Machine learning-based NLOS recognition for UWB positioning using low-dimensional features

Abstract Ultra-Wideband (UWB) positioning is significantly affected by Non-Line-of-Sight (NLOS) errors, which limit its accuracy in complex indoor environments. This paper proposes a lightweight NLOS error recognition method based on two low-dimensional physical features: the power difference between the total received and the first path power and the ranging value. Three datasets were collected under different occlusion conditions (wooden door, wall, glass, and human body) across multiple environments. Four representative algorithms—KNN, RF, BPNN, and CNN—were evaluated in terms of accuracy and robustness. Results show that traditional machine learning methods (KNN, RF) consistently achieve 97%–99% accuracy across all datasets and demonstrate strong robustness under cross-scenario conditions, while deep learning models (BPNN and CNN) exhibit limited performance with low-dimensional features, achieving accuracies generally below 65%.The study indicates that low-dimensional features are better suited for machine learning models, whereas deep learning requires higher-dimensional and richer inputs to realize its potential. These findings provide practical insights for error modeling and compensation in low-cost UWB positioning systems.