Real-Time AI Systems INTRODUCTION KickView creates real-time AI - - PowerPoint PPT Presentation

▶
real time ai systems
SMART_READER_LITE
LIVE PREVIEW

Real-Time AI Systems INTRODUCTION KickView creates real-time AI - - PowerPoint PPT Presentation

Sep 18, 2022 208 likes •351 views

TACKLING THE CROWDED RADIO FREQUENCY (RF) SPECTRUM USING DEEP LEARNING Krishna Karra, March 28 2018 Real-Time AI Systems INTRODUCTION KickView creates real-time AI systems. Intelligent Multi-Sensor Analytics (IMSA) kvMotif - Our

slide-1
SLIDE 1

Real-Time AI Systems

TACKLING THE CROWDED RADIO FREQUENCY (RF) SPECTRUM USING DEEP LEARNING

Krishna Karra, March 28 2018

slide-2
SLIDE 2
  • Intelligent Multi-Sensor Analytics (IMSA)
  • kvMotif - Our intelligent multi-sensor software stack
  • Collect, process and learn from networks of sensors in real-time
  • Directly connect AI to the physical world
  • Real-time AI Processing at Network Scale - Telco, IoT, Industrial, Smart

Cities, Security and Defense

  • Real-time collection, processing, streaming analytics
  • Network optimization and prediction
  • Intelligent video and multi-sensor processing
  • Custom datasets, algorithms and training

2

INTRODUCTION

  • KickView creates real-time AI systems.
slide-3
SLIDE 3

MOVING TOWARDS INTELLIGENT RF SYSTEMS

3

Expert RF Systems, while powerful, do not scale to the IoT

The RF spectrum is increasingly crowded as devices and networks compete for frequency bands Unique cyclostationary features of digitally modulated signals can be extracted for signal detection and discrimination

Tools and theory from DSP can aid and inform the development of AI-powered Intelligent RF systems

Example of demodulation of Bluetooth signal to extract unique header information

!

"

Symbol Duration time Optimal Sample Time #

"

Symbol Rate Band-Edge Spectral component $% $& '(()*+,- $/0-1+-2ℎ frequency Time Domain Frequency Domain 4

5

slide-4
SLIDE 4
  • In-Phase/Quadrature data representation
  • RF signals contain information about both magnitude and phase
  • Massive data rates
  • Many modern communication signals are wideband in order to

transmit data at a higher rate (e.g. WiFi 802.11ac is 160 MHz wide)

  • Datasets
  • Lack of multiple, comprehensive, freely available datasets compared to
  • ther domains (e.g. ImageNet, KITTI, Coco)

4

DEEP LEARNING APPLIED TO RF DATA

The RF modality presents unique challenges

slide-5
SLIDE 5

Signal Detection from Spectrograms

5

  • Actual RF signal environment at KickView office

kvPrelude

  • RF spectrum usage and interference detection
  • KickView created the Signal Processing with

DIGITS tutorial for NVIDIA Deep Learning Institute

DEEP LEARNING APPLIED TO RF DATA

  • Frequency
  • Time
slide-6
SLIDE 6

OFDM Signal Detection - WiFi IEEE 802.11g (1)

DEEP LEARNING APPLIED TO RF DATA

80 samples 80 samples 80 samples 80 samples 80 samples 80 samples 80 samples 80 samples 80 samples

+ +

80 samples 80 samples 80 samples

80-pt DFT

t0 t1 t2

OFDM Symbol CP OFDM Symbol CP OFDM Symbol CP

Channelization transforms the raw time-domain OFDM signal into a complex signal image

Orthogonal Frequency Division Multiplexing (OFDM) is a digital multi-carrier modulation scheme that is employed in many fielded systems; WiFi, cable systems (e.g. DOCSIS 3.1) and cellular networks (e.g. 4G, 5G) either currently deploy or are moving towards OFDM as the PHY-layer standard.

Depiction of OFDM carriers in the frequency domain for IEEE 802.11g, a common WiFi standard

6

slide-7
SLIDE 7

OFDM Signal Detection - WiFi IEEE 802.11g (2)

7

Pre-processing maximizes the structure in the data for feature extraction using deep learning approaches

0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1 1.1

  • 8
  • 6
  • 4
  • 2

2 4 6 8 PROBABILITY OF DETECTION (PD) SIGNAL TO NOISE RATIO (SNR) Noise (Full-band) Noise (5MHz Sub-band) Noise + BT (Full-band) Noise + BT (5MHz Sub-band)

Bluetooth 802.11g WiFi

Generalizes to real-world RF environments and does not require explicit time/ frequency synchronization

DEEP LEARNING APPLIED TO RF DATA

Live RF collection at KickView office

slide-8
SLIDE 8

kvPrelude RF Dataset Generation Tools

DEEP LEARNING APPLIED TO RF DATA

8

Large, representative datasets are required for deep learning approaches to be successful

KickView WiFi Dataset

  • ~20,000 I/Q snapshots
  • OFDM (IEEE 802.11g WiFi)
  • Bluetooth
  • Noise
  • Complex signal images
  • Output of channelizer
  • Includes time/frequency errors

OFDM Noise Bluetooth Stay tuned for the public release of our WiFi dataset along with others on our website Example snapshots kvPrelude RF Dataset Our patent pending kvPrelude tools can automatically generate labeled RF datasets collected over-the-air for any modern communications protocol

slide-9
SLIDE 9

Intelligent Spectrum Monitoring and Anomaly Detection

DOCSIS 3.0 cable modem signal interference and anomaly detection using a combination of DSP and deep learning

Frequency Receiver Frequency Overlap

Frequency Time

Receiver Bandwidth Maximum Bandwidth Signal Fits into Overlap Region

am

Dwell Time

Receiver collection parameters (e.g. dwell time, receiver bandwidth, frequency overlap) all affect a system’s ability to monitor and evaluate RF bands of interest.

9

DEEP LEARNING APPLIED TO RF DATA

slide-10
SLIDE 10

Generalized Modulation Recognition

10

DEEP LEARNING APPLIED TO RF DATA

!" # $ ⃗ & ⃗ $ ⃗ & ⃗

'" '"() '"(* '"(+

$ ⃗ & ⃗

A B C D '" = A '"() = C '"(* = B '"(+ = A Complex Sampling instance(s)

Trajectory through I/Q space

  • vs. time for a QPSK

modulated signal (4 symbols, 5 samples per symbol) Symbol Mapping Depicted Symbol
 Sequence

We are conducting research to develop a generalized modulation recognition recognition scheme using deep learning. We train a deep neural network over a sampling of possible combinations of waveform trajectories for a short number

  • f observed symbols.

Key benefits include fast fixed inference time, robustness to co-channel interference, and no need for explicit time/frequency synchronization.

slide-11
SLIDE 11

Space Applications - RF Emitter Geolocation

11

Differential Delay Differential Doppler

Cross Ambiguity Function computes raw TDOA/ FDOA measurements across a sensor pair Deep neural network learns a manifold that maps all possible TDOA/FDOA measurements and sensor configurations to potential geolocation solutions

KV-Geo Engine Θ Ψ # $ ⃗ Θ Ψ # $ ⃗

= TDOA/FDOA estimates = Satellite Ephemerides = Geolocation Estimates = GDOP

DEEP LEARNING APPLIED TO RF DATA

slide-12
SLIDE 12

12

INTELLIGENT MULTI-SENSOR ANALYTICS

kvPrelude: Real-Time Intelligent Spectrum Monitoring

  • kvPrelude is a scalable software solution for real-

time RF data acquisition and intelligent spectrum monitoring

  • Automated spectrum monitoring and analysis
  • Container-based architecture
  • Processing at the edge, fog, or cloud
  • Common software framework for applications
  • Supports multiple receivers and sample rates
  • Real-time stream processing and visualization
  • Automated training data collection and labelling
slide-13
SLIDE 13

13

Contact Us: krishna.karra@kickview.com

Real-Time AI Systems

See a live RF demo at the ADLink Booth (207)