Denoising-Dynamic Attention Network for Circular and Noncircular Mixed Signals DOA Estimation

In this work, we consider a challenging circular and noncircular signal direction-of-arrival (DOA) estimation problem in arbitrary sparse arrays. The problem is formulated as a regression task. Using second-order statistical properties of mixed signals and contemporary neural networks for arbitrary sparse array receiving data, we propose a denoising-dynamic attention (DDAttention) network to achieve gridless sparse recovery and mixed signal separation. DDAttention consists of a denoising module and a dynamic attention (DAttention) module. The denoising module enhances signal features and provides more accurate signal information for subsequent sparse recovery. Dynamic convolution neural network (DCNN) and dynamic self-attention network (DSAN) in the DAttention module to model the virtual array’s hole measurement data and ensure that depth and spatial features are extracted. By training the network using arbitrary sparse array receiving data, the network exhibits significant performance when different numbers of array elements are missing. We consider that with 1/6, 1/3, and 1/2 of the array elements missing, the network’s performance in terms of the recovered signal and the separated signal is still close to the case without missing array elements. Experimental results on simulated data demonstrate stable recovery and separation performance gains.