Generative Adversarial Network for Variable-Length Sensing Waveform Synthesis

We propose a generative adversarial network (GAN) system for synthesizing multiple families of novel radar waveforms with desirable Ambiguity Function (AF) shapes as well as a constant modulus property. Additionally, we introduce a penalty term to promote a low cross-correlation among synthesized waveforms. Many commonly-used radar code families contain only a limited number of sequences of a certain length. In modern radar applications, such as multifunction, MIMO, and cognitive radars, this may reduce the achievable performance gains. These systems launch multiple waveforms simultaneously in order to deal with low observable targets or large numbers of small targets. Therefore, the synthesis of novel waveforms and waveform families is important. Generating new waveforms on demand is also beneficial in scenarios where an adversary attempts to detect or recognize launched waveforms. We develop a conditional Wasserstein GAN (WGAN) for multiple datasets of complex-valued input data with varying code lengths. For each dataset, the model aims to synthesize waveforms with the same length and AF shape as the waveforms in that dataset. The model is trained to synthesize two classes of waveforms with different lengths using Frank and Oppermann codes. The AF shapes of synthesized waveforms are nearly identical to those of the training data. Additionally, the proposed penalty term allows for a tradeoff between the AF fidelity of synthesized samples and the expected cross-correlation among synthesized waveforms.