CrossNet: Lightweight Transformer-Based CSI Compression Network for Massive MIMO Systems

We consider a frequency division duplex massive multiple input multiple output (mMIMO) system, where efficient acquisition and feedback of channel state information (CSI) are critical for optimal system performance. High-dimensional CSI matrices in mMIMO systems have a significant overhead, which makes compression essential for real-time feedback. Existing deep learning transformer-based CSI compression methods have made progress, but they do not fully exploit the inherent spatial and frequency domain features of CSI, and have a high computational complexity. We propose CrossNet, a dual branch transformer-based autoencoder, which first separately exploits the spatial and frequency domain features of CSI by using self-attention and cross-attention modules, and then fuses them. We show using exhaustive numerical studies that CrossNet not only provides comparable and superior performance over existing models, but also has a significantly lower parameter count and complexity.