Trans-LSTMNet: An Approach of Optimization with Multi-Domain Features for CSI Feedback
Daren Feng, Jin Xu, Xiaofeng Tao · 2024
In frequency division duplexing (FDD) massive multiple-input multiple-output (MIMO) systems, precise acquisition of downlink channel state information (CSI) through CSI feedback is crucial to harnessing the potential advantages of massive MIMO. To alleviate the excessive feedback overhead caused by the proliferation of antennas, a series of deep learning (DL)-based CSI feedback schemes have been proposed. However, most existing solutions have not fully exploited correlations among channel features of multiple domains, leading to insufficient CSI reconstruction accuracy. In this paper, a novel architecture named Trans-LSTMNet is proposed to improve the performance of CSI feedback, in which transformer is utilized to extract spatial and frequency domain features, while a memory block is designed to capture the temporal correlation based on convolutional long short-term memory (ConvLSTM). Additionally, a learnable vector quantization layer is introduced to minimize the quantization error. Results of comparable experiments show that our proposed Trans-LSTMNet could effectively improve the CSI reconstruction accuracy and significantly reduce the feedback overhead. Moreover, Trans-LSTMNet exhibits commendable generalization capability and superior performance with different settings.