Near-field localization promises high‑resolution multi‑user positioning for future wireless networks, yet scattering‑induced coherent propagation often degrades its performance. Existing approaches separate parameter estimation from path/source association, leading to heavy computation, error accumulation, and no reliability guarantees. This work introduces MUSIC‑Net, an end‑to‑end deep‑learning framework that incorporates two‑stage Multiple Signal Classification (MUSIC) into training for mixed line‑of‑sight (LoS) and non‑LoS (NLoS) multipath scenarios. The two‑stage MUSIC objects first isolate the LoS‑related signal subspace and then generate a surrogate distance, enabling direct recovery of multi‑user positions without explicit NLoS parameter estimation or path association. Split conformal prediction (SCP) is further employed to move beyond point estimates, providing statistically guaranteed confidence sets for all users. Simulations show that MUSIC‑Net achieves lower mean positioning error (MPER) than prior benchmarks and yields tighter SCP‑calibrated prediction regions, demonstrating accurate LoS localization and efficient uncertainty quantification in coherent multipath environments.
Review