License-free Use of RF and Infrared Spectrum

A Thesis Presentation by

Md Shaifur Rahman

in Partial Fulfillment of the

Requirements for the Degree of

PhD in Computer Science

October 17, 2024

​

--------------------Dissertation Committee------------------------

​

Aruna Balasubramanian

Committee Chair

Associate Professor

Samir R. Das

Department Chair &

Professor

Himanshu Gupta

Thesis Advisor &

Professor

Petar Djuric

External Member &

Professor

Department of Electrical and Computer Engineering

Department of Computer Science

Stony Brook University

Outline

• Background & Thesis Statement ------------8 minutes
• Contributions
• SpecSense---------------------------------- 20 minutes
• FSONet & FSOVR-------------------------- 15 minutes
• DynoLoc------------------------------------- 15 minutes
• Summary----------------------------------------- 2 minutes
• Q/A----------------------------------------------- 15 minutes

​

Wireless Spectrum is Precious!

Scarcity

Regulation

Interference

Band Size

Location

Propagation

Electro Magnetic Spectrum

Courtesy of NASA

Licensed Band

• Very expensive
• Exclusive access
• Prohibits any interference

1. Unlicensed Band

915MHz ISM Band

2. Optical Spectrum

Free Space Optics

3. Non-interfering Co-existence

4. Dynamic Spectrum Access

Opportunities & Challenges

• Unlicensed Band
• Always available, no prior-allocation
• Interference-prone, low data-rate (up to Mbps)
• Optical Spectrum i.e. Free Space Optics
• Always available, high data-rate (Gbps)
• Line of Sight requirement
• Non-interfering co-existence i.e. UWB
• Always available
• Low data-rate (<1Mbps), short range (<100m)
• Dynamic Spectrum Access (DSA)
• As good as exclusive access, when granted
• Not always available

​

​

​

• Infrastructure-free indoor tracking
• Complementary tracking of HMD

• Cellular backhaul placement
• Reconfigurable Data-center network
• Dynamic Link to HMD

• Where & When available:

Spatio-temporal spectrum maps

Thesis Statement

• License-free, non-interfering, concurrent and dynamic access to wireless spectrum can be achieved via the use of
• accurate spatio-temporal spectrum maps and
• simultaneous wideband and narrowband transmission
• For infrared spectrum, it can be achieved via reconfigurable network by ensuring line-of-sight

Contributions

Contribution 1: �Accurate Spatio-temporal Spectrum Map Generation

Spectrum Occupancy Query

• Goal: Build infrastructure to respond to spectrum occupancy query
• Spectrum occupancy query: Is channel f at location (x, y) in use?
• Why?
• To identify spectrum sharing opportunities
• Can help in spectrum patrolling
• Deeper understanding of spectrum usage for policy formulation
• How?: Crowdsensing
• Enables large-scale monitoring via many low-cost low-power sensors
• Incentive mechanism

Spectrum Sensing Trade-off

VS.

High Accuracy Expensive Sensor

ThinkRF Realtime Spectrum Analyzer

Low Accuracy Inexpensive Sensor

RTL-Dongle connected to Cellphone

Spectrum Sensing Trade-off

• Challenge: Cost of large-scale deployment of spectrum sensors

For a given budget:

Small no. of high-accuracy expensive sensors

Or

Large no. of low-accuracy inexpensive sensors?

High Level Overview

Query for

Channel: f

Location: (x, y)

Estimated signal value at (x, y)

SpecSense System

Sensor Selection

Interpolation

Sensor

Query Location

Interpolation Techniques

• Inverse Distance Weighting (IDW): Predicted value = distance-weighted average of neighboring values

​

• Ordinary Kriging (OK):
• Takes into accounts structure of the spatial correlation via a “Variogram” function
• Predicted value = a weighted average of neighboring values, where the weights come from the variogram
• Minimizes the prediction variance

​

• In this work, we improve OK for our context in two ways:
1. Detrending
2. Partitioning

Interpolation: Ordinary Kriging

• Predicted value = weighted average of neighboring values
• Considers spatial correlation

​

Distance: ~ 1 meter

Difference: ~ 1 unit

Distance: ~ 1 meter

Difference: ~ 1 unit

Variogram

1

2

3

30

5

15

y₁₃ = (30 – 15)² = 225

x₁₂ = 20

x₁₃ = 50

x₂₃ = 65

y₁₂ = (15 – 5)² = 100

y₂₃ = (30 – 5)² = 625

h

X

Y

y_ij

(0, 0)

(20, 100)

(50, 225)

(30, 625)

OK Interpolation

1

2

3

30

5

15

x₁₄ = 50

x₃₄ = 20

x₂₄ = 45

4

?

Improving OK by Detrending

• OK requires mean of signal values to be constant across all locations
• But that is not the case in our context

​

Improving OK by Detrending

• Solution: Decompose sensed signal values

​

Signal value = Path-loss + Shadowing

Path-loss Estimation

1. Assume TX at the sensor with max. signal
2. Estimate path loss exponent α at sensors
3. Estimate path-loss at query using α.

1. Has zero-mean
2. Use OK interpolation

Improving Ordinary Kriging by Partitioning

Improving Ordinary Kriging by Partitioning

• Issue with OK: It assumes that “variance” depends only on distance. Not true in our context (i.e. signal variance depends on terrain)
• Solution: Partition the region into sub-region such that:
• each sub-region has similar path-loss characteristics
• Algorithm (OK with Partitioning):
• Divide the region into small cells
• Initially, each cell is a separate sub-region
• Iteratively merge neighboring sub-regions with similar estimated path-loss exponent.

Sensor Selection

• Problem: Given locations of sensors and potential queries, select at most k sensors to minimize “interpolation prediction error”
• NP-hard.

Two algorithms:

• Nearest Query Cover: In each iteration, for each query, select the closest remaining sensor.
• Iterative Query Cover: In each iteration, select the sensor that “covers” the most number of queries. Here, “coverage” is based on the variogram.
• Above generalize to weighted versions (e.g., weight may be battery-level, precision etc. )

​

System Architecture

Results

• Dataset: Synthetic and real (Cellular, WiFi & Local TX)
• Lower prediction error
• Lower # of selected sensors, hence lower cost
• More opportunistic Spectrum use

Spatio-temporal Map

Observation Vector

Clustering Problem

• Given a set of observation vector with missing values i.e.
• o(1):[v(1,1), X, X, v(1,4), X, X, X, v(1,8),…., v(1,L)]
• o(2):[X, v(2,2), X, X, v(2,5), X, v(2,7), X,…., v(2,L-1), X]
• o(3):[X, X, v(3,1), X, X, X, X, v(3,7), X,…., X]
• …………
• o(T):[X, X, X, X, X, X, v(T,10),….,v(T,L)]
• find the accurate spectrum occupancy map for each configuration i.e.
• c(1): [v(1,1), v(1,2), ….., v(1,L)]
• c(2): [v(2,1), v(2,2), ….., v(2,L)]
• …………….
• c(N): [v(N,1), v(N,2), ….., v(N,L)

Correlation-Based Merging(CBM)

• Step 1 Diffusion: Using interpolation, populate some of the missing values in the vectors
• Step 2 Distance-calculation: Calculate distance metric between each pair of vectors
• Step 3 Merging: Merge the min-distance pair
• Step 4: Update Vectors

CBM Algorithm Contd.

• Termination condition of merging:
• When distance metrics is < S( 1 - exp(i) )
• where S = std. deviation of known values
• where i=(# of known values in the merged vectors /# of known values of all given observation vectors)
• Complexity: O(DxH^3)
• where D is the dimension of the vectors i.e # of locations
• H is the given number of observation vector i.e snapshot
• Performance Metric
• Average Error
• Predicted Error
• Adjusted Rand Index (ARI): -1<=ARI<=+1
• -1 worst merging, +1 best merging

Results

• Compared with K-Means and EM-based algorithms
• Data: Simulated and Real Wi-Fi sensed data

Crowdsensed Shared Spectrum System

PU

PUR

PUR

PUR

SS

SS

SS

SS

SS

SU

Spectrum Manger

Observed Path-loss

Sensing Report

Allocation Request

Spectrum Allocation

SUR

PU: Primary User

PUR: Receiver of PU

SU: Secondary User

SUR: Receiver of SU

SS: Spectrum Sensor

Problem: Based on the SS reports, how to allocate maximum power to SU without interfering with any of the PURs?

Spectrum Allocation Principle

PU

PUR

PUR

PUR

SU

Spectrum Manger

SUR

Tolerable Interference i1

1. Allocates power s.t. created Interference < i1, i2 and i3
2. Update tolerable interference i1, i2 and i3
3. Consider the next SU request

Tolerable Interference i2

Tolerable Interference i3

How to calculate tolerable interference without SU transmission?

Our solution: Estimating pathloss between SU and PUR

SU-PUR Pathloss Estimation

• Requires 3 steps:
• Power Splitting: At each SS, determine PU-SS path-loss after splitting aggregate received power.
• Interpolation: Determine PU-SU pathloss using interpolation from PU-SS values derived above.
• Log-normal Distance Heuristics: Calculate SU-PUR pathloss using PU-SU pathloss and distances.

​

Power Splitting at SS

PU1

PUR

PUR

PUR

SS

SS

SS

SS

SS1

SU

SUR

PU2

Received Power due to PU1, R(PU1, SS1)

= Total Receive Power x (R1)/(R1+R2)

Where R1 = T(PU1)/d(SS1, PU1)²

And

R2 = T(PU2)/d(SS1, PU2)²

Pathloss P(PU1, SS1)=T(PU1)/R(PU1, SS1)

Log-normal Distance Heuristics

PU

PUR

PUR

PUR

SU

P (PU, SU)

P(SU, PUR) = P(PU, SU) X d(PU, SU)²/d(SU, PUR)²

Power Allocation

PU

PUR3

PUR2

PUR1

SU

P(SU, PUR1)

P(SU, PUR2)

P(SU, PUR3)

Allocated Power T(SU) s.t

1. T(SU)/P(SU, PUR1) < t(PUR1)
2. T(SU)/P(SU, PUR2) < t(PUR2)
3. T(SU)/P(SU, PUR3) < t(PUR3)

Update thresholds

1. t(PUR1) 🡨 t(PUR1) –T(SU)xP(SU, PUR1)
2. t(PUR2) 🡨 t(PUR2) –T(SU)xP(SU, PUR2)
3. t(PUR1) 🡨 t(PUR3) –T(SU)xP(SU, PUR3)

​

Results

• Less prediction error
• More SU’s are served
• More efficient spectrum use

Contribution 2: Steerable FSO in Indoor and Outdoor

Optical Communication

Free Space Optics

Free Space Optics (FSO)

39 of 12

Collimator

Laser Source

Photodiode

​

​

​

​

Modulation

​

​

​

​

Demodulation

1550nm Wavelength 10 Gbps link using off-the-shelf devices

1. Class I eye-safe
2. Small divergence (compared to RF)
3. zero EMI

Known Issues

1. Precise LOS
2. Beam Divergence and loss
3. Environmental effect: fog, rain, background noise39

Max. TX power

Min. Sensitivity

Maintaining LOS

TX

RX

Perfectly Aligned

Linear Movement

TX Angular Movement

RX Angular Movement

Galvo Mirror (GM)

• GM corrects TX/RX movement by beam-steering

TX

RX

GM

Angular Tolerance

TX

RX

Feedback

Feedback latency = t second

RX lateral motion = v meter/second

RX-TX distance = d meter

TX angular tolerance Ɵ >= 2*v*t/d radian

d meter

Ɵ

Contributions on Steerable FSO

• Indoor Application: Steerable FSO link for HMD
• Outdoor Application: Cellular backhaul network

Movement of Headsets

Testbed for emulating Tracking and Pointing (TP)

TP-feedback Loop

Simulation using Zemex

TP Latency = 20 mili-sec

VR Lateral Movement = 14 cm/sec

TX Angular Tolerance:

VR Angular motion = 20 degree/sec

RX Angular Tolerance: 14 mrad

FSO-VR System

SFP+ Host

SFP+

VR-Headset

Collimating lenses

2 - 10 meter

Downstream Link

Simpler GM

GM

Ceiling-mounted

TX

Low bandwidth RF link for

TP-Feedback + Upstream

SFP+

Moving RX

Steerable FSO for Backhaul Network

Outdoor Challenges

Robust 100m Link with TP

Effect of TP

50u and 200u Multimode Fiber

FSONet: DYNAMIC BACKHAUL-NETWORK DESIGN

• Candidate Links: Pairs of FSO having LoS within each others coverage cone
• Dynamic Network: Set of all candidate links
• Realizable Topology: Only one candidate link per FSO is active
• Short vs Long links : 100m robust link vs the rest
• Backbone Network: a realizable topology i) consists only short links ii) nodes render full coverage iii) each node has a path to a gateway

Find a backbone network that maximizes the average flow from each possible subset of nodes to the gateway

4-step Heuristics

• The problem is NP-hard (reduction from set-cover or Steiner tree problem)
• Solution: 4-step heuristics
• Select min. # of nodes to cover the area (greedy set cover)
• Connect them using only short links (Steiner-forest problem)
• Reduce max. degree of the nodes iteratively (Furer-Raghavchari’s trick)
• Add additional nodes and links to increase the avg. flow (flow maximization algorithm)

​

FSONet Results

• Simulation settings: All major US cities, building corners of 2D open-street maps, LOS calculation
• Smaller number of backbone nodes
• High degree of coverage
• Higher avg. flow
• Higher flow completion rate

Contribution 3: Ultra Wide Band for Indoor Tracking

Infrastructure-free RF Tracking in Dynamic Indoor Environments

Why

-Free

Damaged

Non-existent

Impractical to deploy

Infrastructure

Infrastructure-free RF Tracking in Dynamic Indoor Environments

How

DynoLoc

Infrastructure-free RF Tracking in Dynamic Indoor Environments

What

Beacon 2

Beacon 1

Beacon 3

Anchor

Controller

Visualizer Tab

LoRa/WiFi

UWB

Motivation for RF

Problem Statement

Localize a set of nodes with high accuracy:

• In GPS-denied environment
• With node mobility
• In dynamic and unknown environments
• Without any pre-deployed infrastructure

1

2

3

4

6

5

7

Controller

Peer-to-peer TOF-based ranging

Relative and Absolute Localization

1

2

3

5

4

1

2

3

5

4

1

2

3

5

4

3

Rigid Graph

Non-rigid Graph

Relative Localization

Absolute Localization

Which edges to choose to form the rigid graph?

Core decomposition

1

2

3

5

4

6

7

8

1-Core

1

2

3

5

4

2-Core

2

3

5

4

3-Core

An (n+1)-core graph forms a rigid graph in n-dimension

Mobility Metric and Core Maintenance

Mobility Metric Priority Queue: 2, 4, 3, 5

2

3

5

4

Joint Solver for P2P Ranges

Matrix Completion

• Sequential Multi-lateration & lateration order

​

​

Range Measurement

UWB Ranging

• Bandwidth 500MHz ~ 1GHz
• Measures Time of flight using 40-bit timestamp
• Each bit of timestamp = 15.64 picoseconds
• 1 nanosecond major tap, 15.64 nanoseconds minor tap
• Few Centimeters of error in LOS

​

DecaWave DW1000

Two-way ranging

LOS vs NLOS Detection

2

3

5

4

1

6

DynoLoc Algo in Nutshell

• Run core decomposition
• Form the rigid body
• Order nodes by node mobility metric
• Maintain 3-core rigid body by selecting potentially LOS links
• After each measurements, run matrix completion and CMDS
• For 2-core nodes, use overheard neighbor list to resolve location
• For 1-core nodes, temporarily use IMU

Instrumentations

Absolute localization

Periodic Sync Frame

15.64 Picosecond Clock Tick

AoA

TOF-diff

Absolute localization

AoA-1

AoA-2

AoA-3

GUI

Vertical Detection

Multi-storied Operation

• Challenges
• Un-calibrated Pressure Sensor
• Un-mapped Floors
• Changing Values

​

Elapsed Time (min)

Pressure (mBar)

0

10

20

Floor 1 Anchor

Beacon (Floor 1🡪 2)

101

Results

• Up to 1.5 meters of errors for 16 nodes, 50% node mobile with 2m/sec speed in 100mx100m large indoor area with soft, hard, metallic and concrete partitions

​

• Micro-benchmarking shows mobility metric and link-quality metric increases accuracy at least two times

​

• Aggregate and concurrent ranging and joint solving increase refresh rates at least two times

Applications

In the Wild Demo

• Person in distress extraction time:

35 seconds vs

10 minutes

Conclusion

• This thesis presents License-free use of Spectrum in three distinct paradigms
• Dynamic Spectrum Access
• Infrared Spectrum via LOS
• Non-interfering Wideband
• Each paradigm resulted in novel system implementation that addressed real-world problems
• Each implementation has its limitations and promises that might be basis for future research

Questions and Suggestions?

Contributors

Himanshu Gupta

Professor, Dept. of CS Stony Brook University

Samir Das

Chair & Professor, Dept. of CS, Stony Brook University

Ayon Chakraborty

Assistant Professor

Dept. of CSE, IIT Madras

Max Curran

Software Developer

Google Inc.

Vyas Sekar

Professor, Dept. of ECE

Carnegie Melon University

Kai Zheng

Researcher

Apple Inc.

Jon Longtin

Professor,

Dept. of Mechanical Engr.,

Stony Brook University

Karthik Sundaresan

Professor, Dept. of ECE

Georgia Institute of Tech.