Neural Network-Based Orthogonal Waveform Design for Multiple-Input-Multiple-Output Radar

To address the limitations of conventional Multiple-Input-Multiple-Output (MIMO) radar orthogonal waveform design methods-including high computational complexity, lengthy optimization processes, and degraded orthogonality performance-this paper proposes a residual neural network-based approach for designing waveforms with enhanced orthogonality. The proposed method specifically achieves lower autocorrelation sidelobes and cross-correlation energy. The methodology comprises three key steps: (1) establishing a waveform design framework guided by peak sidelobe level (PSL) minimization principles; (2) formulating a composite loss function integrating autocorrelation and cross-correlation suppression objectives; and (3) training the network through iterative parameter optimization. The network iteratively optimizes its parameters through gradient-driven loss minimization, ultimately generating a set of orthogonal waveforms that approach Pareto optimality in balancing autocorrelation and cross-correlation performance. Simulation results demonstrate that the designed waveforms achieve an autocorrelation PSL below <tex xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">$-16dB$</tex> and cross-correlation peaks as low as <tex xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">$-15dB$</tex>, significantly outperforming conventional optimization methods. These waveforms exhibit robust orthogonality, meeting the stringent requirements of MIMO radar systems for high-resolution target detection and interference mitigation.