to bind techniques presented in [16] and [33]. We aso assume that the noise power σw 2 is known or at east estimated thanks to bind methods such as those detaied in Section 2.1 or in [13,34] Estimation agorithm The estimation of the time-frequency activity rate τ is equivaent to detect the active sots from the non-active ones τ = k,n I(ε k,n =1), M s N sc (5) where I(A) is the indicator function of any event A and M s is the number of observed OFDM symbos. Intuitivey, considering that σw 2 is known, a cassic detector structure coud be used so that k,n ˆτ = I( Y k,n >θ(σ w )), (51) M s N sc where θ(s w )isathreshodfunctionandy k, n is the signa observation on the sot (k, n). nm N 1 sc 1 2iπ Y k,n = y[k(n sc + D)+D + m]e N sc, (52) N sc m= = ε k,n a k,n H k,n Es + W k,n, (53) where H k, n and W k, n are, respectivey, the channe frequency response at subcarrier n and the noise at subcarrier n of the kth received symbo. The imitation of such approach is that the performance is strongy impacted by the choice of a threshod. In order to avoid this constraint, we hereafter propose compementary aternative method. The proposed technique reies on the absoute vaue of the first- and second-order moments of the observed sampes. These moments are indeed dependent of the activity rate τ. For a (k, n) such that ε k, n =, the observations are made of noise-ony sots such that they satisfy Y k,n CN(, σw 2 ). Therefore, in this case the absoute vaue Y k, n has a Rayeigh distribution and its expectation is given by π E[ Y k,n /ε k,n =]= 2 σ w, (54) where E[./.] defines the conditiona expectation. When the observations are made of signa pus noise sampes (i.e., ε k, n = 1), the distribution of Y k, n is hard to define. Indeed, actua systems are using the adaptive

12 Page 12 of 25 moduation and coding (AMC) scheme, and the consteationcanbedifferentfromasottoanother.thea k, n may have a distribution corresponding to BPSK, QPSK, 16-QAM, or 64-QAM [35]. According to the principe of maximum entropy [36], the state of ignorance on the consteation distribution is here modeed by an uniform aw. Hence, without prior information, we assume that the probabiity to get each consteation equas 1/4. (Note that the impact of this assumption is discussed in Section 4). Consequenty, the expectation of Y k, n when ε k, n = 1 can be written as E[ Y k,n /ε k,n =1]=E[ a k,n H k,n Es + W k,n ], = 1 4 ] E [ a k,n H k,n Es + W k,n /a k,n C Mj, 4 j=1 (55) where the C Mj consteations are M j -QAM such that for j = 1,..., 4, M j is equa to 2,4,16,64. Assuming a Gaussian noise, a Rayeigh fading channe and a known a k, n, the distribution of the observed sots is Gaussian: Y k,n /a k,n, ε k,n =1 CN(, L 1 σh() 2 E s a k,n 2 + σw 2).It then foows that the absoute vaue Y k, n /a k, n, ε k, n =1 has a Rayeigh distribution. After performing integration over a the possibe vaues of a k, n in each C Mj consteation, we find that π 1 E[ Y k,n /ε k,n =1]= M j=1 j p=1 M j σh() 2 E s c p 2 + σw 2, (56) where c p is the pth symbo of te consteation C Mj, and consequenty, π E[ Y k,n /ε k,n =1]= 5 σh() Es + σ w σ 2 E s h() σ w σh() Es + σ w σ 2 E s h() σ w σh() Es + σ w σh() Es + σ w σh() Es + σ w σh() Es + σ w σh() Es + σ w σh() Es + σ w σh() Es + σ w ] 7 σh() Es + σ w 2, ( ) = ϕ σh() 2 Es, (57) where is a function that associate with each σ 2 h() E s the expectation E[ Y k,n /ε k,n =1], whenσ 2 w is assumed to be known. Since τ% of the sots are active and (1 - τ)% are not, the expectation of the modue of the observed sampes is expressed as ( ) E[ Y k,n ]=τϕ σh() 2 E s π +(1 τ) 2 σ w. (58) Moreover, the second-order moment E[ Y k,n 2 ] is given by E[ Y k,n 2 ]=σw 2 + τ σh() 2 E s, ε k,n. (59) It foows that σh() 2 E s = E[ Y k,n 2 ] σw 2, (6) τ If we denote by μ 1 = E[ Y k,n ] and μ 2 = E[ Y k,n 2 ], then M 1 1 N 1 ˆμ 1 = Y k,n, (61) M s N sc ˆμ 2 = k= 1 M s N sc 1 n= Y k,n 2. (62) M 1 N 1 k= n= Substituting this vaue in Equation (58), an estimate of the channe occupancy rate ˆτ is obtained by soving the foowing equation ( ˆμ2 σw 2 ˆτϕ ˆτ ) +(1 ˆτ) π 2 σ w ˆμ 1 =. (63) This equation has no anaytica soution. We propose to sove it by a binary search agorithm. The whoe corresponding technique is presented in Agorithm 3. The computationa cost of the proposed agorithm is negigibe compared to the FFT, and thus the C.C is O(N sc og N sc ). Agorithm 3 Moments method Observe M s OFDM symbos; Estimate σ 2 w ; Compute Y k, n ; Compute ˆμ 1 and ˆμ 2 thanks to (61) and (62); Deduce ˆτ soving (63) thanks to the binary search agorithm. 4 Architecture of the proposed detector The current design of cognitive receivers is based on software defined radio (SDR) technoogy that enabes through software, dynamic reconfiguration of a protocos stacks incuding the physica ayer. In other words, frequency band, air-interface protoco, and functionaity can be upgraded with software downoad and update instead of a compete hardware repacement. SDR provides an efficient and secure soution to the probem of

13 Page 13 of 25 buiding muti-mode, muti-band, and muti-functiona wireess communication devices [7]. A cognitive radio (CR) is an SDR that additionay senses its environment, tracks changes, and reacts upon its findings. The main components of a cognitive radio transceiver are the radio front-end and the baseband processing unit. In the RF front-end, the received signa is ampified and mixed and is anaog to digita converted [6]. The output of the digita front-end is then fed into the baseband processing engine. Each component must be abe to be reconfigurabe via a contro bus. Note that a baseband processing engine can service mutipe RF front-ends, each of which supports specific airinterface standards. The baseband processing unit has first to detect the presence of a signa by any weknown techniques in the iterature [25,37],... and then identify the systems corresponding to the detected signa. The identification of OFDM systems has been addressed in many papers, with different approaches. The reader can refer to [38-41] for exampe. Once the system has been identified, according to the protoco used by this system, the baseband processing unit wi start and estimation of the reevant metrics using our proposed agorithms in Sections 2 or 3. When the metrics are estimated, an interaction needs to be performed with the higher ayers to decide whether to trigger a vertica handover or not. A bock diagram of the receiver is iustrated in Figure 5. 5 Simuation and experimenta resuts 5.1 Metrics for CSMA/CA based networks In this section, we present computer simuations resuts that show the agorithms performance SNR estimation In this section, the performance of the proposed estimator is assessed on WiFi signas. WiFi signas are OFDM signas with 64 subcarriers and a guard interva of ength equa to 16. The propagation channe {h()} =,..., L - 1has an exponentia decay profie for its non-nu component (i.e., E[ h() 2 ]=Ge /μ for =,..., L-1), G is chosen such that L = E[ h k() 2 ]=1. Thechanneis assumedtobetimevariantwithadopperfrequency equa to 1 Hz for WiFi signas and a root-mean-square deay spread of 25% of D. The SNR is processed as described in Section 2.1. In Figure 6, we pot the normaized mean square error (NMSE) of the SNR estimation versus the true SNR for ) ] 2 σ 4 S 2 [ different M s, NMSE = E (Ŝ/ ˆσ 2 w S/σ w 2.Our method is compared with the approximate maximum ikeihood (AML) estimator described in [14]. This estimator reies on an empirica threshod a that is used to determine the channe ength which is required to estimate the SNR. The choice of this threshod, as described in [14], is subjective. If apha is too sma, the channe ength wi be overestimated, resuting in a poor efficiency of the estimator. If it is too arge, signa sampes are incuded in the noise variance estimator, eading to an underestimation of the SNR. a is here set to.5; this choice is empirica in our agorithm [13] and has been compared to the one in [14] for many vaues of a and aways outperforms it. The reader can refer to [13] for more detais on the impact of a. Figure 6 highights two imitations of the AML agorithm. First, as previousy expained, this method depends on the subjective threshod a, which has a strong impact on the Receiver Radio Frequency (RF) RF Front end Anaog to Digita Converter (A/D) Detection Unit Baseband Processing Unit Identification Unit Metrics Estimation Unit To higher Layers Figure 5 Bock diagram of the proposed detector. Contro Bus

14 Page 14 of M s =24 Proposed method M s =48 Proposed method M s = 24 Cui et a. M s = 48 Cui et a. NMSE (db) Figure 6 NMSE on the estimation of the SNR vaue. SNR (db) performance. Then, as the signa power and noise variance estimations are not independent, the SNR estimation gets deteriorated at ow and high SNR. Moreover, Figure 6 reveas that the agorithm presented in this paper gobay outperforms the AML. WiFi supports a arge number of moduation and forward error correction coding schemes and aows to change it based on the channe conditions (adaptive moduation and coding (AMC)). The objective of AMC is to maximize the throughput in a time-varying channe. Since the adaptation agorithm typicay cas for the use of the highest moduation and coding scheme that can be supported for the current SNR, it is possibe to know the used data rate. In Figure 7, we pot the probabiity of estimating the SNR within the range of ± 1 db of the true vaue. It ceary indicates that our SNR M s =24 Proposed method M s =48 Proposed method M s =24 Cui et a. M s =48 Cui et a. Pr SNR (db) Figure 7 Probabiity of estimating the SNR within ± 1 db of the true vaue.

15 Page 15 of 25 NMSE on Ĉor (db) Smoothed Φ(u) Parzen estimator Energy detector P fa =.1 CFAR P fa = SNR (db) Figure 8 NMSE of the channe occupancy rate versus SNR. estimator gives a reiabe measure that can be used for vertica handoff decision. Note that this probabiity becomes greater than 8% for M s =24 and a SNR db if the toerated range is increased to ± 2 db Channe occupancy rate In Figure 8, we show the NMSE (normaized mean square error) of the estimation of the channe occupancy rate versus the SNR. The resuts are averaged over 5 Monte Caro runs, and the NMSE is here [ ) 2/C ] defined as E (Ĉor,k C 2 or or, where Ĉ or,k is the channe occupancy rate estimated at the kth reaization and C or is the true channe occupancy rate. In this figure, we pot the performance of the estimator based on a smoothed F(u) criterion and a Parzen-based estimator. The Parzen estimator is aso smoothed. The proposed method is compared with the CFAR (constant fase aarm rate) method with a probabiity of fase aarm P fa =1-4 and with the energy detector proposed by Urkowitz [25], with a P fa =1-4. The cognitive termina is supposed to have N = 2 antennas. We can ceary see that the proposed approach outperforms the other methods. Figure 9 shows the NMSE of the C or estimated with a smoothed F(u) for different SNR versus the spectra occupancy rate. We can ceary see that the performance of the proposed method depends on the channe occupancy rate vaue. However, even for ow C or, the method is very accurate (-49 db). As stated previousy, the criterion has vaidation imits, andforacertainrangeofthenoisevariance,itbehaves bady. To fix this probem, we proposed the Parzen estimator and stated that it does not depend on the noise variance. Figure 1 shows the NMSE of the three proposed methods versus the noise variance vaue, the SNR is fixed to 15 db, and the channe occupancy rate is equa to 64%. For this SNR vaue, the criterion shoud be vaid for: <σw 2 < Inthefigure, the ower bound corresponds to 1/πe 1+g = and the upper bound to 1/πe = We can ceary see that ony the Parzen estimator-based method is not affected by the noise variance vaue Coision detection Figure 11a and 11b show the performance of the proposed method versus SNR. We ceary see that for both AIC and MDL, we get a good probabiity of detection for a SNR greater than 1 db, which is the usua operating range of the WiFi. Note that there is no motivation to trigger a vertica handover toward an access point that does not satisfy the signa strength condition. The simuations were done with an observation window of 4 μs ength. We observe that AIC behaves better than MDL. The simuations were processed on frames whose starting, and ending points are supposed to be perfecty known Experimenta resuts The proposed bind agorithm for the estimation of the channe occupancy rate of a WiFi AP is evauated using the RAMMUS RF patform deveoped in the Signa & Communications department of TELECOM Bretagne. The aim of the experiments was not to highight the precision of the agorithm since the true C or is not avaiabe but to highight the efficiency of the proposed

16 Page 16 of SNR=1 db SNR=15 db Cor NMSE (db) Channe occupancy rate Figure 9 NMSE of the smoothed F(u) proposed method versus the channe occupancy rate. metric in different scenarios. Experiments were investigated on the Channe 6 (2.437 GHz) using the IEEE g norm. We tested different schemes with different number of users for different maximum bit rate aocated to each user. The schemes are based on Cient/ Server systems using the User Datagram Protoco (UDP) as presented in Figure 12. The physica ayer signa is captured thanks to an USRP2 device (Universa Software Radio Periphera [42]). The samping rate is set to 2 Mega-sampes/s. The traffic rate is controed thanks to J-Perf which is a software for UDP/TCP traffic generation, and the ist of used equipments is iustrated in Tabe 1 c. The observation window varies from 1 to 1 ms, and the presented resuts were averaged over 5 non-correated experiments. We test three scenarios varying the 1 Smoothed Φ(u) Parzen based Estimator Cor NMSE (db) Lower bound Upper bound σw 2 Figure 1 NMSE versus noise variance vaue, constant SNR.

17 Page 17 of antennas 4antennas 6antennas 8antennas Probabiity of detection SNR (db) (a) Probabity of detection with AIC versus SNR antennas 4antennas 6antennas 8antennas Probabiity of detection SNR (db) (b) Probabity of detection with MDL versus SNR Figure 11 Infuence of the SNR on the probabiity of detection. (a) Probabiity of detection with AIC versus SNR; (b) Probabiity of detection with MDL versus SNR. number of C/S systems from one to three. Each C/S coupe is exchanging data at a 1 Mbps rate. The resuts areshowninfigure13a.wecearyseethatasthe number of users increases, the channe occupancy rate increases too. In Figure 13b, we pot the variance of the estimated channe occupancy rate for one and three C/S systems. It is obvious that for the shortest the observationwindow,thevarianceisthehighest.therefore,to

18 Page 18 of 25 Figure 12 Configuration of the used network for the experiments. have the minimum variance, the observation window shoud be as ong as possibe, but for a seamess and a minimum atency handover, this window shoud be taken as short as possibe. Concerning the seection of the observation window, it depends on the degree of accuracy desired by the user. The onger the observation window, the more accurate the estimator. However, in a vertica handover context, the user does not need to have an accurate estimation of the metric. He just needs to know approximatey in which range is it and prefers certainy to decrease the scanning time, since using a ong observation window increases the goba scanning time, which is a crucia parameter that needs to be reduced to ensure a seamess and proactive handover. In Figure 14, we show the infuence of the data rate on the channe occupancy rate. For three Cient/Server systems, we pot the channe occupancy rate for Tabe 1 Configurations of the experiments Equipment Function Quantity NETGERAR RangeMax WNR35L Router and access point, DHCP server De Laptop Mobi Stations Cients 6 De Laptop PHY Scanner PHY Scanning and processing 1 USRP2 Scanning PHY open hardware 1 card NETGEAR RangeMax Wireess USB adapter 3 WNDA31 Inte(R) WiFi Link 53 AGN Integra wireess card 3 J-Perf Software Traffic generator 6 1 different data rate. Each system uses the same data rate. We observe that as the data rate increases, the channe occupancy rate increases in the same way. We aso notice that the variance is ower for systems using higher data rates. As expained previousy, the aim of the agorithm is to trigger a vertica handoff toward the access point where the trafficisower.accordingtothefigures,wecearyseethat the channe occupancy rate is ower in the configurations where a ower bit rate is required by users and increases as the required bit rate and number of users increases. In Figure 15, we show the channe occupancy rate for different bit rates, the number of C/S systems is set to one and the presented vaues are measured with a 4-ms observation window duration. We observe that for high data rates, the C or reaches a certain vaue and does not change. This is due to the backoff intervas. More precisey in Figure 16, we can see that the C or for three users is ower than the one for two users, this is due to the fact that for three users, the probabiity of coisions increases and then the used backoff are onger and the measured C or decreases. In such a case, the C or is not a good metric to trigger a VHO, and the more appropriate metric is the one that we proposed for coision detection in Section Metrics for OFDMA-based networks OFDMA SNR estimation In this section, the performance of the proposed estimators is assessed on WiMAX signas. The configuration tested is a partia usage of subchannes configuration