Abstract: To address the non-line-of-sight (NLOS) errors caused by people moving in indoor positioning using UWB, as well as the issue of immovable base stations after training conventional neural networks, this paper proposes an Chan algorithm based on bidirectional long short-term memory neural networks (BiLSTM-Chan). This algorithm processes UWB timing data through a bidirectional LSTM to provide error correction values for UWB timing data, and then calculates the final 3D coordinates using the Chan method. The bidirectional LSTM is capable of synthesizing past and future information, capturing the characteristics in the timing data more effectively. Integrating attention mechanisms into the network can aid in analyzing key features extracted by the BiLSTM layer, making the neural network's predictions more accurate. This paper compares the proposed method through simulation experiments and real-world experiments with the BiLSTM algorithm, Chan algorithm, and the LS algorithm. The real-world experiments demonstrate that compared to the BiLSTM, Chan, and LS algorithms, the accuracy of BiLSTM-Chan algorithm respectively improved by 30.66%, 61.78%, and 61.96%.

Key words: indoor positioning, ultra-wideband, bidirectional long short-term memory neural network, attention mechanism, deep learning