Machine Learning for Classification of RF Signals

Machine Learning for Classification of RF Signals is a university project for Automatic Modulation Classification, also known as AMC. The system classifies radio-frequency signals into 24 modulation types across 26 SNR levels, from -20 dB to +30 dB, using the RadioML 2018.01A dataset.

The project compares several deep learning architectures under a controlled protocol. Every model uses the same stratified 60/20/20 train, validation, and test split, the same random seed, optimizer, learning rate, batch size, epoch budget, and evaluation procedure. This makes the comparison focused on model behavior instead of accidental training differences.

What It Solves

Automatic Modulation Classification identifies the modulation scheme of a received RF signal, such as BPSK, QPSK, QAM, or OFDM, from raw I/Q samples without manual feature engineering.

This matters in spectrum monitoring, cognitive radio, interference or jamming detection, and adaptive communication systems, where the receiver needs to understand the signal type before demodulation or channel adaptation.

Dataset

The project uses RadioML 2018.01A, a large labeled RF dataset with 2.55 million I/Q signal samples. The split pipeline preserves the relationship between 24 modulation classes and 26 SNR levels, producing 624 stratified groups for fair training and evaluation.

Models Compared

The final results focus on three project-specific PyTorch architectures:

CNN1D
CNN-LSTM
ResNet1D

The strongest portfolio result currently comes from ResNet1D, with an overall test accuracy of about 59%. The accuracy-vs-SNR plot shows the expected RF behavior: performance is weak at very low SNR values and improves as the signal becomes cleaner.

ResNet1D accuracy versus SNR for automatic modulation classification — ResNet1D accuracy by SNR level, used as the main portfolio thumbnail for the project.

Annotated ResNet1D confusion matrix for RF modulation classes — Annotated confusion matrix showing where the model separates or confuses modulation classes.

Pipeline

The project includes an end-to-end training and evaluation workflow:

HDF5 dataset loading
Stratified train, validation, and test split
YAML-based experiment configuration
Model factory pattern
PyTorch training with early stopping
Test-set evaluation with accuracy, precision, recall, and F1
Per-SNR and per-class metrics
Confusion matrix generation
Cross-model KPI comparison
MATLAB visualization export
Docker and Docker Compose setup for reproducible GPU runs

Architecture Snapshots

These diagrams show the three project-specific PyTorch baseline architectures. CNN1D focuses on local RF signal features, CNN-LSTM adds temporal sequence modeling, and ResNet1D uses residual one-dimensional blocks to support deeper feature extraction.

CNN1D architecture for automatic modulation classification — CNN1D baseline architecture.

CNN-LSTM architecture for automatic modulation classification — CNN-LSTM baseline architecture.

ResNet1D architecture for automatic modulation classification — ResNet1D baseline architecture.

My Work

I built the end-to-end PyTorch training and evaluation pipeline, including src/train.py, src/evaluate.py, src/dataset.py, and src/utils.py.

I also implemented the stratified RadioML split pipeline, custom PyTorch baselines, YAML experiment configs, cross-model comparison tooling, MATLAB visualization scripts, and Docker-based reproducible training.

This work supports my academic thesis, Machine Learning for Classification of RF Signals, developed at UNIZA in 2026.

Technologies

Python, PyTorch, NumPy, HDF5/h5py, scikit-learn, Matplotlib, Seaborn, YAML, tqdm, Docker, Docker Compose, MATLAB, CUDA/GPU, and the RadioML 2018.01A dataset.