Machine Learning Model for Reconfigurable Intelligent Surface Channel Estimation

The patent introduces a machine learning model as an intermediary between the received pilot signals and the channel coefficient estimates. The model learns the mapping from pilot signals to channel coefficients during training, and during runtime, it directly outputs channel coefficients without requiring exhaustive estimation procedures. This intermediary approach reduces the overhead while maintaining estimation accuracy.

The patent uses a trained machine learning model that has learned the channel characteristics during a training phase. Instead of performing heavy channel estimation during runtime, the system copies the learned patterns from the training phase and applies them during operation, significantly reducing the computational overhead and signaling requirements while maintaining estimation precision.

The patent enables the passive RIS to effectively participate in channel estimation by having the base station configure the RIS reflection coefficients based on learned patterns. The system self-organizes the estimation process by training the ML model to understand how passive RIS configurations affect the channel, eliminating the need for active RF elements at the RIS while maintaining estimatability.

The patent changes the approach from direct physical measurement to parameter-based estimation. Instead of trying to directly measure channel coefficients through complex physical procedures, the system uses ML to learn the relationship between RIS configuration parameters and channel characteristics, then uses these learned parameters to infer channel state information, making estimation feasible with passive elements.