Local Power Control and Precoding in Cell-Free Massive MIMO Systems with Distributed Learning

Cell-free massive multiple-input multiple-output (CF-mMIMO) has garnered significant attention due to its ability on reducing the path loss and providing uniform service quality. As the cell boundary eliminates, the inter-user interference is dominant. Hence, power control and precoding which can effectively surpass the inter-user interference become crucial in CF-mMIMO. However, the practical application of them needs centralized information on the central processing unit which leads to challenges from the fronthaul payload and delays constraints. To address this issue, we propose a distributed learning (DL) framework to design local power control and precoding. Specifically, we deploy distributed deep neural network (DNN) models at each access point, which takes the optimal max-min power control as the label to train the local power control, and take the centralized zero-forcing precoding as label to train a local precoding. This approach do not need channel state information exchange during the entire process, resulting in a CF-mMIMO system with fully local signal processing. The simulation results show that the DL-aided power control and precoding can significantly improve the spectral efficiency over the existing local methods. Its advantage is more pronounced for a larger scale of CF-mMIMO.