ORCA-PROJECT.EUORCA-PROJECT.EU

ORCHESTRATION AND RECONFIGURATION CONTROL ARCHITECTURE

Crowncom

Ghent, Sept 18th 2018

PresenterUSING DEEP LEARNING AND

RADIO VIRTUALISATION FOR

EFFICIENT SPECTRUM

SHARING AMONG

COEXISTING NETWORKS

Wei Liu

IMEC – Ghent University

WWW.HUB4NGI.EU

ORCA-PROJECT.EU

• Motivation

• Machine learning and spectrum monitoring

• Radio virtualization

• Proposed solution

• Proof of concept – showcase experiment

2

Outline

WWW.HUB4NGI.EU

ORCA-PROJECT.EU 3

Motivation

frequency

time

power

throughput (Kbps/km2) delay (ms)

cells-links (per km2)

Radio Degrees of Freedom

WWW.HUB4NGI.EU

ORCA-PROJECT.EU 4

Motivation for spectrum sharing driven by

context awareness• Current approaches

• Vertical sharing -> licensed bands, opportunistic access of spectrum• Horizontal sharing -> unlicensed bands, technologies each has its own strategy

to share the medium, some MAC are compatible (eg, different CSMA technologies could in theory coexist)

• Disadvantages• Only local vision of spectrum availability,

• can’t assess QoS performance from a global level: eg some services can’t tolerate

much delay, can’t coexist with traffic with long burst -> low latency control loop in

industrial application + youtube video stream application

• Advantages with machine learning aided spectrum sharing• Learn context information from ongoing traffic over air-(ie. Waveform, MAC, the amount of

transmission opportunity, how frequently opportunities arrive)

• Make decisions for the best combinations to insert new traffic in limited physical resource

APPAPP

APPAPP

Good

services

Bad

services

WWW.HUB4NGI.EU

ORCA-PROJECT.EU

• Trend of future: RATs need to coexist, instead of having exclusive spectrum bands -> need spectrum sensing

• Feature Detection vs Energy Detection

• Pure energy detection has limited applicability besides CCA

• Feature detection is impractical in very diverse environments

• Needs substantial redesign and redeployment for every change in the environment

• Expert-based costly design despite its low generality (companies tend to shy away from complicated approaches)

• Sensitive to impairments (e.g. multipath, offsets, non-linearities, collisions)

5

Signal classification - Challenges

Expert

Feature

Detection

Energy

Detection

generality

context

awarenessMachine

learning

WWW.HUB4NGI.EU

ORCA-PROJECT.EU

• Advantage of machine learning:

• Generic, less sensitive to changes and impairments, and can be aware of more complex environment

• Concrete examples:

• recognize persistent vs bursty channel activity (e.g. Wi-Fi vs LTE DL or DVBT)

• predict neighbours’ spatial deployment topologies (e.g. mesh vs centralized)

• predict neighbours’ duty cycles / inter frame durations

• use gathered information for interference cancellation/avoidance

6

Signal Classification - Motivation

WWW.HUB4NGI.EU

ORCA-PROJECT.EU 7

Motivation for radio virtualization

• Definition• A virtualized radio is a logical radio instance created from a physical device,

we may create multiple logical instances operating in homogeneous or

heterogeneous modes• Similar to CPU virtualization, a core can do multi-tasking

• Use cases• Use a single Wi-Fi access point as APs on multiple Wi-Fi channels (frequency

slicing), or serve multiple space (antenna slicing), simultaneously • Multi-standard IoT gateway, use one device to support multiple IoT standards

eg, SigFox, LoRa

WWW.HUB4NGI.EU

ORCA-PROJECT.EU 8

Motivation for radio virtualization

• Definition• A virtualized radio is a logical radio instance created from a physical device,

we may create multiple logical instances operating in homogeneous or

heterogeneous modes• Similar to CPU virtualization, a core can do multi-tasking

• Use cases• Use a single Wi-Fi access point as APs on multiple Wi-Fi channels (frequency

slicing), or serve multiple space (antenna slicing), simultaneously • Multi-standard IoT gateway, use one device to support multiple IoT standards

eg, SigFox, LoRa

WWW.HUB4NGI.EU

ORCA-PROJECT.EU 9

Motivation for radio virtualization

• Definition• A virtualized radio is a logical radio instance created from a physical device,

we may create multiple logical instances operating in homogeneous or

heterogeneous modes• Similar to CPU virtualization, a core can do multi-tasking

• Use cases• Use a single Wi-Fi access point as APs on multiple Wi-Fi channels (frequency

slicing), or serve multiple space (antenna slicing), simultaneously • Multi-standard IoT gateway, use one device to support multiple IoT standards

eg, SigFox, LoRa

WWW.HUB4NGI.EU

ORCA-PROJECT.EU

• Current SoA• Share the chip in a time division manner, one function at a time, eg, Wi-Fi access

point broadcast a message for channel switching

• Duplicate chips, one for each standard -> increase form factor + complexity, potential waste of resources (not all functionalities required at all time)

• Advantage with virtualized radio• No need to duplicate chips, virtual radios are created on demand

• Always available for multiple concurrent services from application point of view

10

Motivation for radio virtualization

WWW.HUB4NGI.EU

ORCA-PROJECT.EU

• Motivation

• Machine learning and spectrum monitoring

• Radio virtualization

• Proposed solution

• Proof of concept

11

Outline

WWW.HUB4NGI.EU

ORCA-PROJECT.EU 12

N slices initiated from the same HW, services are mapped to channels

according to the classified properties

Proposed solution

ML classifier

CH1=?

CH1: duty cycle,

burst length, modulation, etc

Classification output

N x Virtualized Radio

Hypervisor

S1:CHX

U1

S1 SNCH2=?

CHN=?

CH2: ….

CHN: ….

S2:CHY

SN:CHY

S1

U2 Un

WWW.HUB4NGI.EU

ORCA-PROJECT.EU 13

Proof of concept

CH1=? CH2=?

TCD’s classifier

CH1=very occupied,

short burst

CH2=less occupied, long burst

Classification output

S1 low latency

S2 high TP

IMEC BS

Hypervisor

SDR platform creates two slices from same HW and RF frontend

Slice 1:Short burst

access in CH1

Slice 2:Long frame

access in CH2

UE1 UE2

WWW.HUB4NGI.EU

ORCA-PROJECT.EU 14

Experiment set-up

BS

UE1 UE2

CH1 traffic generator

CH2 traffic generator

TCD Rx1 TCD Rx2 Spectrum Monitoring+GUI

Zynq SoC

WWW.HUB4NGI.EU

ORCA-PROJECT.EU 15

How is radio virtualization achieved?

DDC filterbank

Rx antenna

DUCfilterbank

Tx antenna

Channelselection

OFDMreceiver40Msps/40MHz

40Msps/40MHzOFDM

transmitter

Processor

(MAC and upper)

PL PS

10 channels:2*20MHz/20Msps4*2MHz/2Msps

10 channels:2*20MHz/20Msps4*2MHz/2Msps

Channelselection

WWW.HUB4NGI.EU

ORCA-PROJECT.EU 16

Show time!

WWW.HUB4NGI.EU

ORCA-PROJECT.EU 17

Take away message