An Optimized Neural Network Framework for Designing Spectrally Compatible Radar Waveforms

To improve the radar target detection performance in the presence of interference, this paper addresses spectrally compatible waveform design for multiple-input multiple-output (MIMO) radar systems. An optimized neural network, called WaveNet, is proposed to design constant modulus waveforms with minimal stopband energy and precise control over integrated sidelobes level (ISL) or peak sidelobes level (PSL), which overcomes the shortcomings of previous works in implementing the PSL constraint. Leveraging WaveNet, the NP-hard waveform design problem is solved iteratively. More specifically, the update of the waveforms relies on the nonlinear mapping of the neural network, which is synchronously tuning during the iterative process based on a carefully designed loss function. Such an updating scheme exhibits good convergence in numerical simulations. Additionally, simulation results also show that compared with the other waveform designs, the proposed scheme can achieve deeper energy notches in stopbands while maintaining a lower sidelobes level, benefiting from its precise control over PSL. As a result, the designed waveforms enable the radar to improve the anti-interference ability and reduce the false alarm probability or missed detection probability.