1076IEICE TRANS. COMMUN., VOL.E95–B, NO.4 APRIL 2012

PAPER Special Section on Cognitive Radio and Heterogeneous Wireless Networks in Conjunction with Main Topics of CrownCom2011

Energy Detection Based Estimation of Channel Occupancy Ratewith Adaptive Noise Estimation∗

Janne J. LEHTOMAKI†a), Member, Risto VUOHTONIEMI†, Nonmember, Kenta UMEBAYASHI††, Member,and Juha-Pekka MAKELA†, Nonmember

SUMMARY Recently, there has been growing interest in opportunis-tically utilizing the 2.4 GHz ISM-band. Numerous spectrum occupancymeasurements covering the ISM-band have been performed to analyze thespectrum usage. However, in these campaigns the verification of the cor-rectness of the obtained occupancy values for the highly dynamic ISM-band has not been presented. In this paper, we propose and verify chan-nel occupancy rate (COR) estimation utilizing energy detection mechanismwith a novel adaptive energy detection threshold setting method. The re-sults are compared with the true reference COR values. Several differenttypes of verification measurements showed that our setup can estimate theCOR values of 802.11 traffic well, with negligible overestimation. The re-sults from real-time real-life measurements also confirm that the proposedadaptive threshold setting method enables accurate thresholds even in thesituations where multiple interferers are present in the received signal.key words: occupancy measurements, cognitive radio, wireless LAN

1. Introduction

In energy detection (ED), the received energy (power) ismeasured in a channel and compared to a threshold γ to as-sess if signal(s) are present or if only noise was received [1].Spaulding and Hagn [2] defined spectrum occupancy for EDas the fraction of the measurement time that the detectedpower P in a channel exceeds a threshold level. Spectrumoccupancy measurements are important to verify the avail-ability of the white space for a cognitive radio system andthey are also important for radio regulators to see the effec-tiveness of the frequency allocations. The important issuesin these measurements are noise uncertainty, resistance tooutliers and sensitivity.

Spectrum occupancy measurements have been per-formed, e.g., in [3]–[8]. One of the bands of interest for oc-cupancy measurements is the 2.400–2.500 GHz Industrial,Scientific and Medical (ISM) band [3]. In this band, sev-eral systems exist such as the popular 802.11 Wireless Lo-cal Area Network (WLAN) networks and Bluetooth. Therecan also be microwave ovens causing interference. Most of

Manuscript received August 17, 2011.Manuscript revised November 7, 2011.†The authors are with the Centre for Wireless Communications

(CWC), University of Oulu, Finland.††The author is with the Department of Electrical and Com-

puter Engineering, Tokyo University of Agriculture and Technol-ogy, Koganei-shi, 184-8588 Japan.

∗This paper was presented in part at the International Confer-ence on Cognitive Radio and Advanced Spectrum Management(CogART), Barcelona, Spain, 2011.

a) E-mail: jannel@ee.oulu.fiDOI: 10.1587/transcom.E95.B.1076

the previous measurements have measured a much broaderbandwidth and the ISM-band has been just one of the fre-quency bands swept over [4]. In [5], extensive spectrummeasurements in the ISM-band were performed in an of-fice environment and statistical characterization of spectrumutilization was done. Directional spectrum occupancy mea-surements in the ISM-band have been performed in [6] andit was found that there can be significant differences betweendifferent directions.

The ED is easy to implement and it can make detectiondecision even in a short time. However, the noise uncer-tainty can be a significant problem with the ED [9]. It mayeven lead to a phenomena called “SNR wall”, where noiseuncertainty prevents detection of weak signals [10]. Rea-sons for the noise uncertainty include thermal noise levelchanges due to variable temperature and ambient noise (dueto atmospheric noise, galactic noise and man-made noiseamong others). The man-made noise is caused, for example,by unintended radiation from electrical machinery, electricaland electronic equipment and power transmission lines. Thepower level of this man-made noise varies with the time ofday, polarization and frequency. A variable level of ambientnoise has been observed even at 4 GHz [8].

Using adaptive approaches for threshold setting in EDleads to more robustness against the fluctuation of noisefloor compared to the typical ED approach with a fixedthreshold. In [11], adaptive threshold was found using atwo-step process. First, a reference set was formed fromsome of the previously received energy values in differentchannels and/or time intervals. As there is no guaranteethat all of the received energy values are due to only noise,censoring techniques (such as order statistics or the blindand iterative forward consecutive mean excision (FCME) al-gorithm) were used to remove the corrupted energy values(containing also signal components). The found supposedlysignal-free energy values were used to calculate the detec-tion threshold. In [12], a longer term background processwas used to identify frequency bands where only noise ispresent and the noise variance is estimated from these bands.In [13], the FCME algorithm [14] was used for adaptivethreshold setting. Although being effective and computa-tionally simple, a problem with the FCME algorithm is thatsometimes it can give too large thresholds.

The true channel occupancy rate (COR) for WLANtraffic can be defined to be the actual time domain occu-pancy. Most of the previous measurements have used stan-

Copyright c© 2012 The Institute of Electronics, Information and Communication Engineers

LEHTOMAKI et al.: ENERGY DETECTION BASED ESTIMATION OF CHANNEL OCCUPANCY RATE WITH ADAPTIVE NOISE ESTIMATION1077

dard spectrum analyzers. Our own experiments with a stan-dard spectrum analyzer (Agilent E4446A) showed that a sig-nificant COR bias can be caused, e.g., by parameters relatedto successive sweep averaging and sweep time. Since in theprevious measurements verification of the used spectrum an-alyzer parameters for the 2.45 GHz ISM-band has not beenpresented, it is uncertain how accurate the estimated occu-pancies are.

In addition to measurement campaigns in the ISM-band, accurate WLAN COR estimation is needed by cog-nitive radio networks that can protect other 802.11 users,considered as primary users, by adaptively changing theirminimum contention window size according to the COR ofthe incumbent user(s) [15].

In this paper, we apply ED with adaptive thresholdingand verify by measurements the COR estimation accuracyfor 802.11 traffic. Our main contributions are that:

1. We propose and apply an improvement to the FCMEcalled median (MED) FCME algorithm. The MED-FCME enables accurate noise floor estimation with sig-nificantly improved robustness against large outliers.The resistance to outliers is very important in real-lifemeasurements (where we do not control the use of theradio spectrum) as otherwise incorrect results can beobtained due to too large threshold. This method isgeneric and not limited to the ISM-band.

2. We verify the validity of our FFT-based frequencysweeping ED setup for WLAN COR measurements bycomparing results with the reference (true) COR infor-mation. This has not been presented earlier. In fact, wenoticed overestimation with averaging, so we turnedaveraging off.

3. We also measure COR by a WLAN card in the monitormode, where the card will receive all WLAN packets.The results showed that COR estimates by our flexibleED setup were more accurate than WLAN card (whichcould be considered as a primary receiver) based CORestimates when the signal strength was low.

2. Measurement Setup

We consider the following measurement approaches;

1. 802.11 WLAN card in the monitor mode,2. ED with fixed threshold and frequency sweeping,3. ED with FCME for adaptive threshold setting and fre-

quency sweeping,4. ED with MED-FCME for adaptive threshold setting

with median filtering and frequency sweeping.

There are total of fourteen overlapping WLAN chan-nels each with 22 MHz channel bandwidth in the 2.4 GHzISM-band. The approach #1 is only valid in the specificchannel where the WLAN card is tuned to, and only forWLAN signals. Approach #1 can be used to assess per-formance of a WLAN card for receiving WLAN packets.The approaches #2–#4 employ frequency sweeping to cover

the whole 100 MHz ISM-band and they can detect all kindsof signals (e.g. Bluetooth). The used resolution bandwidth(RBW) sets some limits on the resolvable signals. However,in principle, there are no limitations on the bandwidths ofthe signals to be detected.

The approach #2 requires manual work to set the noiselevel and has validity only if the actual noise floor is invari-ant. The FCME algorithm is used for the threshold setting inthe approach #3 and this adaptive approach has robustnessfor noise floor fluctuation (e.g. when temperature is chang-ing). The aim of our proposed median filtering in the ap-proach #4 is to have more robustness against outliers andto reduce variance of noise floor estimation. The rationalefor the approach #4 is that the noise floor is usually static atleast for a few minutes [16] so it is not needed to form anindependent noise floor estimate in every frequency sweep.

2.1 Energy Detection

The ED setups for approaches #2 – #4 and the used param-eters are listed in Table 1 and the measurement process isillustrated in Fig. 1. Due to a high frequency sweep rate ofthe Agilent RF sensor, the revisit time between measure-ments in the same bin was just ∼20 ms; so that there are lessgaps between the measurements in the same bin as com-

Table 1 Measurement configuration.

Instrument Agilent N6841A RF sensorCenter frequency fc = 2450 MHzFrequency span 100 MHzResolution bandwidth 242.27 kHzFrequency bin separation 109.3750 kHzNumber of frequency points N = 916Sweep time 19.5 msAverage type OffThreshold 10 dB above noise floor

Antenna Linksys ISM-band dipole antennaFilter Creowave ISM-band filterLow noise amplifier Mini-Circuits ZRL-3500

FCME parameter TCME = 4.6052MED-FCME filtering length M = 3000 (around one minute)Record duration T = 200 sweeps (3.9 s)

Signal generation Agilent E4438C ESG+ Agilent Signal Studio software

Fig. 1 Energy detection measurement process.

1078IEICE TRANS. COMMUN., VOL.E95–B, NO.4 APRIL 2012

pared many of the previous measurement campaigns. Sincewe noticed that even root-mean-square (RMS) averaging oftwo samples caused COR overestimation with our FFT (FastFourier Transform) based RF sensor, we did not use averag-ing. Thus, the frequency domain samples in the noise-onlycase follow the (scaled) chi-squared distribution with twodegrees of freedom or equivalently the exponential distribu-tion [17].

We define a record to have duration of T = 200 sweepsover the whole ISM-band. Detection threshold 10 dB abovethe noise floor (NF) was used, i.e., γi = NFi + 10 [dBm],where NFi is the noise floor estimate for the sweep within arecord with index i and γi the threshold for the sweep withindex i. This is a rather sensitive threshold, for example theFluke Networks AnalyzeAir Wi-Fi Spectrum Analyzer usesa threshold 20 dB above the noise floor for duty cycle mea-surements. In [5], the ED measurements had L = 8 degreesof freedom corresponding to two times the time-bandwidthproduct. Because one benefit of averaging is reduced out-put variance, they could use as a threshold a value just 6 dBabove the noise power.

The measurement process is illustrated in Fig. 2, show-ing one record from a measurement performed close to theRotuaari shopping street in the Oulu city centre. The re-ceived frequency domain power for a record Pk

i,c is com-pared to the adaptive detection threshold γi, resulting intobinary data Dk

i,c, i.e.,

Fig. 2 Illustration of the ED processing for one record k: a) raw fre-quency domain data Pk

i,c [dBm]; b) binary data Dki,c (1 or 0); c) occupancy

data S kc [%]. Measurement performed close to the Rotuaari shopping street

in the Oulu city centre in June 2011.

Dki,c =

{1, Pk

i,c > γi

0, otherwise(1)

where k is the record index, i = 1, 2, · · · , T is the sweep in-dex, and c = 1, 2, · · · , 916 is the frequency bin index. An ex-ample of the power data for one record is shown in Fig. 2 a)and an example of the corresponding binary data is shown inFig. 2 b). The estimated occupancy value S k

c for a given fre-quency bin c was found by calculating the number of timesthat the received signal power exceeded the detection thresh-old (number of ones in the binary data), divided by the totalnumber of sweeps in the record, i.e.,

S kc =

∑Ti=1 Dk

i,c

T. (2)

An example of the per-bin occupancy data S kc is shown in

Fig. 2 c). These processes were implemented in real-timewith a program made with the C programming language.The sweep speed did not have to be reduced due to storageissues since with real-time processing the raw frequency do-main data Pk

i,c does not need to be saved. Instead, the full bi-nary data Dk

i,c for each sweep, frequency bin, and record wasrecorded with run-length encoding (RLE) compression in abit-packed format. Due to these, one week measurementswith around 20 billion data points were possible.

Since each WLAN channel has a nominal bandwidth ofaround 22 MHz, the number of frequency bins inside eachWLAN channel c2 − c1 + 1 =∼201. The per-bin occupancyvalues corresponding a frequency subband can be combinedwith averaging [18] and we refer to this as AV (Average)-combining. However, we use instead the maximum of theper-bin occupancy estimates (MAX-combining), i.e., theCOR Lk

s corresponding to the WLAN channel s is given by

Lks = max

c=c1,···,c2S k

c = maxc=c1,···,c2

∑Ti=1 Dk

i,c

T(3)

where is s = 1, 2, · · · , 14 is the WLAN channel number andthe c1 and c2 are the first and the last frequency bins insidethe WLAN channel s, respectively. The rationale for usingthe MAX-rule is that even if a part of the channel is occu-pied, the channel should be counted as occupied for oppor-tunistic radio use (since otherwise we may harm the otherusers). Also, the MAX-rule could achieve frequency do-main diversity gain so that a few channel nulls caused bymultipath do not significantly affect the estimation. In ad-dition, when using the MAX-rule it is not required to as-sess signal’s sidelobe widths; with AV-rule including the(possibly weak) sidelobes may cause underestimation. Fig-ure 3 shows a sample COR result from energy detection forWLAN channel 13 over a time period of around 26 min-utes (400 records). The records were measured continu-ously, without time gaps between successive records. Themeasurement was performed in an anechoic chamber andthe reference occupancy was stepped from 10 percent to 90percent. The accuracy of the MAX-combining of frequencybins inside the studied WLAN channel can be confirmed.

LEHTOMAKI et al.: ENERGY DETECTION BASED ESTIMATION OF CHANNEL OCCUPANCY RATE WITH ADAPTIVE NOISE ESTIMATION1079

Fig. 3 An example of anechoic chamber measurement results Lks by ED

for WLAN channel 13. Measurement duration 26 minutes. The referenceoccupancy is stepped from 10 percent to 90 percent. The signal generatoroutput power β = 0 dBm.

For details about the measurement arrangement, see Sect. 6.The reference occupancy refers to the true COR (taking intoaccount the preamble duration, packet size and the data rateused to transmit the packet) and it was controlled by sig-nal generator parameters. Let us denote the signal generatoroutput power with β. In this sample result, the signal gener-ator output power was set to β = 0 dBm.

Since the occupancy estimates are saved per-record, itis possible to combine by averaging in post-processing per-bin occupancy estimates S k

c or CORs Lks from K records (e.g.

records k from k1 to k2 with K = k2 − k1 + 1) to form alsoestimates (in addition to the per-record ones) with less vari-ability and higher resolution (but also with less dynamicity).The number of records K needed to be combined for a givenaccuracy level can be estimated from [2], for both indepen-dent and dependent sampling.

2.2 Packet Analyzer

As a measurement reference/primary user receiver weused a Linux laptop with external D-Link DWL-AG650802.11a/b/g Wireless Cardbus Adapter. We also used amodified D-Link DWL-AG650 card with antenna connec-tor. The card was set to the monitor mode. The monitormode allows all traffic of the studied wireless channel tobe captured and monitored. A modified Wireshark networkprotocol analyzer was used to record the occupancy infor-mation, taking into account the factors affecting the occu-pancy (such as the preamble). Time resolution of 1 secondwas used for recording the obtained occupancy values. Fig-ure 4 shows a sample result of the CORs from the packetanalyzer. Sometimes the packet analyzer failed to receiveall packets after 70 percent reference occupancy. However,since our verification measurements used smaller occupancyvalues (20 and 40 percent), this did not affect them.

Fig. 4 An example of anechoic chamber measurement results by apacket analyzer tuned to WLAN channel 13. Measurement duration 26minutes. The reference occupancy is stepped from 10 percent to 90 per-cent. The signal generator output power β = 0 dBm.

3. Noise Floor Estimation

For many cases, practical noise floor estimation should bebased on the actual received samples. With this approach,noise floor estimation can be done simultaneously with theactual measurements. However, the significant problemwith this approach is that it is difficult to find which of thereceived samples correspond to just noise (the samples wewant to locate) and which samples have also signal com-ponents (undesired samples). This is important as if the setfrom which we estimate the noise floor from has signal com-ponents, the estimated noise floor can be significantly toolarge even due to one very high power signal sample.

3.1 FCME

The FCME algorithm [14] is a multipurpose iterativemethod that can be used for finding which of the receivedsamples correspond to just noise. The details of the FCMEwill not be explained here but we briefly explain the out-line of the FCME. The FCME first rearranges the frequencydomain samples (that are in linear scale) in an ascendingorder according to their energies. After that, the FCMEalgorithm calculates mean of a small initial set which isassumed to be free of interference. The larger the initialset, the simpler the algorithm is because less iterations areneeded. However, the possibility that the initial set is notclean increases. On the other hand, the smaller the initialset is, the higher the possibility that the initial threshold istoo small. In that case, it may be possible that the algorithmdoes not converge at all. Usually, the size of the initial setis 10% of all the samples. The threshold is the mean multi-plied with the threshold parameter TCME. The value of thethreshold parameter is calculated based on the desired clean

1080IEICE TRANS. COMMUN., VOL.E95–B, NO.4 APRIL 2012

sample rejection rate (CSRR) in the noise-only case. Basi-cally, the CSRR means how many clean (noise-only) sam-ples are (on average) wrongly classified by the FCME asoutliers having signal components. Commonly used valuesare TCME = 4.6052, 6.9078 and 9.2103 [17]. The new valuefor mean is calculated from the samples below the thresh-old. The FCME algorithm iteratively calculates a new valuefor mean and a new threshold until there are no new samplesbelow the threshold. The FCME algorithm does not requireknown vacant channels. Drawback of this approach is thatif there are not enough noise-only samples in the current setof samples (such as the frequency domain samples from thestudied band from the current frequency sweep), the noisefloor estimation will be too high, even by a large amount.

3.2 MED-FCME

The proposed improvement to the FCME algorithm calledMED-FCME uses a median filter on the previous noisefloor estimates from the FCME algorithm. Let us denotethe sorted (in ascending order) FCME noise floor estimatesfrom the last M sweeps (including the current sweep withindex i) with bk, where k = 1, 2, · · · ,M. Assuming even M,the median of these estimates, i.e., the MED-FCME noisefloor estimate, is given by

NFi =bM/2 + bM/2+1

2. (4)

Although this concept is a very simple modification to theFCME, it can significantly improve noise floor estimationaccuracy. It is obvious that by using median filtering,the variability of the estimates is reduced (Benefit 1). Asthe FCME algorithm is mainly a narrowband signal detec-tion/excision method, it can catastrophically fail when a sig-nificant portion of the studied band is currently occupied bysignal(s). However, the median will not fail when some ofthe FCME noise floor estimates it is based on are too large(Benefit 2). This is because the median is a robust measurewith 50% breakdown point. This mean that the proportionof incorrect observations can be even 50% before completefailure. If standard averaging would be used, even muchsmaller proportion of incorrect observations can lead to fail-ure. It could be said the proposal can take advantage of notonly signal-free areas in frequency (as the FCME algorithm)but also signal-free time intervals. For example, packetbased transmissions, even if wide-band, usually have gapsbetween the packets. Specially, in the CSMA/CA basedWLAN, there could be noise free samples intermittently be-cause of Inter Frame Spacing (IFS) or back-off time. In thiscase, the MED-FCME algorithm can estimate the noise floorfrom the time intervals between packets. Both the MED-FCME and FCME algorithms can handle even continuousnarrowband signals. The boundary issues when the medianfiltering is applied to FCME are dealt with by using the con-ventional FCME thresholds in the first M sweeps.

4. Analysis in the Noise-Only Case

The used TCME = 4.6052 corresponds to target CSRR = 1%[17] so that one out of hundred clean (noise-only) samplesare rejected as outliers. However, from [19] the actual ob-tained CSRR is approximately CSRR = 1.3% (this was alsoverified by simulations). The exact number of rejected sam-ples in any given sweep is random. However, let us approx-imate the FCME as always censoring �CSRR ∗ N� largestsamples (since the FCME performs ordering of the sam-ples). The average bias in the estimated noise floor with rcensored clean samples can be obtained by using the theoryfor the weighted sum of order statistics of exponential ran-dom variables [20]. By using this average bias and the targetthreshold (10 dB above the estimated noise floor), we canestimate the obtained per-bin false alarm probability PFA.The per-bin false alarms cause a bias level in the final CORestimates Lk

s since due them the estimation output will benon-zero. The average value of this bias level in COR esti-mates due to false alarms can be obtained by using the the-ory of order statistics of binomial populations [21]. Fromthe above considerations, the average theoretical COR biasφ is obtained as

φ = 1/T×

T∑x=0

⎛⎜⎜⎜⎜⎜⎜⎜⎜⎜⎜⎜⎜⎜⎜⎜⎜⎜⎜⎜⎜⎝x

⎡⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎣

F

⎛⎜⎜⎜⎜⎜⎜⎝x, T, e⟨−10

N−r∑m=1

(N−r−m+1)(N+1−m)(N−r)

⟩⎞⎟⎟⎟⎟⎟⎟⎠c2−c1+1

−F

⎛⎜⎜⎜⎜⎜⎜⎝x − 1, T, e

⟨−10

N−r∑m=1

(N−r−m+1)(N+1−m)(N−r)

⟩⎞⎟⎟⎟⎟⎟⎟⎠c2−c1+1

⎤⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎦

⎞⎟⎟⎟⎟⎟⎟⎟⎟⎟⎟⎟⎟⎟⎟⎟⎟⎟⎟⎟⎟⎠(5)

where r = �CSRR∗N�, F(x, T, p) is the binomial cumulativedistribution function with T trials and success probabilityp. The above result is an approximation since the medianfiltering does not exactly correspond to averaging. By usingc2− c1+ 1 = 201 (the number of bins in a WLAN channel),CSRR = 1.3%, T = 200, and N = 916, results into φ =0.493% (theory). The average COR-bias with MED-FCMEbased on 193200 CORs was φ = 0.489% (measurement).

5. Cable Measurements

The purpose of the cable measurements was to verify andto find the sensitivity values of the ED and packet analyzersystems in the case where signal is fed directly to the in-put. The results shown are averages from the records duringwhich the reference COR and β were kept the same (around2 minutes, i.e., K = 30 records).

5.1 Arrangement

The signal was generated with Agilent E4438C ESG Vec-tor Signal Generator and the Agilent Signal Studio softwarefor WLAN. Different reference occupancy levels were ob-tained by changing the idle time between packets. The sig-nal was transmitted using a short cable to the filter input of

LEHTOMAKI et al.: ENERGY DETECTION BASED ESTIMATION OF CHANNEL OCCUPANCY RATE WITH ADAPTIVE NOISE ESTIMATION1081

Fig. 5 Cable measurement results by energy detection and by a 802.11receiver in the monitor mode, 802.11b signal with 11 Mbit/sec data rate,total packet size 1508 bytes.

energy detection system or to the external antenna connec-tor of the WLAN card. Due to the short high-quality cable,the received signal powers (denoted with α) were practicallyequal to the signal generator’s output power, so that α = β.

5.2 Results

The results in Fig. 5 show that when the signal power fromthe signal generator was β = −70 dBm or more, the esti-mated occupancy values were almost exactly the real val-ues for both the energy detection and the packet analyzer.When β = −77 dBm, the signal was still found but the esti-mated occupancy values were now around 85–90 percent ofthe real values for the ED and around 58 percent of the realvalue for the WLAN card. With β = −80 dBm, packet ana-lyzer failed to receive the packets (only 8 percent of themwere received). The estimated occupancy values by EDwere still more than 70 percent of the real values. When thesignal power was further reduced by using β = −85 dBm,the estimated occupancy values by energy detection also de-graded and were less than 40 percent of the real values. En-ergy detection failed to find the signal with β = −90 dBmas occupancies were 10 percent or less as compared to thereal values. The gain from ED compared to the WLAN carddid not come from false alarms. Based on above, we maycall −77 dBm (or −75 dBm) the sensitivity of the WLANcard and −80 dBm the sensitivity of the ED system, whenantenna gains are not included.

6. Anechoic Chamber Measurements

The purpose of the anechoic chamber measurements was tofurther verify the ED system in a more realistic case. Also,we compared the whole ED system to the whole WLANcard based system, including antenna gains. The resultsshown are averages from the records during which the refer-

Fig. 6 Anechoic chamber measurement results by energy detection,802.11b signal with 11 Mbit/sec data rate, total packet size 1508 bytes.

ence COR and β were kept the same (around 2 minutes, i.e.,K = 30 records).

6.1 Arrangement

The anechoic chamber of the University of Oulu was usedto carry out the measurements. The chamber has frequencyrange of 30 MHz to 20 GHz and it is 11 m long, 6.5 m wide,and 5.5 m high. The same arrangement as with the cablemeasurements was used except that a transmit antenna wasused instead of a cable and the receivers used their antennasto receive the signal. The cable loss from the signal gener-ator to the transmit antenna was measured to be 10 dB. Thehigh loss value was due to a long cable length between theantenna and the signal generator. The distances from thetransmit antenna to both the 802.11 card used for reference(its internal antenna was used) and the antenna of the en-ergy detector were both 3.6 meters, so the free space pathloss formula gives 51.4 dB.

6.2 Results

For signal generator output power β = 0 dBm (−10 dBm atthe transmit antenna input), the received power α at the Ag-ilent RF Sensor was −62 dBm, so the combined loss includ-ing antennas from the signal generator to ED was around62 dB, which agrees relatively well with the calculated freespace attenuation of 51.4 dB combined with the cable lossof 10 dB. Thus, for the ED system α ∼ β − 62 dBm.

The results for the energy detector in Fig. 6 show thatwhen β = −10 dBm or more so that received signal powerα = −72 dBm or more, the estimated occupancy values werealmost exactly the true values. With β = −20 dBm (receivedpower α = −82 dBm), the signal was still found but the esti-mated occupancy values were now around 65 percent of thereal values. When β was further reduced to −25 dBm so that

1082IEICE TRANS. COMMUN., VOL.E95–B, NO.4 APRIL 2012

α = −87 dBm, the estimated occupancy values were only 20percent of the real values. Energy detection failed to find thesignal when β = −30 dBm (received power α = −92 dBm)or less. These results agree well with the cable measure-ments. The combined loss to the WLAN card was muchhigher than for the ED, around 76 dB. For optimized orien-tation, we managed to reduce the loss to 67 dB. This is still5 dB higher loss than with the ED system.

7. Corridor Measurements

The purpose of the corridor measurements was to ver-ify the estimation performance in the presence of a realWLAN traffic including acknowledgment (ACK) packetsand Carrier Sense Multiple Access With Collision Avoid-ance (CSMA/CA). Another purpose was to compare noisefloor estimation accuracy of the FCME and MED-FCMEapproaches.

7.1 Noise Floor Estimation

7.1.1 Arrangement

We performed a separate corridor measurement where wegenerated a significant WLAN activity on two WLAN chan-nels. These WLAN signals act as interfering signals to thenoise floor estimation, to test the robustness of the estima-tion. At around record index k = 50, we switched offthe transmitters to get only noise. In this particular mea-surement, we focused on the first half of the ISM-band toget higher relative occupancies. This measurement usedslightly different parameters as compared to the other mea-surements, so the noise floor level was somewhat higher.However, this does not affect the validity of the comparisonbetween the FCME and MED-FCME.

7.1.2 Results

Although the situation was challenging, the results in Fig. 7show that the MED-FCME could still estimate the noisefloor with enough accuracy as compared to the true noisefloor value of ∼ −89 dBm. The reduction in variability dueto MED-FCME (Benefit 1) and the outlier resistance due toMED-FCME (Benefit 2) can be clearly seen in the record in-dices 1 < k < 50. Even after the transmitters were switchedoff (k > 50), the Benefit 1 is still significant.

7.2 COR Estimation

7.2.1 Arrangement

The corridor measurements were performed in a standardoffice environment in the University of Oulu. The signalwas generated with a transmitter-receiver pair composed ofa laptop and Linksys WRT54G access point. Traffic wasgenerated using the Bwping program. The target occupancy

Fig. 7 The FCME and MED-FCME algorithms, 2400–2450 MHz band(half of the ISM-band), TCME = 4.6052, 916 frequency bins.

Fig. 8 Time domain corridor measurement results by energy detectionand by a 802.11 receiver in the monitor mode, 802.11b signal.

was approximately obtained by controlling the Bwping pa-rameters. The measurement was performed in the evening,so there was not much other traffic present. The transmitter-receiver pair was moved to different measurement pointsfurther away from the fixed location of the ED and thepacket measurement laptop using WLAN card in a officeroom (with door closed).

7.2.2 Results

Figure 8 shows the estimated COR as a function of timefor the measurement performed just 16.6 meters away fromthe sensors in the office room, i.e., signal strength was high.It can be seen that the per-record estimates by ED (beforecombining several records) have a higher variance, due toeach occupancy value shown being based on T = 200 mea-surements for each 3.9 s record. The WLAN card is con-

LEHTOMAKI et al.: ENERGY DETECTION BASED ESTIMATION OF CHANNEL OCCUPANCY RATE WITH ADAPTIVE NOISE ESTIMATION1083

Fig. 9 Corridor measurement results by energy detection and by a 802.11receiver in the monitor mode, 802.11b signal.

tinuously receiving in the studied channel. However, if thesweeping would not be employed in the ED system, vari-ance for the same measurement period would be much lesssince then it is possible to get more measurements from thestudied channel in the same time (3.9 s). Also, even with theWLAN card, the variance increased significantly when thereference COR was increased from 20 percent to 40 percent.

The record averaged results in Fig. 9 show that alsoin this case with a real transmitter-receiver pair, the EDhad a superior COR estimation performance. Both ED andWLAN card could estimate the COR well when the distancewas less than d < 20 m. After that point, there was a thickdoor and wall and so the estimation performances were re-duced. However, even with 40 meters separation (includ-ing several thick walls and doors) between the transmitter-receiver pair and the ED antenna, the estimated COR wasstill roughly half of the true value. At this point, the WLANcard COR was only ∼ 25% of the true COR. Some variabil-ity can be seen in the results for the WLAN card, likely dueto the antenna orientation issues and/or local shadowing.

8. City-Center Measurements

The purpose of the city-center measurements was to verifythe noise floor estimation accuracy of the MED-FCME in areal situation.

8.1 Arrangement

One real-life week measurement was performed close tothe Rotuaari shopping street in the Oulu city centre in June2011. The results shown in Fig. 2 correspond to one partic-ular record from this measurement.

8.2 Results

Figure 10 shows the distribution of the noise floor estimates

Fig. 10 Noise floor estimation with FCME and MED-FCME algorithms.Measurement performed close to the Rotuaari shopping street in the Oulucity centre in June 2011.

with both the conventional FCME and the proposed MED-FCME. The maximum estimation error with the MED-FCME was 1 dB as the real noise floor measured without an-tenna connected was ∼ −91.5 dBm. Instead, the estimationerror with the FCME ranged from −1 to 4 dB. Both under-estimation and overestimation of the noise floor can causeproblems. The MED-FCME did not underestimate.

9. Conclusions

A flexible FFT-based frequency sweeping energy detectionsetup for ISM band COR measurements was introduced.Also, a blind and effective noise floor estimation methodbased on the FCME algorithm was proposed. The pro-posed method is based on utilizing the already receivedfrequency domain samples so that no dedicated time peri-ods are needed in the measurements to estimate the noisefloor. The measurement results showed that the proposalyields significantly improved performance as compared tothe baseline FCME algorithm. Several different kind ofmeasurements verified that the utilized energy detection sys-tem (with the proposed adaptive threshold setting) can esti-mate the channel occupancy rates (CORs) with high accu-racy. The sensitivity of the introduced setup was exceedingthe performance of a WLAN card based system. Due to thefull coverage frequency sweeping of the ISM-band in theED system, the resultant variance was found to be slightlyhigher compared to the WLAN card based system whereonly fraction of the whole ISM-band could be monitored si-multaneously. The results indicate that the proposed systemand methodology can be used for efficient frequency bandoccupancy monitoring in the future cognitive radio systems.

Acknowledgments

The research leading to these results was derived from the

1084IEICE TRANS. COMMUN., VOL.E95–B, NO.4 APRIL 2012

European Community’s Seventh Framework Programme(FP7) under Grant Agreement number 248454 (QoSMOS).

References

[1] H. Urkowitz, “Energy detection of unknown deterministic signals,”Proc. IEEE, vol.55, pp.523–531, 1967.

[2] A.D. Spaulding and G.H. Hagn, “On the definition and estima-tion of spectrum occupancy,” IEEE Trans. Electromagn. Compat.,vol.EMC-19, no.3, pp.269–280, Aug. 1977.

[3] M. Biggs, A. Henley, and T. Clarkson, “Occupancy analysis of the2.4 GHz ISM band,” IEE Proc. Commun., vol.151, no.5, pp.481–488, Oct. 2004.

[4] M. Lopez-Benıtez and F. Casadevall, “On the spectrum occupancyperception of cognitive radio terminals in realistic scenarios,” Proc.CIP, pp.99–104, Elba, Italy, June 2010.

[5] L. Stabellini, “Quantifying and modeling spectrum opportunities ina real wireless environment,” Proc. IEEE WCNC, Sydney, April2010.

[6] M. Matinmikko, M. Mustonen, M. Hoyhtya, T. Rauma, H.Sarvanko, and A. Mammela, “Distributed and directional spectrumoccupancy measurements in the 2.4 GHz ISM band,” Proc. ISWCS,pp.976–980, York, UK, Sept. 2010.

[7] M.H. Islam, C. Koh, S. Oh, X. Qing, Y. Lai, C. Wang, Y.-C. Liang,B. Toh, F. Chin, G. Tan, and W. Toh, “Spectrum survey in Singapore:Occupancy measurements and analyses,” Proc. CROWNCOM, Sin-gapore, May 2008.

[8] M. Wellens and P. Mahonen, “Lessons learned from an extensivespectrum occupancy measurement campaign and a stochastic dutycycle model,” Proc. TridentCom, pp.1–9, April 2009.

[9] A. Sonnenschein and P.M. Fishman, “Radiometric detection ofspread-spectrum signals in noise of uncertain power,” IEEE Trans.Aerosp. Electron. Syst., vol.28, no.3, pp.654–660, 1992.

[10] R. Tandra and A. Sahai, “SNR walls for signal detection,” IEEE J.Sel. Top. Signal Process., vol.2, no.1, pp.4–17, 2008.

[11] J.J. Lehtomaki, M. Juntti, and H. Saarnisaari, “CFAR strategies forchannelized radiometer,” IEEE Signal Process. Lett., vol.12, no.1,pp.13–16, 2005.

[12] D. Panaitopol, A. Bagayoko, P. Delahaye, and L. Rakotoharison,“Fast and reliable sensing using a background process for noise es-timation,” Proc. CROWNCOM, Osaka, Japan, June 2011.

[13] J. Vartiainen, J. Lehtomaki, T. Braysy, and K. Umebayashi, “Spec-trum sensing in public safety applications: The 2-D LAD ACCmethod,” Proc. CROWNCOM, Osaka, Japan, June 2011.

[14] H. Saarnisaari, P. Henttu, and M. Juntti, “Iterative multidimensionalimpulse detectors for communications based on the classical diag-nostic methods,” IEEE Trans. Commun., vol.53, no.3, pp.395–398,March 2005.

[15] K. Umebayashi, T. Kazmi, Y. Kamiya, Y. Suzuki, and J. Lehtomaki,“Dynamic selection of CWmin in cognitive radio networks for pro-tecting IEEE 802.11 primary users,” Proc. CROWNCOM, Osaka,Japan, June 2011.

[16] D. Torrieri, “The radiometer and its practical implementation,” Proc.MILCOM, San Jose, CA, USA, 2010.

[17] J. Vartiainen, Concentrated Signal Extraction Using ConsecutiveMean Excision Algorithms, Ph.D. thesis, Acta Univ Oul TechnicaC 368. Faculty of Technology, University of Oulu, Finland, Nov.2010.

[18] M. Matinmikko, M. Mustonen, M. Hoyhtya, T. Rauma, H.Sarvanko, and A. Mammela, “Directional and cooperative spec-trum occupancy measurements in the 2.4 GHz ISM band,” Interna-tional Journal of Autonomous and Adaptive Communications Sys-tems (IJAACS), in press.

[19] J.J. Lehtomaki, J. Vartiainen, M. Juntti, and H. Saarnisaari, “CFARoutlier detection with forward methods,” IEEE Trans. Signal Pro-cess., vol.55, no.9, pp.4702–4706, 2007.

[20] M. Lops, “Hybrid clutter-map/L-CFAR procedure for clutter rejec-tion in nonhomogeneous environment,” Proc. Inst. Elec. Eng.-Radar,Sonar and Navigation, vol.143, no.4, 1996.

[21] S.S. Gupta and S. Panchapakesan, “Order statistics arising from in-dependent binomial populations,” Mimeograph Series no.120, De-partment of Statistics, Purdue University, West Lafayette, 1967.

Janne J. Lehtomaki graduated with anM.Sc. (Tech.) and Ph.D. (Tech.) in Telecom-munications from the University of Oulu in1999 and 2005, respectively. Currently, he isa Project Research Fellow at the Centre forWireless Communications (CWC), Universityof Oulu. In 2008, he was a visiting researcherat the University of Agriculture and Technology,Tokyo, Japan. His research interests are in spec-trum measurements, energy detection, and cog-nitive radio networks. He is a Member of IEEE.

Risto Vuohtoniemi received his M.Sc.(Tech) and Licentiate of Technology degreesin Telecommunications from the University ofOulu. He is now working as a senior assistant inthe Telecommunication Laboratory and as a re-search scientist in the Centre for Wireless Com-munications (CWC) at the University of Oulu.His research interests are in RF technology es-pecially in future wireless communication sys-tems, cognitive radio, passive radar applications.

Kenta Umebayashi received his B.E.,M.E. and Ph.D. degrees from Yokohama Na-tional University in 1999, 2001 and 2004, re-spectively. From 2004 to 2006 he was with theCentre for Wireless Communications, Univer-sity of Oulu. He is currently an Assistant Pro-fessor in Tokyo University of Agriculture andTechnology. In 2010, he was a visiting profes-sor at the University of Oulu. His research in-terests lie in the areas of statistical signal pro-cessing and detection and estimation theory in

wireless communication such as cognitive radio. He is a Member of IEEE.

Juha-Pekka Makela received his M.Sc.degree in electrical engineering, and Licentiatesand Dr.Sc (Tech.) degrees in Telecommunica-tions from University of Oulu, Finland, in 1997,2002 and 2008 respectively. He is now workingas a Postdoctoral Research Fellow in Telecom-munication Laboratory, University of Oulu. Hisresearch interests include wireless ad hoc andsensor networks. He is a Member of IEEE.