Probabilistic Transfer Learning for Cognitive LOS/NLOS Identification in Indoor UWB Networks

Abstract

Accurate perception of the propagation environment is a fundamental requirement for cognitive wireless systems operating in complex indoor scenarios. In ultra-wideband (UWB) networks, reliable line-of-sight (LOS) identification plays a critical role in enabling adaptive ranging, localization, and context-aware network operation, as non-line-of-sight (NLOS) propagation introduces severe uncertainty and systematic ranging biases. However, strong multipath propagation, partial obstructions, and structural variability in indoor environments lead to intrinsically ambiguous channel conditions, challenging conventional hard-decision LOS/NLOS classification approaches. This paper proposes a probabilistic transfer learning framework for cognitive LOS/NLOS identification in indoor UWB networks. Each indoor environment is interpreted as a distinct cognitive domain, and a shared deep encoder pretrained across heterogeneous environments is employed to learn transferable and domain-invariant channel representations. To capture complementary propagation characteristics, the encoder output is processed through a multi-branch architecture that enables specialized latent feature learning while preserving generalization across environments. Instead of producing binary decisions, the proposed framework estimates calibrated posterior LOS probabilities, providing uncertainty-aware environmental awareness suitable for cognitive networking applications. The proposed approach is evaluated using a global training and testing protocol on the publicly available EWINE indoor UWB dataset, comprising seven indoor environments with diverse propagation characteristics. Experimental results demonstrate strong and consistent performance, achieving a maximum classification accuracy of 92.49%, a ROC–AUC of 0.976, and a PR–AUC of 0.978. Beyond discrimination performance, the framework exhibits excellent probabilistic reliability, as reflected by a low Brier score of 0.057 and an expected calibration error (ECE) of 0.010. Robustness analysis under additive feature noise further confirms stable performance in challenging indoor propagation conditions. These results indicate that probabilistic transfer learning enables reliable environment cognition in indoor UWB networks. By delivering calibrated confidence estimates rather than hard LOS/NLOS decisions, the proposed framework directly supports adaptive ranging, uncertainty-aware sensor fusion, and intelligent decision making in cognitive communication and networking systems.