Dynamic Channel Charting: Integrating Online Sample Selection with Continual Learning for Streaming CSI Data

Wireless channel charting consists in the application of dimensionality reduction methods to the channel state information (CSI) collected during the operation of wireless communication systems. In the context of streaming data and due to hardware limitations, it is desirable to minimize the stored information during training; hence we propose online sample selection to perform dynamic channel charting. More specifically, we introduce a dual-memory approach consisting of a short-term buffer for recently received CSI samples and a longterm memory that retains key samples accumulated over time. We present novel unsupervised replay-based continual learning strategies for triplet-based channel charting from streaming CSI data. In particular, we develop three sample selection strategies for memory updates that are compatible with real-time operation and based on the channel chart, the loss function, or its gradient. Validation using real data demonstrate that our methods outperform existing replay-based unsupervised approaches.