On the Measurement of Duty Cycle and Channel ... - Semantic Scholar

Accepted to the IEEE Journal on Selected Areas in Communications For the published version please go to http://ieeexplore.ieee.org/xpl/articleDetails.jsp?arnumber=6518478 DOI = 10.1109/JSAC.2013.131114 Questions/comments to [email protected]

1

On the Measurement of Duty Cycle and Channel Occupancy Rate Janne J. Lehtom¨aki, Risto Vuohtoniemi, and Kenta Umebayashi

Abstract Duty cycle (DC) and channel occupancy rate (COR) measurements are important to characterize the availability of white space for cognitive radio systems (CRSs). The COR is related to the real occupancy of channels considered by a CRS but despite its importance it has not yet been widely considered for spectrum use measurements. Spectrum use measurements have typically used detection with a single threshold. In order to improve sensitivity, we utilize the localization algorithm based on double-thresholding (LAD) with the adjacent cluster combining (ACC) to spectrum use measurements and propose a modified version of it. We theoretically analyze the probability of false alarm after the LAD processing in a realistic case with correlation among samples and theoretically analyze the probability of detection after the LAD ACC processing. We propose COR estimation algorithm based on hard decision fusion of the frequency domain decisions. Measurement and theoretical results confirm the accurate DC and COR estimation and the significant sensitivity gains with the proposed algorithms.

Index Terms Cognitive radio, channel occupancy rate, double-thresholding, duty cycle, measurements

Manuscript received April 15, 2012, revised August 31 and November 27, 2012. This work was supported by the Research Council of the University of Oulu and the research leading to these results was derived from the European Community’s Seventh Framework Programme (FP7) under Grant Agreement number 248454 (QoSMOS). This paper was presented in part at the 7th International Conference on Cognitive Radio Oriented Wireless Networks and Communications (CROWNCOM), Stockholm, Sweden, 2012 [1]. J. Lehtom¨aki and R. Vuohtoniemi are with the Centre for Wireless Communications (CWC), Department of Communications Engineering, University of Oulu, Oulu, Finland (e-mail: [email protected]). K. Umebayashi is with the Department of Electrical and Computer Engineering, University of Agriculture and Technology, Tokyo, Japan (e-mail: ume [email protected]).

© 2013 IEEE. Personal use of this material is permitted. Permission from IEEE must be obtained for all other uses, in any current or future media, including reprinting /republishing this material for advertising or promotional purposes, creating new collective works, for resale or redistribution to servers or lists, or reuse of any copyrighted component of this work in other works.

2

I. I NTRODUCTION Spectrum use measurements are important to characterize and model the availability of white space for cognitive radio systems (CRSs). The main issues in these measurements are sensitivity, flexibility, and accuracy. Most of the previous spectrum use measurement campaigns have been power measurements performed with conventional spectrum analyzers and the classification into signal (1) or noise (0) has been done with a single (power) threshold [2], [3]. The sensitivity of the classification is reduced due to frequency selective channel fading and receiver noise. The localization algorithm based on double-thresholding (LAD) with the adjacent cluster combining (ACC) uses upper and lower thresholds and can combine close-by clusters [4] for improving sensitivity. In channel occupancy rate (COR) measurements [5]–[8], the purpose is to measure the occupancy of a CRS channel which typically corresponds to several frequency bins. We define the COR as the fraction of the time that the investigated CRS channel is considered to be not available for transmissions by a CRS given that the CRS shall protect co-existing users no matter which part of a CRS channel they occupy. We use the term ”co-existing user” as a general term to refer to both primary users and any users already sharing the channel. The assumption is that we are not allowed or do not want to harm the co-existing users. Despite its importance, the COR has not yet been widely considered for spectrum use measurements. Instead, the spectrum utilization has been evaluated by the means of the duty cycle (DC), which is based on frequency bin measurements associated to the frequency resolution of the employed measurement devices, instead of the COR, which is related to the real occupancy of primary radio channels. A typical approach to detect if a signal is present in a channel is to use energy detector (ED) [9]. In [10], both ED and waveform-based sensing have been applied for extracting idle and active times. In [11], the ED has also been applied; for extracting channel statistics in the land mobile radio band for modeling purposes. Several references and discussion on different sensing approaches including cyclostationarity feature detection are given in [12]. We emphasize that most of the previous spectrum use measurement campaigns have used frequency bin power measurements with single power threshold and not the actual ED corresponding to a CRS channel. In generic spectrum use measurement campaigns we usually cannot assume any fixed channels and it should be possible to study simultaneously arbitrary channels

3

also after the measurement with different bandwidths and center frequencies so that the results are useful for different kinds of CRSs. This can be challenging with the ED due to the data storage requirements. Our main contributions in this paper are: 1) We propose a modified version of the LAD ACC applying m-dB method based on median (MED) filtered forward consecutive mean excision (FCME) with a novel threshold correction factor β, leading to improved flexibility, accuracy, and robustness. 2) We analyze the probability of false alarm (PFA) after the LAD processing in a realistic case with correlation among samples. In stark contrast to the previous results [13], we find that the LAD processing does increase the PFA. 3) We present novel analysis which shows clearly why the double threshold methods results into significantly higher detection performance, and therefore, higher COR accuracy. The theoretical COR estimation results agree remarkably well with the actual measurements. 4) We propose a new algorithm for accurate COR estimation based on hard decision fusion of the frequency domain decisions from the same sensor. It can be used for generic spectrum use measurements since the frequency domain decisions (binary data) can be easily stored. II. W HY COR

MEASUREMENTS ARE NEEDED ?

Usually in spectrum use measurements, the average (in frequency domain) DC is calculated [2], [3]. However, the average DC is not necessarily sufficient to accurately characterize the availability of channels for a CRS. For example, different parts of a IEEE 802.11 channel may be occupied by not only the users in that channel, but also the users from adjacent channels. Additionally, there may be several co-existing users using different channelization. For example, one 802.11 channel contains several 802.15.4 channels. Finally, the verification of the availability of spectrum white space for CRSs should be flexible and not limited to channels utilized by some co-existing users. Consider the cases in Fig. 1. In all of these cases the average DC would significantly underestimate the real channel occupancy. Sensing during the actual operation of a CRS is usually done in the current CRS operating channel during silent periods. During the idle times of the CRS, all possible CRS channels could be examined using simultaneous COR measurement techniques. The COR measurement results could be used for selecting the best channel to use. It may happen that by reducing the current CRS channel bandwidth and/or by adjusting its center frequency much more white space can be utilized.

4

III. M EASUREMENT SETUP The measurement setup is summarized in Table I. The studied band is the 2400–2500 MHz industrial, scientific and medical (ISM) band. The number of frequency points was N = 916. The frequency sweep time was ∼10 ms. We define a record to correspond to T = 200 complete k sweeps. We denote the received frequency domain power data with Pi,c [dBm], where k is the

record index, c = 1, 2, · · · , N is the frequency bin index and i is the sweep index within the record (i = 1, 2, · · · , T ). The instrument is based on fast Fourier transforms (FFTs). The time required to collect the samples for one FFT was around 10 µs for a 20 MHz frequency segment with FFT size 256. We did not apply power spectrum averaging. It would require calculating more than one windowed FFT in the same 20 MHz frequency segment before moving to the next segment, leading to longer measurement times than the 10 µs. We emphasize that the time required to gather the power data for all frequency bins in the 100 MHz range was 10 ms, however the 10 µs refers to the actual time utilized for measuring each of the five 20 MHz frequency segment within the sweep. The significance of the utilized measurement time of 10 µs is that this short time avoids overestimation and also it roughly corresponds to the time utilized by 802.11 devices for clear channel assessment (CCA) (≤ 15 µs [IEEE Std 802.11-2007, Sect. 15.3.3]). Sampling with too much averaging would not anymore give the instantaneous state of the channel, but the sampling result would be affected by channel state over a wider time interval. Avoiding time domain overestimation in addition to other aspects such as mentioned in [2] enables meaningful comparison of different measurements campaigns. IV. C ONVENTIONAL DC

ESTIMATION WITH A SINGLE THRESHOLD

k In single threshold detection, the received power levels Pi,c [dBm] are compared with detection k k threshold γi,c [dBm] to make binary decisions Di,c about the signal presence or absence [2],   1, P k > γ k i,c i,c k Di,c = (1)  0, otherwise. ) ( k = 1|H0 , where H0 denotes the situation with only noise. Let The PFA is given by Prob Di,c ˙ . For each frequency bin c, a us denote the target PFA for all time-frequency elements with PFA

DC estimate Sck based on the binary spectral occupancy matrix for record k is [2], [3] T ∑ k k 1 Sc = T . Di,c i=1

(2)

5

V. T HRESHOLD SETTING Threshold setting can be done directly or through estimating the noise floor (NF), which refers to the mean of the noise values. The uncertainty about NF can be a significant issue [14]. Additionally, the NF can change as a function of time [15]. Thus, the NF estimate should be updated regularly. Perfect sensing during long sensing periods has been assumed for noise power estimation in [16]. In [17], the authors propose a method to estimate flat or non-flat NF in the presence of unknown signals by using morphological image processing operations. Information theoretic criteria based spectrum sensing methods based on Akaike’s information criterion (AIC) or minimum description length (MDL) have been proposed in [18]. They can be applied without knowledge of the noise power. Assuming a sufficiently static situation, the correct threshold ˙ ) can be found out by using the PFA method proposed for each frequency bin (for given PFA by L´opez-Ben´ıtez in [2]. In this method, we collect a sufficient number of samples that are guaranteed to be noise-only. Then we simply find out the detection threshold that is exceeded ˙ . The assumption of known noise-only samples can limit the with probability equal to the PFA practicality of this approach. However, no knowledge about the noise distribution is required. The PFA method threshold is independent of time but varies with the frequency bin index. A. FCME threshold The FCME algorithm [19] first sorts the input into ascending order. Although not necessarily true, it is assumed that I smallest samples are clean from signal components. Usually, I = ⌈0.1 · N ⌉, where ⌈·⌉ is the ceiling function. The mean of these I samples is calculated and it is multiplied by TCME in order to get the first threshold. The FCME parameter TCME determines the properties of the FCME algorithm. Every observation under the first threshold not already present in the clean set is added to it. Then the second threshold is calculated as the mean of the assumed clean set multiplied by the TCME . Every sample under the second threshold not already present in the clean set is added to it. This process is continued as long as there are new samples under the latest threshold. The FCME threshold is the threshold from the final iteration. B. m-dB method based on MED-FCME with a correction factor β In each sweep, the m-dB method based on MED-FCME [8], [20] uses the FCME algorithm. We do not use the FCME threshold, instead we estimate the NF by calculating the mean of the

6

final assumed clean set. The NF estimates are filtered with a median filter of length M , leading k to the MED-FCME NF estimate NˆF where k is the record index and i is the sweep index i

within the record. The estimated NF is used to obtain the decision threshold based on the m-dB k k criterion of [2], i.e. γi,c = NˆF i + m. Since windowed magnitude-squared FFT data follows ˙ )) [dB]. We the exponential distribution (see e.g. [21]), the proper m is m = 10 log10 (− ln(PFA propose correction factor β for the m-dB method based on MED-FCME so that ( ( )) ˙ m = 10 log10 − ln PFA − β,

(3)

where β is given by first finding the smallest value of ~ ∈ {0, 1, 2, · · · , N } that satisfies ( )   N −~ e  

−TCME

∑

j=1

(N −~−j+1) (N +1−j)(N −~)

N  = ~. 

(4)

The smallest value of ~ corresponds to the smallest average number of censored samples by the FCME that will not, after using new biased mean to estimate the number of censored samples, change the estimate of the number of censored samples (see also [8]). The biased mean for a given average number of censored samples is found using the theory of weighted sum of order statistics [22]. Finally, the correction factor β [dB] is given by utilizing the found value of ~ in (N −~ ) ∑ (N − ~ − j + 1) β = 10log10 . (5) (N + 1 − j) (N − ~) j=1 Fig. 2 shows the ratio between obtained and target PFAs for the FCME method and the m-dB method based on MED-FCME with a correction factor β. For some values, the FCME gave even more than five times too large values. The significantly improved accuracy of the m-dB method based on MED-FCME with a correction factor β can be confirmed. The FCME independently finds the clean frequency bins for each frequency sweep. Thus, the NF can be estimated even with frequency hopping signals covering the whole band. Even if the whole band corresponding to several services/channels would temporarily be fully utilized, the MED-FCME can still estimate the NF from the previously occurred gaps between transmissions. C. Comparison of the m-dB method based on MED-FCME and the PFA method We performed a measurement in an anechoic chamber in order to guarantee noise-only samples. The PFA method threshold was set based on all received data during the hour. Fig. 3 shows thresholds given by the m-dB method based on MED-FCME and the PFA method

7

˙ = 10−4 and 0.03. The theoretical values for m corresponding thresholds derived with target PFA ˙ given by (3) are 9.9 dB and 5.7 dB, respectively. However, we used slightly higher to these PFA values of 10 dB and 6 dB to protect against the non flat NF. Since the m-dB method based on MED-FCME threshold varies with time, Fig. 3 shows the range of thresholds. It can be seen that m-dB method based on MED-FCME thresholds were within ±1 dB of the values given by the PFA method. The main advantage of m-dB method based on MED-FCME is that it estimates the NF directly from the signal while simultaneously performing regular measurements. It can also react rapidly to changes in the NF. However, in the case of wider bands the difference between the PFA method and the m-dB method based on MED-FCME may be too large. The reason is that the noise floor of a radio receiver tends to increase with frequency. The threshold provided by the PFA method is able to follow this trend as it is computed on a frequency bin basis, while the m-dB method based on MED-FCME method provides a single numerical value for the whole band. In those cases, we recommend segmenting the band into smaller pieces for which the NF estimation is done separately. This results into staircase function for the NF estimate. VI. LAD METHODS In LAD [4] signals are detected with the lower threshold γL . All contiguous groups of detected signals need to have at least one sample also exceeding the upper threshold γU . In the ACC processing [4] contiguous groups that are separated by n or less samples are joined. For example, assume that using γL leads to frequency bin decisions ”011001110”. We notice two clusters of 1s. These clusters are accepted or rejected based on whether they contain at least one sample also exceeding γU . Decisions corresponding to rejected clusters are replaced with zeros. If both of the clusters in the above example are accepted and n ≥ 2, the gap between the two clusters is filled and the finally estimated binary frequency bin occupancy sequence would be ”011111110”. A. MED-LAD ACC We apply the LAD ACC framework to spectrum use measurements and propose a modified version called MED-LAD ACC. The conventional LAD ACC uses thresholds derived from the FCME thresholds. From [23] and results in Fig. 2, we know that the FCME has significant ˙ ≥ 10−2 . To address this issue, the MED-LAD ACC problems at controlling the PFA when PFA uses the m-dB method based on MED-FCME with the proposed correction factor β in (5) and

8

˙ . Let us use mU and mL to denote TCME selected completely independently from the desired PFA values of m used for the upper and lower thresholds in the MED-LAD ACC, respectively. As an additional benefit the m-dB method is flexible as threshold levels can be easily and intuitively set just by changing the used value of m. In many previous measurement studies the threshold has been specified as m dB above the noise floor, corresponding to the m-dB method. Although the FCME is a significantly robust algorithm, it can catastrophically fail in peak load condition, giving much too large thresholds [20]. The median (used in the MED-FCME) is a robust measure with 50 % breakdown point leading to improved robustness. Even during our stress testing in [20] involving heavy use of the frequency bands we never had this kind of failure with the MED-FCME. We will evaluate the improved robustness of the MED-LAD ACC by simulations later in this paper by comparing the COR estimation performance of conventional LAD ACC to that of the MED-LAD ACC in a peak load condition. B. Analysis of the per-bin PFA after LAD processing with correlated samples We assume that only noise is present and the noise is circularly symmetric white Gaussian noise with mean zero and variance of σ 2 per component (real and imaginary parts). The time domain samples ri for one FFT segment with FFT size NFFT are collected into a vector r = [ri ]i=0,1,··· ,NFFT −1 . The windowed FFT operation can be represented with y = FWr, ( ) where F = e−2πjki/NFFT j,k=0,1,··· ,N −1 is the discrete Fourier transform matrix and the FFT

diagonal matrix W = diag (w0 , w1 , · · · , wNFFT −1 ) contains the real-valued window coefficients. Due to invariance to linear transformations with circular symmetry, the y is also circularlysymmetric with covariance matrix Ψ = 2σ 2 FWWH FH , where (·)H denotes the Hermitian transpose. The frequency domain samples yi are magnitude squared leading to output vector NFFT ∑−1 2 P = [|yi |2 ]i=0,1,··· ,NFFT −1 . The NF is 2S2 σ 2 , where S2 = wk . Since the ACC has negligible k=0

effect on the probability of false alarm [13], the analysis of the LAD method is also valid for LAD ACC in the noise-only case. We assume that the NF is perfectly estimated. 1) Joint distribution of two frequency bins’ events: Let us denote the event that the upper threshold is exceeded in a bin with ”D”, the event that the output is between upper and lower thresholds with ”B”, and the event that output is below the lower threshold with ”A”. Additionally, let event ”C” indicate that either event ”A” or ”B” occurred. Let us denote with ξv (e1 , e2 ) the probability that a frequency bin has event e1 (A, B, C, or D) and the frequency bin with offset v

9

has event e2 . By using the joint distribution of two magnitudes from a joint complex Gaussian distribution derived by Miller in [21] and by evaluating and inverting the subset of correlation matrix Ψ corresponding to the considered frequency bins, we get υ(e ∫ 1)

) ( S2 D12 2 +r 2 4r1 r2 − 2ψσ r ( ) 2 1 2 dr1 dr2 , e I0 2r1 r2 |4ψσ 4 | 2ψσ 2

υ(e ∫ 2)

ξv (e1 , e2 ) =

(6)

r1 =ℓ(e1 ) r2 =ℓ(e2 )

where ψ = |D12 |2 − S22 , I0 is the modified Bessel function of the first kind and order zero, NFFT ∑−1 2 2νπik/N √ and the integration limits are ℓ (A) = 0, υ (A) = γL , ℓ (B) = wk e D12 (v) = k=0 √ √ √ √ γL , υ (B) = γU , ℓ (C) = 0, υ (C) = γU , and ℓ (D) = γU , υ (D) = ∞. 2) Distribution of runs: Runs are consecutive sequences of 1s, where 1s refer to sensing observations where the result is a busy channel (1) instead of an idle channel (0). Let us denote the number of runs with length k in a sequence with length n as En,k . For example, ”001100111011” has E12,2 = 2 and E12,3 = 1. To find distribution of runs in the binary data processed only with the γU , we model the binary data processed with γU as Markov chain with states ”C” (below the γU ) and ”D” (above the γU ). By using (6) we get the 2 × 2 transition probability matrix

P= C D

C

D

ξ1 (C,C) ξ1 (C,C)+ξ1 (C,D) ξ1 (D,C) ξ1 (D,C)+ξ1 (D,D)

ξ1 (C,D) ξ1 (C,C)+ξ1 (C,D) ξ1 (D,D) ξ1 (D,C)+ξ1 (D,D)

,

(7)

where, for example, the first element is the transition probability from state ”C” (in a frequency bin) to state ”C” in the next frequency bin. By using the distribution theory of runs [24], we can get the probability P (ENFFT ,k = x) for the matrix (7). The resulting distribution is approximation since we only use the joint distribution of two consecutive samples. 3) Increase of run lengths due to the LAD processing: Let us assume the LAD processing increases each run length with ε. This assumption is an approximation, since the increase depends on the original run length. The PFA after LAD processing is theoretically obtained with ∑∑ LAD = P (ENFFT ,k = x) x (k + ε) /NFFT , PFA k

(8)

x

where the multiplication with x is due to the number of runs being x and the multiplication by k +ε is due to increased the run length. The probability that a sample with offset v has event ”B” given that the original sample exceeded the upper threshold, is Pv (B|D) = Pv (B ∩ D)/P (D) =

10

ξv (D, B)/ (ξ1 (D, D) + ξ1 (D, C)). We approximate ε with ε =

ℓ ∑

Pv (B|D) +

v=1

−1 ∑

Pv (B|D),

v=−ℓ

where ℓ corresponds to maximum offset where the probability of ”B” is still affected by the considered sample having event ”D”. We used as ℓ the maximum v that still satisfies Pv (B|D) ≥ P (B) + 0.01, where P (B) is the probability of event ”B” in a randomly picked sample. 4) Verification: Fig. 4 shows comparison between simulated and theoretical PFAs. It can be said that analysis is valid. We can observe that the frequency domain correlation due to the FFT window leads to an increase in the PFA. The main reason for this is that if a sample exceeds the γU , it is more likely that due to correlation that at least the γL is exceeded in adjacent samples. If there is no correlation, the increase in PFA is typically negligible [13]. We performed a real measurement with the setup described in Sect. III. The PFA with only the γU was 1.05 · 10−4 . The PFA after MED-LAD ACC with Gausstop window was 3.08 · 10−4 . Details of the window are not disclosed by the manufacturer. The Kaiser window we used has similar equivalent noise bandwidth (2.2830 vs 2.2). The theoretical PFA with that window was 3.00 · 10−4 . C. Analysis of the per-bin detection probabilities after LAD ACC processing 1) Signal model: We model the received signal yc in a frequency bin c as circularly symmetric white Gaussian with mean zero and variance of σc2 per component (real and imaginary parts). The variance including also the receiver noise is σc2 = (PS,c + PN )/2, where PS,c denotes the signal power within the frequency bin c and PN denotes the noise power within a frequency bin. The frequency domain samples are magnitude squared leading to output vector P = [|yc |2 ]c=1,2,··· ,N . For the IEEE 802.11b signal we use the classical sinc shape [25] within the 802.11 channel,  2  τ sinc((f − f |fc − f802.11,c | < T1c c 802.11,c ) Tc ) (9) PS,c =  0 otherwise where Tc is the chip duration, fc is the frequency corresponding to the frequency bin c, f802.11,c is the signal center frequency, and the scaling constant τ is selected to normalize the total received power to γ + ρ, where γ is the total signal power [dBm] and ρ is the receiver gain [dB]. 2) Analysis of single threshold detection: The magnitude of a zero-mean complex Gaussian follows the Rayleigh distribution [21]. Thus, the probability of detection for frequency bin c for any threshold level m dB above the noise power PN can be obtained with PD,c (m) = ( / ) exp −10m/10 (PS,c /PN + 1) .

11

U L 3) Analysis of double thresholds detection: Let PD,c = PD,c (mU ) and PD,c = PD,c (mL ). For

simplicity, we assume independent samples. Let us use the same events (A, B, C, D) as in Sect. VI-B. By considering all possible sequences of events leading to binary output of 1 for a frequency bin c, we get the probability of detection after LAD ACC processing with ( ) ( L ) L U LAD ACC U PD,c (mU , mL ) = PD,c + 1 − PD,c s1,c (0) + PD,c − PD,c (1 − (1 − z1,c ) (1 − z2,c )) + | {z } | {z } ACC LAD ) ( L U PD,c − PD,c s1,c (1), | {z } ACC

(10)

where

n−1 ∑

s1,c (ν) =

z3,c (x) z4,c (x) z7,c (x)

x=ν

 ( c−1 ) ) ) ( ∏ (  L U  x>1 1 − PD,c−x 1 − PD,c′    c′ =c−x+1 z3,c (x) =

    

z4,c (x) =

c−x−1 ∑

L 1 − PD,c−1

x=1

1

x=0 )]

[ U PD,x 2

( c−x−1 ∏ (

z7,c (x) =

 

n−x−1 ∑

[

cN ∑

z6,c (x3 )

x2 =c+x3 +1

x3 =ν

)

c′ =x2 +1

x2 =c1

  

L U PD,c ′ − PD,c′

U PD,x 2

x∏ 2 −1 c′ =c+x3 +1

(

L U PD,c ′ − PD,c′

)

] n−x−1≥ν

0

otherwise

 ( ) ( ) c+x∏3 −1 ( )  L U  1 − PD,c+x3 1 − PD,c′ x3 > 1    c′ =c+1 z6,c (x3 ) =

    

z1,c =

c ∑

z2,c =

x=c

x3 = 1

1

x3 = 0

U PD,x−1

x=c1 +1 c∑ N −1

L 1 − PD,c+1

U PD,x+1

( c−1 ∏(

(

L U PD,c ′ − PD,c′

)

)

c′ =x x ∏ (

L U PD,c ′ − PD,c′

)

)

c′ =c+1

and c1 and cN are LAD combining limits such as c1 = 1 and cN = N . In (10), the 1st term is the single threshold detection probability. The remaining terms give the improvement from the

12

double threshold processing. The 2nd term is the probability of detection due to ACC processing for frequency bin c given that event ”A” occurred in that frequency bin. In this case, we loop over the variable x which indicates the number of zeros in the input (to ACC processing) in the left side of c. For example, with x = 4, the events in the left side of the considered sample are ”ACCC”, leading to 4 zeros. The one in the binary input before these x zeros can be caused by sample at c − x − 1 exceeding the upper threshold or by LAD processing which leads to sample at c − x − 1 being detected. Thus, we loop over variable x2 which indicates where the upper threshold was exceeded. Then all the samples from x2 + 1 to c − x − 1 must have event ”B” as otherwise the LAD processing does not cause the sample at c − x − 1 being detected. Now, given that there was x zeros in the input data after LAD processing we can have maximum n − x − 1 zeros in the right size of the considered frequency bin c as otherwise ACC processing cannot cause detection for frequency bin c. The third term in (10) is the probability that LAD processing leads to detection for the considered frequency bin c given that the frequency bin c had event ”B”. To get this probability we find the probability that in left side or in the right side (or in both sides) there was continuous chain of samples with events ”B” leading to sample with event ”D”. The last term in (10) indicates the additional detection probability from ACC processing on top of the benefit from LAD processing, given that frequency bin c had event ”B”. Here we use ν = 1 to remove the sequences where the LAD processing alone is sufficient. 4) Fading channel: Typically frequency selective fading has a high correlation between frequency bins and the nulls are several MHz wide. Thus for a frequency bin, we may consider the channel around it to be Rayleigh flat fading. Based on this, we get the probability of detection after LAD ACC processing in fading channel with ∫ LAD ACC LAD ACC PˆD,c (γ; mU , mL ) = PD,c (γ + x; mU , mL ) f (x)dx,

(11)

where we have written the detection probability for the non-fading channel as a function of the signal power γ, x corresponds to offset in [dB] to γ caused by the channel fading and f (x) is the density function of squared Rayleigh random variable in [dB] which is easily obtained from [26]. This process can also be used for evaluating single threshold detection in a fading channel.

13

VII. C ONSIDERED APPROACHES FOR DC/COR ESTIMATION A. Single threshold detection 1) DC estimation: The DC estimates Sck are found with (2) using a spectrum occupancy matrix found with a single threshold detection (1) with threshold by m-dB method based on MEDFCME. This method was proposed by us in [8] and it is used as a reference method corresponding to the conventional single threshold detection (however with adaptive NF estimation). 2) COR estimation: It was shown in [8] that maximum combining of the DC estimates Sck by is suitable for COR estimation. The COR estimate for a CRS channel s is [8] Lks = max Sck = max T1 c∈Θs

c∈Θs

T ∑

k Di,c ,

(12)

i=1

where Θs is the set of frequency bins belonging to the CRS channel s. The maximum combining with single threshold detection was proposed by us in [8]. We use this method as a reference method for the proposed COR estimation methods (double thresholds and hard decision fusion). B. Double thresholds detection 1) DC estimation: We apply the MED-LAD ACC method with two thresholds for DC k measurements by calculating the binary matrix Di,c with MED-LAD ACC and by using (2).

2) COR estimation: The COR estimation is performed with maximum combining of the DC estimates Sck given by the MED-LAD ACC similar as with single threshold detection. C. COR estimation with MED-FCME CORε We utilize hard decision fusion with r-out-of-|Θs | fusion rule [27] for the individual binary sensing decisions of a sensor within a CRS channel s. In the proposed MED-FCME CORε, fused ∑ k decision for sweep i within a record k is Λki,s = 1 if c∈Θs Di,c ≥ r and Λki,s = 0 otherwise. The k are made using the m-dB method based on MED-FCME threshold. We point out that ORDi,c

gating of outputs of a bank of frequency domain detectors has been utilized in [28] for detecting frequency hopping signals. The COR estimate for the CRS channel s in record k is given by T ) ( ∑ Lks = T1 Λki,s . The PFA for a channel is defined by Prob Λki,s = 1|H0 . For the MED-FCME i=1

˙ , the number of frequency bins CORϵ, the channel PFA comes from per frequency bin PFA PFA ˙ , the control of the channel PFA must come from in each channel (|Θs |) and r. Since we fix PFA

14

r. The larger r is, the smaller the channel PFA is since with larger r more frequency bins need to have local decision of ”1” in order to yield global decision of ”1”. However, for maximum sensitivity, we should use the smallest possible r. Thus the parameter r is selected as follows: ”Select the minimum r such that the probability of the CRS channel being wrongly classified as ←→ ←→ occupied in the case of noise-only observations ≤ PFA ”, where PFA is the target channel PFA. D. COR estimation with Channel Power (ED) The ED corresponding to channel power can be obtained from frequency domain samples by summing the frequency bin power samples corresponding to the CRS channel s [5], [11]. In [11], power spectrum values obtained with FFT are integrated for each studied channel in order to obtain channel powers. These channel powers were compared to suitable thresholds in order to classify the channels as idle or occupied. This is the approach utilized herein so that the ED ∑ k decision for sweep i within a record k is Λki,s = 1 if c∈Θs Pi,c ≥ α, and Λki,s = 0 otherwise. For this calculation, the received power levels are assumed to have been converted from dBm to linear scale. If the CRS channel covers several FFT segments, we recommend partitioning to avoid overestimation (Sect. VII-E). To set the α, we recorded a large number of ED decision ←→ variables in the noise-only case and found the α that is exceeded with the probability PFA . E. Partitioning of CRS channels for COR estimation Let us denote the set of frequency bins belonging to the FFT segment j with Aj . Now, the intersection Θs ∩ Aj is the subset of the CRS channel s corresponding to the FFT segment j. ∪ Please note that j (Θs ∩ Aj ) = Θs . If Θs ∩ Aj = ∅ the subset is empty and it can be ignored. If a CRS channel s covers several FFT segments, the COR estimation is done separately for each non-empty subset Θs ∩ Aj . This is required since data in each FFT segment is taken at slightly different time instants even in the same sweep due to the frequency step time and the data collection time so that measurements within a CRS channel could be from different time instants. Without partitioning, this could lead to significant overestimation since there are more chances for signal to present within a sweep if the sweep has possibility of covering several different packets. The variance of COR estimates for the subsets can be reduced by combining several records. Finally, the COR results from each subset can be combined using the maximum of the COR estimates in order to estimate the COR for the CRS channel s.

15

VIII. A NALYSIS OF COR

ESTIMATION WITH THE MAXIMUM BASED COMBINING

The probability mass function (PMF) for T Sck is simply binomial random variable with success probability equal to the detection probability of the considered detection method for the frequency bin c and the number of trials is TH1 . The TH1 denotes the number of sweeps wherein the signal was present in the record. Noise only sweeps are assumed to have negligible impact. Now, we get the PMF of maxc∈Θs T Sck , by assuming independence and by using the result that cumulative distribution function (CDF) of the maximum of independent random variables is the product of the individual CDFs. From this CDF, we can get the PMF by subtracting two consecutive CDF values. To find the average COR, the PMF is weighted by the success count divided by the total number of sweeps. TH1 ∑ x COR (ℵ, Θs ) = T x=1

(

∏

(

)

ℵ F x, TH1 , PD,c −

k∈Θs

∏

(

ℵ F x − 1, TH1 , PD,c

)

) ,

(13)

k∈Θs

where ℵ is the method (such as ”LAD ACC” or ”U” corresponding to the single threshold detection) and F is the binomial PMF and TH1 = θT , where θ is the actual COR. We could also assume that TH1 is a binomial random variable with T trials and success probability θ and use weighted summing of the individual COR estimates. IX. M EASUREMENT RESULTS A. Measurement arrangement The studied signals were generated with Agilent E4438C ESG signal generator. The IEEE 802.11b signal had 11 Mbit/sec data rate, random payload, total packet size 1508 bytes, and center frequency 2472 MHz. The reference COR 40 % was obtained by controlling the idle time between packets. The signal was transmitted using a short cable to the channel emulator (EB Propsim F8) input and the channel emulator output was connected to the measurement system input. By utilizing a signal generator and a channel emulator, measurements can be controlled with a high precision. In our previous work [8] we have also measured real signals. We considered both the additive white Gaussian noise (AWGN) channel and time-variant and frequency selective ETSI BRAN WLAN channel model A [29] corresponding to a typical office environment. We simultaneously estimated the COR of the studied CRS channels and compared the values to the true value. The studied CRS channels are illustrated in Fig. 5. The CRS channels

16

were within one 20 MHz FFT segment of the Agilent RF sensor. Thresholds given by the m-dB method based on MED-FCME with mU = 10 dB and mL = 6 dB were used as the upper and lower threshold. The threshold given by m = 10 dB was used for single threshold detection and ←→ for the MED-FCME CORε. We utilized PFA = 0.005, TCME = 4.6052 and M = 800. B. DC estimation Fig. 6 shows the mean of Sck (over 46 records) denoted with S¯ck for conventional approach with a single threshold (represented by the m-dB method based on MED-FCME) and double thresholds detection (MED-LAD ACC) for frequency bins between 2450–2490 MHz (the measurement actually covered the full ISM band). This result confirms the improvement by the MED-LAD ACC, see (10) for theoretical analysis related to this improvement. C. COR estimation 1) 802.15.4 channels: Experimental validations with a real 802.15.4 node have shown that its CCA cannot sense 802.11b signals at all when 802.11b signal power is -81 dBm or less [30]. They studied a 802.15.4 channel with center frequency having 2 MHz offset to the 802.11b signal center frequency corresponding to our CRS channel B. Fig. 7 shows measurement results for CRS channels B (with 2 MHz offset) and C (with 7 MHz offset). It can be seen that the double thresholds detection had excellent performance as compared to a real 802.15.4 node. The frequency domain hard decision fusion had somewhat better performance and the ED had the best performance. The differences are not radical and all three methods were within 5 dB and were significantly better than conventional single threshold detection. It can be seen that the performance reduces the larger the offset is due to spectral shape of the 802.11b signals. Fig. 7 shows that there is no significant overestimation as compared to the true COR (40 %). 2) 15 MHz CRS channel: Fig. 8 show measurement results and also theoretical results found with (13) for the CRS channel A. It is remarkable that the theoretical results are mostly within ± ∼ 1 dB from the actual measurement results which include the real-life aspects such as frequency domain correlation, frequency offset, preambles, and receiver imperfections. For reference, a energy detection threshold specified in the IEEE 802.11 standard for CCA is -76 dBm [IEEE Std 802.11-2007, Sect. 15.4.8.4]. It can be seen the double thresholds detection had sufficient performance as compared to standard requirements. It also gives the DC estimates

17

with full frequency resolution. It can also be seen that dedicated COR estimation approaches MED-FCME CORε and ED had even much better performance. Fig. 8 shows that there is no significant overestimation as compared to the true COR (40 %). 3) FM signals: Fig. 9 shows measurement results with a FM signal with bandwidth around 400 kHz for CRS channel A. The true COR was 100 %. It can be seen that the double thresholds detection had the best performance and the ED had the worst performance. X. F URTHER DISCUSSION A. Relative performances Table II is based on the results in Figs. 7–8 and results from further measurements with AWGN channel. It shows the minimum channel emulator output level in order to get COR estimation result of 40±1 % when to co-existing signal is 802.11b signal with 40 % real COR. It can be seen that the sensitivity gain from using double thresholds as compared to single threshold approach was ranging between 7.5–11.5 dB (mean gain 9.4 dB). The sensitivity gain of the MED-FCME CORϵ as compared to double thresholds was rather small (1–2.7 dB) for the CRS channels B and C but the difference was more than 8 dB for the CRS channel A. The sensitivity gain of the MED-FCME CORϵ to single threshold approach was ranging between 9.7–15.9 dB. In fact, the sensitivity gain would have been even higher if we would have used the average DC with single threshold approach instead of the maximum combining. Finally, the mean gain of the ED to the proposed MED-FCME CORϵ was 2.6 dB. However, its data storage requirements are much higher than with the MED-FCME CORϵ. Results in Table II are for 802.11b signal. The ED had lower sensitivity than the MED-FCME CORϵ with the FM signal. B. FM Signal The ED compares the total power (including also noise power) from the whole CRS channel to a threshold. The other studied methods measure power in frequency bins and compare the individual frequency bin powers to a threshold(s). In the case of narrowband signals, much of the signal energy may be in one frequency bin. Since increasing the measurement bandwidth increases the noise power, the signal-to-noise ratio may be much higher in a frequency bin than in the whole CRS channel. The reason for the small difference between MED-FCME CORϵ and maximum combining based methods in Fig. 9 is that due to the large bandwidth of the studied

18

CRS channel (15 MHz), the number of frequency bins inside the channel was large. Thus we had to use r = 2 instead of the more sensitive r = 1. This led to some performance loss even with the correlation between frequency bins caused by using a resolution bandwidth that is larger than the frequency bin separation. However, the performance was still better than with the ED. C. MED-FCME CORϵ vs ED The ED approach showed better performance than MED-FCME CORϵ for some of the results. However, the main advantage of the proposed MED-FCME CORϵ method as compared the ED approach is much lower data storage requirements for supporting COR estimation for all possible CRS channels also off-line. In the case of the MED-FCME CORϵ, this is very realistic, since the full binary data matrix of the per-bin detection decisions for each frequency sweep can be stored for longer measurements [8], [31], even with fast frequency sweeping. Then, we form (off-line) the fused decision for each sweep by comparing the sum of binary decisions from arbitrary CRS channel with the threshold and proceed. For the ED approach, we need to sum the floating point values corresponding to the frequency bin power data for the bins inside each studied CRS channel in order to get the energy value for the CRS channel which is then compared to a threshold. Floating point values require much more storage than binary data. If a limited set of CRS channels is defined before the measurement or if the frequency sweeping is slow, this approach is practical. In other cases the ED may not be a practical approach to adopt in all circumstances. We note that the storage requirements of the double thresholds approach are minimal as we only need to store once per record the DC values Sck for all N frequency bins. D. Comparison of MED-LAD ACC against the conventional LAD in peak load condition Let us use the same signal model as in Sect. VI-C. As before, there is a WLAN signal with center frequency 2472 MHz and duty cycle 40 %. In order to simulate peak load we added two additional WLAN signals to channels 1 and 6 with -75 dBm and -70 dBm powers and 90 % duty cycle (corresponding to peak usage such as downloading large files). Note that both of these new signals are outside the CRS channel A that is considered here. However, the coexisting signals affect the threshold calculation since the threshold calculation is done based on all received frequency domain samples. Results in Fig. 10 confirm much improved COR estimation performance of the MED-LAD ACC. The MED-LAD ACC gain was around 10 dB.

19

The reason for poor performance of the conventional LAD ACC is the fluctuation of the FCME thresholds when several signals are present [20]. XI. C ONCLUSIONS We proposed and evaluated methodologies for DC and COR measurements. The proposed approaches are generic and not limited to a specific band or signal. Measurement and theoretical results showed that the MED-LAD ACC can be used for accurate DC/COR estimation, with much more sensitivity than with the conventional single threshold approach. The proposed frequency domain hard decision fusion technique called MED-FCME CORε was shown to have even more sensitivity for COR estimation, with much less storage requirements than the ED. In fact, there is no reason why both of the approaches cannot be used simultaneously so that MED-LAD ACC is used for DC estimation and the hard decision fusion is used for COR estimation. XII. ACKNOWLEDGMENTS The authors acknowledge Dr. Miguel L´opez-Ben´ıtez for helpful discussions. The authors would like to thank the anonymous reviewers for their valuable comments and suggestions. R EFERENCES [1] J. Lehtom¨aki, R. Vuohtoniemi, and K. Umebayashi, “Duty cycle and channel occupancy rate estimation with MED-FCME LAD ACC,” in Proc. CROWNCOM, Stockholm, Sweden, Jun. 2012. [2] M. L´opez-Ben´ıtez and F. Casadevall, “Methodological aspects of spectrum occupancy evaluation in the context of cognitive radio,” European Transactions on Telecommunications, vol. 21, no. 8, pp. 680–693, Dec. 2010. [3] M. Wellens, “Empirical modelling of spectrum use and evaluation of adaptive spectrum sensing in dynamic spectrum access networks,” Ph.D. dissertation, RWTH Aachen University, Germany, May 2010. [4] J. Vartiainen, H. Sarvanko, J. Lehtom¨aki, M. Juntti, and M. Latva-aho, “Spectrum sensing with LAD based methods,” in Proc. PIMRC, Athens, Greece, Aug. 2007. [5] R. Bacchus, T. Taher, K. Zdunek, and D. Roberson, “Spectrum utilization study in support of dynamic spectrum access for public safety,” in Proc. DySPAN, Apr. 2010. [6] K. Umebayashi, T. Kazmi, Y. Kamiya, Y. Suzuki, and J. Lehtom¨aki, “Dynamic selection of CWmin in cognitive radio networks for protecting IEEE 802.11 primary users,” in Proc. CROWNCOM, Osaka, Japan, Jun. 2011. [7] A. D. Spaulding and G. H. Hagn, “On the definition and estimation of spectrum occupancy,” IEEE Transactions on Electromagnetic Compatibility, vol. EMC-19, no. 3, pp. 269–280, Aug. 1977. [8] J. J. Lehtom¨aki, R. Vuohtoniemi, K. Umebayashi, and J.-P. M¨akel¨a, “Energy detection based estimation of channel occupancy rate with adaptive noise estimation,” IEICE Trans. Commun., vol. E95-B, no. 04, pp. 1076–1084, Apr. 2012. [9] H. Urkowitz, “Energy detection of unknown deterministic signals,” Proc. IEEE, vol. 55, no. 4, pp. 523–531, 1967.

20

[10] S. Geirhofer, L. Tong, and B. M. Sadler, “Dynamic spectrum access in WLAN channels: Empirical model and its stochastic analysis,” in Proc. TAPAS, Boston, MA, Aug. 2006. [11] T. M. Taher, R. B. Bacchus, K. J. Zdunek, and D. A. Roberson, “Empirical modeling of public safe voice traffic in the land mobile radio band,” in Proc. CROWNCOM, Jun. 2012. [12] T. Yucek and H. Arslan, “A survey of spectrum sensing algorithms for cognitive radio applications,” IEEE Communications Surveys Tutorials, vol. 11, no. 1, pp. 116–130, 2009. [13] J. J. Lehtom¨aki, J. Vartiainen, M. Juntti, and H. Saarnisaari, “Analysis of the LAD methods,” IEEE Signal Processing Letters, vol. 15, pp. 237–240, 2008. [14] A. Sonnenschein and P. M. Fishman, “Radiometric detection of spread-spectrum signals in noise of uncertain power,” IEEE Transactions on Aerospace and Electronic Systems, vol. 28, no. 3, pp. 654–660, 1992. [15] D. Torrieri, “The radiometer and its practical implementation,” in Proc. MILCOM, San Jose, CA, USA, Oct./Nov. 2010. [16] A. Mariani, A. Giorgetti, and M. Chiani, “Effects of noise power estimation on energy detection for cognitive radio applications,” IEEE Transactions on Communications, vol. 59, no. 12, pp. 3410–3420, Dec. 2011. [17] M. J. Ready, M. L. Downey, and L. J. Corbalis, “Automatic noise floor spectrum estimation in the presence of signals,” in Proc. ASILOMAR, vol. 1, 1997, pp. 877–881. [18] S. Liu, J. Shen, R. Zhang, Z. Zhang, and Y. Liu, “Information theoretic criterion-based spectrum sensing for cognitive radio,” IET Communications, vol. 2, no. 6, pp. 753–762, Jul. 2008. [19] H. Saarnisaari, P. Henttu, and M. Juntti, “Iterative multidimensional impulse detectors for communications based on the classical diagnostic methods,” IEEE Transactions on Communications, vol. 53, no. 3, pp. 395–398, Mar. 2005. [20] J. Lehtom¨aki, J. Vartiainen, R. Vuohtoniemi, and H. Saarnisaari, “Adaptive FCME-based threshold setting for energy detectors,” in Proc. CogART, Barcelona, Spain, Oct. 2011. [21] K. S. Miller, “Complex Gaussian processes,” SIAM Review, vol. 11, no. 4, pp. 544–567, Oct. 1969. [22] M. Lops, “Hybrid clutter-map/L-CFAR procedure for clutter rejection in nonhomogeneous environment,” Proc. Inst. Elec. Eng.-Radar, Sonar and Navigation, vol. 143, no. 4, 1996. [23] J. J. Lehtom¨aki, J. Vartiainen, M. Juntti, and H. Saarnisaari, “CFAR outlier detection with forward methods,” IEEE Transactions on Signal Processing, vol. 55, no. 9, pp. 4702–4706, Sep. 2007. [24] J. C. Fu and W. Y. W. Lou, Distribution Theory of Runs and Patterns and Its Applications: A Finite Markov Chain Imbedding Approach. River Edge, NJ, USA: World Scientific, 2003. [25] E. G. Villegas, E. Lopez-Aguilera, R. Vidal, and J. Paradells, “Effect of adjacent-channel interference in IEEE 802.11 WLANs,” in Proc. CROWNCOM, Aug. 2007, pp. 118–125. [26] B. Rivet, L. Girin, and C. Jutten, “Log-rayleigh distribution: A simple and efficient statistical representation of log-spectral coefficients,” IEEE Transactions on Audio, Speech, and Language processing, vol. 15, no. 3, pp. 796–802, 2007. [27] M. Schwartz, “A coincidence procedure for signal detection,” IRE Transactions on Information Theory, vol. 2, no. 4, 1956. [28] R. A. Dillard, “Detectability of spread-spectrum signals,” IEEE Transactions on Aerospace and Electronic Systems, vol. AES-15, pp. 526–537, Jul. 1979. [29] “Channel models for HIPERLAN/2 in different indoor scenarios,” ETSI EP BRAN 3ER1085, Mar. 1998. [30] W. Yuan, X. Wang, J.-P. Linnartz, and I. G. Niemegeers, “Experimental validation of a coexistence model of IEEE 802.15.4 and IEEE 802.11b/g networks,” International Journal of Distributed Sensor Networks, vol. 2010, 2010. [31] J. Naganawa, K. Hojun, S. Saruwatari, H. Onaga, and H. Morikawa, “Distributed spectrum sensing utilizing heterogeneous wireless devices and measurement equipment,” in Proc. DySPAN, Aachen, Germany, 2011.

21

100 90

CRS CH #1 (COR = 20 %)

80 70

Combined DC

CRS CH #2 CRS CH #3 (COR = 60 %) (COR 80−100 %) S3 with DC 80 %

DC [%]

60 50

S2 with DC 40 %

40 30

S1 with DC 20 %

20 10 0

Variation of DC within CRS channels and the resulting CORs.

Ratio of obtained and target false alarm probabilities

Fig. 1.

Fig. 2.

6 FCME m−dB method based on MED−FCME with a correction factor

5.5 5 4.5 4 3.5 3

0.005

2.5

0.0001

2 1.5 1 0.5

0

0.02

0.04 0.06 0.08 Target false alarm probability

0.1

Ratio of obtained and target false alarm probabilities for FCME and m-dB method based on MED-FCME with a

correction factor β and TCME = 4.6052, N = 916.

22

−80 −81

dBm / 242 kHz

−82 −83

m-dB method based on MED-FCME m=10 dB (γ and γU )

−84 m-dB method based on MED-FCME m=6 dB (γL )

−85 −86 −87 −88 2400

Fig. 3.

˙ = 0.03 PFA method with PFA ˙ = 10−4 PFA method with PFA 2420

2440 2460 Frequency [MHz]

2480

2500

Thresholds by the m-dB method based on MED-FCME and the PFA method, TCME = 4.6052 and M = 800.

−2

10

Simulation LAD ACC Theory LAD

Probability of false alarm

−3

10

−4

10

−5

10

False alarm probability corresponding to the upper and lower threshold [1E−3 0.05]

[1E−4 0.03]

[1E−5 0.1]

[1E−6 0.07]

−6

10 Hamming

Fig. 4.

Hanning

Kaiser

Flattop

Simulated probability of false alarm after LAD ACC processing and the theoretical result for LAD.

23

Co−existing user’s (802.11b) channel (22 MHz wide) Center frequency

CRS channel A (15 MHz wide) CRS channel B (2 MHz wide) CRS channel C (2 MHz wide) 2460

Fig. 5.

2465

2470 2475 Frequency [MHz]

2480

2485

Illustration of the studied CRS channels and the 802.11 channel.

40 35 30

double thresholds

single threshold

¯k [%] S C

25 20 15 10 5 0 2450

Fig. 6.

2455

2460

2465 2470 2475 Frequency [MHz]

2480

2485

2490

Measurement results (S¯ck ), AWGN channel, channel emulator output level −80 dBm. True DC within the WLAN

channel 13 40 %. Single threshold refers to DC estimates of m-dB method based on MED-FCME and double thresholds refers to DC estimates with MED-LAD ACC.

24

50 single threshold double thresholds MED-FCME CORǫ Channel Power (ED)

45 40

COR [%]

35

CRS channel B (802.15.4 channel with 2 MHz offset to 802.11b)

30 25 CRS channel C 20 (802.15.4 channel 7 MHz offset) 15 10

After this 802.15.4 cannot sense 802.11b 5 with 2 MHz offset [Yuan2010] 0 −60 −65 −70 −75 −80 −85 −90 Channel emulator output level [dBm]

Fig. 7.

−95

−100

Measurement results, 802.11b signal, ETSI BRAN WLAN channel model A, device speed 5 km/h, CRS channels B

and C (both 2 MHz), r = 1 (MED-FCME CORε). Single threshold refers to maximum processing of DC estimates of m-dB method based on MED-FCME and double thresholds refers to maximum processing of DC estimates with MED-LAD ACC.

50 45 40

COR [%]

35 30 25 20

IEEE 802.11−2007 [Sect. 15.4.8.4] single threshold (meas) single threshold (theory) double thresholds (meas) double thresholds (theory) MED−FCME COR ε meas) Channel Power/ED (meas)

15 10 5 0 −60

Fig. 8.

−65

−70 −75 −80 −85 −90 Channel emulator output level [dBm]

−95

−100

Measurement and theoretical results, 802.11b, ETSI BRAN WLAN channel model A, device speed 5 km/h, CRS

channel A (15 MHz), r = 2 (MED-FCME CORε). Single threshold refers to maximum processing of DC estimates of m-dB method based on MED-FCME and double thresholds refers to maximum processing of DC estimates with MED-LAD ACC. For theory block Rayleigh fading channel, receiver gain + cable loss 17.3 dB, noise power -95.1 dBm.

25

100 90 80

COR [%]

70 60 50 40 30 20

single threshold double thresholds MED-FCME CORǫ Channel Power (ED)

10 0 −80

Fig. 9.

−85 −90 −95 −100 Channel emulator output level [dBm]

−105

Measurement results, wideband FM signal, AWGN channel, CRS channel A (15 MHz), r = 2 (MED-FCME CORε).

Single threshold refers to maximum processing of DC estimates of m-dB method based on MED-FCME and double thresholds refers to maximum processing of DC estimates with MED-LAD ACC.

50 LAD ACC (simul) MED−LAD ACC (simul)

45 40

COR [%]

35 30 25 20 15 10 5 0 −60

Fig. 10.

−65

−70

−75 −80 −85 −90 WLAN Signal power [dBm]

−95

−100

Simulated results in peak load condition, 802.11b signal in AWGN channel, CRS channel A (15 MHz), actual COR

40 %. LAD ACC refers to maximum processing of DC estimates of the conventional LAD ACC and MED-LAD ACC refers to maximum processing of DC estimates with MED-LAD ACC, receiver gain + cable loss 17.3 dB, noise power -95.1 dBm. Additional co-existing WLAN signals present in channels 1 and 6.

26

TABLE I M EASUREMENT SETUP

Instrument

Agilent N6841A RF sensor

Center frequency

2450 MHz

Frequency span

100 MHz

Resolution bandwidth

242.27 kHz

Frequency bin separation

109.3750 kHz

Window type

Gausstop window

FFT size (in a frequency segment)

256

Digital IF bandwidth

20 MHz

Number of frequency points

N = 916

Sweep time

∼10 ms

Power spectrum average type

No averaging

Antenna

Linksys ISM band antenna

Frequency range

2400–2500 MHz

Gain

7 dBi

Filter

Creowave ISM band filter

Band-pass frequency range

2400–2500 MHz

Insertion loss (pass-band)

0.8 dB

Rejection bands

DC–2300 MHz, 2600–5500 MHz

Rejection at rejection bands

≥90 dB

Low noise amplifier

Mini-Circuits ZRL-3500

Frequency range

700–3500 MHz

Gain (2.4 GHz ISM band)

19.5 dB

Noise Figure (2.4 GHz ISM band)

2.5 dB

Sensitivity

around −100 dBm/242 kHz

27

TABLE II M INIMUM CHANNEL EMULATOR OUTPUT LEVEL [dBm] TO GET COR ESTIMATION RESULT OF 40±1 % (39–41 %) IN THE STUDIED

CRS CHANNELS WHEN THE REAL COR OF THE 802.11 B SIGNAL IS 40 %, FADING REFERS TO ETSI BRAN

WLAN CHANNEL MODEL A WITH DEVICE SPEED 5 km/h. S INGLE THRESHOLD REFERS TO MAXIMUM PROCESSING OF DC ESTIMATES OF

m-dB METHOD BASED ON MED-FCME AND DOUBLE THRESHOLDS REFERS TO MAXIMUM PROCESSING OF DC ESTIMATES WITH MED-LAD ACC.

Channel Power (ED)

MED-FCME CORϵ

double thresholds

single threshold

fading

AWGN

fading

AWGN

fading

AWGN

fading

AWGN

CRS CH #A

-85.7

-90.3

-82.2

-87.2

-73.8

-82.4

-66.3

-73.0

CRS CH #B

-77.8

-86.6

-74.3

-84.8

-73.3

-82.1

-64.6

-72.4

CRS CH #C

-66.8

-76.1

-65.2

-74.2

-62.1

-73.2

(>-60)

-61.7