用生成模型跨频段预测信道指纹，提升多频段大规模MIMO性能。

Multi-band massive multiple-input multiple-output (MIMO) communication can promote the cooperation of licensed and unlicensed spectra, effectively enhancing spectrum efficiency for Wi-Fi and other wireless systems. As an enabler for multi-band transmission, channel fingerprints (CF), also known as the channel knowledge map or radio environment map, are used to assist channel state information (CSI) acquisition and reduce computational complexity. In this paper, we propose CF-CGN (Channel Fingerprints with Cycle-consistent Generative Networks) to extrapolate CF for multi-band massive MIMO transmission where licensed and unlicensed spectra cooperate to provide ubiquitous connectivity. Specifically, we first model CF as a multichannel image and transform the extrapolation problem into an image translation task, which converts CF from one frequency to another by exploring the shared characteristics of statistical CSI in the beam domain. Then, paired generative networks are designed and coupled by variable-weight cycle consistency losses to fit the reciprocal relationship at different bands. Matched with the coupled networks, a joint training strategy is developed accordingly, supporting synchronous optimization of all trainable parameters. During the inference process, we also introduce a refining scheme to improve the extrapolation accuracy based on the resolution of CF. Numerical results illustrate that our proposed CF-CGN can achieve bidirectional extrapolation with an error of 5-17 dB lower than the benchmarks in different communication scenarios, demonstrating its excellent generalization ability. We further show that the sum rate performance assisted by CF-CGN-based CF is close to that with perfect CSI for multi-band massive MIMO transmission.