Channel Fingerprints Extrapolation via Generative Model for Multi-band Massive MIMO Transmission

The application of multi-band massive multiple-input multiple-output (MIMO) technology efficiently addresses the challenge of spectrum scarcity. It ensures robust data transmission, bringing new opportunities to advance 6G communication systems. Channel fingerprints (CF), alternatively referred to as the channel knowledge map or radio environment map, are now proposed as an enabler for multi-band transmission to facilitate the acquisition of channel state information (CSI) and reduce pilot overhead. In this paper, we propose an efficient generative model for CF extrapolation capitalized on the cycle consistency. Specifically, we first establish a CF model based on a relationship between the spatial covariance matrix and channel power angle spectrum (PAS). Then, we emphasize a specific correlation of PAS at different frequencies for multiband systems and formulate the extrapolation problem as an image translation task, which aims to transform CF from one band to another while preserving their similar characteristics of statistical CSI. Based on these similarities, we propose a pair of generative networks to extract features and complete CF transfer between different frequencies. In order to couple the paired networks together, we then introduce the cycle consistency loss so that the reciprocal nature in CF extrapolation can be well simulated. Numerical results demonstrate that our proposed method achieves bidirectional extrapolation with an error of 4 17 dB lower than the benchmark methods. We also evaluate the proposed approach in different communication scenarios, illustrating its good generalization ability.