You are on page 1of 6

Spectrum Segmentation for Wideband Sensing

of Radio Signals

Doris Bao, Luca De Vito, MEMBER IEEE, Sergio Rapuano, SENIOR MEMBER IEEE
Department of Engineering, University of Sannio,
Corso Garibaldi, 107, 82100, Benevento, Italy.
{doris.bao, devito, rapuano}@unisannio.it

Abstract - Spectrum sensing is a key factor for either the In this scenario, the spectrum sensing process is essential:
development of Cognitive Radio or the implementation of (i) for the development of CR systems, and (ii) for the
instruments able to automatically monitor the radio spectrum. realization of instruments able to automatically monitor the
Within the process of spectrum sensing, segmentation aims to radio spectrum.
detect the boundaries of the transmission channels in a wide
frequency range. The paper proposes a new spectrum As shown in Fig.1, the spectrum segmentation task in the
segmentation method characterized by low processing time. The CR systems consists in locating the active radio transmission
developed histogram-based method is described in the paper. channels (channel boundaries detection, red limits in Fig.1). It
Several experimental tests on emulated and actual signals are is the preliminary step to allow the detection of not busy
presented to prove the effectiveness of the method in terms of spectrum segments or, at large, the spectrum sensing.
success percentage, processing time.
The same segmentation task (Fig.1) can be pursued in
Keywords - Spectrum Sensing/Segmentation, Cognitive Radio, instruments able to monitor the radio spectrum in order to
Opportunistic Spectrum Access. estimate the number and the type of the signals in the radio
spectrum under analysis.
I. INTRODUCTION
In this paper, a new spectrum segmentation method is
Advances in radio technologies allow to overcome the way proposed. A histogram-based method has been designed and
current regulations allocate the radio spectrum. Exclusivity in implemented to meet the detection speed and accuracy
the allocation of spectrum was necessary in the past to prevent requirements of CRs and measurement instruments over a wide
multiple users from interfering each other. Today, exclusive frequency range.
allocation of channels causes bandwidth scarcity, as a relevant
part of the radio spectrum is reserved even if it is not busy. Although the implementation and evaluation is devoted to a
specific monitoring application, the described segmentation
In order to support the increasing demand for wireless approach can be considered effective and useful for CR
services it is necessary to design a more flexible way to systems in general. Specifically, the design of the proposed
manage the radio spectrum, in which different services and method has been motivated by the need of improving the pre-
technologies can share the radio spectrum at the same time. processor block of an automatic modulation classifier. The
Smart radios, properly called Cognitive Radios (CRs) [1], automatic modulation classifier presented in previous papers
have been conceived to enable such flexible spectrum [2-3] works already properly on single signals. The instrument
allocation. CRs, basically, are transceivers able to look for recognizes the modulation of a single input signal but it is not
available frequency bands and communicate using any channel able to consider a wideband radio signal and to distinguish the
that is currently not busy. CRs not only shift the carrier different transmission channels and signals automatically. A
frequencies but also change modulation protocol dynamically preprocessing block has been introduced in [4, 5] to divide the
in order to exploit various not busy spectrum sections. wide band in equally spaced channels, to check the presence of
a signal in each channel and to select the channels where a
Spectral detection and estimation are essential features of signal is found.
the instruments able to automatically monitor the radio
spectrum in order to ensure both the telecommunication
technology advances and the regulatory policy development
and control. These are instruments using signal classification
and sensing capabilities in order to perform automatic signal
detection, synchronization, and provide the automatic
recognition of the modulation format (automatic modulation
classification) without any a priori knowledge of the signal.
Figure 1. Assuming oportunistic allocation of the radio spectrum, the
segmentation process locates the transmission channels in a specified time.

978-1-4577-0457-4/11/$26.00 ©2011 IEEE


The current aim is to complete the preprocessing block to Let [ f 0 , f
] be the observed frequency range of the radio
automatically select the channels, which can be even of spectrum and suppose that such range is available for
different widths, by observing a wideband spectrum. supporting heterogeneous wireless communications. Different
This preprocessor has, therefore, the tasks of: transceivers may adopt different wireless technologies for
transmission over different bands within the observed
• dividing the whole observed spectrum in consecutive frequency range. The segmentation process has to locate the
sub-bands ( B1 , B 2 ,..., B
) by locating their frequency f 0 , f1 ,..., f
boundaries of the
frequency intervals.
boundaries ( f 0 , f1 ,..., f
);
The proposed segmentation method consists of two
• detecting the presence of signals in each sub-band and functional blocks, as described in next sub-sections A and B,
estimate their carrier frequencies. which deal with (i) the estimation of the signal PSD, and (ii)
the spectrum segmentation, respectively.
The first task is the subject of the method presented below.
The second problem has been already dealt with and an A. PSD Estimate
innovative solution has been reported in [4, 5]. Handling the The estimate of the PSD of the radio signal in the analysed
two problems separately has been considered more effective to frequency band is a key aspect to assure the correct boundary
enhance the accuracy of the whole spectrum sensing process. detection.
II. HISTOGRAM-BASED SEGMENTATION METHOD The considered ways to estimate the PSD are:
The problem of detecting a signal in a wide band has been • acquire directly the trace from a signal analyzer;
faced by many researchers and several proposals can be found
in literature. Some researchers try to solve the detection • acquire the IQ (In-Phase and Quadrature) components
problem in the whole observed spectrum. In [5] and [6], for from a signal analyzer and estimate the PSD by means
example, a narrowband signal detection is implemented of digital processing.
assuming known the width of each sub-channel. In [7] the In order to assure better performance of the proposed
sensing problem is formulated in a Bayesian framework in segmentation method, the PSD of the signal has to be quite
which the unknown parameters, such as the frequency smooth in each sub-band (transmission channel) and has to
boundaries and spectral density of each sub-band are treated as show a discontinuity between contiguous sub-bands.
random variables with known distribution. The wavelet
transform is adopted in [8] to provide a multi-resolution The first condition can be achieved: (i) using the zero phase
spectrum sensing. filtering, and (ii) averaging several acquisitions.
In other papers the problem of wideband detection is When the IQ components are acquired, the PSD can be
divided into two sub-problems. First, the spectrum is properly evaluated by nonparametric methods such as the
segmented by searching the boundaries of each channel; then a Welch estimator [10].
narrowband detection method is applied to each channel. In [9] B. Segmentation method
the Continuous Wavelet Transform is used to detect the
frequency boundaries of the signal assuming its Power Spectral The proposed segmentation method (Fig.2) executes the
Density (PSD) smooth and almost flat in each sub-band. following steps:

In this paper, the second approach is followed in order to • Step 1; to build an amplitude histogram of the
obtain an easier and less complex algorithm, able to be smoothed PSD and to detect of its local maxima
implemented in CRs or measurement instruments. M1,…,M
(Fig.3b). The histogram of the smoothed
PSD allows to identify the average PSD level of each
According to the above specifications the segmentation step sub-band. In fact, the number of Mi exceeding a
has to be accurate in the frequency boundaries location. calibration threshold m, matches the number of sub-
Moreover, it could give qualitative information about the bands while the Mi values are proportional to the sub-
spectral density within each frequency interval useful to the band widths. The threshold on maxima detection, m,
subsequent detection stage. For example, only three PSD depends on the minimum bandwidth considered for the
amplitude classes could be provided, black, gray or white, to transmission channels.
quantify the presence of signals in a given frequency band. The
detection stage could use the qualitative information about each • Step 2; to square the PSD within a given tolerance ±δ
channel to take different operating decisions such as starting around each Mi . All the amplitudes of the PSD within
the detection from the channels with higher spectral density. the tolerance regions are equalized to the nearer Mi.
The two steps collaborate to improve the performance of the The tolerance δ is set by a ripple calibration phase. The
whole spectrum sensing process. estimate of the PSD of the radio signal in the
considered frequency range affects the δ tolerance
The method proposed in this paper looks at the irregular choice and the performance of the method. Fig.3.c
edges in the PSD amplitude of the incoming signal in order to shows in grey the regions where the method squares
identify candidate boundaries of the spectra of existing the PSD and detects a transmission channel. The
modulated signals. histogram maxima are reported with the considered
tolerance threshold δ.
PSD of the RF signal, P(i) (a)

Build Histogram, H(i)


1
m, threshold
on maxima Detect H(i) Maxima, Mj
detection
Square PSD
δ , tolerance If (Mj -δ ≤ P(i) ≤ Mj +δ), 2
around Mj Ps(i) = Mj
Else Ps(i) = P(i)
Squared PSD, Ps (i)

Detect channels, Ck
l, min channel
(b)
slope Ps(i)=0, for
width s1≤ i ≤ s2 , with s2 - s1> l

Detect boundaries, Bt 3
after Ck
Ps(i) slope transition
from negative to positive
or from 0 to positive

Detected boundaries, Bt
Figure 2. Steps of the histogram based segmentation method. Step 1:
evaluation of the histogram, built on the PSD of the considered RF signal,
and its maxima. Step 2: squaring of the PSD. Step 3: identification of the (c)
boundaries of the transmission channels.

• Step 3; looking for the slope changes of the squared


PSD to identify the frequency boundaries
f 0 , f1 ,..., f
(Fig.3d). The analysis of the slope of the
squared PSD between two recognized channels. A
transmission channel (grey bands in Fig.3c) is found
only if the corresponding frequency bandwidth is more
than l (minimum channel width).
Thresholds m and l depend on the minimum bandwidth (d)
considered for the transmission channels. That is, m is the
number of frequency bins having the same PSD amplitude and
is expressed as the number of occurrences of that amplitude in
the histogram. Such bins could not be contiguous. The second
threshold, l, is expressed as the minimum number of
contiguous bins which constitute a channel (squared segments
which can be considered a transmission channel). These
thresholds can be configured separately thus increasing the
flexibility of the method.
III. EXPERIMENTAL VALIDATION
After a successful investigation using simulated signals, the Figure 3. Assuming (a) the PSD of multiple modulated signals as input of
validation process of the proposed segmentation method has the histogram based segmentation method: the histogram built on the
been performed on emulated signals. original signal after the zero phase filtering is reported in (b); in (c) the
signal is squared by means of the maxima and the threshold δ, the regions
Accuracy, complexity and processing time are the main in grey are the detected channels; the study of the slope beetwen two grey
terms of survey. bands allows the boundaries detection (d).

The sub-section A introduces some considerations about the A. Test scenarios


choice of the scenarios considered in the phase of validation of
the proposed method. In sub-section B, some experimental A suitable test bench to characterize the proposed
results are reported and evaluated. Further experimental results segmentation method with actual signals has been realized
and comparison with other segmentation methods will be (Fig.4) at the Laboratory of Signal Processing and
shown in sub-section C. Measurement Information of the University of Sannio in
Benevento [11].
Fig.6 shows the frequency boundaries detection on the
R&S SMIQ03B squared PSD. The preset and the measured frequency
Signal Generator boundaries are reported in Table II.
R&S FSQ8 The reported results refer to 100 consecutive acquisitions of
Signal Analyzer the RF signals in the considered scenario, when the PSD is (i)
LAN acquired directly from the Signal Analyzer (resolution
bandwidth = 0,03 MHz), and (ii) estimated from IQ vectors
using Welch estimator (frequency resolution = 40 kHz, FFT
Spectrum length = 1024, overlap = 50%). The proposed segmentation
Segmentation method obtained these results in T = 2,87 ± 0,51ms and
Software
MATLAB
GPIB T = 195,9 ± 4,2ms in the two cases, respectively. The data
transfer times from the Signal Analyzer to the PC have not
been considered in the time evaluations.
Figure 4. Test bench for emulated a) and actual b) signals.
The segmentation method succeeded in each acquisition.
The frequency boundaries have been correctly located and the
Several modulated signals have been generated with M-ary reached accuracy allowed to recognize the transmission
Phase Shift Keying (MPSK), Frequency Shift Keying (FSK) channels.
and Quadrature Amplitude Modulation (QAM), in different
combinations. C. Comparison with other segmentation methods
The emulated signals generated by a R&S SMIQ03B The paper [8] proposes two segmentation methods based on
Signal Generator are the inputs of the R&S FSQ8 Signal the wavelet transform, both of them have been taken into
Analyzer to obtain the PSD in the considered frequency band account to characterize the presented method in terms of
(Fig.4). comparison with other approaches.
0
A MATLAB virtual instrument has been developed to (i)
-10
implement the proposed method, (ii) control the SMIQ03B,
through the GPIB interface, (iii) control the FSQ8, through the -20
PSD [dBm]

LAN interface, and (iv) automate the test execution. -30

The behavior of the segmentation method has been studied -40


(i) in an ideal case (no noise), and (ii) when each signal is -50
corrupted, within its bandwidth, by white Gaussian noise also
-60
generated by means of the SMIQ03 signal generator. 890 892 894 896 898 900 902 904 906 908 910
Frequency [MHz]
Each considered test scenario has been evaluated using the Figure 5. Original signal PSD from the Signal Analyzer: four frequency
proposed segmentation method and the alternate wavelet based sub-bands with piecewise smooth PSD.
method in [9]. The comparison aimed to demonstrate the
effectiveness of the histogram based segmentation method in 0
terms of success percentage and processing time.
-10
Several tests have been realized to study the performance of
-20
the method for the different ways of obtaining the PSD of the
PSD [dBm]

RF signal. Both the possibilities described in sub-section II.A -30


have been evaluated in the experimental tests. The aim is to
-40
examine a wideband spectrum in a fast and precise manner and
to improve the reliability. -50

The investigation among the different options wants to find -60


890 892 894 896 898 900 902 904 906 908 910
a balance between (i) simple and fast computations but less Frequency [MHz]
accurate PSD estimate, and (ii) computationally complex Figure 6. Boundaries detection.
procedures but enhanced PSD estimate.
B. Results on emulated signals TABLE I. CONFIGURATION FOR A TEST SCENARIO IN CASE OF
A very simple test scenario is presented in Fig.5. EMULATED SIGNALS
Specifically, the reported RF signal is composed by four sub- Raised-
Carrier
bands and four 4-PSK modulated signals. Table I presents the Modulation frequency
Amplitude Symbol rate cosine
configuration parameters. The test involves a frequency range [dBm] [Msymbol/s] filter
[MHz]
of 20 MHz, from 890 MHz to 910 MHz with the four sub- Roll-off
1. 4-PSK 893 -10 40/9 0.22
bands (channels) in the case of an SNR value of 30 dB. 2. 4-PSK 897 -30 1.6 0.22
3. 4-PSK 903 -20 8 0.22
4. 4-PSK 909 0 1.6 0.22
TABLE II. DETECTED BOUNDARIES ON 100 ACQUISITIONS. band and to preserve discontinuities between adjacent sub-
bands.
f1 [MHz] f2 [MHz] f3 [MHz] f4 [MHz]
The results reported in Table III refer to a RF signal
Preset composed by Orthogonal Frequency Division Multiplexing
frequency 896,00 898,00 908,00 910,00 (OFDM) modulated signals. The aim is to emulate OFDM
boundaries based signal systems such as Digital Video Broadcasting –
CASE 1 Terrestrial (DVB-T). The SNR value is 30 dB.
PSD from the Spectrum Analyzer
Mean 895,65 898,00 907,32 909,92 The mean and the standard deviation for each frequency
Standard
0,11 0,02 0,72 0,25 boundary have been evaluated on 100 acquisitions when the
Deviation methods succeeded and the percentage of success is reported in
CASE 2 the table, too. One more detected boundary or one less detected
Welch Spectrum Estimator
Mean 895,66 897,92 907,44 910,84 boundary is enough to consider failed the method on a specific
Standard acquisition.
0,01 0,04 0,02 0,50
Deviation
The performed tests aimed to evaluate the applicability of
the methods for CR systems in order to locate the transmission
Such methods, that in the original paper were referred as: channels in real time. It is central to consider the percentage of
(i) wavelet modulus maxima, and (ii) multi-scale wavelet success as well the processing time.
product, respectively, have been implemented in MATLAB to
Table IV reports processing times for the three methods
compare the segmentation results in different scenarios.
both in the case of PSD from the spectrum analyzer and Welch
When it is used, the preprocessing step of zero-phase estimator. The histogram method shows a significant
filtering is common to the three methods, to achieve the improvement in the processing time as compared to the
condition of smoothness of the radio signal PSD in each sub- wavelet-base methods.

TABLE III. COMPARISON RESULTS WITH OFDM MODULATED SIGNALS.

f1 [MHz] f2 [MHz] f3 [MHz] f4 [MHz] f5 [MHz] f6 [MHz] f7 [MHz] f8 [MHz]


Preset frequency
892,00 894,00 896,00 897,00 902,00 904,00 908,00 910,00
boundaries
CASE 1
PSD from the Spectrum Analyzer

Histogram method Mean 892,06 893,96 896,41 897,65 902,11 903,91 908,15 909,88

91% of success Standard Deviation 0,04 0,11 0,05 0,14 0,04 0,10 0,04 0,07

Wavelet modulus maxima Mean 892,33 893,55 896,40 897,51 902,35 903,56 908,33 909,58

100% of success Standard Deviation 0,01 0,01 0,03 0,03 0,01 0,01 0,01 0,04

Multi-scale wavelet product Mean 892,34 893,55 896,40 897,51 902,35 903,55 908,34 909,57

100% of success Standard Deviation 0,01 0,01 0,03 0,03 0,01 0,02 0,01 0,02

CASE 2
Welch Spectrum Estimator

Histogram method Mean 891,78 893,91 896,36 897,66 902,18 903,90 907,41 910,41

90% of success Standard Deviation 0,13 0,08 0,02 0,05 0,01 0,08 0,17 0,22

Wavelet modulus maxima Mean 892,22 893,67 896,33 897,57 902,25 903,67 908,24 909,68

91% of success Standard Deviation <0,01 <0,01 0,01 <0,01 0,01 <0,01 <0,01 <0,01

Multi-scale wavelet product Mean 892,22 893,71 896,32 897,57 902,22 903,71 908,20 909,73

100% of success Standard Deviation <0,01 <0,01 <0,01 <0,01 <0,01 <0,01 <0,01 <0,01
TABLE IV. COMPARISON OF THE PROCESSING TIMES. Communications) and UMTS (Universal Mobile
Telecommunication System) bands.
CASE 1
ACKNOWLEDGMENT
PSD from the Spectrum Analyzer
The authors wish to thank Prof. Pasquale Daponte for his
Mean 2,09 ms
Histogram method helpful suggestions and active interest during all the phases of
Standard Deviation 0,25 ms the research activity.
Mean 293,7 ms The whole project has been funded by the Italian Ministry
Wavelet modulus maxima of Scientific and Technological Research in the context of the
Standard Deviation 4,2 ms
PRIN 2008 programme.
Multi-scale wavelet Mean 164,1 ms
product REFERENCES
Standard Deviation 2,3 ms
[1] M.Matinmikko, M.Mustonen, H.Sarvanko, M.Hoyhtya, A.Hekkala,
CASE 2 A.Mammela, M.Katz, M.Kiviranta, “A motivating overview of cognitive
radio: Foundations, regulatory issues and key concepts”, Proc. of first
Welch Spectrum Estimator Int. Workshop on Cognitive Radio and Advanced Spectrum
Management, Feb. 2008, pp.1-5.
Mean 202,08 ms [2] L.De Vito, D.Grimaldi, S.Rapuano, “An automatic digital modulation
Histogram method
classifier for measurement on telecommunication networks”, IEEE
Standard Deviation 0,57 ms
Trans. on Instrum. and Meas., vol.56, No.4, Oct. 2007, pp.1711-1720.
Mean 792,1 ms [3] N.Björsell, P.Daponte, L.De Vito, S.Rapuano, “Automatic signal
Wavelet modulus maxima recognition for a flexible spectrum managment”, Proc. of XIX IMEKO
Standard Deviation 5,1 ms World CongressSept. 2009, pp.568-573.
Multi-scale wavelet Mean 483,1 ms [4] L.De Vito, D.Napolitano, S.Rapuano, M.Villanacci, “Eigenvalue-based
product signal detector for an automatic modulation classifier”, Proc. of Int.
Standard Deviation 6,3 ms Instrum. and Meas. Tech. Conf., May 2010, pp.1131–1136.
[5] N.Björsell, L.De Vito, S.Rapuano, “A GNU Radio-based signal detector
for cognitive radio systems”, Proc. of IEEE Int. Instrum. and Meas.
IV. CONCLUSIONS Tech. Conf., May 2011, pp. 1045-1049.
[6] Z.Quan, S.Cui, A.H.Sayed, and H.V.Poor, “Wideband spectrum sensing
The paper proposes a new spectrum segmentation method in cognitive radio networks”, IEEE Int. Conf. on Commun., May 2008,
adopting a histogram based algorithm to provide speed and pp.901–906.
accuracy over a wide frequency range. [7] S.Tascioglu, O.Ureten, "Bayesian wideband spectrum segmentation for
cognitive radios", Proc. of 18th Int. Conf. on Computer Communications
The method has been validated using emulated and actual and Networks, Aug. 2009, pp.1-6.
signals and has been compared with two methods following a [8] Y. Hur, J. Park, W. Woo, K. Lim, C.-H. Lee, H.S. Kim, and J. Laskar,
wavelet approach. “A wideband analog multi-resolution spectrum sensing (MRSS)
technique for cognitive radio (CR) systems”, Proc. of IEEE Int. Symp.
The use of the segmentation method over unknown signals on Circuits and Syst., May 2006, pp. 4090–4093.
in the real radio spectrum is under evaluation and will be [9] Z.Tian and G.B.Giannakis, "A wavelet approach to wideband spectrum
reported in further publications. In fact, a second test phase sensing for cognitive radios", Proc. of Int. Conf. Cognitive Radio
aims at evaluating the performance of the segmentation method Oriented Wireless Networks and Communications, 2006, pp.1-5.
in an unknown scenario: different frequency bands will be [10] P.D.Welch, "The use of Fast Fourier Transform for the estimation of
scanned, including the DVB-T (Digital Video Broadcasting - power spectra: a method based on time averaging over short, modified
periodograms", IEEE Trans. Audio Electroacoustics, vol.AU-15, June
Terrestrial), GSM (Global System for Mobile 1967, pp.70-73.
[11] LESIM website: http://lesim1.ing.unisannio.it

You might also like