Deep Learning-Based Location to Precoding Mapping in Massive MIMO Systems

Channel acquisition for precoding design in massive multiple-input multiple-output (MIMO) systems faces increasingly prominent challenges due to the huge overhead caused by channel feedback and transmission of pilot signals. In response, directly mapping location information to precoding without channel feedback has emerged as a feasible and efficient solution. However, existing deep learning-based methods often struggle to map low-dimensional positional data to high-dimensional precoding vectors accurately. To address this challenge, we propose a spatially adaptive mapping network (SAM-Net) that enhances feature extraction and representation by leveraging transposed convolution and incorporating spatial information for fine-grained adaptive adjustments. While SAM-Net improves mapping performance, its complexity also increases. Therefore, we introduce a lightweight spatially adaptive mapping network (LSAM-Net) that combines average pooling and max pooling to reduce the number of parameters and computational complexity while it maintains near-optimal performance. Evaluation results demonstrate that SAM-Net achieves superior and stable mapping performance, while LSAM-Net offers a more lightweight alternative with minimal loss in performance.