UWB NLOS Identification and Mitigation Based on Gramian Angular Field and Parallel Deep Learning Model

Ultrawideband (UWB) wireless localization technology has been widely applied in the field of indoor localization due to its good ability of noise resistance, strong penetration, and high measurement accuracy. However, the performance of UWB-based localization technology becomes poor when suffering from nonline-of-sight (NLOS) propagation conditions. Thus, it is necessary to identify NLOS propagation and mitigate the NLOS error. In this article, a novel NLOS identification and mitigation method based on multiinputs parallel deep learning model and Gramian angular field (GAF) is proposed. We utilize GAF to transform 1-D channel impulse response (CIR) signal into 2-D colored images, which adds additional high-level abstract features to the CIR signals. In the model training phase, the convolutional neural network (CNN) is used to extract temporal features from original CIR signals, and the residual network (ResNet) is used to extract visual features from GAF-encoded images. Besides, the received signal strength (RSS) information is also considered as an auxiliary feature to assist in identifying some NLOS scenarios with similar CIR features and further reduce the NLOS error. The experimental results show that our method has good ability in both line of sight (LOS) and NLOS binary classification and NLOS multiclassification, with accuracy over 96%. Additionally, based on the identification results, the proposed method can reduce the mean absolute error (MAE) and root mean square error (RMSE) of the range error from 65.61 and 96.82 cm to 4.19 and 6.95 cm, respectively. In the real indoor localization experiment, the proposed method can improve the localization accuracy by over 80%.