Geometric Consensus Convolutional Neural Network for NLOS Identification via Range-Only Measurements in High-GDOP Conditions

Abstract

Non-line-of-sight (NLOS) identification is essential for accurate range-only localization (ROL) but remains challenging in high geometric dilution of precision (high-GDOP) environments. This letter proposes a geometric consensus convolutional neural network (GC-CNN) framework for simultaneous multi-anchor NLOS identification. The proposed method systematically evaluates pairwise anchor configurations to construct a geometric consensus matrix that transforms discrete range inconsistencies into a structured topological representation. A convolutional neural network is then employed to learn discriminative spatial features from this matrix and jointly predict the LOS/NLOS states of all anchors. An anchor-wise shared classifier enables a single model to accommodate different anchor counts. Across 15 test scenarios covering variations in geometry, anchor count, noise, and NLOS-error distribution, GC-CNN achieves an average Macro-F1 of 86.9%, compared with 81.3% for the best baseline. Generalization is also retained for surrounding layouts and unseen anchor counts. After predicted NLOS measurements are filtered, the localization RMSE is reduced by 52.0% on average, compared with 42.3% for the best baseline.