Coprime Array DoA Estimation Based On Improved Multi-label cGAN

An improved multi-label conditional Generative Adversarial Network (mcGAN) is proposed to obtain Direction-of-arrival (DoA) estimates for coprime phased arrays. First, data preprocessing is performed. The real and imaginary part data of a normalized coprime array covariance matrix are used as inputs to the mcGAN generator with its outputs as the covariance matrix of a uniform array. The DoAs are treated as label information, and the label dimension equals the number of signals. The multi-dimensional labels are only used as the inputs to the discriminator and not fed into the generator by avoiding information leakage. Finally, estimation of signal parameters via rotational invariant techniques (ESPRIT) is applied to the recovered uniform array covariance matrix to yield the DoA estimates. Simulation results have verified the efficiency and accuracy of mcGAN. It is revealed that mcGAN can effectively control the generation of expected patterns without pattern mismatch or pattern collapse; moreover, the generator shows good generalization abilities. Based on the reconstructed data, the DoAs can be estimated quickly and accurately, and closely spaced spatial signals can be differentiated properly as well.