A Lightweight Multi-Attention Network With Denoising Block Designed for Massive MIMO CSI Feedback

In massive multiple-input multiple-output (MIMO) systems, the frequency division duplex (FDD) mode significantly enhances data transmission capabilities. However, it requires user equipment (UE) to feedback the channel state information (CSI) to the base station (BS) for efficient precoding. As the number of BS antennas increases, the complexity of CSI feedback grows, making CSI compression techniques essential. Deep learning (DL)-based approaches have shown great potential in addressing this challenge. Given the storage and computational limitations at the UE, it is critical to design networks that deliver high reconstruction performance with minimal floating-point operations per second (FLOPS). We propose a novel method that leverages dilated convolutional attention and multi-head attention mechanism to improve CSI feedback quality without increasing system complexity. To enhance its practical application, we introduce a denoise block based on multi-head attention to correct errors that occur during transmission. Simulation results show that our method achieves lower FLOPS compared to existing models, while maintaining comparable or even better performance. Additionally, the attention-based denoising block exhibits strong error-correction capabilities, further improving system robustness.