A deep learning-based approach to lightweight CSI feedback

Some deep learning-based CSI feedback models have high computational and storage requirements, which limit their feedback efficiency on mobile devices, making them challenging to deploy on a large scale. Therefore, to address the poor feasibility of existing deep learning-based CSI feedback methods in practical deployment on user devices, a lightweight CSI feedback network suitable for mobile terminals is proposed to reduce the demand for computational and storage resources. This network enables efficient feedback on mobile devices. It leverages the design concept of a multi-resolution network to enhance feedback performance while reducing the number of parameters and computational load of the feedback network. Additionally, it employs dynamic convolution to effectively capture the contextual information of CSI. Through simulation comparison, it is found that compared with other lightweight CSI feedback networks based on deep learning, the feedback accuracy in each scenario is improved by 8.57% on average. [Display omitted] •This chapter investigates the CSI feedback problem for massive MIMO in FDD mode, proposing LWNet, a lightweight network that incorporates a multi-resolution encoder and full-dimensional dynamic convolution (ODConv) in the decoder for enhanced feature extraction.•A comparison of LWNet with four other lightweight networks (CCA-Net-L, ACRNet, CLNet, CRNet) in both indoor and outdoor environments shows LWNet outperforms the others at most compression ratios, except for a slight performance drop compared to CLNet at a 1/4 compression ratio. This indicates LWNet improves feedback accuracy while maintaining efficiency.•As compression ratio increases, NMSE also rises, indicating that larger compression ratios degrade feedback performance by reducing the CSI information provided in the codeword. Feedback is more accurate in indoor environments than outdoor ones at the same compression ratio.•To validate the experimental results, all ODConv layers in the LWNet decoder were replaced with standard convolutions, creating a new network, NoODConvNet. LWNet outperforms NoODConvNet in terms of NMSE under all compression ratios, both indoors and outdoors. The ODConv mechanism is less effective in outdoor conditions, leading to slightly better feedback from LWNet.