SDR Smart Modem

Introduction

The Smart Modem is designed to receive any signal from a USRP2, recognize the modulation scheme, and demodulate the signal. It also can be given an analog or digital signal, modulate it using a given scheme, and send it to a USRP2. To find this project, please visit ​the project GitHub.

Background

This project utilizes machine learning algorithms to recognize the modulation schemes of incoming signals. We first generated data using GNURadio to collect representative sample vectors of signals modulated with various modulation schemes. Then, we trained a convolutional neural network with this data. The results of the training are shown below:

Performance of Modulation Scheme Recognition

The neural network can always detect a signal modulated with a QAM scheme but has trouble determining the specific QAM scheme. Therefore, we use a support vector machine to accompany the neural network when it detects a signal modulated with QAM to find the specific scheme. This SVM determines ​the 2nd and 4th k-statistic of the QAM signal to better determine the scheme.

To modulate and demodulate the signals, GNURadio scripts are used according to the desired modulation schemes.

Tools Used

USRP2: Software defined radio

Quadro K5000: high-end GPU

GNURadio: SDR Toolkit

TensorFlow: Neural Network Library

Keras: High level Neural Network API

Scikit-learn: Machine Learning Library

Presentations

​Week One

​Week Two

​Week Three

​Week Four

​Week Five

​Week Six

​Week Seven

​Week Eight

​Week Nine

​Week Ten

​Week Eleven

​Week Twelve

​Poster

The Team

Avanish Mishra	Brendan Bruce