A Structured Bidirectional LSTM Deep Learning Method For 3D Terahertz Indoor Localization

High-accuracy localization technology has gained increasing attention in gesture and motion control and many diverse applications. Due to the shadowing, multi-path fading, blockage effects in indoor propagation, 0.1m-level precise localization is still challenging. Promising for 6G wireless communications, the Terahertz (THz) spectrum provides ultra-broad bandwidth for indoor applications. Applying to indoor localization, the channel state information (CSI) of THz wireless signals, including angle of arrival (AoA), received power, and delay, has unprecedented resolution that can be explored for positioning. In this paper, a Structured Bidirectional Long Short-term Memory (SBi-LSTM) recurrent neural network (RNN) architecture is proposed to solve the CSI-based three-dimensional (3D) THz indoor localization problem with significantly improved accuracy. As a two-level structure, the features of individual multi-path ray are first analyzed in the Bi-LSTM network at the base level. Furthermore, the upper level residual network (ResNet) of the constructed SBi-LSTM network extracts for the geometric coordinates. Simulation results validate the convergence of our SBi-LSTM method and the robustness against indoor non-line-of-sight (NLoS) blockage. Specifically, the localization accuracy in the metric of mean distance error is within 0.27m under the NLoS environment, which demonstrates 60% enhancement over the state-of-the-art techniques.