Deep Learning-Aided Spatial Discrimination for Multipath Mitigation

Abstract

A deep learning-aided spatial discriminator for multipath mitigation is developed. The proposed system compensates for the limitations of conventional beamforming approaches, especially at the stages of: prefiltering, model order estimation (MOE), and direction-of-arrival (DOA) estimation. Three environments are considered to design and train the proposed deep neural networks (DNNs): indoor office buildings, indoor open ceiling, and outdoor urban area. The performance of the proposed DNN-based MOE is compared to the conventional approaches of minimum description length (MDL) criterion and Akaike information criterion (AIC). The proposed DNN-based MOE is shown to significantly outperform existing approaches and to increase the degrees-of-freedom. Four experiments are presented to assess the performance of the proposed system in multipath-rich environments corresponding to indoor pedestrian navigation and ground vehicle urban navigation with cellular long-term evolution (LTE) signals. The proposed system exhibited a position root mean-squared error (RMSE) of 1.67 m, 3.38 m, 1.73 m, and 2.16 m.