MIMO Radar Waveform Synthesis Using Generative Adversarial Networks

Modern radars such as MIMO and multifunction radars may launch multiple waveforms simultaneously to perform different radar tasks or resolve more targets. As existing radar code families are limited in size and number, novel waveform synthesis methods are needed. Machine learning methods offer an alternative to traditional code design approaches. We propose a deep learning method for the synthesis of diverse waveform families for MIMO radars, with an emphasis on orthogonality properties. To this end, a Generative Adversarial Network (GAN) method and associated penalty terms promoting diversity are developed. GANs are generative deep learning models that can learn a variety of data distributions without explicit formulation. We structure the GAN latent space using input labels and proposed penalty terms to promote orthogonality among the generated waveforms. Promoting diversity properties such as orthogonality makes the developed approaches applicable to modern radar applications, such as MIMO systems, fully digital antenna arrays and multifunction radars. The proposed system is trained using Oppermann codes, and diversity properties of the synthesized codes are studied. The proposed penalty term is demonstrated to successfully produce varying batch sizes of waveforms that are close to orthogonal with each other.