AI for RF Signal Modulation Classification

2024 AI Challenge by Ball Aerospace

This repository is a submission for the AI for RF Signal Modulation Classification Challenge. The challenge aims to develop an automated signal classifier for use as a payload skill on a next-generation satellite mission. The primary goal is to autonomously classify radio frequency (RF) signals into predefined classes based on sampled waveforms.

RF signal classification is critical for various applications, including: • RF interference detection • RF intelligence gathering • Spectrum management • Jammer detection • Spectral compliance verification

Introduction

Radio frequency signal classification is a critical task in wireless communication systems, allowing for efficient and accurate interpretation of signals. This project aims to classify RF signals into various modulation types using a neural network architecture optimized for accuracy.

---

Features

Input Format: Time-series data of complex-valued IQ samples at a sampling rate of 100 MHz. Supported Modulation Types: 1. BPSK 2. QPSK 3. 8PSK 4. MSK 5. FSK 6. PAM4 7. GMSK 8. GFSK 9. 16QAM 10. 64QAM 11. 128QAM Training Dataset: • 297,000 examples across 11 modulation types • Labels provided as one-hot encoded vectors

---

Installation

Grab the training data: https://www.icloud.com/iclouddrive/04dGShJz9KTKeWditkLtJyeDw#AI_Challenge_Training_Data

1. Clone the repository:

   git clone https://github.com/your-username/RF-Signal-Classifier.git
   cd RF-Signal-Classifier

2. Set up a Python virtual environment:

   python3 -m venv env
   source env/bin/activate  # On Windows: env\Scripts\activate

3. Install the dependencies:

   pip install -r requirements.txt

---

Usage

Training the Model

Run the following command to start training the model:

python rf_signal_classifier/train.py

Inference

Run the model on the test dataset python run_inference.py --test_data_path

Evaluation

Score = (accuracy / 100) + (10 / log10(Nparams)) + β - ε

Results

Model Performance

THe model achieved a validation accuracy of 78.27% with a drastically reduced parameter count of 68,923 parameters, making the model lightweight and efficeint while maintaining strong performance across most modulation types.

Classification Report

          precision    recall  f1-score   support

    BPSK       0.97      0.96      0.97      2765
    QPSK       0.83      0.87      0.85      2765
    8PSK       0.82      0.85      0.84      2743
   16QAM       0.49      0.72      0.58      2758
   64QAM       0.44      0.18      0.26      2715
  128QAM       0.46      0.46      0.46      2659
    PAM4       0.93      0.98      0.95      2662
     FSK       0.93      0.70      0.80      2648
     MSK       0.79      0.94      0.86      2661
    GMSK       0.95      0.97      0.96      2663
    GFSK       0.94      0.96      0.95      2661

accuracy                           0.78     29700

macro avg 0.78 0.78 0.77 29700 weighted avg 0.78 0.78 0.77 29700

Future Improvements

1. Enhance data augmentation: Apply more targeted automation for underperforming classes likes 64QAM ans 128QAM