This is my Master's Degree thesis repository. It's about Signal Modulation Classification using Deep Learning.
Inbound signals are encoded using different kind of modulations, so it's mandatory to know (or predict) the modulation used in order to decode the signal content. AMR (Automatic Modulation Recognition) can be used to predict the modulation of an inbound signal. Deep Learning tecniques can be used for AMR, and that's what I did with my Master Degree thesis.
Paper: An Effective Convolution Neural Network for Automatic Recognition of Analog and Digital Signal Modulations for Cognitive SDR Applications
I collaborated at writing a paper (with Luca Pallotta and Gaetano Giunta from Roma Tre University in Rome) which resumes all the work done in this repository. In that work, architecture shown and explained is called FreeHand, but it's corresponding to the FreeHand v3 implemented in this repository. The paper is available at https://doi.org/10.1109/IWSDA50346.2022.9870449.
The project contains different neural network trainings. Each directory contains the trainings for a different data version. Inside each directory there are the .ipynb files which are the neural network trainings for the data version specified by the directory where they are. The schema below describes an example of the directory tree structure used in this project.
.
├── ...
│
├── DatasetTransformation 1
│ │
│ ├── neural_network_1_training.ipynb
│ │ ...
│ └── neural_network_M_training.ipynb
│
├── ...
│
├── DatasetTransformation N
│ │
│ ├── neural_network_1_training.ipynb
│ │ ...
│ └── neural_network_M_training.ipynb
│
└── ...
The project contains different neural network trainings. Inside the IQ directory there are the ones with the raw dataset, the other ones are using a pre-processing of the dataset before the training. All the directories content is shown in the table below.
| Directory | Content description |
|---|---|
| DFT + IQ | Datapoints transformed using DFT (Discrete Fourier Transform) concatenated with raw datapoints |
| DFT | Datapoints transformed using DFT (Discrete Fourier Transform) |
| IQ - Data Augmentation | Raw dataset enlarged using data augmentation tecniques |
| IQ | Raw dataset |
| MP + IQ | Datapoints transformed in Module/Phase (signals can be represented as complex numbers) concatenated with raw datapoints |
| MP | Datapoints transformed in Module/Phase (signals can be represented as complex numbers) |
I wrote some libraries to employ in the neural network trainings I did. A short description of those libraries content is shown in the list below:
- dataaugmentationlib.py enlarging the dataset using data augmentation tecniques
- datasetlib.py reading the dataset
- dftlib.py transform the dataset using DFT (Discrete Fourier Transform)
- evaluationlib.py evaluating models after training
- modulephaselib.py transforming the datapoints in Module/Phase
- neural_networks.py neural network implementations
- trainlib.py training neural networks
- traintestsplitlib.py split the dataset in training set and test set
Follow those steps to get all running.
Dataset can be downloaded from here. Extract the dataset in a directory named data/. Dataset file should has to be named RML2016.10a_dict.pkl.
You need to have Python (at least 3.9.x) and pip (included with Python) installed on your system.
You need to run pip install -r requirements.txt to install this project dependencies.
Check if your GPU is CUDA enabled here.
I had a problem with CUDA in the beginning, problem/solution couple available here on StackOverflow.
You need to install CUDA Toolkit in your System: download.
You need to install cuDNN Library in your System: download and follow NVDIA installation guide
| Name | Ref |
|---|---|
| The paper which resumes the work done in this repository | link |
| Background, history and an introduction to Automatic Modulation Recognition with Deep Learning | link |
| Dataset RadioML2016.10a | link |
| CNN2 neural network | link |
| RadioML2018.01a, an extended version of the dataset used in my thesis | link |
| Baseline neural network, relabeling and ensamble learning | link |
| Batch normalization | link |