dof: a local wireless information plane
DESCRIPTION
DOF: A Local Wireless Information Plane. Steven HongSachin Katti. Stanford University. August 17, 2011. 1. Unlicensed spectrum (e.g. ISM Band - 2.4 GHz) has historically been managed “socially ”. Problem. - PowerPoint PPT PresentationTRANSCRIPT
DOF: A Local Wireless Information Plane
Stanford University
Steven Hong Sachin Katti
1
August 17, 2011
ProblemUnlicensed spectrum (e.g. ISM Band - 2.4 GHz) has historically been
managed “socially”
How can we design a smart radio which maximizes throughput while causing minimal
harm to coexisting radios?2
Can we use current mechanisms to design these smart radios?
Current coexistence mechanisms • Carrier Sense, RTS/CTS• Rate Adaptation• Adaptive Frequency Hopping• ...
Current mechanisms are not sufficient for designing high performance smart radios
3
How would we build a smart radio which coexists with legacy devices?
Microwave
Smart Transmitter Smart Receiver
1. The protocol types operating in the local vicinity 2. The spectrum occupancy of each type3. The spatial directions of each type
Knowledge of
WiFi AP Heart Monitor
AoA
Freq2.3 GHz
2.5 GHz
0 °18
0°
AoA
Freq2.3 GHz
2.5 GHz
0 °18
0°
Freq2.3 GHz
2.5 GHzFreq
2.3 GHz
2.5 GHz
4
DOF(Degrees Of Freedom)
Local wireless information plane which provides all 3 of these quantities (type, spectral occupancy, spatial directions) in a single framework
DOF Performance Summary• DOF is robust to SNR of detected signals
• Accurate at received signals as low as 0dB
• DOF is robust to multiple overlapping signals• Accurate even when three unknown signals are present
• DOF is relatively computationally inexpensive• Requires 30% more computation over standard FFT
5
DOF: High Level Architecture
FeatureExtraction
{𝑻𝒚𝒑𝒆 , �⃗� (𝒊)}𝒏=𝟏𝑵
DOF Estimation(AoA Detection)
ADCSignal
Time Samples
Classification
F( i )⃗
MAC
DOF
{𝑻𝒚𝒑𝒆 ,𝑭 𝒄 ,𝑩𝑾 }𝒏=𝟏𝑵
{𝚯 }𝒏=𝟏𝑵
{𝚯 ,𝑻𝒚𝒑𝒆 ,𝑭𝒄 ,𝑩𝑾 }𝒏=𝟏𝑵
DOF Estimation(Spectrum Occupancy)
6
DOF operates on windows of raw time samples from the ADC
Raw samples are processed to extract feature vectors
Feature Vectors are used to detect1. Signal Type2. Spectral Occupancy3. Spatial Directions
The MAC layer utilizes this mechanism to inform its coexistence policy
For almost all “man-made” signals – there are hidden repeating patterns that are unique and necessary for operation
Key Insight
CP CP CPData Data Data …………………….
Repeating Patterns in WiFi OFDM signals
Repeating Patterns in Zigbee signalsTime
Leverage unique patterns to infer 1) type, 2) spectral occupancy, and 3) spatial directions
7
Pattern Frequency (α)Delay (τ)
Advantages• Robustness to noise, • Uniqueness for each
protocol
Extracting Features from Patterns
8
If a signal has a repeating pattern, then when we• Correlate the received signal against itself delayed by a fixed amount, the
correlation will peak when the delay is equal to the period at which the pattern repeats.
𝑅𝑥𝛼 (𝜏 )=∑
𝑛
∞
𝑥 [𝑛 ] [𝑥∗ [𝑛−𝜏 ] ]𝑒− 𝑗2 𝜋𝛼𝑛
Pattern Frequency () – The frequency at which the pattern repeats
Disadvantage: Computationally expensive to calculate the patterns in this manner
Cyclic Autocorrelation Function (CAF)
Feature Extraction: Efficient Computation
The CAF can be represented using an equivalent form called the Spectral Correlation Function (SCF)
• SCF can be calculated for Discrete Time Windows using just FFTs
9
𝑆𝑥𝛼 ( 𝑓 )= ∑
𝜏=−∞
∞
𝑅𝑥𝛼 (𝜏 )𝑒− 𝑗 2𝜋 𝑓 𝜏¿ 1
𝐿∑𝑙=0
𝐿−1
𝑋 𝑙𝑁 ( 𝑓 ) 𝑋 𝑙𝑁∗ ( 𝑓 −𝛼)
Pattern Frequency (α)Frequency (f)
WiFi Spectral Correlation Function
Feature Vectors are calculated by computing at different values of
Classifying Signal Type
10
• Single signals are well separated in the feature vector space,
Works well when there is a single signal but fails when there are multiple interfering signals
Feature Dimension 1:
Fea
ture
Dim
ens
ion
2:
• Support Vector Machines (SVM) can be used to classify signal type,
Multiple interfering signals are not straightforward to classify
• Multiple signals are made up of components and features of single signals, making them difficult to distinguish
Need a robust algorithm to determine the number of interfering signals
Feature Dimension 1:
Fea
ture
Dim
ens
ion
2:
Feature Dimension 1:
Fea
ture
Dim
ens
ion
2:
11
1) Real signal packets are asynchronous
12
Inferring the number of signals: Exploiting Asynchrony
ZigBee
t
Overlapping PacketsWiFi
Nonzero Components in F( i )⃗
Received Signal
2) This asynchrony shows up in as an increase or decrease in the number of non-zero components
F( i )⃗
Measuring differences in is more robust than differences in energy
DOF: High Level Architecture
FeatureExtraction
{𝑻𝒚𝒑𝒆 , �⃗� (𝒊)}𝒏=𝟏𝑵
DOF Estimation(AoA Detection)
ADCSignal
Time Samples
Classification
F( i )⃗
MAC
Asynchrony Detector/
Power Normalization
SVM-1
SVM-N
Counter++
. . .
If ΔL0>Threshold
Counter--If ΔL0<-Threshold
Sig1 Class
Sig i Class
1 Signal
N Signals
SigN Class
. . .
. . .
While
DOF = Active
DOF
{𝑻𝒚𝒑𝒆 ,𝑭 𝒄 ,𝑩𝑾 }𝒏=𝟏𝑵
{𝚯 }𝒏=𝟏𝑵
{𝚯 ,𝑻𝒚𝒑𝒆 ,𝑭𝒄 ,𝑩𝑾 }𝒏=𝟏𝑵
The signal types can be leveraged along with the feature vectors to estimate 1) Spectrum Occupancy
2) Spatial DirectionsDOF Estimation
(Spectrum Occupancy)
13
Estimating Spectrum Occupancy
14
• Communication signals are sequences of periodic pulses
𝑠 (𝑡 )=𝑏𝑐𝑜𝑠 (2𝜋 𝑓 𝑏𝑡 )𝑒 𝑗 2𝜋 𝑓 𝑐𝑡
• These pulses are patterns embedded within the signal which repeat at a particular frequency
• These frequencies at which these patterns repeat tell us the bandwidth and carrier frequency of the signal
1
0
Bit Sequence b
Amplitude modulated Pulse
Pulse multiplied by Carrier Wave
Estimating Spectrum Occupancy
15
Modulated Zigbee Signal
Time
• Because these patterns repeat, they are natural components of the feature vector
Relationship between feature vector and Bandwidth/Carrier Frequencies
Signal Type Feature Vector Frequencies
WiFi
Bluetooth
ZigBee
all between
DOF leverages this relationship to compute the spectral occupancy of each signal type
Pattern Frequency (α)Frequency (f)
ZigBee Spectral Correlation Function
Estimating Angles of Arrival
. . .
1 2 M
Incoming Signal
d
Array Elements
• Each array element experiences a delay of τ relative to the first array element, which is a function of the Angle of Arrival (AoA)
. . .
16
Θd si
nΘ
• This unique delay induces a particular characteristic on the feature vector which can be computed
DOF uses the same feature vector to infer 1)type, 2)spectral occupancy, 3)spatial directions
Implementation
RX2
RX 1RX 3
• Channel traces were collected using a modified channel sounder with a frontend bandwidth of 100MHz spanning the entire ISM band.
• Wideband Radio Receiver placed at 3 different locations while transmitter was placed randomly in the office
• Raw Digital Samples are collected and processed offline on a PC with Intel Core i7 980x Processor and 8GB RAM
17
Compared ApproachesIdentifying Protocol Types
• RF Dump (CoNEXT 2009) – Energy Detection + Packet Timing
Estimating Spectrum Occupancy• Jello (NSDI 2010) – Edge Detection on Power Spectral Density
Estimating Angles of Arrival• Secure Angle (HOTNETS 2010)– MUSIC (subspace based approach)
Experimental Setup
18
Comparison Setup• Each testing “run” consists of 10 second channel traces. • Random Subset of 4 different radios are selected in each “run” (WiFi,
Bluetooth, ZigBee, Microwave) with varying PHY parameters • 30 Different “runs” for each signal combination
Evaluation: Classification
-5 -4 -3 -2 -1 0 1 2 3 4 5 6 7 8 9 10 11 12 13 14 150
0.2
0.4
0.6
0.8
1
1.2 Single Signal Classification
DOFRFDump
Acc
ura
cy
SNR
19
DOF achieves greater than 85% accuracy when the SNR of the detected signal is as low as 0dB
Evaluation: Classification
00.02
0.06 0.10.14
0.180.22
0.26 0.30
0.2
0.4
0.6
0.8
1
1.2
Probability of Missed Classification
Cu
mu
lati
ve F
ract
ion
1 Signal2 Signals3 Signals
DOF: Multiple Signal Classification
20
DOF classifies all component signals with greater than 80% accuracy, even with 3 interfering signals
Evaluation: Spectrum Occupancy
-2 -1 0 1 2 3 4 5 6 7 8 9 10 11 12 130
0.1
0.2
0.3
0.4
0.5
0.6 Single Signal Spectrum Occupancy EstimationN
orm
aliz
ed E
rro
r
SNR
DOFJello
21
DOF’s spectrum occupancy estimates are at least 85% accurate at SNRs as low as 0dB
Evaluation: Spectrum Occupancy
0 0.05 0.1 0.15 0.2 0.25 0.3 0.35 0.4 0.45 0.5 0.55 0.6 0.65 0.7 0.750
0.2
0.4
0.6
0.8
1
Cu
mu
lati
ve F
ract
ion
DOFJello
Normalized Error
22
Multiple Signal Spectrum Occupancy Estimation
DOF’s spectrum occupancy estimates are robust in the presence of multiple overlapping signals
Evaluation: Angle of Arrival
DOFSecureAngle
(MUSIC)
4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 210.0
0.2
0.4
0.6
0.8
1.0
1.2Multiple Signals: AoA Detection Accuracy
Absolute Difference per Angle (Deg)
Cum
ulati
ve F
racti
on
23
In addition to being accurate, DOF can also associates each AoA with each signal type
Smart Radio Prototype: DOF- SRDOF-SR (Policy Aware Smart Radio)
• Policy 0 – Only use unoccupied spectrum
WiFi
Microwave
Smart Tx
Ao
A
Freq2.3 GHz
2.5 GHz
Frequency
2.5 GHz
Smart Rx
Ao
AFreq
2.3 GHz
PSD
• Policy 1 – Use all unoccupied spectrum. Further use spectrum occupied by microwave ovens.
• Policy 2 – Use all unoccupied spectrum + microwave occupied spectrum. Further compete for spectrum occupied by WiFi radios and get half the time share on that spectrum.
24
Heart Monitor(ZigBee Based)
DOF-SR Performance
0.0 0.2 0.4 0.6 0.8 1.00.0
0.2
0.4
0.6
0.8
1.0
DOF-SR Policy 0 and Jello
Normalized Harm
Nor
mal
ized
Thr
ough
put
0.00 0.50 1.000.00
0.20
0.40
0.60
0.80
1.00
DOF-SR Policy 1 and Jello
Normalized Harm
Nor
mal
ized
Thr
ough
put
0.00.20.40.60.81.00.0
0.2
0.4
0.6
0.8
1.0
DOF - SR Policy 2 and Jello
Normalized Harm
Nor
mal
ized
Thr
ough
put
DOF-SRJello
Legend
DOF-SR enables users to decide how aggressive their policy should be
25
Conclusion
26
DOF exploits repeating patterns to infer 1) type, 2) spectral occupancy, and 3) spatial directions
DOF Performance Summary• DOF is robust to SNR of detected signals
• Accurate at received signals as low as 0dB
• DOF is robust to multiple overlapping signals• Accurate even when three unknown signals are present
• DOF is relatively computationally inexpensive• Requires 30% more computation over standard FFT