Data-driven approach to wireless spectrum crunch

Transcription

1 Data-driven approach to wireless spectrum crunch Aakanksha Chowdhery Princeton University Collaborators: Mariya Zheleva, Ranveer Chandra, Ashish Kapoor, Paul Garnett

2 Outline Wireless spectrum crunch Data-driven approach Characterizing transmitters Unsupervised machine learning RGMM Time-frequency analysis Evaluation Conclusion & Future work - M. Zheleva, R. Chandra, A. Chowdhery, A. Kapoor, P. Garnett, TxMiner: Identifying transmitters in real-world spectrum measurements, IEEE Dyspan, Oct A. Chowdhery, R. Chandra, P. Garnett, and P. Mitchell, Characterizing Spectrum Goodness for Dynamic Spectrum Access, IEEE Allerton (Invited paper), Oct

3 Wireless spectrum crunch 3

4 Wireless spectrum crunch Spectrum assignment (e.g. FCC dashboard) Assignments are granted to 88 unique entities in Washington. 50% of all licenses are owned by 10 companies. PCS Cellular MHz Cellular UHF TV MHz MHz Broadband and Educational Radio Services (BRS and EBS) Spectrum assignment in Washington state Source:

5 Solving wireless spectrum crunch Measure spectrum occupancy What % of time is the measured power above a threshold? Most spectrum is unused most of the time Yet policy regulations continue to limit access to otherwise unused spectrum 5

6 Solving wireless spectrum crunch Measure spectrum occupancy What % of time is the measured power above a threshold? Quantify spectrum utilization Number of transmitters Usage patterns Transmissions: - Center freq. - # transmitters - Bandwidth - TDMA/FDMA - Mobility - Direction 6

7 Outline Wireless spectrum crunch Data-driven approach Characterizing transmitters Unsupervised machine learning RGMM Time-frequency analysis Evaluation Conclusion & Future work 7

8 Data-driven approach: Spectrum Observatory Goal: large-scale continuous spectrum measurements over time, frequency at various locations RF Sensor Cloud Storage 8

9 Data-driven approach: Spectrum Observatory Challenges Heterogeneous sensors limited upload bandwidth Cloud storage 60MB data per hour per station Needs TB storage over 1000 stations or a year Processing costs data points per 25 MHz window Must process 534,528 data points per minute for 4350 MHz spectrum 9

10 Data-driven approach: Spectrum Observatory Data stored as blob storage Limited upload bandwidth Extract relevant transmitter features 10

11 Outline Wireless spectrum crunch Data-driven approach Characterizing transmitters Unsupervised machine learning RGMM Time-frequency analysis Evaluation Conclusion & Future work 11

12 Characterizing transmitters: Key insight Measured signal distributions tell us about channel occupancy. Idle TV channel Mean -108dBm Occupied TV channel Mean -70dBm Two occupied TV channels Bimodal distribution Bluetooth Long tail at high PSD Mobile transmitter Large variation Stationary: Δ=10dBm Mobile: Δ=25dBm

13 Characterizing transmitters: Unsupervised machine learning Mixture model: Probabilistic model to represent multiple transmitters in the received signal (RF sensor) RF Sensor

14 Characterizing transmitters: RF signal model RF signals affected by multipath Rayleigh distribution Background noise Gaussian distribution At RF sensor, measured power modeled as a mixture of multiple Rayleigh distributions and Gaussian distribution

15 Unsupervised machine learning: Rayleigh-Gaussian Mixture Models (RGMM) Transmitters modeled as Rayleighs Noise modeled as a Gaussian RGMM: k p MM (s) = w i R(s;µ i )+ w n N(s;µ n,σ n 2 ) i=1 RGMM: received signal s probability distribution is weighted sum of individual probability distributions. Weight: contribution to the mixture k Rayleigh distributions; one for each transmitter. One Gaussian distribution for noise.

16 Unsupervised machine learning: TxMiner Algorithm EM to maximize the likelihood of mixture model Probability sample s belongs to signal or noise Update parameters 16

17 Unsupervised machine learning: TxMiner Algorithm Measured signal PDF pdf TX > 0 Rayleigh-Gaussian Mixture Model Time Frequency pdf N = 0 PSD Association probability (AP) with the noise component (left) and the transmitter component (right) AP Noise AP TX Compute association probability for each transmitter/noise Time Frequency Time Frequency

18 Unsupervised machine learning: TxMiner Algorithm Measured signal PDF pdf TX > 0 Rayleigh-Gaussian Mixture Model Time Frequency pdf N = 0 PSD Signature Frequency Time P i f P i t = = F f T t R i (s ft,µ i ) R i (s ft,µ i ) F T Time-frequency analysis AP Noise Association probability (AP) with the noise component (left) and the transmitter component (right) Time Frequency AP TX Time Frequency

19 Characterizing transmitters: Time-frequency analysis For each Rayleigh distribution i Temporal signature: Probability distribution over time Association probability # of frequency samples Frequency signature: Probability distribution over frequency Identify key transmitter properties: bandwidth, active time & type (TDMA, FDMA, broadcast, frequency hopping) # of time samples

20 Characterizing transmitters: Additional challenges Mixture model initialization Robust fit depends on initialization Solution: MultiScale initialization Noisy association probabilities Transmitters occupy adjacent time-frequency samples Solution: time-frequency regularization using belief propagation. - M. Zheleva, R. Chandra, A. Chowdhery, A. Kapoor, P. Garnett, TxMiner: Identifying transmitters in real-world spectrum measurements, IEEE Dyspan, Oct. 2015

21 Outline Wireless spectrum crunch Data-driven approach Characterizing transmitters Unsupervised machine learning RGMM Time-frequency analysis Evaluation Conclusion & Future work 21

22 Evaluation Data collected by Microsoft Spectrum Observatory Ground truth: TV & FM band Controlled WiMax & proprietary DSA Artificially mixed Performance Accuracy in spectrum occupancy Bandwidth detection Transmitter type Transmitter count 22

23 Characterizing single transmitter: high accuracy TV-UHF occupancy Bandwidth detection 23

24 Characterizing multiple transmitters Detects multiple transmitters on same frequency when they differ in transmit powers 24

25 Characterizing multiple transmitters Detects simultaneous transmissions with high accuracy Tx 2, Tx 1, Tx 3, Bandwidth Detection Accuracy (in %) 25

26 Characterizing transmitter type Broadcast, TDMA, FDMA, frequency hop Broadcast 26

27 Characterizing transmitter type Broadcast, TDMA, FDMA, frequency hop TDMA 27

28 Practical applications Mapping spectrum occupancy Number of transmitters Transmitter type Dynamic spectrum access (DSA) opportunity Identifying rogue transmitters

29 Mapping spectrum occupancy Analysis of two bands 30MHz 173MHz 700MHz 900MHz

30 Dynamic spectrum access opportunity 24-hour observation of a 6MHz band ( MHz) Identified one transmitter at -94dBm Temporal characteristics 1 Active time duration 2 Active time gap 3 Active time cycle Frequency characteristics 4 Fraction of occupied band Temporal characteristics 4 Frequency characteristics 30

31 Conclusions & Future Work Large-scale spectrum measurement data critical to solve spectrum crunch Spectrum Observatory - robust design to aggregate large-scale measurements Unsupervised machine learning robust in mining transmitter characteristics without prior knowledge Fading follows a Rayleigh distribution Model signal as a mixture of distributions Applications: spectrum maps Improved policy Improved technology Rogue transmitter detection Future outlook Integration with known transmitter characteristics Mobility detection Crowdsourced data