DIRECTION OF ARRIVAL ESTIMATION BASED ON DNCNN IN LOW SNR

The direction of arrival (DOA) estimation is a key issue of array radar. Improving signal-to-noise ratio (SNR) is essential for DOA estimation, making denoising become a necessary step before DOA estimation. In this letter, de-noising convolutional neural network (DnCNN) is introduced into array radar to realize signal denoising. It adopts residual learning to remove latent noise-free signal, and then outputs noise estimation. Considering that the inputs are one-dimensional complex signals, we adjust the DnCNN parameters such as convolutional channels number, convolutional filters number, convolutional kernel size, and discuss the appropriate network depth. The results show that the DnCNN has remarkable effect on noise filtering, so that accurate DOA estimation can be obtained. In addition, DnCNN has quite strong generalization ability for signals with even lower SNR.