system design issues in building a cognitive radio network: ieee 802.22
DESCRIPTION
“What I did in my summer internship”. System Design Issues in building a Cognitive Radio Network: IEEE 802.22. Detection of DTV signals at very low SNR using PN sequences. Joint work with Steve Shellhammer (Qualcomm) & Rahul Tandra (U.C. Berkeley). - PowerPoint PPT PresentationTRANSCRIPT
Suhas Mathur, WINLAB, Rutgers University Summer internship @ Qualcomm 10th October 2006
System Design Issues in building a
Cognitive Radio Network: IEEE 802.22
Detection of DTV signals at very low SNR using PN sequences
Joint work with Steve Shellhammer (Qualcomm) & Rahul Tandra (U.C. Berkeley)
“What I did in my summer internship”
Suhas Mathur, WINLAB, Rutgers University Summer internship @ Qualcomm 10th October 2006
•Apparent scarcity of spectrum•Spectrum use model flawed – allocated but not utilized•J. Mitola coins ‘cognitive radio’ -Ph.D. thesis (May 2000)
•FFC issues NPRM for TV bands (May 2004)
•802.22 is born (Sept 2004)
•How to protect the ‘incumbent’ ?•May 2006: v0.1 of the 802.22 Draft is shipped•Sept. 2006: FCC Docket: Retail by Feb 2009.•Sensing specs. still not met, broadcasters dissatisfied•Unlicensed economic model still attractive
–3G auction: $17 Bn , $34 Bn , $46 Bn
A brief history of Cognitive Radio and IEEE 802.22
3Suhas Mathur, WINLAB, Rutgers University Summer internship @ Qualcomm 10th October 2006
802.22 deployment scenario
802.22: Fixed wireless broadband access network operating in TV bands (Channel 2 -51)
Cellular system, central BS, CPEs (Consumer Premise Equipment)
Targeted at but not limited to rural deployment
Strict requirements on protections of existing licensed services
4Suhas Mathur, WINLAB, Rutgers University Summer internship @ Qualcomm 10th October 2006
Incumbent protection numerology
Parameter TV broadcasting
Part 74 devices
Channel detection time < 2 sec < 2sec
Incumbent detection threshold
-116 dBm -107dBm
Required prob(detection) 90% 90%
Max. False alarm allowed 10% 10%
5Suhas Mathur, WINLAB, Rutgers University Summer internship @ Qualcomm 10th October 2006
Why do we need to detect at such low SNR?
6Suhas Mathur, WINLAB, Rutgers University Summer internship @ Qualcomm 10th October 2006
Why do we need to detect at such low SNR?
Shadow faded CPE
7Suhas Mathur, WINLAB, Rutgers University Summer internship @ Qualcomm 10th October 2006
Why do we need to detect at such low SNR?
Shadow faded CPE
INTERFERENCE
8Suhas Mathur, WINLAB, Rutgers University Summer internship @ Qualcomm 10th October 2006
‘Hidden incumbent’ problem• An in-band Incumbent
appears at a location where BS cannot sense it but CPE can.
• CPE wishes to inform BS about incumbent but may loose sync to BS due to interference from incumbent
• BS continues to transmit on the channel occupied by incumbent, causing interference to it.
9Suhas Mathur, WINLAB, Rutgers University Summer internship @ Qualcomm 10th October 2006
‘Hidden cell’ problem
10Suhas Mathur, WINLAB, Rutgers University Summer internship @ Qualcomm 10th October 2006
Sharing licensed spectrum
Vertical sharing
Horizontal sharing
11Suhas Mathur, WINLAB, Rutgers University Summer internship @ Qualcomm 10th October 2006
Horizontal spectrum sharing• An unlicensed user must
adhere to FCC guidelines for protection of incumbents.
• But other unlicensed systems on the TV bands do not enjoy the same protection.
• This means a Cognitive Radio must not only detect, but also classify signals.
• Simple narrowband power detection [Tandra & Sahai] does not achieve this objective
Suhas Mathur, WINLAB, Rutgers University Summer internship @ Qualcomm 10th October 2006
Part II
Detection of DTV signals at very low SNR using PN sequences
13Suhas Mathur, WINLAB, Rutgers University Summer internship @ Qualcomm 10th October 2006
Outline of Part II
• The ATSC standard and PN sequences• Optimal & approximate detectors• Problem with detector a proposed detector• Our detection algorithms• Effects of multipath fading on PN
correlation• Fundamental limits in PN seq. detection• Sensitivity to pilot estimate• A comparison of PN and pilot detection• Pilot spectral line detection
14Suhas Mathur, WINLAB, Rutgers University Summer internship @ Qualcomm 10th October 2006
ATSC DTV numerology
• Symbol rate: 10.76 Msymbols/sec
• Data Segment SYNC: A ‘1001’ pattern at the beginning of each segment
• Data Field SYNC: An entire segment containing PN sequences: PN511 + PN63 + PN63 + PN63
15Suhas Mathur, WINLAB, Rutgers University Summer internship @ Qualcomm 10th October 2006
ATSC framing structure
A single VSB Data Segment
The Data Field SYNC
16Suhas Mathur, WINLAB, Rutgers University Summer internship @ Qualcomm 10th October 2006
Typical ATSC DTV spectrumPilot tone
6 Mhz = 1 DTV channel
f_IF = 5.38 Mhz
VSB spectrum at IF
Nyquist Rolloff = 0.1152
17Suhas Mathur, WINLAB, Rutgers University Summer internship @ Qualcomm 10th October 2006
Simulation Methodology
DTV signal captures
(50 DTV signals captured at high SNR,
provided to us by MSTV Corp.)
+X
Scale
Gaussian Noise
Pull out reqd. # of samples Detection
algorithm
Threshold calibration
Signal at desired SNR
Multipath fading, noise
Test locations in NY, Washington
18Suhas Mathur, WINLAB, Rutgers University Summer internship @ Qualcomm 10th October 2006
The optimal PN seq. detector
• Assumption: exactly 1 PN sequence present
• Generalized Likelihood Ratio Test (GLRT) in favor of H1 iff:
• is the MLE given by:
• GLRT is:
19Suhas Mathur, WINLAB, Rutgers University Summer internship @ Qualcomm 10th October 2006
Approximate detector• Neglect the signal in alternate hypothesis (low
SNR)
• is the MLE given by
• GLRT is:
• Threshold can be found easily by using:
simply correlate and compare max
value with
20Suhas Mathur, WINLAB, Rutgers University Summer internship @ Qualcomm 10th October 2006
A detector proposed by France Telecom
• Correlate the received signal, r(n), with the known PN sequence
1
0
)()(N
nkn nxrkp
• Use a simple low pass filter to estimate the mean and the variance of the correlator output
• An ATSC DTV is declared detected when
2))()()(1()1()(
)()1()1()(
kpkkk
kpkk
)'()'( 2 kck
( and c are constants set by the BS)
21Suhas Mathur, WINLAB, Rutgers University Summer internship @ Qualcomm 10th October 2006
The Problem with this Detector
• The estimate of the mean approach zero even for noise only input
• This implies that the test statistic can approach zero even for a noise only input signal
• This results in false alarms. The detector threshold ‘c’ cannot be calibrated for a given false alarm probability pFA.
2( ) ( ) / ( )T k k k
( ) [ ( )] ( ) ( ) 0k E p k E r n k x n
22Suhas Mathur, WINLAB, Rutgers University Summer internship @ Qualcomm 10th October 2006
Test statistic with noise-only input
Output of the correlator when only filtered receiver
noise is fed as input.
Ratio shows false peaks because mean goes very close to
zero.
)(/)( 2 kk
Correlator output Detector output
23Suhas Mathur, WINLAB, Rutgers University Summer internship @ Qualcomm 10th October 2006
Detector operation with DTV signal + Noise (SNR = 0 dB)
Output of the correlator over duration of 1 ATSC field when DTV signal + noise is fed as input. The
peak indicated the position of the PN sequence.
Ratio again shows false peaks because mean goes very close to zero
)(/)( 2 kk
Correlator output Detector output
24Suhas Mathur, WINLAB, Rutgers University Summer internship @ Qualcomm 10th October 2006
A simple ‘fix’
• Take absolute value of the correlator output before feeding as input to the detector.
• Mean cannot go arbitrarily close to zero
• The ratio is random for noise input.
)()1()1()( kpkk
1
0
)()(N
nkn nxrkp
)(/)( 2 kk
25Suhas Mathur, WINLAB, Rutgers University Summer internship @ Qualcomm 10th October 2006
Modified Detector with noise input
Absolute value of the correlator output when only filtered receiver
noise is fed as input.
Ratio is random)(/)( 2 kk
Correlator output Detector output
26Suhas Mathur, WINLAB, Rutgers University Summer internship @ Qualcomm 10th October 2006
Modified detector with DTV Signal + Noise (SNR = 0 dB)
Absolute value of correlator output over duration of 1 ATSC field when DTV signal
+ noise is fed as input. The peak indicated the position of the PN
sequence.
Ratio shows peak at location of PN sequence
)(/)( 2 kk
Correlator output Detector output
27Suhas Mathur, WINLAB, Rutgers University Summer internship @ Qualcomm 10th October 2006
A simple test statistic• Correlate received signal with the following
sequence:
[PN511 + PN63 + 63 zeros + PN63]
• Find the tallest peak in the correlator output• Compare its magnitude with threshold
28Suhas Mathur, WINLAB, Rutgers University Summer internship @ Qualcomm 10th October 2006
Performance of a simple correlator
29Suhas Mathur, WINLAB, Rutgers University Summer internship @ Qualcomm 10th October 2006
The effect of multipath
Phase reversal
48.4 ms
30Suhas Mathur, WINLAB, Rutgers University Summer internship @ Qualcomm 10th October 2006
The effect of multipath
Multiple peaks
6 s
~ 2km dominant echo
31Suhas Mathur, WINLAB, Rutgers University Summer internship @ Qualcomm 10th October 2006
The effect of multipath
Multiple peaks
6 s
Can we utilize multipath in a positive way ?
(Lessons from CDMA ?)
~ 2km dominant echo
32Suhas Mathur, WINLAB, Rutgers University Summer internship @ Qualcomm 10th October 2006
Fundamental limits in PN detection
• Using a larger received signal for detection does not always improve performance.
multipath fading
noise enhancement
Multipath, jitter and small variations in clock frequency cause timing offsets resulting in misaligned peaks by +/-1,2 samples
33Suhas Mathur, WINLAB, Rutgers University Summer internship @ Qualcomm 10th October 2006
Using the Data Segment SYNC• The DATA Segment SNYC is a 1001 pattern• Very non unique but much more frequent (every 77.3 s=1 seg.)
• Present in noise and data quite often, but can we use its periodicity?
• A long sequence of the following form can be used for correlation:
[1001 (828 zeros) 1001 (828 zeros) … N times ]
34Suhas Mathur, WINLAB, Rutgers University Summer internship @ Qualcomm 10th October 2006
Performance of ‘1001 correlator’
35Suhas Mathur, WINLAB, Rutgers University Summer internship @ Qualcomm 10th October 2006
Aligning peak polarities helps! (sometimes)
36Suhas Mathur, WINLAB, Rutgers University Summer internship @ Qualcomm 10th October 2006
The issue of Multiple antennas
• The height and polarity of a PN correlation peak depends on the instantaneous fade.
• Two or more independent fades would provide different correlation peaks.
• If multiple antennas can provide statistically independent fading, they can help.
• A simple system would be to run 2 parallel PN detectors and OR their decisions. i.e.
Miss detection = (Miss detection 1) AND (Miss detection 2)
PN sequence detection
• Power detector is affected by shadowing
• Therefore multiple sensors located in INDEPENDENT shadow fades would help
• Existence of such independent CPEs not guaranteed (incumbents not happy with the idea)
• Multiple antenna on a single CPE do not help (same shadow fade)
Power detection
37Suhas Mathur, WINLAB, Rutgers University Summer internship @ Qualcomm 10th October 2006
Multiple antennas
• Running 2 parallel detectors
• Assumption: Fading processes at two antennas are statistically independent
• We use fields from widely time-separated part of a DTV signal capture to satisfy independent fade assumption
• At 600 Mhz, /2 = 25 cm. A practical antenna array can be built.
38Suhas Mathur, WINLAB, Rutgers University Summer internship @ Qualcomm 10th October 2006
The effect of pilot estimation error
• The pilot frequency is used in downconverting a passband signal to recover the baseband transmit signal.
• How sensitive is PN detection to the estimate of the pilot frequency?
• We attempt to measure the effect of pilot estimation error on a PN correlation peak
39Suhas Mathur, WINLAB, Rutgers University Summer internship @ Qualcomm 10th October 2006
The effect of pilot estimation error
40Suhas Mathur, WINLAB, Rutgers University Summer internship @ Qualcomm 10th October 2006
The effect of pilot estimation error
• At low enough SNR even 0.2% error in pilot frequency estimate can be disastrous for PN correlation
• Pilot estimation needs to be accurate
• Why not use the pilot line as a means for DTV signal detection?
Suhas Mathur, WINLAB, Rutgers University Summer internship @ Qualcomm 10th October 2006
Part III
Spectral line detection of a DTV pilot[Extension of ‘Narrowband pilot energy detection’ done by
Rahul Tandra]
42Suhas Mathur, WINLAB, Rutgers University Summer internship @ Qualcomm 10th October 2006
Power spectrum of a DTV signal at -22 dB
43Suhas Mathur, WINLAB, Rutgers University Summer internship @ Qualcomm 10th October 2006
Performance of pilot spectral line detection
• Signal shows a strong pilot• Detector performance is
very good even at -25 dB !
strong pilot
Pilot energy detection
[Tandra]
44Suhas Mathur, WINLAB, Rutgers University Summer internship @ Qualcomm 10th October 2006
Performance of pilot spectral line detection
• Signal shows a moderately strong pilot
• Detector begins to fail at -21dB
Moderate pilot
45Suhas Mathur, WINLAB, Rutgers University Summer internship @ Qualcomm 10th October 2006
Performance of pilot spectral line detection
• Pilot is almost completely absent from signal
• Detector begins to fail at -16 dB
Very weak pilot
46Suhas Mathur, WINLAB, Rutgers University Summer internship @ Qualcomm 10th October 2006
Good, moderate and bad pilots
47Suhas Mathur, WINLAB, Rutgers University Summer internship @ Qualcomm 10th October 2006
Doubling the listening time
Very weak Pilot
48Suhas Mathur, WINLAB, Rutgers University Summer internship @ Qualcomm 10th October 2006
Doubling the listening time
Moderate Pilot
49Suhas Mathur, WINLAB, Rutgers University Summer internship @ Qualcomm 10th October 2006
Some gains from listening longer
50Suhas Mathur, WINLAB, Rutgers University Summer internship @ Qualcomm 10th October 2006
Noise Uncertainty
• For signals with weak pilots, uncertainty in the noise power estimate can drastically change performance.
• We assume = +/- 1 dB.
• How does performance change?
51Suhas Mathur, WINLAB, Rutgers University Summer internship @ Qualcomm 10th October 2006
Noise Uncertainty
• Signal shows a strong pilot• Detector performance worse
by approx 2.5 dB. • Detector begins to fail at -24
dB
strong pilot
52Suhas Mathur, WINLAB, Rutgers University Summer internship @ Qualcomm 10th October 2006
Noise Uncertainty
• Signal shows a moderately strong pilot
• Worse by ~2 dB• Detector begins to fail at -17
dB
Moderate pilot
53Suhas Mathur, WINLAB, Rutgers University Summer internship @ Qualcomm 10th October 2006
Noise Uncertainty
• Pilot is almost completely absent from signal
• Worse by ~2 dB• Detector begins to fail at -12
dB
Very weak pilot
54Suhas Mathur, WINLAB, Rutgers University Summer internship @ Qualcomm 10th October 2006
Noise Uncertainty
• Pilot is almost completely absent from signal
• Worse by ~2 dB• Detector begins to fail at -12 dB• If we listen for double the time
detector begins to fail at -14 dB
Very weak pilot
55Suhas Mathur, WINLAB, Rutgers University Summer internship @ Qualcomm 10th October 2006
A comparison of PN detection and pilot energy detection
PN Sequence Pilot Energy Detection
Occurs once in 24.2 ms Always present in time
Frequency content over entire DTV spectrum
Present at a specific frequency (14 known pilot freq. location)
Power in PN sequence is 1/313rd of the total energy in signal (25 dB below)
Power in pilot is 11 dB below avg signal power.
Need to sense for longer to average out noise
Need to sense for shorter duration.
Cannot improve performance by sensing for infinitely long duration.
Noise uncertainty results in an SNR wall effect below which signal cannot be sensed.
Requires knowledge of pilot frequency Does not require knowledge of PN seq.
Achieves signal classification. Does not achieve signal classification
56Suhas Mathur, WINLAB, Rutgers University Summer internship @ Qualcomm 10th October 2006
Proposed dual mode DTV detector
• Pilot detection detects pilot and triggers PN detection for signal classification (confirms DTV)
• Multiple antennas (help both pilot and PN)• Detection quiet periods are kept small.
57Suhas Mathur, WINLAB, Rutgers University Summer internship @ Qualcomm 10th October 2006
Conclusions and future directions
• PN Correlation detector performance limited by multipath fading, noise and jitter.
• Difficult to accurately estimate multipath at low SNR.
• Multiple antennas on a single detector can help (statistically independent fading processes)
• Pilot spectral line detection is promising but does not classify signal as DTV.
• A ‘dual mode’ detector utilizing – Pilot spectral line detection for detection and – PN correlation (with multiple antennas) for classification.
can achieve quick and reliable DTV incumbent protection
58Suhas Mathur, WINLAB, Rutgers University Summer internship @ Qualcomm 10th October 2006
The internship experience
• Worked in the standards division of Corporate R&D
• Some exposure to systems, IEEE standards procedures (tedious).
• Focus on practical feasibility. • Economic & political side of technology.• Overall – a satisfying, busy 3 months (no time to work on WINLAB research
simultaneously )
Suhas Mathur, WINLAB, Rutgers University Summer internship @ Qualcomm 10th October 2006
Thank you
Suhas Mathur, WINLAB, Rutgers University Summer internship @ Qualcomm 10th October 2006
Backup slides
61Suhas Mathur, WINLAB, Rutgers University Summer internship @ Qualcomm 10th October 2006
Incumbent protection numerology
Required prob(detection) = 90 % False alarm allowed = 10%
62Suhas Mathur, WINLAB, Rutgers University Summer internship @ Qualcomm 10th October 2006
Hidden wireless microphones
• An example of hidden incumbents
• Microphones are low power devices and can appear and disappear on a finer time scale.
• Wireless microphones are likely candidates for hidden incumbents.
• Propagation curves show that it Is virtually impossible for a BS to detect a microphone at the edge of a WRAN cell.
Distance from wireless mic (km)
Rec
eive
d po
wer
(dB
m)
63Suhas Mathur, WINLAB, Rutgers University Summer internship @ Qualcomm 10th October 2006
The effect of pilot estimation error
64Suhas Mathur, WINLAB, Rutgers University Summer internship @ Qualcomm 10th October 2006
The effect of pilot estimation error
65Suhas Mathur, WINLAB, Rutgers University Summer internship @ Qualcomm 10th October 2006
Do multiple antennas help?
• The height and polarity of a PN correlation peak depends on the instantaneous fade.
• Two or more independent fade would provide different correlation peaks.
• If multiple antennas can provide statistically independent fading, they can help.
• A simple (suboptimal) system would be to run 2 parallel PN detectors and OR their decisions. Therefore:
Miss detection = (Miss detection 1) AND (Miss detection 2)
66Suhas Mathur, WINLAB, Rutgers University Summer internship @ Qualcomm 10th October 2006
Multipath Fading
Source: Gorka Guerra, Pablo Angueira, Manuel M. Vélez, David Guerra, Gorka Prieto, Juan Luis Ordiales, and Amaia Arrinda ‘Field Measurement Based Characterization of the Wideband Urban Multipath Channel for Portable DTV Reception in Single Frequency Networks’ IEEE TRANSACTIONS ON BROADCASTING, VOL. 51, NO. 2, JUNE 2005 171