Reducing pilot overhead with distributed neural networks in RIS-aided downlink channel estimation

In advanced communication systems using reconfigurable intelligent surfaces (RIS), estimating the downlink channel is crucial but often demands lots of pilot signals. This can hinder efficiency. We present a new way to cut down those pilot needs by using distributed neural networks for the estimation. Leveraging neural networks' learning power, we've crafted a machine learning technique that figures out high-dimensional channels from just a few pilot signals at the user's end. Since single-user training data might have limited channel scenarios, we've also developed a distributed machine learning framework. This lets the base station and multiple users team up to train a shared channel estimation network, making it adaptable to various channel conditions. Plus, we've introduced a hierarchical neural network structure to pick out different channel features for different scenarios, improving accuracy. Our simulation studies reveal that the proposed approach improves normalized mean square error (NMSE) by up to 25% and reduces pilot signal overhead by 35% when compared to conventional single-user training methods. These results highlight the effectiveness of the proposed method in achieving efficient channel estimation for RIS-assisted wireless communication systems. • Novel DNN-based channel estimation to reduce pilot overhead in RIS-aided systems. • ML method estimates high-dimensional channels using minimal pilot signals. • Distributed ML enables collaborative BS-user training for adaptive estimation. • Hierarchical neural structure captures diverse channel features efficiently.