UWB NLOS Identification and Mitigation based on Bidirectional Encoder Representations from Transformer (BERT) Deep Learning

The Non-Line-of-Sight (NLOS) phenomenon can hinder signal propagation and significantly reduce the accuracy of UWB for indoor positioning and navigation. The Channel Impulse Response (CIR) sequence generated during UWB ranging is widely used for channel identification. However, existing deep learning algorithms struggle to balance the local and global features of the CIR sequence effectively. To address this, this paper constructs a Bidirectional Encoder Representations from Transformers (BERT) channel identification model using the self-attention mechanism to improve the NLOS identification. The identification Accuracy, LOS recall, and F2 scores in multiple scenarios are 96.65%, 97.13%, and 0.9703 respectively. Comparing to state-of-art algorithms such as LS-SVM, CNN, and LSTM, our algorithm outperformed by 17.9%, 11.86%, and 10.80% respectively. For NLOS ranging errors, a fine-grained classification model is constructed with error correction model based on BERT. In multiple scenarios, the average NLOS ranging error is reduced by 41.97% and outperforms LS-SVM, CNN, and LSTM by 34.04%, 31.99%, and 16.81% respectively. In the overall positioning experiment, our algorithm achieves better performance than the existing algorithms by 32.13%.