CANSOFCOM

Problem

The Canadian Special Operations Forces Command wants to classify drones based on the RADAR signals that bounce off of them. The drone actually distorts the returned frequency: the blades are moving fast, so some micro-doppler frequencies are returned. Each drone has a unique return signal pattern, determined by its dimensions, number of rotors, radial speed of the rotors, number of blades per rotor, etc.

Parameters

Drone parameters

$N$: The number of blades in a single rotor.
$f_{rot}$: (1/s) the frequency of rotation of the rotor. This is drone-dependent, but could be [50, 18].
$L_1$: (meters) The distance of the blade roots from the center of rotation of the rotor. For now, assumed to be 0.
$L_2$: (meters) The distance of the blade tips from the center of rotation of the rotor. Differs depending on drone, generally 0.05 to 0.25.

Scenario parameters:

$A_r$: a scale factor, arbitrary constant. Set to 1 for now.
$R$: (meters) The distance from the radar to the center of rotation (the rotor). Could be in the 1000 to 5000 range (a few kilometers). Assumed to be 0 for this hackathon.
$\theta$: (radians) the positive angle between the plane of rotation of the rotor and the line of sight from the radar to the center of rotation (of the rotor). In the range [0, $pi/2$].
$V_{rad}$: (rad/s) the radial velocity of the center of rotation (of the rotor) with respect to the RADAR. Assumed to be 0 for this hackathon, but this would affect doppler considerations.

RADAR parameters:

$f_c$: (1/s) the transmitted RADAR frequency. In the range [10GHz, 94 GHz].
$\lambda$: (meters) the wavelength of the transmitted signal. lambda = c/f_c, where c is the speed of light in m/s.
$t$: (s) the time, our real-valued continuous input.
$f_s$: (1/s) the sampling frequency (how often we sample our RADAR return signal). 10 KHz for X-Band RADAR, 26KHz for W-Band RADAR.

Overview

CANSOFCOM.ipynb - annotated overview of the whole repository, as well as results.
signalgenerator.py - the home for the function that builds $\psi$, and for generating data from sampling $\psi$.
classifier.py - definitions of the Convolutional Classifier Neural Networks.
train.py - training script for training a convnet on the generated data.
dataset_generation.py - script that generates the filesystem of SW STFTs for all the different frequencies, SNRs, etc.
configurations.py - place for storing dictionaries about scenarios and drones
helpers.py - analysis tools
test.py - big function for testing specific neural networks (TODO: make more general)
visualize.py - example for plotting a fourier transform, time-domain signals, etc.
fourier.py - provides a function to generate short and long-window STFTs.

Dependencies

scipy
numpy
pytorch
matplotlib
scikit-learn
mlxtend

If you're using conda, just run this and it will install everything:

conda env create -f environment.yml

History

This repository was built for the CANSOFCOM Drone Classification Machine Learning challenge of Hack The North 2020++.
You can see the original challenge in CANSOFCOM_Challenge.pdf.

TODO

These are some ideas that I don't have time to implement. (in order of predicted impact on performance)

Finding the ideal STFT parameters based on the f_s and such.
Focal Loss instead of crossentropy to focus on classes that are harder
Small architecture changes. Deeper? Larger receptive field?
Hyperparam sweep

SinclairHudson/CANSOFCOM