Beyond Sensing: Multi-GHz Realtime Spectrum Analytics
Beyond Sensing: Multi-GHz Realtime Spectrum Analytics Lixin Shi (MIT) Victor Bahl (Microsoft), Dina Katabi (MIT)
Spectrum Sensing • Measures the usage of each spectrum band • Important because: – Guides the FCC’s policies – Enables dynamic spectrum access
• Two decades of work on spectrum sensing, but little understanding of how the spectrum is used
Today's Spectrum Sensing Reports Microsoft Spectrum Observatory (08/03/2013 – 08/08/2013)
Today's Spectrum Sensing Reports Microsoft Spectrum Observatory (08/03/2013 – 08/08/2013)
(Radar Band)
Today’s Spectrum Sensing Reports Microsoft Spectrum Observatory (08/03/2013 – 08/08/2013)
(Air Force)
Today’s Spectrum Sensing Reports Microsoft Spectrum Observatory (08/03/2013 – 08/08/2013)
Today’s spectrum sensing systems miss important signals! (Air Force)
Why?
Today: Sequential sensing with MHz BW Freq
Ideally: Realtime sensing with GHz BW Freq
Time
Time
Cheap, practical
Capture all signals
Miss important signals
Costly, and impractical
Can we use MHz radios but capture all signals?
SpecInsight • Uses MHz radios to accurately sense GHz spectrum • Evaluated in 7 US cities • Captures very low occupancy signals which are missed by past work
How does it work?
Intuition: Scan Bands to Maximize the Probability of Detecting Signals 638 MHz
Random Check
632 MHz 3500 MHz
3600 MHz 436.0 MHz
Always-on (TV signal)
Brief Check
Periodic (Radar Signal)
Brief Check Use all of the saved time
Dynamic (Amateur Radio)
435.5 MHz
Time
SpecInsight Architecture Learning Spectrum Patterns
Scheduling Based on the Patterns
Pattern is a representative time-frequency chunk 𝑓
𝑓 𝑡
𝑡
SpecInsight Architecture Learning Spectrum Patterns
Scheduling Based on the Patterns
Learning Patterns Extract the patterns FCC Band
𝑓
Detect the distribution of inter-arrival
PDF 𝑡
Pattern 1 𝑡
𝑓 𝑡
Pattern 2
Learning Patterns Extract the patterns FCC Band
𝑓
Detect the distribution of inter-arrival
PDF 𝑡
Pattern 1 𝑡
𝑓 𝑡
Pattern 2
Extracting Patterns f
f
t
Input signal
t
Dividing
Cluster!
Chunks
Identifying Patterns f
f
t
Input signal
Cluster 1
t
Chunks
Dividing
Noise
Cluster 2
Clustering
Learning Patterns Extract the patterns FCC Band
𝑓
Detect the distribution of signal inter-arrival
CDF 𝑡
Pattern 1 𝑡
𝑓 𝑡
Pattern 2
Learning Patterns Extract the patterns FCC Band
𝑓
Detect the distribution of signal inter-arrival
CDF 𝑡
Pattern 1 𝑡
𝑓 𝑡
Pattern 2
Pattern Inter-Arrival Distribution
𝝉(𝟎)
…
𝝉(𝟏)
𝑡
PDF(𝝉) 𝜇
𝜎
𝑡
Distribution Parameters: • 𝜇 Period • 𝜎 Dynamism
SpecInsight Architecture Learning Spectrum Patterns
Scheduling Based on the Patterns
Scheduling Sensing Based on Patterns Expected occurrence
Sensing Schedule
𝑡 Proportional to 𝝈
𝝁
When should we schedule the next sensing? CDF(𝝉)
𝜇
𝜎
𝑡
Exploration
Exploitation
Learning Spectrum Patterns
Scheduling Based on the Patterns
How do we divide time between exploring new patterns We map the problem to the known multi-armed and sensing the spectrum the learned patterns?
bandit game to find the optimal tradeoff
Performance
SpecInsight’s Implementation Outdoor Antenna
Indoor USRPs
Frequency Range: 50MHz-4.4GHz Instant BW: 40MHz
Evaluated in Seven Locations Amherst, MA Boston, MA
Redmond, WA
New York City, NY
San Francisco, CA
One week of data
Upper Arlington, OH Maui, HI
Compared algorithms • SpecInsight • Sequential Scanning
Ground Truth • 10 USRPs to continuously monitor a subset of the bands; and repeat for different subsets of bands
Percentage of Occupancy Error (%)
Accuracy 4 3 2 1 0 Always-On
Periodic
Dynamic
Percentage of Occupancy Error (%)
Accuracy 4
4.081%
Sequential Scanning
3 2 1 0
0.001%
Always-On
0.342%
Periodic
Dynamic
Percentage of Occupancy Error (%)
Accuracy 4 3
4.081%
Sequential Scanning Specinsight
On2 average, our error is 10x smaller than today’s sequential scanning. 0.999%
1 0
0.001% 0.001%
Always-On
0.007%
0.342%
Periodic
Dynamic
Understanding why SpecInsight is more accurate Percentage of Time Spent in Each Class (%)
100 80
96.097%
SpecInsight Sequential Scanning
SpecInsight saved more than 95% of the time to 60 be spent on the dynamic classes
40
33.333%
33.333%
33.333%
20 0
0.642%
Always-On
3.261%
Periodic
Dynamic
Spectrum Report
Spectrum Analytics Chart
Spectrum Analytics Chart
Fixed Freq, Fixed Cycle
Wide-Band, Fixed Cycle
Frequency Hopping, Always On
Frequency Hopping, Dynamic
Fixed Frequency, Always On
Fixed Frequency, Dynamic
Fixed Frequency, Fixed Cycle
Wide-Band, Dynamic
Wide-Band, Fixed Cycle
Spectrum Analytics Chart
Spectrum Analytics Chart
Frequency Hopping, Always On Fixed Always Onis 38%Frequency, of spectrum
Frequency Hopping, Dynamic
reportedFixed empty by past Frequency, Dynamic systems used Fixed Frequency, Fixed while Cycle it is actually Wide-Band, Dynamic Wide-Band, Fixed Cycle
Conclusion • SpecInsight can sense multi-GHz spectrum using cheap, MHz radios • Provides deep understanding of spectrum utilization • Key primitive for future dynamic usage of the spectrum
Thanks!
Spectrum Sensing • Measures the usage of each spectrum band • Important because: – Guides the FCC’s policies – Enables dynamic spectrum access
• Two decades of work on spectrum sensing, but little understanding of how the spectrum is used
Today's Spectrum Sensing Reports Microsoft Spectrum Observatory (08/03/2013 – 08/08/2013)
Today's Spectrum Sensing Reports Microsoft Spectrum Observatory (08/03/2013 – 08/08/2013)
(Radar Band)
Today’s Spectrum Sensing Reports Microsoft Spectrum Observatory (08/03/2013 – 08/08/2013)
(Air Force)
Today’s Spectrum Sensing Reports Microsoft Spectrum Observatory (08/03/2013 – 08/08/2013)
Today’s spectrum sensing systems miss important signals! (Air Force)
Why?
Today: Sequential sensing with MHz BW Freq
Ideally: Realtime sensing with GHz BW Freq
Time
Time
Cheap, practical
Capture all signals
Miss important signals
Costly, and impractical
Can we use MHz radios but capture all signals?
SpecInsight • Uses MHz radios to accurately sense GHz spectrum • Evaluated in 7 US cities • Captures very low occupancy signals which are missed by past work
How does it work?
Intuition: Scan Bands to Maximize the Probability of Detecting Signals 638 MHz
Random Check
632 MHz 3500 MHz
3600 MHz 436.0 MHz
Always-on (TV signal)
Brief Check
Periodic (Radar Signal)
Brief Check Use all of the saved time
Dynamic (Amateur Radio)
435.5 MHz
Time
SpecInsight Architecture Learning Spectrum Patterns
Scheduling Based on the Patterns
Pattern is a representative time-frequency chunk 𝑓
𝑓 𝑡
𝑡
SpecInsight Architecture Learning Spectrum Patterns
Scheduling Based on the Patterns
Learning Patterns Extract the patterns FCC Band
𝑓
Detect the distribution of inter-arrival
PDF 𝑡
Pattern 1 𝑡
𝑓 𝑡
Pattern 2
Learning Patterns Extract the patterns FCC Band
𝑓
Detect the distribution of inter-arrival
PDF 𝑡
Pattern 1 𝑡
𝑓 𝑡
Pattern 2
Extracting Patterns f
f
t
Input signal
t
Dividing
Cluster!
Chunks
Identifying Patterns f
f
t
Input signal
Cluster 1
t
Chunks
Dividing
Noise
Cluster 2
Clustering
Learning Patterns Extract the patterns FCC Band
𝑓
Detect the distribution of signal inter-arrival
CDF 𝑡
Pattern 1 𝑡
𝑓 𝑡
Pattern 2
Learning Patterns Extract the patterns FCC Band
𝑓
Detect the distribution of signal inter-arrival
CDF 𝑡
Pattern 1 𝑡
𝑓 𝑡
Pattern 2
Pattern Inter-Arrival Distribution
𝝉(𝟎)
…
𝝉(𝟏)
𝑡
PDF(𝝉) 𝜇
𝜎
𝑡
Distribution Parameters: • 𝜇 Period • 𝜎 Dynamism
SpecInsight Architecture Learning Spectrum Patterns
Scheduling Based on the Patterns
Scheduling Sensing Based on Patterns Expected occurrence
Sensing Schedule
𝑡 Proportional to 𝝈
𝝁
When should we schedule the next sensing? CDF(𝝉)
𝜇
𝜎
𝑡
Exploration
Exploitation
Learning Spectrum Patterns
Scheduling Based on the Patterns
How do we divide time between exploring new patterns We map the problem to the known multi-armed and sensing the spectrum the learned patterns?
bandit game to find the optimal tradeoff
Performance
SpecInsight’s Implementation Outdoor Antenna
Indoor USRPs
Frequency Range: 50MHz-4.4GHz Instant BW: 40MHz
Evaluated in Seven Locations Amherst, MA Boston, MA
Redmond, WA
New York City, NY
San Francisco, CA
One week of data
Upper Arlington, OH Maui, HI
Compared algorithms • SpecInsight • Sequential Scanning
Ground Truth • 10 USRPs to continuously monitor a subset of the bands; and repeat for different subsets of bands
Percentage of Occupancy Error (%)
Accuracy 4 3 2 1 0 Always-On
Periodic
Dynamic
Percentage of Occupancy Error (%)
Accuracy 4
4.081%
Sequential Scanning
3 2 1 0
0.001%
Always-On
0.342%
Periodic
Dynamic
Percentage of Occupancy Error (%)
Accuracy 4 3
4.081%
Sequential Scanning Specinsight
On2 average, our error is 10x smaller than today’s sequential scanning. 0.999%
1 0
0.001% 0.001%
Always-On
0.007%
0.342%
Periodic
Dynamic
Understanding why SpecInsight is more accurate Percentage of Time Spent in Each Class (%)
100 80
96.097%
SpecInsight Sequential Scanning
SpecInsight saved more than 95% of the time to 60 be spent on the dynamic classes
40
33.333%
33.333%
33.333%
20 0
0.642%
Always-On
3.261%
Periodic
Dynamic
Spectrum Report
Spectrum Analytics Chart
Spectrum Analytics Chart
Fixed Freq, Fixed Cycle
Wide-Band, Fixed Cycle
Frequency Hopping, Always On
Frequency Hopping, Dynamic
Fixed Frequency, Always On
Fixed Frequency, Dynamic
Fixed Frequency, Fixed Cycle
Wide-Band, Dynamic
Wide-Band, Fixed Cycle
Spectrum Analytics Chart
Spectrum Analytics Chart
Frequency Hopping, Always On Fixed Always Onis 38%Frequency, of spectrum
Frequency Hopping, Dynamic
reportedFixed empty by past Frequency, Dynamic systems used Fixed Frequency, Fixed while Cycle it is actually Wide-Band, Dynamic Wide-Band, Fixed Cycle
Conclusion • SpecInsight can sense multi-GHz spectrum using cheap, MHz radios • Provides deep understanding of spectrum utilization • Key primitive for future dynamic usage of the spectrum
Thanks!
