Deep learning aided CSI feedback optimization with robust error recovery

Driven by the increasing demand for efficient data transmission, massive Multiple-Input Multiple-Output (MIMO) systems have emerged as a key technology for future communication systems. However, effective utilization of MIMO relies heavily on accurate Channel State Information (CSI) that is fed back to the base station, which poses significant challenges due to the overhead associated with CSI feedback, especially with the increasing number of antennas. To overcome these drawbacks, this paper proposes a Deep Learning (DL) scheme to improve the CSI feedback, presenting a network named CsiDNet, which compresses CSI at the user end and decompresses it at the base station side. In addition, an auxiliary module is designed to restore CSI information under error-prone scenarios, enhancing the robustness of the system. Extensive performance analysis and simulations demonstrate that CsiDNet achieves an improvement of 2.7 dB and 0.1 dB in terms of Normalized Mean Square Error (NMSE) and Square Generalized Cosine Similarity (SGCS) respectively compared to other models, while significantly reducing computational complexity. The auxiliary module further improves the NMSE and SGCS performance by 4 dB and 0.1 dB respectively, reflecting its effectiveness in recovering error-prone CSI components. Overall, our research improves the accuracy and efficiency of CSI feedback while enhancing the system's robustness against real-world transmission challenges.