SIABR: A Structured Intra-Attention Bidirectional Recurrent Deep Learning Method for Ultra-Accurate Terahertz Indoor Localization

High-accuracy localization technology has gained increasing attention in gesture and motion control and many diverse applications. Due to multi-path fading and blockage effects in indoor propagation, 0.1m-level precise localization is still challenging. Promising for 6G wireless communications, the Terahertz (THz) spectrum provides multi-GHz ultra-broad bandwidth. Applying the THz spectrum to indoor localization, the channel state information (CSI) of THz signals, including angle of arrival (AoA), received power, and delay, has unprecedented resolution that can be explored for positioning. In this paper, a Structured Intra-Attention Bidirectional Recurrent (SIABR) deep learning method 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 rays are first analyzed in the recurrent neural network with the attention mechanism at the lower level. Furthermore, the upper-level residual network (ResNet) of the constructed SIABR network extracts hidden information to output the geometric coordinates. Simulation results demonstrate that the 3D localization accuracy in the metric of mean distance error is within 0.25m. The developed SIABR network has very fast convergence and is robust against THz indoor line-of-sight blockage, multi-path fading, channel sparsity and CSI estimation error.