Resilient Channel Charting Under Varying Radio Link Availability

TL;DR

This work tackles the challenge of channel charting when radio link availability is variable, such as during antenna outages or handovers. It introduces AdaPos, a transformer-based CC architecture with a shared signal encoder and learnable antenna embeddings that can fuse arbitrary subsets of channel state information inputs without retraining. Using Siamese training and novel antenna-dropout strategies, AdaPos achieves state-of-the-art or near-state-of-the-art localization accuracy while replacing roughly 57 configuration-specific models with a single unified model. Evaluations on two real-world datasets (Dichasus and Fraunhofer 5G) demonstrate strong resilience to missing antennas and scalable performance for large antenna arrays, with training strategies that further boost robustness. The approach significantly improves practical deployability of RF-based localization in dynamic environments, enabling reliable localization under partial or complete antenna failures.

Abstract

Channel charting (CC) has become a key technology for RF-based localization, enabling unsupervised radio fingerprinting, even in non line of sight scenarios, with a minimum of reference position labels. However, most CC models assume fixed-size inputs, such as a constant number of antennas or channel measurements. In practical systems, antennas may fail, signals may be blocked, or antenna sets may change during handovers, making fixed-input architectures fragile. Existing radio-fingerprinting approaches address this by training separate models for each antenna configuration, but the resulting training effort scales prohibitively with the array size. We propose Adaptive Positioning (AdaPos), a CC architecture that natively handles variable numbers of channel measurements. AdaPos combines convolutional feature extraction with a transformer-based encoder using learnable antenna identifiers and self-attention to fuse arbitrary subsets of CSI inputs. Experiments on two public real-world datasets (SISO and MIMO) show that AdaPos maintains state-of-the-art accuracy under missing-antenna conditions and replaces roughly 57 configuration-specific models with a single unified model. With AdaPos and our novel training strategies, we provide resilience to both individual antenna failures and full-array outages.

Figures (6)

• Figure 1: Overview of the adapos architecture. Information flows from bottom to top. We first pass individual channel impulse responses (CIRs) through a signal encoder (1D-ResNet) that creates one embedding per cir. Next, the signal combiner adds an antenna embedding to each cir embedding so the model identifies the source antenna. A transformer encoder then combines the set of cir embeddings through self-attention. Its output is average-pooled and passed through an MLP head to return the CC pseudo 2D positions.
• Figure 2: MAE of ResNet (left), AdaPos (right) on the Fraunhofer 5G dataset.
• Figure 3: MAE of ResNet (left), AdaPos (right) on the Dichasus dataset, w. single antennas masked.
• Figure 4: ResNet (left), AdaPos (right) trained with full arrays masked and evaluated with single elements missing.
• Figure 5: MAE of ResNet (left), AdaPos (right) on the Dichasus dataset, w. arrays masked.