Self-Attention-Assisted TinyML With Effective Representation for UWB NLOS Identification
Yifeng Wu, Xu He, Lingfei Mo, Qing Wang
Abstract
Ultra-Wide Band (UWB) Non-Line-of-Sight (NLOS) identification is a crucial task in wireless localization systems. Various Deep Learning (DL) solutions have demonstrated promising outcomes in UWB NLOS identification by utilizing Channel Impulse Response (CIR) and channel characteristics. However, effective and robust UWB NLOS identification on resource-constrained edge devices remains a challenge. Hence, this paper presents a self-attention-assisted Tiny Machine Learning (TinyML) solution that offers an effective representation for UWB NLOS identification. To overcome computational limitations, a feature selection method is devised for the proposed data-driven DL-based approach. By leveraging feature selection, the self-attention mechanism enhances the representation capability of a pre-trained model for UWB NLOS identification. The proposed method is evaluated on both personal computer (PC) and edge platforms, and compared against multiple baselines. The evaluation demonstrates its effective representation and optimal performance on both PC and edge platforms, as indicated by various metrics. Thanks to the effective representation, the proposed method also enables the quantized model to achieve State-of-The-Art (SOTA) in UWB NLOS identification, while significantly accelerating inference efficiency at the edge.