Electrical Engineering Department 2008 Annual Research Review
28 January 2008

Spectrum Sensing Design for Cognitive Radios
Prof. Danijela Cabric danijela@ee.ucla.edu Spectrum Utilization
Power Spectrum Density (dBm/Hz) -100 -110 -120 -130 -140 -150

0

1

2

3

4

5

6 GHz

Freq (GHz) Utilization(%)

0~1 54.4

1~2 35.1

2~3 7.6

3~4 0.25

4~5 0.128

5~6 4.6

Even though the spectra is allocated it is almost unused Cognitive Radios can make better use of the spectrum

How Does a Cognitive Radio Operate?
PU1 PU3 CR1 PSD PU2 CR2
Primary Cognitive

PU4 Frequency

Sense the spectral environment over a wide bandwidth Transmit in “white space” & Adapt bandwidth and power Detect if primary user appears Move to new white space

Cognitive Radio System Design

Network Management Sensing MAC Sensing Signal Processing Sensing radio Spectrum Allocation

Network Link Layer

Wideband signaling Wideband radio Physical Layer

Spectrum sensing is the key enabling functionality How do we implement spectrum sensing in a system?

Spectrum Sensing Problem
Primary User Cognitive Radio users must guarantee non-interference requirement

Tx Rx

CR CR

Decoding SNR

Sensing SNR

distance

Distance and channel not known

Cognitive radio can only observe (sense) primary system Tx signals Need to sense signals in highly negative SNR Sensing SNR < Decoding SNR – worst case channel Sensing SNR < [5dB to 20 dB] – [20 dB to 40 dB] = [-35 to 0 dB]

Designing Spectrum Sensors
– Sensing Requirements set by Primary User system ● Signal level (dBm) ● Maximum detection time (s) ● Interference protection (%) – Can we use standard detection techniques?
● Energy detection ● Pilot detection ● Feature detection

– Can a radio sense primary signals robustly and guarantee noninterference to primary users in negative SNR regimes?

Sensing Signal Presence
PSD
DTV channels (8 VSB) Cellular CDMA, GSM WiFi 802.11b/g (OFDM, CDMA)

…
Frequency

Use N samples to detect signal with required probability of detection and false alarm

Energy
Requires knowledge of - carrier and bandwidth - noise power Non-coherent processing, Sensing time scales as:

Pilot
Requires knowledge of - pilot data - perfect synch Coherent processing, Sensing time scales as:

Feature
Requires knowledge of - carrier and bandwidth - modulation type Statistical processing, Sensing time scales as:

N~

1 | SNR |

2

N~

1 p | SNR |

?
Used for modulation classification in positive SNR regimes [Gardner1986]

[Urkovitz1967]

p is % of signal energy in pilot [Middleton 1957]

Sensing Time vs. SNR Trade-off
Achievable region by different spectrum sensing techniques Log (Sensing Time)

Minimum achievable sensing time

Signal Level (dBm) SNR (dB) < 0

Spectrum Sensing Implementation
Signal Sensing DSP Hardware testbed

BEE2 platform

Reconfigurable wireless modem

Energy Detection: Sensing Time vs. SNR
4 MHz QPSK @ 2.485 GHz (-94 to -110 dBm) Pd=80% Pfa=10% 10*log10(Sensing time) (s) measured theoretical

SNR(dB)

Theoretical prediction:
Implies if we allow N to be large, arbitrarily weak signals can be detected

Experimental result:
Below -21 dB detection is not possible regardless of sensing time duration

Why Does Energy Detection Break Down?
Noise power varies due to temperature and adjacent interference
45.6

Noise level [dB] Noise level

45.4 45.2 45 44.8 6 p.m. 10 p.m. 2 a.m. 6 a.m. 10 a.m. 2 p.m 6 p.m.

Time

Theoretical assumption that noise is Gaussian and stationary fails in negative SNR Threshold T set based on estimated noise power if S + Nactual < T (Nactual +x) then detection becomes impossible Estimation error x sets sensing SNR limit (0.03 dB -> SNRmin-21 dB)

SNRmin = 10log10[10(x/10)-1] dB

Sahai, Tandra [2005]

Correlator-based Pilot Detection
10Hz offset
10*log10(Sensing time) (s)

~ SNR-1
100Hz offset

1KHz offset

Input Level (dBm)

Ideal radio (0 Hz freq. offset): Very weak signal detection possible Practical radio (offsets in the order of KHz) – Signal detection below -110 dBm not possible (802.22 requires -122 dBm) Can we design pilot detection that relaxes frequency offset requirements?

Feature Detection
Feature is a known statistical property
Noise spectrum correlation

S x (α , f ) = E{ X ( f ) X * ( f + α )}
Signal and noise features separated Sensor estimates features for a finite number of FFT averages z ncy MH Freque

1 N ~ S x (α , f ) = ∑ X (i, f ) X * (i, f + α ) N i=1
(10,000 spectral avgs) SNR≈0 dB

Frequency MHz

(500,000 spectral avgs) SNR≈-15 dB

Feature Detection under Clock Offset
Time offset modifies Fourier transform

x(t ) = x(t − t ) ⎯ X ( f ) = X ( f )e − j 2πft ⎯→
F 0
0

0

Constant timing offset and ideal SCF

S x (α , f ) = S x (α , f )e − j 2παt
Perfect sampling

N ~ ~ S x (α , f ) = S x (α , f )i∑ e − j 2παt =1

Clock offset and estimated SCF

i

100 Hz sampling clock offset (25 ppm)

Feature smeared

Partial Coherent Feature Detector
For a fixed number of spectral averages N

δ N N / 2) − j 2π δ N ( N +1) / 2 ~ ~ α S x (α , f ) = S x (α , f ) e α FFT δ sin( 2π N / 2) α sin ( 2π
FFT FFT

Feature are cancelled if

NN

FFT

>

α δ

where

δ

is a clock sampling offset

Proposed feature detector exploits partial coherent gain and overcomes the sampling offset problem:

1 M M ~ S x (α , f )' = ∑ ∑ X ( k + mM , f ) X ( k + mM , f + α ) T m=1 k =1
2 1

*

2

2

FFT

M1 determines coherent processing gain, chosen based on maximum sampling clock offset

Feature Detector: Sensing Time vs. SNR

~1/SNR2

Sampling offsets can be overcome with partial coherent feature processing

Feature Energy
Feature gain depends on spectral redundancy in pulse shaping filter Q(f)

S x (α , f ) =

α α 1 Q( f + )Q( f − ) * for α = 1 / Ts 2Ts 2 2
1− β 1+ β < f < 2Ts 2Ts

Pulse shaping filter with roll-off β = 0.5

Pulse shaping filter with roll-off β = 0.01

Advantages of Feature Detectors
Strong adjacent band primary user 802.11g
50

Feature detector robust to non-stationary noise

0 20 10 0 25

-5

0

5

10

15

20

Energy detector

Feature detector

PGED ~

SNRin N 1 + x(1 + N )

PGFD ~

Kα SNRin N 1+ x x – adjacent band leakage

PG - Processing gain

Choosing the Right Spectrum Sensor
Energy Detector
Sensing time Sensing time

Pilot Detector
Sensing time Partial coherent pilot detector

Feature Detector
Partial coherent feature detector

Noise wall Theoretical prediction

Clk. offset Freq. offset SNR SNR K1 K2 SNR

N~1/SNR2
Noise wall Noise estimation 18 mult, 320Kb BRAM 10, 353 slices Coarse sensing candidate band selection

N~1/(p*SNR)
Freq. offset ~ 0.01 ppm Partial coherent pilot processing 40 mult, 64Kb BRAM 22,683 slices Fine sensing if pilots/preambles available

N~K/SNR2
Clk offset ~ 1 ppm Partial coherent feature processing 82 mult, 3200Kb BRAM 21, 200 slices Fine sensing with increased sensitivity

Cross Layer Design for Spectrum Sensing
Network Management Control Channels Sensing MAC Sensing Signal Processing Sensing radio Spectrum Allocation Network Link Layer

Wideband signaling Wideband radio Physical Layer

• Sensing radio:
- Improving radio sensitivity and dynamic range using multiple antenn - Use of spatial filtering for blocker suppression

• Sensing MAC:
- Network cooperation for sensing and sensing protocols - Use of control channels for network management