Multipoint Channel Charting for Wireless Networks

Abstract

Multipoint channel charting is a machine learning framework in which multiple massive MIMO (mMIMO) base-stations (BSs) collaboratively learn a multi-cell radio map that characterizes the network environment and the users' spatial locations. The method utilizes large amounts of high-dimensional channel state information (CSI) that is passively collected from spatiotemporal samples by multiple distributed BSs. At each BS, a high-resolution multi-path channel parameter estimation algorithm extracts features hidden in the acquired CSI. Each BS then constructs a local dissimilarity matrix based on the extracted features for its collected samples and feeds it to a centralized entity which performs feature fusion and manifold learning to construct a multi-cell channel chart. The objective is to chart the radio geometry of a cellular system in such a way that the spatial distance between two users closely approximates their CSI feature distance. We demonstrate that (i) multipoint channel charting is capable of unravelling the topology of a Manhattan-grid system and (ii) the neighbor relations between CSI features from different spatial locations are captured almost perfectly.