AI for Improved Indoor Positioning via Multi-Band Channel Charting

Abstract

Channel charting is an unsupervised positioning method that learns a low-dimensional spatial representation from channel state information (CSI), which describes how wireless signals propagate between a user and base stations. Unlike fingerprinting, it does not require ground-truth position labels during training. This thesis investigates whether using CSI from two frequency bands improves channel charting compared with conventional single-band CSI. Simulated CSI was generated in a streetcanyon environment at 3.5 GHz and 12 GHz using Sionna RT. Three dual-band channel charting methods were evaluated: averaging dissimilarities between CSI samples, multiplying similarity scores from both bands, and aligning two separately trained networks. The methods were compared with single-band channel charting and with supervised fingerprinting baselines. The results show that dual-band fusion improves channel charting performance across both chart-quality metrics and positioning accuracy. The best channel charting result was obtained with similarity multiplication, reducing the mean absolute error from 6.67 m for the best single-band reference to 5.86 m. Dual-band fingerprinting also improved performance, reducing the mean absolute error from 1.09 m to 1.02 m, although the relative improvement was smaller than for channel charting. The gains were strongest in non-line-of-sight conditions, where the best channel charting error decreased from 7.64 m for the best single-band reference to 6.54 m with dual-band similarity multiplication. These results indicate that multi-band CSI provides complementary spatial information and is especially useful for unsupervised channel charting in challenging propagation environments.