DiscoCSINet: Dissymmetric Convolution Neural Network for CSI Feedback in FDD Massive MIMO System

Channel state information (CSI) is an essential aspect of the frequency division duplex (FDD) massive multiple-input multiple-output (MIMO) system since there is no reciprocity between the bidirectional channels. However, the CSI transmission often requires significant channel resources because there may be hundreds of antennas transmitting and receiving data simultaneously. In this paper, we design an dissymmetric convolution neural network for CSI feedback (DiscoCSINet). Specifically, we utilize the dissymmetric convolution blocks (Disco-Blocks) to address the CSI compression and decompression issue, where convolution's feature extraction capability can be enhanced. To improve the storage efficiency of the receiver, we also employ a lightweight approach of the DiscoCSINet. Furthermore, we explore the fusion strategies of multi-rate and multi-scenario, respectively, and strengthen the generalization capability of the DiscoCSINet in practical settings. Experiment results indicate that the proposed DiscoCSINet can notably enhance the NMSE and cosine similarity rho, especially in outdoor scenarios. Additionally, the proposed lightweight approach and multi-model fusion strategies can greatly decrease the parameter amounts by over 80% and 89%, respectively, but the performance are only slightly decayed.