Design of Phase-Quantized Unimodular Waveforms on Neural Networks for MIMO Radar Systems

Multiple-input, multiple-output radar systems have had great success in communication systems and in probe-based sensing with multiple orthogonal waveforms. In implementations of these systems, a set of analog-to-digital converters quantizes the waveforms, forcing the waveform phase to lie within a finite alphabet, which degrades auto- and cross-correlation properties. Therefore, it is usually desirable that the sequence entries are from a finite alphabet. Recently, the design of orthogonal waveforms using neural networks (NNs) rather than coding theory or exhaustive searches has received much attention. However, the design of phase-quantized waveforms using NNs has been delicate because of waveform differentiability. This paper proposes a framework for designing phase-quantized orthogonal waveforms with any codelength and any finite alphabet using the ResNet autoencoder, which has been applied to the design of sequences in wireless communications with resistance to vanishing gradients when handling high-dimension sequences. Numerical results show that sequences designed using the proposed framework have better correlation properties than those designed using existing algorithms.