Fusion-Based Off-Grid Direction of Arrival Estimation via Deep Learning with Array Imperfection

In practical applications, array imperfections frequently arise and significantly degrade the precision of directionof-arrival (DOA) estimation. Currently, with the development of deep learning, Data-Driven (DD) algorithms have certain advantages over Model-Driven (MD) methods in some aspects. This paper proposed a fusion-based DOA estimation network, namely IQ-ENSCM-Net, which integrates the array signal and Sample Covariance Matrix (SCM) in the presence of array imperfections for multiple sources. First, time features of inphase (I) and quadrature (Q) components of the array signal are extracted using the Transformer encoder. Then, the SCM is enhanced based on Toeplitz structure and spatial features are extracted using the residual connection network. Mostly, features of the signal and enhanced SCM (ENSCM) are fused using the weighting network to get DOA estimation results. And this paper designs a Pseudo-Spatial Spectrum (PSS) whose amplitude represents the fractional information of the DOA, thus enabling off-grid DOA estimation. In addition, the Binary CrossEntropy (BCE) loss function is improved by regular function and category weighting coefficients. The weighting coefficients are based on the difference between the estimation results and the true values, which improve the contribution of the sample with large error to the loss. Finally, simulations demonstrate the effectiveness of the proposed PSS and improved BCE. Results also demonstrate superior DOA estimation performance compared to other methods and strong robustness against array imperfections.