Scaling Network- based Spectrum Analyzer with Constant Communica<on Cost

Transcription

1 Scaling Network- based Spectrum Analyzer with Constant Communica<on Cost Youngjune Gwon H. T. Kung Presented at the 32 nd IEEE Interna6onal Conference on Computer Communica6ons (INFOCOM 13) in Turin, Italy April 16, 2013

2 Introduc<on Dynamic spectrum access (DSA) with cogni6ve radios Alleviates inefficient spectrum alloca6on and licensing Accurate low- latency spectrum sensing most important to maximize DSA benefits Conven6onal spectrum analyzer Can be ideal spectrum sensor Measures amplitude of signals over!me and converts to power magnitudes across frequency FFT at the heart of modern spectrum analyzer equipment Expense of FFT true boyleneck is to keep up with Nyquist sampling» E.g., 1- MHz channel: sample size bps 40 Mbps (if 20- bit sample)

3 Network- based Spectrum Analysis Distributed spectrum sensing Spectrum analyzer & measurers separate en66es but networked Simple, in situ compression of measurement data at acquisi6on Compressive sensing encode In- network processing of data Combine mul6ple compressed measurements Recovery of original data Undo in- network data processing & compressive sensing decode

4 Case for Distributed Spectrum Sensing (1) Geographically distributed networked sensors Sensor node y GHz Network node... Network Recovery & analysis Network node Sensor node xxx MHz Improve on spa<al diversity Not our Focus

5 Case for Distributed Spectrum Sensing (2) Geographically distributed networked sensors Sensor node y GHz Network node... Network Recovery & analysis Network node Sensor node xxx MHz Measure disjoint bands and aggregate Our Focus

6 Challenges Naïve FFT spectrum analysis boylenecked by high data rate of Nyquist sampling Use of mul6ple spectrum sensors each monitoring a sub- band, we have mi6gated this problem How to minimize network communica6on cost of sensor measurements propaga6ng network Fine- grained spectral analysis of wideband spectrum

7 Problem Statement What is the size- reducing opera<on θ that makes network- based analyzer feasible? arg min θ J i=1 dim(y i = θ(x i )) s. t. X(f k ) ˆX(f k ) 2, J: # of par66ons in the spectrum x i : raw measured data from par66on i y i : compressive measurement of x i X: frequency response of original x = {x i } i X: frequency response of restored x ε: some small error requirement

8 Problem Illustrated y GHz? x 1?????? x 2 x Recovery & analysis x J xxx MHz

9 Solu<on Approach (1) y GHz x 1 x 1 {x 1,x 2 } {x 1,x 2,x 3,x J } {x 3,x J } x 3 x 2 x 2 x Recovery & analysis x J No in- network data reduc3on x J xxx MHz

10 Solu<on Approach (2) y GHz Compressive sensing (CS) y i = Φx i y 1 x 1 y i y 1 +y 2 y 3 +y J y 2 x 2 x 3... y 3... y J x J Recovery & analysis xxx MHz Compressive sensing (CS) encoding and in- network combining of compressed data

11 Reminder: Compressive Sensing Decode Encode Ψ x = s (K-sparse) y = Φ x N N N 1 M 1 M N N 1 y = Φ Ψ - 1 s Encode: C N C M Simple & data- blind N:M compression (M << N) for sparse signal Decode Available sparsifying basis (Ψ) determines M c K log(n/k) Sparsity K revealed by Ψ L1- minimiza6on (e.g., linear programming): min s 1 s.t. y = ΦΨ 1 s

12 How to Separate Sum of Compressed Measurements? Generalized P- way sum y = Φx 1 + Φx Φx P Joint Decoding Algorithm y 1 = Φx 1 y 2 = Φx 2... y P = Φx P s 1 y = Φ Ψ 1 1 Ψ Ψ 1 P s 2... Must solve for P N unknowns in one- shot Ψ 1 (Overcomplete basis) Compressive sensing decode on y = (Φ Ψ 1 ) s to solve for s 1, s 2,..., s P jointly Can we do be@er? s P

13 Ini<al Approxima<on by Least Squares Keep only several leading eigenvectors of Q i y = Φ [ Q 1 Q2 QP ] s 1 s 2. s P Equivalent to keeping variables at leading posi6ons only (others are mul6plied by zeros) Require {Q 1, Q 2,..., Q P } dis!nct sparsifying bases for each channel Q can be es6mated from R x = E[xx H ] = Q Λ Q H Leading components have largest eigenvalues Remove non- leading components un6l we have overdetermined system More equa6ons than unknowns: dim(y) > # of unknowns Least squares does this job well

14 Itera<ve Refinement by CS Decode y = Φ [Q 1 Q 2 Q P ] s 1 ˆ s 2. ˆ s P Relax s 1 ~ 1. Back- subs6tute: yʹ = y Φ[Q ~ ^ ~ 2 s Q ~ ^ P s P ] using ini6al approximates of s 2,..., s P 2. Do compressive sensing decode with yʹ = (Φ Q 1 ) s 1 to obtain refined s 1 Compressive sensing decodes underdetermined system More unknowns than equa6ons Relax s i s in descending order of their L1- norm Compressive sensing works beyer on largest- first decoding principle No need to solve for more than N unknowns at once N length of original, uncompressed measurements (x i s) on channel i Can be repeated in another stage

15 Evalua<on in Lab Testbed of SW- defined Radios Spectrum analyzer Collects in- network combined, compressed measurements point FFT Network node (simulated) Combines mul6ple compressed measurements in- network 100BaseT (f 1,B) (f 5,B) (f 2,B) (f 6,B) (f 3,B) (f 7,B) (f 4,B) (f 8,B) Four sensor nodes (USRP2/USRP- N200) with WBX RF daughterboards Measure 8 channels from UHF white space f i = {512.5, 537.5, 562.5, 587.5, 612.5, 637.5, 662.5, 687.5} MHz Each channel with B = 25 MHz bandwidth

16 Some Details Sensing & recovery methods 1. Compressive sensing only (no combining) 2. P- way combined compressed measurements for P = 2, 4, 8 M = # of compressed measurements (per channel) Varied from 26 (20x compression) to 308 (1.67x) Error metric ξ = 1 L L k=1 X(f k ) ˆX(f k ) 2 X(f k ) 2 Average normalized frequency response error per sample L = = 4096 f k [500,700) MHz

17 Error Performance Number of measurements transmi]ed Total # of measurements transmiyed communica6on cost P- way in- network combining could reduce measurements up to P- fold Given error budget, # of measurements can remain constant un6l some limit This limit depends on sparsity of channels in spectrum Proposed algorithm achieves similar accuracy performance as joint decoding while requiring P 6mes less unknowns to solve concurrently

18 Improvement at Refinement Stages 8- way combined 20x compression Error (ξ) Number of CS refinements Error improvement more significant with smaller M tot Small gain on accuracy axer 2 stages

19 Summary Network- based spectrum analysis Distributed spectrum sensors employed by distant analyzer operate over network Key is to overcome network communica6on cost to move spectrum measurements Our approach Compressive sensing encoding at sensor nodes Simple in- network summing of mul6ple compressed measurements to further reduce overhead at network nodes New recovery algorithm Least squares on leading principal components to separate individual measurements from the sum Itera6ve relaxa6on by compressive sensing decode on each individual data Conclusion: sensors can be added without addi<onal communica<on cost Hold true un6l some limit determined by sparsity Sparsity = true measure for channel informa6on content

20 Suppor<ng Materials

21 Remark on Sparsity and Discre<za<on Discrete measurements performed by sensors preserve or bring out sparsity of original signal in frequency domain or in a custom basis This is fundamental premise of our approach Design of beyer sparsity- inducing discre6za6on schemes is challenging but can hugely enhance our approach

22 CS Recovery of Complex Signals y = Φx = (ΦΨ - 1 )x x = N 1 complex- valued y = M 1 complex- valued Φ = M N real- valued Ψ = N N complex- valued y = Φ[x R +j x I ] = ΦΨ - 1 [X R +j X I ] = Φ[Ψ R - 1 +j Ψ I - 1 ][X R +j X I ] j = sqrt(- 1) We want: yʹ = AXʹ! y'= y R # " y I $ & A = % $ & & %&!" #1 #!" #1 R I!" #1 I!" R #1 ' ) ) ()! X '= X R # " X I $ & %

23 Decode yʹ = AXʹ X R y R = ΦΨ R -1 ΦΨ I -1 y I ΦΨ I -1 ΦΨ R -1 2M 1 2M 2N X I 2N 1