Deep Learning-Based Multipath DoAs Estimation Method for mmWave Massive MIMO Systems in Low SNR

Abstract

To realize the direction-of-arrivals (DoAs) estimation without the prior information about the multipath number, a novel method using the deep learning is introduced for the millimeter (mmWave) massive multiple-input and multiple-output (MIMO) systems. In particular, the DoAs estimation is decomposed into three sub-problems, which are solved by the corresponding convolutional neural network (CNN), respectively. The estimation of multipath number is achieved by a multi-label classification model using the proposed CNN-I. Then, the proposed CNN-II is introduced for the DoA estimation of the line-of-sight (LOS) path. Based on the predicted multipath number obtained by the CNN-I, a regression model using the proposed CNN-III is applied for the DoAs estimation of non-line-of-sight (NLOS) paths. The proposed DoAs estimation method is implemented by learning the non-linear relationship between the sample covariance matrix of the received signal and the angles, thus reconstructing the mapping model. A series of results validate the superiority of the proposed DoAs estimation method in low signal-to-noise-ratio (SNR) regimes. The CNN-I is capable of achieving the good multipath number estimation performance across a range of SNRs. The classification model based on the proposed CNN-II and the regression model using the proposed CNN-III achieve the lower DoAs estimation root mean square error (RMSE) compared with some deep learning-based methods, estimation of signal parameters via rotational invariance techniques (ESPRIT) and Root-MUltiple SIgnal Classification (Root-MUSIC) methods. Remarkably, the proposed method using the CNN-II exhibits the good robustness in the DoA estimation of the LOS path. Furthermore, the low DoAs estimation RMSE is also achieved in the uniform planar array.