Neural Network Generative Models for Radio Frequency Data

Abstract

Neural Networks (NN) provide great flexibility in modeling non-linear relationships and have recently proved to be very valuable in Radio Frequency (RF) domain, e.g. NN based Digital Signal Processing (DSP) provide equal or better performance than traditional DSP function blocks. One key requirement of NN models is their need for large amounts of training data to mitigate model overfitting. If this data requirement is satisfied it leads to models that generalize well, but does not hold in cases of limited training data. Generating high fidelity data can help to mitigate data scarcity. In this paper we demonstrate the application of Generative Adversarial Network (GAN) based RF data generation capability for training interference removing NN models. We developed two GAN models that learn to generate RF data with tone jammers. The first is a "Sequence to Sequence GAN model", which learns the jammer data distribution by learning to map clean RF signals to impaired RF signals with a jammer present. The second is a physics driven "Tone GAN" model, which learns the frequency, phase, and magnitude of the interferer from a limited data distribution and applies this impairment to the clean signal to generate impaired waveforms similar to the GAN training data. Experimental results show that the new data generated from the trained GAN models can be used to augment the original NN model, by fine-tuning them via the GAN generated data, leading to improved performance and generalization.