Model-assisted channel charting for 5G-Advanced localization with sparse data

Achieving robust 5G-Advanced indoor localization under sparse data is an active research topic. Pure Channel Charting (CC) preserves local neighborhood structure but yields unitless embeddings with potentially nonlinear global distortions, complicating metric alignment. Model-based localization degrades under Non-Line-of-Sight (NLoS) propagation and unknown noise statistics. We propose Model-assisted Channel Charting (MCC), which fuses: (i) masked-autoencoder pretraining on Channel State Information (CSI); (ii) triplet-based manifold learning driven by geodesic dissimilarity; and (iii) a link-wise heteroscedastic pseudorange likelihood that jointly predicts per-link residual means and uncertainties. A multi-head regressor outputs positions, residual biases, and noise variances, and uncertainty weighting balances the losses. Unlike prior model-assisted CC that relies on pairwise distance scaling, MCC anchors the chart directly to metric coordinates via the physics-consistent likelihood while learning data-driven link reliability for reweighting. Evaluations on a challenging 5G Indoor Factory (InF) dataset with severe NLoS show that MCC achieves state-of-the-art accuracy. With 18 base stations and a labeled density of 1/2 points/m 2 , the proposed Model-based Fingerprinting (MFP) and Model-based Triplet (MT) achieve the position error of 0.62 m and 0.69 m for 90% of cases, respectively. Under very sparse labels ( < 1/20 points/m 2 ), MT outperforms competing baselines. Calibration results further indicate that the predicted uncertainties monotonically correlate with empirical residual errors, supporting reliability-aware localization in harsh NLoS conditions.