Multi-scene Association Training Deep Neural Network Based Channel Estimation for RIS Assisted Wireless Communications

In communication systems that utilize RIS (reconfigurable intelligent surfaces), estimating the downlink channel demands a substantial amount of pilot signals. However, this challenge can be mitigated by employing machine learning (ML) techniques, which leverage the neural network's capacity for learning to estimate high-dimensional channels using minimal pilot signals received by the user. Nevertheless, the data for training gathered by an individual user only reflects a subset of channel scenarios present within a cell. Consequently, the neural network trained on this limited data may not perform effectively when the user transitions between different channel scenarios. To address this issue, we introduce a novel solution that the Multi-scene Association (MSA) training method to obtained the reliable channel estimation. More precisely, we firstly construct a SENet based shared neural network for the downlink channel estimation, which the BS and all users can train with the help of MSA technology. Secondly, we further construct a deep neural network (DNN) to improve the accuracy of channel estimation. The simulation outcomes highlight that in contrast to the neural network calibrated for a solitary user, our proposed MSADNN architecture delivers superior estimation accuracy. This enhancement is notably achieved while concurrently minimizing the pilot signal overhead.