547

Adaptive Energy Detection for Cognitive Radio: An

School of Engineering and Advanced Technology, Massey University, Auckland, New Zealand
[email protected], [email protected]

Abstract
A cognitive radio (CR) is able to sense spectral envi-
ronment over a wide range of frequencies, and provide
opportunistic access to frequency bands temporarily
unoccupied by an incumbent. Accurate channel sensing
is the first important task for a CR, and energy detector
is often used for this purpose. While a normal energy
detector works well with well chosen window size based
on prior knowledge about possible primary users, it
often fails with signals that are narrow compared to the
detector window, or if only a fraction of the signal is
inside the detector window. We propose an adaptive
energy detector that can adjust its detection window,
and evaluate such detectors performance using expe-
rimental results obtained through a real time implemen-
tation.
Keywords: channel sensing, cognitive radio, energy
detection, energy detector model, labview
I. INTRODUCTION
Unlicensed spectrum is becoming increasingly scarce,
especially those under 3 GHz. The Federal Communica-
tions Commission (FCC)s spectrum allocation chart
shows that many frequency bands are being allocated to
multiple incumbents, overlapping each other [1]. While
most frequency bands are licensed, studies from [1] and
[2] suggest utilization rate between 15% and 85%. Cog-
nitive radio (CR) has been highlighted as a possible
candidate in improving spectrum utilization by provid-
ing opportunistic spectrum access [3, 4]. A cognitive
radio can be defined as a radio that is able to sense the
spectral environment over a wide frequency, and exploit
this information to opportunistically provide wireless
links that best meet the users communication require-
ments [1]. In order to provide access to a frequency
band as a secondary user, a CR must first be able to
sense and identify frequencies (spectrum sensing) that
are temporarily unoccupied by an incumbent. These
free areas are termed as holes in [5].
For spectrum sensing, energy detector is a popular tech-
nique [1, 6]. In this paper, we focus on the detection of
signals that are narrow compared to the detector win-
dow (narrow signals), and on the detection of signals
where only a small portion of the signal is captured in
the detector window. A real time detector with adaptive
detection window is proposed and implemented. The
performance of such detector is then investigated using
data obtained from a real time experiment.
The paper is organized as follows: Section 2 briefly
discusses related work done by other research groups. In
Section 3, we provide an overview of the energy detec-
tor model, and our problem statement. Section 4 de-
scribes our adaptive detection strategy in detail. Section
5 presents our testbed set up and experimental data. And
Section 6 presents some limitations of our detector and
energy detectors in general. Finally Section 7 summa-
rizes the work done and concludes the paper.
II. RELATED WORK
There are many researches being done on energy detec-
tors. The hidden terminal problem in detection was dis-
cussed in [1], where the CR is behind a building with
high penetration loss, and may not see the primary user.
However as the CR starts to transmit, the primary user
will be interrupted. A suggestion of an extra 30-40dB
SNR sensitivity requirement was made for the CR to
combat this effect. The article also defined a term local
spectrum sensing, where a CR must independently
detect primary users, due to the lack of a direct way of
measurement between the primary transmitter and re-
ceiver. In [6], theoretical analysis of pilot detection and
energy detection are presented. The models for the de-
tectors were derived, and experimental results were pre-
sented to support the theory. Detailed discussions of the
limitations of each detection strategy were also in-
cluded. In [8], a practical energy detector was presented.
The article starts by detailing the model of the energy
detector, and then moves on to the derivation of thre-
shold calculations. Finally some experimental data were
used to evaluate the performance of the detector under
various conditions. In [9], the concept of an SNR wall
for systems trying to detect low power signals (<< 0dB
SNR) was derived as a fundamental limit of an energy
detector. For a detector with noise uncertainty, there
exist a under
which the signal becomes impossible to detect. In [7], a
complete system using software radio is presented. The
radio uses a Genetic Algorithm (GA) that is able to
search through a large combination of possible system
parameters, and find the one that is best for the current
channel conditions. The demonstration system is able to
stream video continuously under heavy interference by
adjusting the operating parameters of the radio.
978-1-4244-6284-1/09/$26.00 2009 IEEE
Proceedings of 2009 12
th
International Conference on Computer and Information Technology (ICCIT 2009)
21-23 December, 2009, Dhaka, Bangladesh
548
III. ENERGY DETECTION AND PROBLEM
STATEMENT
A. ENERGY DETECTOR MODEL
An energy detector can be used to determine if a given
window contains only noise (1), or signal plus noise (2).
This goal can be simplified to the following hypotheses
test [6, 8]:

(1)

(2)
where is the signal received by the CR, is the
transmitted signal, and is presumed to be Additive
White Gaussian Noise (AWGN). The test statistic for an
energy detector is:
(3)
With large numbers of samples (M number), by central
limit theorem this can be modeled as:

(4)

(5)
where B is the window bandwidth of the detector, is
the noise power spectral density, and is the power of
the signal. The probability of false alarm is the probabil-
ity that noise is large enough to be detected as a signal.
With estimated noise power, we can derive a model for
the probability of false alarm:

(6)
Where Q is the normal cumulative distribution function.
Using (6) we can specify a target false alarm rate, and
calculate the corresponding threshold [8]:

(7)
Given a threshold, the energy detector itself can be im-
plemented easily by squaring the magnitude data of an
FFT, and normalizing over bin size to obtain the power
spectral density (PSD). The detector then integrates over
the entire window, obtaining total power within the
window. Comparing this with the threshold from (7),
the detector then makes a decision, indicating whether
the current window is free or not. When using this me-
thod, we can modify the frequency resolution of the
FFT by increasing the number of points for the FFT.
The tradeoff for this is increased processing time. We
can also average a number of PSDs to smooth out the
effect of noise. This improves detection, again at a cost
of processing time.
B. PROBLEM STATEMENT
Regarding the energy detector from Section 3.A, expe-
riments from [6, 8], and our preliminary experimental
data have shown that this energy detector model is accu-
rate. The detector described is set to work at a particular
window size. First a frequency band is selected and fil-
tered. An FFT is computed, and the energy detection is
performed subsequently. In practice, various operation
parameters are met by changing the number of averages
taken, and sometimes by changing the frequency resolu-
tion [6].
A cognitive radio is often blind, in the sense that it
has no prior knowledge about incumbents. This often
means that the window size chosen for the energy detec-
tor described in Section 3.A may not be appropriate. As
a result, the energy detector often fails if the signal is
narrow (compared to the detector window), or if only a
small fraction of the signal is captured within the win-
dow. We can alleviate this problem by adjusting the
window used for energy detection. But an increased
number of energy detection also leads to an increase in
sensing time. An energy detector that can selectively
subdivide large windows into finer ones can trade the
minimum detectable bandwidth with possible increase
in detection time. This is more effective than decreasing
the detection window across the board, especially in
areas where a small number of narrow signals are ex-
pected.
IV. ENERGY DETECTOR WITH ADAPTIVE
WINDOW SIZE

A weakness of the conventional energy detector when
detecting narrow or partial signals is demonstrated in
Fig. 1. The PSD in Fig. 1 is captured in real time by our
workbench, and shows a frequency band of 195 MHz to
203 MHz (TV3 in New Zealand).

Fig. 1. Energy detector performance on narrow band
signals

Assuming no prior knowledge of the incumbents, the
system uses a 1 MHz window for energy detection, and
the results suggest that windows 1, 3, 4, 5 and 6 are free.
However, we can clearly see that there are signals
present in those window positions. While these signals
are present, they occupy a band that is narrow compared
to the energy detector. As the integration in (3) com-
pletes for each window, the total power in window is
not enough to rise above the threshold. Three possible
ways to combat this problem are to lower the threshold,
to choose a window similar in size to the signal under
detection, or to narrow the detector window. Lowering
the threshold increases significantly according to Eq.
549
(6), resulting in an unusable radio system. Choosing a
window size closer to the signal under detection (6-8
MHz in case of a TV signal) requires prior knowledge
of incumbents, something the CR often does not have.
Using a narrow detector window alleviates the problem,
at a cost of increased sensing time.
We propose an adaptive energy detector based on the
third method. The detector normalizes the threshold
obtained in Eq. (7) to dBm/Hz, does a preliminary scan
of the PSD within the detector window, and performs a
window subdivision if necessary. Fig. 2 describes the
workflow of this adaptive energy detector.

Fig. 2. Adaptive energy detector workflow

If there are peaks above the normalized threshold, and
wider than the minimum detectable bandwidth, then the
detector automatically subdivides the current window
into numbers of sub windows. We note that the min-
imum detectable signal bandwidth is limited by the fre-
quency resolution ( ) of the FFT, and thus the subdivi-
sion is limited also by the frequency resolution. Our
proposed detector sets minimum detectable narrow sig-
nal to the frequency resolution of the FFT, and subdi-
vides the current window into:

(8)
The decision variable then becomes:

(9)
Where is the final decision variable, and is the
decision on the ith window.

V. EXPERIMENTAL RESULTS
A. TESTBED DESCRIPTION
Fig. 3 showcases our CR work bench. The bench con-
sists of: a National Instruments PXI-5661 Vector Signal
Analyzer (consisting of PXI-5600 RF Down converter
and PXI-5142 OSP Digitizer) housed in a PXI chassis
[11]; two computers with LabVIEW [12] for interfacing
with the VSA (including one PXI-8106 embedded con-
troller in the PXI chassis); a Universal Software Radio
Peripheral (USRP) with both RX and TX daughter
boards; an Agilent 33120 Signal Generator; an HP
54600B Oscilloscope; and an Agilent 35670A Dynamic
Signal Analyzer.
For this experiment, only the PCs and the NI VSA were
used. The VSA has a range of 30 kHz to 2.7 GHz, with
20 MHz of real time bandwidth. It uses a 3 stage super
heterodyne design, and can return data in both I/Q and
Frequency domain through the on board signal proces-
sor. The PCs are running LabVIEW for interfacing with
the VSA, through the CRLibs library we developed es-
pecially for cognitive radio applications.
Currently our library supports various spectrum mea-
surements, channel sensing and energy detection with a
flexible bandwidth from 1 kHz to 1GHz. Spectrum av-
eraging, windowing and other signal processing func-
tions are also included. It is used for a range of applica-
tions from wideband 802.22 channel sensing similar to
[10], to UWB communication system simulations.

Fig. 3. Cognitive radio work bench
B. ADAPTIVE ENERGY DETECTOR SETUP
The adaptive energy detector is implemented complete-
ly in software under LabVIEW, based on periodogram
[1]. First, the VSA is initialized with system parameters
such as resource handles and application specific para-
meters such as frequency scan range. Once in acquisi-
tion mode, we configured the onboard signal processor
to return a PSD waveform in dBm/Hz. A noise floor is
estimated, and the appropriate threshold is computed
using (7). The detector then follows the steps described
in Fig. 2, finally reaching a decision for the current win-
550
dow. This adaptive energy detector is now a part of the
CRLibs, and can be used as a subVI in LabVIEW. It
offers programmable FFT frequency resolution, number
of averages, and windowing functions. The detectors
sensing time performance can be seen in Fig. 5.
C. EXPERIMENTAL SETUP
Using our CRLibs with the VSA, we can perform the
energy detection correctly on signals from 9 kHz to 2.7
GHz, with 1 kHz to 1 GHz of bandwidth. We decided to
investigate the VHF/UHF range due to the frequency
bands connection to 802.22, and the readily available
public broadcast signals at known locations. The VSA is
connected with a TV antenna, and has the following
parameters:
Table I. Experimental system parameters
FFT resolution bandwidth 15 kHz
Energy detector window size 1 MHz
Number of averages 60

We then collected sets of data for both the normal ener-
gy detector, and the adaptive detector. To provide accu-
rate experimental data, each measurement was repeated
10,000 times.

Fig. 4. Detection performance


Fig. 5. Sensing time performance
D. RESULTS
First, we evaluate the cumulative distribution function
(CDF) of the probability of detection for our adaptive
detector. While the signal power within each of the de-
tector windows may be small, the regions that rise
above the threshold are enough for the adaptive energy
detector to adjust the window size, and make a more
accurate decision. Fig. 4. shows that for the same prob-
ability of detection, the adaptive energy detector re-
quires 1 to 2dB less SNR. We note that none of the de-
tectors were able to make correct decisions with signals
that are lower than -3dB SNR. This is expected from a
detector using a relatively small number of samples (M
= 60) and is in line with results obtained by [8].
Then, we compare our detectors sensing time performance
against a normal energy detector. While the PC based detec-
tors do not have fast response times, Fig. 5. still demonstrates
the relatively longer sensing time our adaptive method takes
compared to a normal detector. This is to be expected, as our
adaptive method has to first make a preliminary scan of the
window, and then sub divide the window and perform mul-
tiple energy detections when necessary.
VI. LIMITATIONS
For a general energy detector with dB of noise un-
certainty, there exists a wall
under which a
signal is impossible to detect [6, 9]. Also, studies have
shown that weak signals with power << 0dB require
O(1/SNR
2
) samples to detect, which can increase sens-
ing time significantly [6, 9].
Apart from the fundamental limits discussed above, our
adaptive detector also has a limitation in terms of the
minimum detectable signal bandwidth, and further in-
crease in detection time. In Section 5 we noted that the
minimum detectable signal bandwidth decreases with a
decrease of frequency resolution of the FFT. This limits
what we can detect, as decreasing the frequency resolu-
tion increases sensing time. Together with the time re-
quired to do a preliminary scan of the window, the total
extra time is compounded on top of the time needed to
gather the number of samples required to detect weak
signals. The sensing time performance of our adaptive
detector can be seen in Fig. 5.
While FPGA implementations are much faster than our
generic PC setup, our CRLibs subVI library provides a
flexible testbed that can be configured to do most spec-
tral measurement tasks with minimum set up and tear
down time. For example, the set up can be used to test
ultra wideband CR systems, M-ary modulated commu-
nication systems and energy detector systems all using
the CRLibs with no code modification. A theoretical
model can be evaluated quickly using our setup, and
once it is confirmed to be working, deployment onto a
FPGA can start.

VII. CONCLUSION
Without a properly chosen window size, energy detec-
tors often fail to detect signals that are narrow compared
to the detector window, or if only a fraction of the signal
551
is captured within the detector window. We proposed an
adaptive energy detector that can subdivide its detection
window into sub windows, and evaluated such detec-
tors performance. We carried out experiments on the
VHF/UHF band using a real time implementation, and
found that given the same parameters, our detector can
meet a probability of detection with 1 to 2 dB lower
SNR compared to the normal energy detector.
We are now evaluating other models for channel sens-
ing such as entropy detection, and looking at stochastic
channel selection strategies. Interested readers can fol-
low the project at: http://cr.jamesyxu.com.
VIII. ACKNOWLEDGEMENT
The authors would like to thank Dr. James Chang and
Dr. Tom Moir for their support and technical know-
ledge.
REFERENCES
[1] D. Cabric, et al., Implementation issues in spectrum
sensing for cognitive radios, Signals, Systems and
Computers Conference, 2004, pp. 772-776.
[2] FCC, Spectrum Policy Task Force Report, ET
Docket No. 02-155, 2002
[3] FCC, Notice of Proposed Rule Making and Order,
ET Docket No. 03-322, 2003
























[4] D. Chen, et al., Cooperative Spectrum Sensing un-
der Noise Uncertainty in Cognitive Radio, Wireless
Communications, Networking and Mobile Compu-
ting Conference, 2008, pp. 1-4.
[5] Z. Tian, et al., Performance Evaluation of Distri-
buted Compressed Wideband Sensing for Cognitive
Radio Networks, Information Fusion Conference,
2008, pp. 1-8.
[6] D. Cabric, et al., Spectrum Sensing Measurements
of Pilot, Energy and Collaborative Detection, Mili-
tary Communications Conference, 2006, pp. 1-7.
[7] B. Fette, Cognitive Radio Technology, Newnes
Elsevier, 2006
[8] S. Shellhammer, Performance of the Power Detec-
tor, IEEE Doc: 802.22-06/0075r0, 2006
[9] A. Sahai, et al., Some Fundamental Limits on Cog-
nitive Radio, Allerton Conference on Communica-
tion, Control, and Computing, 2004, pp. 1-5.
[10] V. Blaschke, et al., A Cognitive Radio Receiver
Supporting Wide-Band Sensing, International Con-
ference on Communications, 2008, pp. 499-503.
[11] National Instruments, NI PXI-5661, July 2009,
http://sine.ni.com/nips/cds/view/p/lang/en/nid/2030
38
[12] National Instruments, LabVIEW, July 2009,
http://www.ni.com/labview/