Use of ResNet Autoencoders for Designing Phase-Quantized Sequences With Good Correlation for MIMO Radar Systems

Multiple-input multiple-output (MIMO) radar technologies can improve radar detection capabilities and share frequencies with adjacent radar sites by transmitting nearly uncorrelated waveforms. Under certain system constraints, a set of finite-resolution digital-to-analog converters (DACs) can reduce hardware cost and power consumption. However, the waveform quantization process through DACs forces a continuous phase to lie within a discrete phase, which degrades auto- and cross-correlations. Therefore, it is usually desirable that the sequence has a finite alphabet achieving good correlation properties. Recently, uncorrelated waveform design by applying neural networks (NNs) in place of coding theory has received much attention. However, the design of phase-quantized sequences using NNs has been delicate because of differentiability with sequences modulated by discrete phase. This article proposes a framework for designing phase-quantized sequences using an NN. Numerical results show that sequences designed using the proposed framework have better correlation properties compared with those designed using existing algorithms.