부산대학교 도서관

The accuracy and overhead of Channel State Information(CSI) feedback are important for massive MIMO systems. As the scale of antennas increases, the contradiction between CSI accuracy and feedback overhead will become more severer. This paper first studies three training mechanisms for the popular transformer-based CSI feedback algorithm. Then, the algorithm is compared with the Type II codebook in 3GPP Release 16 in terms of both the Squared Generalized Cosine Similarity(SGCS) and User Perceived Throughput(UPT) under three training methods. Furthermore, we simulate its performance under various channel environments to verify its generalization capability, and analyze the complexity and feasibility of different training methods. The results show that with the same feedback overhead, the transformer-based CSI feedback can obtain about 6.5% SGCS gain and up to 15% UPT gain than the Release 16 Type II codebook. By choosing a reasonable training method, transformer-based CSI feedback can achieve good generalization with low training complexity.