DNN-Driven Single-Snapshot Near-Field Localization for Hybrid Beamforming Systems
Suhwan Jang, Chungyong Lee
Abstract
This paper introduces a novel deep neural network (DNN)-driven framework for single-snapshot near-field source localization in hybrid beamforming systems. Existing state-of-the-art methods lack the joint optimization of the analog beamformer and the localizing function. We address this issue by deviating from the conventional DNN approach, where the output of the network is merely extracted. Instead, we additionally leverage the weights of the network for the beamformer design. In the training stage, the DNN is trained under specific constraints with the argument of the received signal, which is unknown in the inference stage. The trained network is then split into early and later parts for localization using the combined signal in the inference stage. We extract optimized real-valued weights from the early part and combine them for the design of the complex-valued beamformer. Simultaneously, the latter part becomes the localizing function that returns the position with the combined signal argument. Consequently, joint optimization is achieved via DNN. The inference network performs the equivalent localization task with the trained network without awareness of the received signal. Simulation results demonstrate the superiority of the proposed method over existing methods.