Eugenia Yu. Felagereva, Nelly R. Rudoman, Marina V. Kuzyakina

Kuban state university, Kuban state university of technology

This study examines the problem of applying classical radio signal recovery methods. The use of neural networks, namely autoencoders, is proposed as a solution. The architecture of a variational autoencoder has been developed to solve the problem of radio signal recovery from an unmanned aerial vehicle. To train and test the model, a large dataset containing radio signals of various types of modulations and noise levels was used, DeepSig Dataset: RadioML 2018.01A. The aim of the study is to develop an architecture that shows the best metrics for radio signal recovery than classical methods. The quality metrics used are PSNR, MSE, and MAE, and the Adam optimizer. The Kalman filter was also applied to the dataset, which showed two orders of magnitude worse results in all quality metrics. The data obtained show that classical algorithms for restoring distorted radio signals may not always show decent quality of work.

variational autoencoder, radio signal, UAV, Kalman filter