Channel Charting: Locating Users Within the Radio Environment Using Channel State Information

Abstract

We propose channel charting (CC), a novel framework in which a multi-antenna network element learns a chart of the radio geometry in its surrounding area. The channel chart captures the local spatial geometry of the area so that points that are close in space will also be close in the channel chart and vice versa. CC works in a fully unsupervised manner, i.e., learning is only based on channel state information (CSI) that is passively collected at a single point in space, but from multiple transmit locations in the area over time. The method then extracts channel features that characterize large-scale fading properties of the wireless channel. Finally, the channel charts are generated with tools from dimensionality reduction, manifold learning, and deep neural networks. The network element performing CC may be, for example, a multi-antenna base-station in a cellular system and the charted area in the served cell. Logical relationships related to the position and movement of a transmitter, e.g., a user equipment (UE), in the cell can then be directly deduced from comparing measured radio channel characteristics to the channel chart. The unsupervised nature of CC enables a range of new applications in UE localization, network planning, user scheduling, multipoint connectivity, hand-over, cell search, user grouping, and other cognitive tasks that rely on CSI and UE movement relative to the base-station, without the need of information from global navigation satellite systems.

Siamese Neural Networks for Wireless Positioning and Channel Charting

“…Channel charting (CC), as proposed in [16], avoids extensive (and expensive) measurement campaigns for applications that do not require absolute positioning capabilities (e.g., for hand-over, cell search, and user grouping). The principle of CC is to exploit the fact that CSI is high-dimensional, but strongly dependent on UE position, which is low-dimensional.…”

“…All of these methods are supervised and require extensive measurement campaigns to generate large databases consisting of CSI measurements and accurate position information at densely sampled locations in space (often at wavelength scales). Channel charting (CC), as proposed in [16], is unsupervised and uses dimensionality reduction to perform relative positioning solely from CSI measurements, without the need of ground-truth position information. Recent extensions of CC include multi-point CC [24] for systems with simultaneous connectivity to multiple BSs and semisupervised CC with autoencoders [22], which enables the inclusion of partially-annotated datasets.…”

“…The parameters w n,m are used to de-weight the importance of pairs of vectors that are dissimilar in high-dimensional space. The common choice for Sammon's mapping is w n,m = x n − x m −1 for n = m. As demonstrated in [16], Sammon's mapping performs exceptionally well for CC tasks. However, this technique is nonparametric as it does not explicitly learn a function f : R D → R D that maps a new high-dimensional vector x n ∈ R D into a corresponding vector y n ∈ R D in low-dimensional space.…”

“…In [16], [22], [24], Sammon's mapping, Laplacian eigenmaps, and autoencoders have been used to learn channel charts. While Sammon's mapping [16] and Laplacian eigenmaps [24] have been shown to perform well in LoS and non-LoS scenarios, their nonparametric nature limits their use in practice. Very recently, out-of-sample extensions for Laplacian eigenmaps have been proposed in [31].…”

Reducing the Complexity of Fingerprinting-Based Positioning using Locality-Sensitive Hashing

“…To evaluate the efficacy of our approach, we consider a massive MU-MIMO-OFDM localization scenario in LoS and non-LoS scenarios with a single basestation containing 32 antennas operating at 2.68 GHz with a bandwidth of 20 MHz and localizing 2000 transmitters distributed uniformly at random in an area of 40, 000 m 2 ; the noisy channel vectors are generated using channel models from [28]. The CSI features are D = 256 dimensional (32 antennas and 8 maximallyspaced subcarriers) and correspond to the absolute value of beamspace/delay domain channel vectors as in [16], [19].…”