Abstract
Angle-based positioning systems have emerged as critical technologies for precise indoor localization across robotics, healthcare, and industrial automation applications. Ultrawideband (UWB) phase-based angle measurements offers theoretical sub-degree accuracy, but practical implementations suffer from channel inconsistency errors that significantly degrade performance. A dual-layer Bayesian neural network fusion framework (DBNNFF) was presented that effectively addresses these systematic errors through an innovative combination of physical constraints and uncertainty-aware modeling. Experiments were conducted in a microwave anechoic chamber using a customized 5-channel UWB base station and single-channel tags. Data was collected across seven azimuth angles between ± 30°, with 30s cold-start cycles per angle. The DBNNFF framework reduced the angle errors by 94.7% to 0.1036° ± 0.0182°, outperforming many existing algorithms by 25-42.1%. The framework's dual-network architecture-combining channel correlation model and cold start state distribution estimator-with uncertainty-weighted Bayesian fusion provides well-calibrated confidence intervals and exceptional noise robustness. Experiments conducted in multi-path environments such as office and hallway demonstrated that the DBNNFF algorithm exhibited robust performance, with errors maintained within 0.17°.