Adapt Tokenization and Patching for Wireless Time-Series in Physical-Layer Foundation Models

Develop tokenization and patching strategies for physical-layer wireless time-series data (for example, IQ samples, Channel State Information, chirps, and FFT-derived statistics) that can robustly accommodate varying entropy, sampling rates, and sequence lengths across diverse wireless technologies and use cases, and that yield a consistent common embedding space suitable for transformer-based wireless foundation models.

Background

The paper envisions a wireless physical-layer foundation model trained with self-supervised objectives, which requires transforming heterogeneous wireless time-series into a common embedding or tokenization space. Unlike NLP, wireless signals exhibit highly variable entropy and representations across technologies (e.g., LTE, 5G-NR, Wi‑Fi) and tasks, complicating patch size selection and tokenization.

Existing tokenization methods for general time-series have been explored, but the paper notes a lack of strategies tailored to wireless time-series, where differences in sampling rates, sequence lengths, and data variability are pronounced. A robust solution is needed to enable unified transformer-based modeling across telecom use cases.