Wideband Cyclostationary Spectrum Sensing and Modulation Classification
Eric RebeizAdvisor: Prof. Danijela Cabric
UCLA Electrical Engineering Department
Ph.D. Defense08/19/2013
D. Markovic / Slide 2
Wideband Cognitive Radio Concept & Vision
Cognitive Radios (CRs) opportunistically access the spectrum
2
How can we achieve this goal?
Future Promise of Wideband CR– Increase radio throughput– Support more users
Sensing Layer PHY Layer
MAC Layer
Higher Layers
In order to achieve this vision, practical wideband spectrum sensing challenges should be considered
D. Markovic / Slide 3
Wideband Spectrum Sensing & Classification Design Goal #1: Interference mitigation through sensing and classification Goal #2: Energy efficient processing
3
Sensing Requirements– False alarm rate– Detection probability– Short sensing time– Minimum SNR
Classification Requirements– Classification accuracy– Blind classification– Differentiate among M-QAM,
M-PSK, GMSK, M-PAM
Frequency800 MHz 1.6 GHz 2.5 GHz
CRPo
wer
Spe
ctru
m D
ensi
ty O
bser
ved
by C
Rs
D. Markovic / Slide 4
Algorithmic ChallengesImplementation Challenges
Wideband Sensing & Classification Challenges
4
– Robust sensing and classification to impairments
– Short sensing time
– Energy efficiency
– High computational complexity
– Carrier & sampling offsets
– Front-end nonlinearities
Wideband LO
ADC
LPF
AGC BPF @ f2
y2[n]
LNA
RF Front-end
Digital Back-End
BPF @ f1
y1[n]Spectrum Sensing
BPF @ fK
yK[n]
•••
Carrier frequency & Sampling clock offsets
Spectrum Sensing
Sele
ct B
ands
to S
ense
Spectrum Sensing
YesModulation
Classification
•••
Occupied?
Occupied?
Occupied?
Yes
YesModulation
Classification
Modulation Classification
Use channelNo
Use channelNo
Use channelNo
D. Markovic / Slide 5
Research Contributions
Proposed a sensing and classification method that is robust to carrier and sampling clock offsets– TSP’13, Asilomar’12
Proposed an energy efficient sensing and classification processor in blind scenarios– TCAS’13, Globecom’11, Milcom’11
Analyzed the impact of receiver nonlinearities on the sensing performance and proposed algorithmic solutions– TSP (to be submitted), Crowncom’13
5
D. Markovic / Slide 6
Outline
Cyclostationary feature detection overview
Spectrum sensing and classification under receiver impairments
Blind energy efficient sensing and classification
Impact of nonlinearities and their compensation
Summary of contributions
6
Cyclostationary feature detection overview
Spectrum sensing and classification under receiver impairments
Blind energy efficient sensing and classification
Impact of nonlinearities and their compensation
Summary of contributions
D. Markovic / Slide 7
Cyclostationary Detection (CD) Overview
CD can perform spectrum sensing and modulation classification
CD estimates the Cyclic Auto-Correlation (CAC) function (C-CAC*and NC-CAC)or the Spectral Correlation Function (SCF)
CAC at SCF
7
5
0.30
Cyclic Frequency (MHz)
CAC
Valu
e
0.250.200.150.100.05
010 15 20 25 30 35 40 45 500
0.450.400.35
0.50
Fsym = 6.25 MHzFcar = 12.5 MHz
BPSK
Fsym
2Fcar
2Fcar–Fsym 2Fcar+Fsym
D. Markovic / Slide 8
Cyclic Features of Different Modulation Types
Modulation type determines the present cyclic featuresFeatures at functions of and
8
Class Signals Cyclic Features at f(Fsym, Fcar)
1 M-QAM / M-PSK2 M-PAM / BPSK , , , 3 GMSK , ,
5
0.30
Frequency (MHz)
CAC
Valu
e
0.250.200.150.100.05
010 15 20 25 30 35 40 45 500
0.450.400.35
0.50
Fsym = 6.25 MHzFcar = 12.5 MHz
QAMFsym
5
0.30
Frequency (MHz)
CAC
Valu
e
0.250.200.150.100.05
010 15 20 25 30 35 40 45 500
0.450.400.35
0.50
Fsym = 6.25 MHzFcar = 12.5 MHz
MSKFsym
2Fcar–0.5Fsym
2Fcar+0.5Fsym
D. Markovic / Slide 9
Outline
9
Cyclostationary feature detection overview
Spectrum sensing and classification under receiver impairments
Blind energy efficient sensing and classification
Impact of nonlinearities and compensation
Summary of contributions
- E. Rebeiz, P. Urriza, D. Cabric, Optimizing Wideband Cyclostationary Spectrum Sensing under Receiver Impairments, in IEEE Transactions on Signal Processing, vol. 61, no. 15, pp. 3931-3943, Aug. 2013
- E. Rebeiz, P. Urriza, D. Cabric, Experimental Analysis of Cyclostationary Detectors Under Cyclic Frequency Offsets, in Proc. Asilomar Conference on Signals, Systems and Computers, Nov. 2012
D. Markovic / Slide 10
Feature Detection Under Receiver Impairments
10
What is the impact on sensing and classification?
Frequency
Modulation Type 1 Modulation Type 2 Modulation Type 3
Thermal noisePS
D
BPF
LNA
Wideband Front-End
AGC ADC
LPF
Cycl
ic A
utoC
orre
latio
n
Sampling Clock Offset
Imperfect Cyclic Frequencies (LO mismatch, Doppler)
<>|Rα1(ν)| γ
<>|Rα2(ν)| γ
<>|RαK(ν)| γ
•••
α1
α2
α3 αK
Wideband LO
B1
f1
B2
f2
B3
f3
BK
fK
Dec
isio
n Si
nk
D. Markovic / Slide 11
Ideal feature computed by
Feature under cyclic offset computed at
Signification Degradation in Low SNR
11
Cycl
ic F
eatu
re D
egra
datio
n (d
B)
-15
-10
-5
0
500 1000 1500 2000 2500 3000
Δα = 500 ppm
Δα = 1000 ppm
Δα = 2000 ppm
Sensing Time (N)High SNRs Low SNRs
D. Markovic / Slide 12
Robust Feature Detection
Proposed Multi-Frame CAC
where is the total sensing time
– yields the conventional CAC
Tradeoff: - N reduces effect of CFO - M yields non-coherent integration
Resulting composite relationship of CAC to ideal one
12
Term with N Term with M
D. Markovic / Slide 13
Tradeoff Between N and M
Single frame processing quickly degrades with CFO
Multi-frame processing spreads the energy across SCO and CFO
How to optimize M and N?13
Single Frame (M=1) Multiple Frames (M = 10)
Cyc
lic F
eatu
re
Cyc
lic F
eatu
re
D. Markovic / Slide 14
Optimization Design Strategy
CFO and SCO are non-deterministic circuit impairments– Design strategy is to optimize the average cyclic SNR
Conditional cyclic SNR defined as – Derived in closed form and given by
where , are functions of the pulse shape filter, N and M
14
D. Markovic / Slide 15
Performance Gains over Conventional Feature Detectors CFO and SCO zero mean normally distributed,
Under what ratios can we expect performance gains?
15
1
Aver
age
Cycl
ic S
NR
0
2
3
4
2 4 6 8 10 12
Theoretical
Empirical
Number of Frames (M)
x10-3
0.2
Probability of False AlarmPr
obab
ility
of D
etec
tion
0.10
0.40.3
0.60.5
0.80.7
0.91.0
0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8
M = 1
M = 2
M = 6
SNR = -5 dBNT = 5000
0.9 1
M = 12
D. Markovic / Slide 16
Best and Worst Case Performance Scenarios
Most gains achieved when CFO is more severe than SCO
16
Total Number of Samples NT
Aver
age
Cycl
ic S
NR
0 0.5 1 1.5 2 2.5 3 3.5 4 4.5 5x104
100
101
102
sco/cfo = 1/2
N = Nmin w/ resampling
sco/cfo > 1
Conventional CAC (M=1)
sco/cfo = 1/4sco/cfo = 1/20
D. Markovic / Slide 17
Contributions to Cyclic Feature Detectors
Analyzed performance loss due to carrier and sampling offsets
Proposed a new multi-frame statistic that achieves robust detection with optimum (N,M) pair
Significant improvement obtained when
17
D. Markovic / Slide 18
Outline
18
Cyclostationary feature detection overview
Spectrum sensing and classification under receiver impairments
Blind energy efficient sensing and classification
Impact of nonlinearities and their compensation
Summary of contributions
- E. Rebeiz, F. Yuan, P. Urriza, D. Markovic, D. Cabric, Energy-Efficient Processor for Blind Signal Classification in Cognitive Radio Networks, to appear in IEEE Transactions on Circuits and Systems I
- E. Rebeiz, D. Cabric, Blind Modulation Classification Based on Spectral Correlation and Its Robustness to Timing Mismatch, in Proc. IEEE Military Communications Conference, Nov. 2011
- E. Rebeiz, D. Cabric, Low Complexity Feature-based Modulation Classifier and its Non-Asymptotic Analysis, in Proc. IEEE Global Communications Conference, Dec. 2011
D. Markovic / Slide 19
Blind Sensing and Modulation Classification
What do we mean by blind?– Signals can appear anywhere in the wideband channel– Signals are not standard compliant
Spectrum sensing performed via a cyclic frequency search
Cyclostationary detection is not robust to CFO– Number of CAC computations becomes a burden
Cyclic detection not energy efficient in blind scenarios
19
10
0.30
Cyclic Frequency (MHz)
CAC
Valu
e
0.250.200.150.100.05
020 30 40 50 60 70 80 90 1000
0.450.400.35
0.50
Δα = 100 Hzαr = 0 : Δα : 100MHz
Compute |Rα(ν)|, Compare to γ
10
0.30
Cyclic Frequency (MHz)
CAC
Valu
e
0.250.200.150.100.05
020 30 40 50 60 70 80 90 1000
0.450.400.35
0.50
Δα = 10 KHzαr = 0 : Δα : 100MHz
Compute |Rα(ν)|, Compare to γ
D. Markovic / Slide 20
Proposed Hybrid Detection and Classification Processor
Most processing is in estimating signal parameters– What accuracy is required in the pre-processor?
20
Spectrum Sensing
FFT based Energy
Detection
No Knowledge of Parameters
- Detection of active signals- Coarse bandwidth estimate- Coarse carrier freq. estimate
Modulation Classification
I/Q Samples
Sensing Outputs
- Fine parameter estimation- Modulation type
Classification Outputs
ADC
Pres
ent M
odul
ation
SCF Computatio
nat (αi,fi) Co
ncat
enat
e
arg min|F – Vi|
2F
Fsym and Fcar
estimates
CAC @ at αi V1 V2 V3
1 2 36
Asymptotic Feature Vectors
I/Q
Dat
a
Parameter Estimation
Fsym and Fcar
estimates
0.1
CAC
0
0.2
4.5 5 5.5Cyclic Frequency x 106
arg max |Rx*|
0.5
CAC
0
1
6.35 6.4 6.45Cyclic Frequency x 107
arg max |Rx|
Symbol RateEstimation
Carrier FrequencyEstimation
Coarse estimates can result in CFOs on the order of 105 ppm
D. Markovic / Slide 21
Design Strategy to Minimize Consumed Energy
Discretization required for implementation purposes (– Total # CAC computations: ,
All blocks use CAC minimizing energy = minimizing # samples
21
Symbol-Rate Estimator
(CAC)
Modulation Type Classifier
(CAC)
Modulation Database
T
Carrier Frequency Estimator
(CAC)
NT,ΔαT
Nf,Δαf
fc
Nc
From CIC
Nc = # Samples for mod. type. class.NT = # Samples for symbol rate est.Nf = # Samples for carrier freq. est
ΔαT = symbol rate est. resolutionΔαf = carrier feq. est. resolution
D. Markovic / Slide 22
Tradeoffs Between Pre-Processor Resolutionand Classification Accuracy Feature at is the weakest among all features
– determined by signals in class 1 (M-QAM, M-PSK)
For every SNR, there is a maximum CFO tolerable to meet Pc
22
SNR (dB) max (ppm)
10 1000
8 900
6 800
4 500
2 400
0 100
D. Markovic / Slide 23
Feasible Region for Classification
All triplets that meet form the feasible region
Features at the carrier frequency are stronger: looser requirement
23
400 500 600 700 800 900 10003600380040004200440046004800500052005400
Δ αf (
ppm
)
Nc
= 5
20
Nc
= 5
35
Nc
= 7
50
Nc
= 6
20Nc
= 5
50ΔαT (ppm)
Nc
= 5
00
Nc
= 5
10
Feasible Region
SNR = 10 dB
D. Markovic / Slide 24
Optimum Operating Point in Feasible Region
Operating near the boundary of feasible region is most efficient
24
0 100 200 300 400 500 600 700 800 900 10001000
1500
2000
2500
3000
3500
4000
45005000
Feasible Region
Δ αf (
ppm
)
ΔαT (ppm)
10.3 µJ
24.9 µJ
59.5 µJ
D. Markovic / Slide 25
Blind Modulation Classification Contributions
The tradeoff between parameter estimation and the resulting classification accuracy was analyzed
An optimization strategy has been developed to minimize the total consumed energy while meeting the classification requirement
25
D. Markovic / Slide 26
Outline
26
Cyclostationary feature detection overview
Spectrum sensing and classification under receiver impairments
Blind energy efficient sensing and classification
Impact of nonlinearities and their compensation
Summary of contributions
- E. Rebeiz, A. Shahed, M. Valkama, and D. Cabric, Analysis of Spectrum Sensing under RF Non-Linearities and Compensation Algorithm, in preparation for submission to IEEE Transactions on Signal Processing
- E. Rebeiz, A. Shahed, M. Valkama, D. Cabric, Suppressing RF Front-End Nonlinearities in Wideband Spectrum Sensing, in Proc. IEEE CROWNCOM 2013
D. Markovic / Slide 27
Wideband Receiver Nonlinearities
Typical LNA IIP3 of -10 dBm, LNA linear gain of 35 dB
Mixer nonlinearity are also important, but have less of an impact – IIP2 dBm, IIP3 dBm
Def: IIP3 = Input power at which linear term power equals power of 3rd order term27
0.1Out
put P
ower
(mW
)
0
0.2
0.4
0.6
0.8
Nonlinear LNA
Input Power (mW)0 0.5 1 1.5 2 2.5
Linear LNA
x 10-3
Wideband LO LPF
AGC
RF Front-End
yw[n]LNA ADC
x[n]
D. Markovic / Slide 28
Received Signal with Intermodulation Terms
SOI @ , blockers @ , such that Due to nonlinearity, intermodulation (IMD) term appears at
How does the IMD term affect the sensing performance?
28
Wideband LO
ADCyW[n]
BPF
AGC
BPF @ fIF
y[n]LNA
Spectrum Sensing(CD or ED)
RF Front-end DSP
fb1 fb2 fc=2fb2-fb1
z1
z0z2
fIF
y[n]
f1 f2 fIF=2f2-f1
β1z1
β1z0β1z2
D. Markovic / Slide 29
SIR = -67 dB, SNR = 10 dB, N = 500 Samples
Cyclostationary Detection Energy Detection
Severe degradation in sensing performanceWhat are possible compensation methods?
Degradation Depends on Blocker Modulation
29
Prob
abili
ty o
f Det
ectio
n 1.0
0.2
0.4
0.6
0.05 0.1 0.15 0.2 0.25 0.3Probability of False Alarm
0
z1[n] QPSK, z2[n] QPSK, z0[n] 4PAMIdeal LNA, z0[n] 4PAM
z1[n] 4PAM, z2[n] 4PAM, z0[n] 4PAMz1[n] QPSK, z2[n] 4PAM, z0[n] 4PAMz1[n] 4PAM, z2[n] QPSK, z0[n] 4PAM
Analytical Simulation
Prob
abili
ty o
f Det
ectio
n 1.0
0.2
0.4
0.6
0.05 0.1 0.15 0.2 0.25 0.3Probability of False Alarm
0
z1[n] QPSK, z2[n] QPSK, z0[n] 4PAMIdeal LNA, z0[n] 4PAM
z1[n] 4PAM, z2[n] 4PAM, z0[n] 4PAMz1[n] QPSK, z2[n] 4PAM, z0[n] 4PAMz1[n] 4PAM, z2[n] QPSK, z0[n] 4PAM
Analytical Simulation
D. Markovic / Slide 30
Increasing Sensing Time is not Effective
Recall that under – How do we set the decision threshold?
30
Det
ectio
n Pr
obab
ility 0.9
0.5
0.7
1000 2000 3000 4000 5000Sensing Time (N)
0.3
1
CD - IdealED - Ideal
CD - UncompensatedED - Uncompensated
SNR = 3 dB, SIR = -70 dB
6000
Analytical Simulation
D. Markovic / Slide 31
Impact of Uncertainties on False Alarm
Setting the threshold requires knowledge of – Blocker and noise power– Blocker modulation
Accurate estimation of parameters needed for threshold setting
CD is more robust to uncertainties than ED31
Prob
abili
ty o
f Fal
se A
larm
0.8
0.2
0.4
0.6
-0.6 -0.2 0.2 0.6 1IIP3 Uncertainty (dB)
-10
1
CD – SIR = -75 dBED – SIR = -75 dB
Target False AlarmCD – SIR = -70 dBED – SIR = -70 dB
SNR = 0 dB, N = 500Analytical Simulation
Prob
abili
ty o
f Fal
se A
larm
0.8
0.2
0.4
0.6
-0.6 -0.2 0.2 0.6 1Blocker Power Uncertainty (dB)
-10
1
CD – SIR = -75 dBED – SIR = -75 dB
Target False AlarmCD – SIR = -70 dBED – SIR = -70 dB
SNR = 0 dB, N = 500Analytical Simulation
– Receiver IIP3
D. Markovic / Slide 32
Actual IMD term , Estimated IMD term Compensation method:
Performance of compensation is modulation dependent
Intermodulation Term Compensation
32
0.3
SIR (dB)
Prob
abili
ty o
f Det
ectio
n
0.20.1
0.50.4
0.70.6
0.90.8
1.0
-74 -72 -70 -68 -66 -64 -62 -60
ED CD
Linear LNANonlinear LNA
Compensated
SNR = 3 dB Analytical Simulation
z1, z2, z0 4PAM
0.3
SIR (dB)
Prob
abili
ty o
f Det
ectio
n
0.20.1
0.50.4
0.70.6
0.90.8
1.0
-74 -72 -70 -68 -66 -64
Linear LNANonlinear LNA
Compensated
SNR = 3 dB
-62 -60
Analytical Simulation
ED CD
z1, z2 QPSK, z0 4PAMBPF @ fIF
y[n] CAC + Thresh.
Decisiony[n]~
z[n]^
BPF @ f1
BPF @ f2
fIF=2f2-f1
fIF=2f2-f1
y = β1z0+ 1.5β3z1*z2
2
z[n]^
z1[n]~
z2[n]~
Compute
________
β1 3
z2[n]2~z1[n]*~
3β3/2
D. Markovic / Slide 33
Compensation requires– Modulation type of blockers– Blocker strength
Modulation Dependent Compensation Algorithm
33
– Receiver IIP3
ADC
Estimate Noise Power
Compensated Detector
Choo
se T
hres
hold
w
/Noi
se O
nly
CAC
Com
puta
tion
+ Th
resh
old
Com
paris
on
Sens
ing
Resu
ltDemodulate + Remodulate Blockers + Estimate IMD term
Estimate Blockers’ Powers
Slow Varying Parameters
yW[n]
BPF @ fIF
y[n]
BPF @ f1
BPF @ f2
z1[n]~
z2[n]~
β3 ^
Fast Varying Parameters
z2[n]2~z1[n]*~ /β13
Estimate Blockers’ Modulation
BPF @ fIFy[n]
Adaptive Estimation of β3
LMS Adaptive Algorithm
Delay
D. Markovic / Slide 34
Performance Gains due to Demodulation / Remodulation
34
Prob
abili
ty o
f Det
ectio
n
0.8
0.2
0.4
0.6
-5 0 5 10 15 20SNR (dB)
1.0
-10
SIR = -65 dB
ED – ProposedΔIIP3 = 0.25 dB, Δp = 100.4/20 CD – ProposedΔIIP3 = 0.25 dB, ΔP = 100.4/20
Analytical Simulation
ED CD
Linear LNANonlinear LNA
Compensated Alg#1
Prob
abili
ty o
f Det
ectio
n
0.8
0.2
0.4
0.6
-5 0 5 10 15 20SNR (dB)
1.0
-10
SIR = -65 dB
ED – ProposedΔIIP3 = 0.25 dB, ΔP = 101/20
CD – ProposedΔIIP3 = 0.25 dB, ΔP = 101/20
Analytical SimulationLinear LNANonlinear LNA
Compensated Alg#1
ED CD
When estimates are off, residual term degrades detection performance
D. Markovic / Slide 35
Nonlinearity Contributions
Analyzed the degradation in sensing performance due to IMD term
Showed that impact of IMD is dependent on blockers’ modulation
Proposed a modulation-aware IMD compensation based on demodulation / remodulation + sample-by-sample subtraction
35
D. Markovic / Slide 36
Ph.D. Thesis Contributions
Analyzed the practical challenges in wideband spectrum sensing and modulation classification
Proposed energy efficient algorithmic solutions that are robust to the considered impairments
Future work: analyze additional wideband challenges such as– High sampling rates– Mixer nonlinearities, I/Q mismatch– Blockers suppression through beamforming– Modulation classification of overlapped signals
36
Thank you very much!Questions?
Acknowledgments
DARPA CLASIC ProgramMy advisor & all faculty on my committee
Lab mates, most importantly Paulo & Fang-Li
D. Markovic / Slide 38
Blind Estimation of Receiver IIP3
Objective function given by
Resulting IIP3 offset is 0.15 dB
38
Wideband LO
ADCyW[n]
BPF
AGC
BPF @ fIF
y[n]LNA
RF Front-end DSP
z[n]^
ComputeBPF @ f1
BPF @ f2
f1 f2 fIF=2f2-f1
fIF=2f2-f1
z[n]^
z1[n]~
z2[n]~
z1[n]~ z2[n]~
y[n]
z2[n]2~z1[n]*~
LMS Adaptive Algorithm β3
^
Estim
ation
of β
3 Par
amet
er
0
-6000
-4000
-2000
2 4 6 8 10Iterations
-8000
1000
Actual β3 Q branchI branch
x104
D. Markovic / Slide 39
Proposed Architecture for Uncompensated Detectors
This architecture makes sure that the uncompensated detectors are operating at the right point on the ROC point
39
Wideband LO
ADC
LPF
AGCLNA
RF Front-end
Adaptively Estimate β3
Estimate Blocker Power
Estimate Blockers’ Modulation
Estimate Noise Power
Uncompensated Detector
Estim
ate
unde
r H0
+ Ch
oose
Thr
esho
ld
CAC Computation
Sens
ing
Resu
lt
<>|Rα(ν)| γ
BPF
D. Markovic / Slide 40
Published Articles Journal articles- E. Rebeiz, A. Shahed, M. Valkama, D. Cabric, Spectrum Sensing under RF Non-Linearities: Theoretical Analysis and Compensation Algorithm, in preparation to submission to IEEE Transactions on Signal Processing- E. Rebeiz, P. Urriza, D. Cabric, Optimizing Wideband Cyclostationary Spectrum Sensing under Receiver Impairments, in IEEE Transactions on Signal Processing, 2013- E. Rebeiz, F. Yuan, P. Urriza, D. Markovic, D. Cabric, Energy-Efficient Processor for Blind Signal Classification in Cognitive Radio Networks, in IEEE Transactions on Circuits and Systems I, 2013- P. Urriza, E. Rebeiz, P. Pawełczak, D. Čabrić, Computationally Efficient Modulation Level Classification Based on Probability Distribution Distance Functions, IEEE Communications Letters, 2011- P. Urriza, E. Rebeiz, D. Cabric, Multiple Antenna Cyclostationary Spectrum Sensing Based on the Cyclic Correlation Significance Test, in IEEE Journal on Selected Areas in Communications, 2013- P. Sofotasios, E. Rebeiz, L. Zhang, T. Tsiftsis, S. Freear, D. Cabric, Energy Detection-Based Spectrum Sensing over Generalized and Extreme Fading Channels, in IEEE Trans. Vehicular Technology, 2013- P. Urriza, E. Rebeiz, D. Cabric, Optimal Discriminant Functions Based On Sampled Distribution Distance for Modulation Classification, accepted for publication in IEEE Communications Letters, 2013 Selected Conference articles- E. Rebeiz, A. Shahed, M. Valkama, and D. Cabric, Suppressing RF Front-End Nonlinearities in Wideband Spectrum Sensing, in Proc. IEEE CROWNCOM, 2013- E. Rebeiz, P. Urriza, D. Cabric, Experimental Analysis of Cyclostationary Detectors Under Cyclic Frequency Offsets, in Proc. IEEE Asilomar Conference on Signals, Systems and Computers,2012- E. Rebeiz, V. Jain and D. Cabric, Cyclostationary-Based Low Complexity Wideband Spectrum Sensing using Compressive Sampling, in Proc. IEEE ICC, 2012- E. Rebeiz, D. Cabric, Blind Modulation Classification Based on Spectral Correlation and Its Robustness to Timing Mismatch, in Proc. IEEE MILCOM, 2011- E. Rebeiz, D. Cabric, Low Complexity Feature-based Modulation Classifier and its Non-Asymptotic Analysis, in Proc. IEEE GLOBECOM, 2011
40
D. Markovic / Slide 41
Theoretical Derivation of IMD Impact on Sensing Performance
41
D. Markovic / Slide 42
Maintaining a Constant False Alarm Rate
Typical ACI scenarios only require estimation of blocker strength
Setting the detection threshold requires– Knowledge of modulation of the blockers– Blockers’ power and the noise power– Accurate knowledge of parameter
42
f (Hz)
PSD
D. Markovic / Slide 43
Design Strategy of Compensation Algorithm
We model the imperfections as – , for
Residual IMD term is given by Objective is to achieve
43
Blocker Uncertainty (dB)
IIP3
Unc
erta
inty
(dB)
-1.5-1-0.500.511.52
-2 -1.6 -1.2 -0.8 -0.4 0 0.4 0.8 1.2 1.6 2
1
0.75
0.5
0.25
0
-0.25
-0.5
-0.75
-1
Objective is met
Example of feasible region under SIR = -65 dB