Robust DOA Estimation Using Deep Complex-Valued Convolutional Networks with Sparse Prior
Shulin Hu, Cao Zeng, Minti Liu, Haihong Tao, Shihua Zhao, Yu Liu
Abstract
Mathematically speaking, the direction of arrival (DOA) estimation methods with data-driven deep learning generally exhibit enhanced robustness under unknown clutter and noise scenarios. However, it is worthily noticed that the existing methods mostly suffer from signal model mismatch, due to relying on real-valued operations ignoring actual received signals are complex values. To this regard, a novel DOA estimation method with sparse prior based on deep complex-valued convolutional neural network (CV -CNN) is proposed. Specifically, the complex signal features of covariance matrix firstly are extracted by directly utilizing many complex-valued (CV) layers mainly referring to CV-Conv2d, CV-CELU, CV-FC, CV-BatchNorm2d and CV-Dropout operations. Then, the modified loss function based on sparsity prior constraint is formulated to improve estimation accuracy. Notably, the training process is innovatively divided into two stages, i.e., pre-training and fine-tuning, to further optimize the model parameters. Finally, simulation results indicate that the proposed CV-CNN method possesses superior performance than that of the existing real-valued deep networks on estimation accuracy and robustness under the low SNR or limited snapshots scenarios.