GENERATION OF SYNTHETIC RADIO FREQUENCY SIGNALS FOR TRAINING DEEP LEARNING MODELS IN ELECTRONIC WARFARE SCENARIOS

Abstract

For the development of robust systems for remote localization and analysis of powerful jamming and countermeasure signals, a critical aspect is the availability of large and diverse training datasets. Traditional data collection using professional radio equipment and field testing is expensive, time-consuming, and often limited by technical and regulatory factors. Synthetic generation of radio frequency signals using state-of-the-art generative deep learning models offers an attractive alternative: models learn the statistical structure of real signals and channels, and then generate realistic I/Q sequences with controllable parameters. This article proposes an RF signal synthesis method based on VQ-VAE combined with an autoregressive Transformer. The proposed scheme learns the distribution of real RF signals and generates synthetic sequences conditionally specified by the modulation type, signal-to-noise ratio (SNR), and propagation channel. Experimental results on a dataset of 500,000 complex sequences showed that the VQ-VAE + Transformer architecture achieves a mean squared reconstruction error (MSE) of 0.0245 and a spectrogram correlation of 0.948, outperforming baseline GAN, VAE, WGAN, and diffusion models in terms of quality and training time.