Application of Neural Networks for Radio Signal Classification

The study reported here focuses on evaluating and comparing different neural network architectures applied to the binary classification of radio signals, with a focus on ZALA (Unmanned Aerial Vehicle) and ELRS (drone communication protocol) signal types. Feature vectors were derived from I/Q-format signal samples and included a novel combination of spectral and statistical characteristics such as power spectral density, entropy, skewness, and spectral centroid. Four neural network models were evaluated: three Dense-based architectures with varying depths, and a 1D convolutional neural network (1D-CNN). All models were trained and validated on balanced datasets using Keras in Python, achieving perfect classification metrics (precision, recall, and $\mathbf{F 1}$-score all equal to $\mathbf{1. 0}$). While the mediumcomplexity Dense model offered the best balance between simplicity and performance, the 1D-CNN model demonstrated greater robustness and generalization capabilities, especially under potential signal variation or noise conditions. The results confirm the effectiveness of both architecture types and highlight the relevance of feature design in radio signal classification tasks.