Disentangling Dual-Polarized Channel Representations for Accurate User Localization

Abstract

Indoor localization has emerged as a critical enabling technology for various smart applications, yet its performance is constrained by multipath propagation, signal blockage, and environmental dynamics. While recent deep learning (DL)-based approaches have demonstrated promising improvements over traditional techniques, their effectiveness is often limited by high architectural complexity, limited robustness against electromagnetic-environment variations, and weak interpretability. To address these limitations, this paper proposes a novel indoor localization framework that leverages the structural characteristics of dual-polarized channel state information (CSI) through disentangled representation learning. By decoupling the shared and exclusive latent features embedded in the two polarization channels, the proposed framework is able to isolate location-relevant properties from device-specific and electromagnetic distortions. Furthermore, a three-stage joint optimization strategy is developed to ensure effective feature disentanglement and robust localization performance. Compared with various representative localization models, the proposed framework achieves over 19% enhancement in localization accuracy, demonstrating improved adaptability under imperfect CSI, clock-synchronization impairments, different obstacle densities, dynamic scenarios, different antenna configurations, and reduced training data conditions, with better model interpretability and low inference complexity. Ablation studies and feature visualization further confirm the viability of the proposed network structure design, highlighting the effectiveness of information-theoretic disentanglement in extracting location-relevant features.